Multitemporal Spaceborne Polarimetric SAR Data for Urban ...567158/FULLTEXT01.pdf · Multitemporal...

Multitemporal Spaceborne Polarimetric SAR Data for Urban Land Cover

Mapping

Xin Niu

Doctoral Thesis

October, 2012

Xin Niu TRITA SoM 2012-18

© Xin Niu 2012

Doctoral Thesis

Royal Institute of Technology (KTH)

Department of Urban Planning and Environment

Division of Geodesy and Geoinformatics

SE-100 44 Stockholm

Sweden

TRITA SoM 2012-18

ISSN 1653-6126

ISRN KTH/SoM/12-17/SE

ISBN 978-91-7501-535-4

Printed by e-print, Sweden 2012

Multitemporal Spaceborne Polarimetric SAR Data for Urban Land Cover Mapping

i

ABSTRACT Urban land cover mapping represents one of the most important remote sensing applications in the context of rapid global urbanization. In recent years, high resolution spaceborne Polarimetric Synthetic Aperture Radar (PolSAR) has been increasingly used for urban land cover/land-use mapping, since more information could be obtained in multiple polarizations and the collection of such data is less influenced by solar illumination and weather conditions. The overall objective of this research is to develop effective methods to extract accurate and detailed urban land cover information from spaceborne PolSAR data. Six RADARSAT-2 fine-beam polarimetric SAR and three RADARSAT-2 ultra-fine beam SAR images were used. These data were acquired from June to September 2008 over the north urban-rural fringe of the Greater Toronto Area, Canada. The major landuse/land-cover classes in this area include high-density residential areas, low-density residential areas, industrial and commercial areas, construction sites, roads, streets, parks, golf courses, forests, pasture, water and two types of agricultural crops. In this research, various polarimetric SAR parameters were evaluated for urban land cover mapping. They include the parameters from Pauli, Freeman and Cloude-Pottier decompositions, coherency matrix, intensities of each polarization and their logarithms. Both object-based and pixel-based classification approaches were investigated. Through an object-based Support Vector Machine (SVM) and a rule-based approach, efficiencies of various PolSAR features and the multitemporal data combinations were evaluated. For the pixel-based approach, a contextual Stochastic Expectation-Maximization (SEM) algorithm was proposed. With an adaptive Markov Random Field (MRF) and a modified Multiscale Pappas Adaptive Clustering (MPAC), contextual information was explored to improve the mapping results. To take full advantages of alternative PolSAR distribution models, a rule-based model selection approach was put forward in comparison with a dictionary-based approach. Moreover, the capability of multitemporal fine-beam PolSAR data was compared with multitemporal ultra-fine beam C-HH SAR data. Texture analysis and a rule-based approach which explores the object features and the spatial relationships were applied for further improvement.


ii

Using the proposed approaches, detailed urban land-cover classes and finer urban structures could be mapped with high accuracy in contrast to most of the previous studies which have only focused on the extraction of urban extent or the mapping of very few urban classes. It is also one of the first comparisons of various PolSAR parameters for detailed urban mapping using an object-based approach. Unlike other multitemporal studies, the significance of complementary information from both ascending and descending SAR data and the temporal relationships in the data were the focus in the multitemporal analysis. Further, the proposed novel contextual analyses could effectively improve the pixel-based classification accuracy and present homogenous results with preserved shape details avoiding over-averaging. The proposed contextual SEM algorithm, which is one of the first to combine the adaptive MRF and the modified MPAC, was able to mitigate the degenerative problem in the traditional EM algorithms with fast convergence speed when dealing with many classes. This contextual SEM outperformed the contextual SVM in certain situations with regard to both accuracy and computation time. By using such a contextual algorithm, the common PolSAR data distribution models namely Wishart, G0p, Kp and KummerU were compared for detailed urban mapping in terms of both mapping accuracy and time efficiency. In the comparisons, G0p, Kp and KummerU demonstrated better performances with higher overall accuracies than Wishart. Nevertheless, the advantages of Wishart and the other models could also be effectively integrated by the proposed rule-based adaptive model selection, while limited improvement could be observed by the dictionary-based selection, which has been applied in previous studies. The use of polarimetric SAR data for identifying various urban classes was then compared with the ultra-fine-beam C-HH SAR data. The grey level co-occurrence matrix textures generated from the ultra-fine-beam C-HH SAR data were found to be more efficient than the corresponding PolSAR textures for identifying urban areas from rural areas. An object-based and pixel-based fusion approach that uses ultra-fine-beam C-HH SAR texture data with PolSAR data was developed. In contrast to many other fusion approaches that have explored pixel-based classification results to improve object-based classifications, the proposed rule-based fusion approach using the object features and contextual information was able to extract several low backscatter classes such as roads, streets and parks with reasonable accuracy. Keywords: RADARSAT-2, Spaceborne, Polarimetric SAR, Urban Land cover, Object-based Rule-based Classification, Support Vector Machines, Contextual, Stochastic Expectation-Maximization, Markov Random Field.


iii

ACKNOWLEDGMENT First, I would like to express my gratitude to my father and mother. Thank you for giving me life and bringing me up. Your endless love will stay with me forever. Particularly, I am very grateful to my supervisor, Prof. Yifang Ban of KTH for her tireless guidance, encouragement, support and patience. During my stay in Sweden, Prof. Ban has given me tremendous help on my study and daily life. I would especially like to thank my alma mater NUDT in China for giving me this opportunity to extend my study in KTH. Thanks to my Chinese supervisor Prof. Yong Dou at NUDT for your concern about my study and help on my home affairs in NUDT. Many thanks to all the staff and my fellow PhD candidates in Geodesy and Geoinformatics, KTH. We have shared a good time studying and working together. Your selfless help and solace will be appreciated. Thanks to Dr. Hans Hauska, Dr. Martin Sjöström and Jan Haas especially for their helpful suggestions and comments on improving this thesis. I am also grateful to my other family members and friends for your constant support and encouragement. I would like to greatly thank the China Scholarship Council (CSC), KTH Geoinformatics and the School of Architecture and Built Environment (ABE) at KTH for providing me the scholarship and stipend. This research has been carried out as part of research projects funded by the Swedish National Space Board 'Spaceborne SAR for Analysis of Urban Environment and Detection of Human Settlements' (Grant Holder: Prof. Yifang Ban) and ‘Fusion of Spaceborne SAR and Optical Data for Urbanization Monitoring’ (Grant Holder: Prof. Yifang Ban) as well as a Canadian Space Agency SOAR Project ‘RADARSAT 2 – SAR for Urban Land-cover Mapping and Change Detection’ (Project PI: Prof. Yifang Ban). Xin Niu Stockholm, 30th September 2012


iv

TABLE OF CONTENTS

1. INTRODUCTION ........................................................................................................................ 1 1.1 Background ................................................................................................................... 1 1.2 Research Objectives ..................................................................................................... 4 1.3 Thesis Organization ...................................................................................................... 4

2. LITERATURE REVIEW .............................................................................................................. 7 2.1 SAR Polarimetry........................................................................................................... 7

2.1.1 SAR Polarimetry Fundamentals ................................................................... 7

2.1.2 Statistical Properties of PolSAR Data .......................................................... 9

2.1.3 Polarimetric Decompositions ..................................................................... 13 2.2 Spaceborne PolSAR Systems ..................................................................................... 17 2.3 SAR for Urban Analysis ............................................................................................. 19

2.3.1 Interactions between the SAR Systems and Urban Features ..................... 19

2.3.1.1 Frequency ......................................................................................... 20

2.3.1.2 Polarization ....................................................................................... 20

2.3.1.3 Incidence Angle and Look Direction ............................................... 21

2.3.2 Spatial Resolution ...................................................................................... 22

2.3.3 Multitemporal Observations....................................................................... 22 2.4 Enhancement Approaches ........................................................................................... 23

2.4.1 Speckle Reduction ...................................................................................... 23

2.4.2 Texture Analysis ........................................................................................ 24

2.4.3 Contextual Analysis ................................................................................... 25 2.5 Methodology for Urban Classification using PolSAR................................................ 27

2.5.1 Pixel- and Object-based Classifications ..................................................... 27

2.5.2 Parametric and Non-parametric Classifications ......................................... 29 2.6 Summary ..................................................................................................................... 31

3. STUDY AREA AND DATA DESCRIPTION .................................................................................. 33 3.1 Study Area .................................................................................................................. 33 3.2 Data Description ......................................................................................................... 36

3.2.1 RADARSAT-2 SAR Data.......................................................................... 36

3.2.2 Auxiliary Data ............................................................................................ 37

4. METHODOLOGY ..................................................................................................................... 38 4.1 Object-based Approach ............................................................................................... 38 4.2 Pixel-based Approach ................................................................................................. 41

4.2.1 Parametric Classification by a Contextual SEM Algorithm with FMM .... 41

4.2.2 Model Selection by a Dictionary-based Approach and a Knowledge-based

Approach .................................................................................................................. 44

4.2.3 Contextual Analysis by an Adaptive MRF and Modified MPAC .............. 46

4.2.3.1 Adaptive MRF .................................................................................. 46

4.2.3.2 Modified MPAC ............................................................................... 47 4.3 Hybrid Approach ........................................................................................................ 50

5. RESULTS AND DISCUSSION ..................................................................................................... 53 5.1 Object-based Approach ............................................................................................... 53

5.1.1 Object-based Polarimetric Feature Comparisons ....................................... 53

5.1.2 Multitemporal Combination Comparisons ................................................. 54 5.2 Pixel-based Approach ................................................................................................. 56


v

5.2.1 Convergence and Learning Control of the Contextual SEM Algorithm .... 56

5.2.2 Contextual Analysis by Adaptive MRF and MPAC .................................. 57

5.2.2.1 Beta and Adaptive Scheme in the MRF Analysis ............................ 57

5.2.2.2 Multiscale Effects in the Modified MPAC ....................................... 59

5.2.2.3 Comaprison of Adaptive MRF and MPAC ...................................... 59

5.2.3 Comparison and Selection of Various PolSAR Distribution Models ........ 61

5.2.3.1 PolSAR Model Comparison ............................................................. 61

5.2.3.2 PolSAR Model Selection.................................................................. 63

5.2.4 Comparison and Fusion of PolSAR and C-HH SAR Data by Contextual

SEM with Texture Analysis ..................................................................................... 65 5.3 Hybrid Approach ........................................................................................................ 67

5.3.1 Comparison of Results with and without Texture Analysis and Rules ...... 67

5.3.2 Fusion of Pixel-based Approach and Object-based Analysis by a

Rule-based Method .................................................................................................. 68 5.4 Evaluation of the Pixel-based, Object-based and Hybrid Approaches ....................... 69

6. CONCLUSIONS AND FUTURE RESEARCH ............................................................................... 73 6.1 Conclusions ................................................................................................................ 73 6.2 Future Research .......................................................................................................... 74

REFERENCES .................................................................................................................................. 75


vi

LIST OF ABBREVIATIONS ANN Artificial Neural Network CRF Conditional Random Field EM Expectation Maximum FMM Finite Mixture Model GLCM Grey Level Co-occurrence Matrix GoF Goodness-of-Fit GTA Greater Toronto Area LULC Land Use/Land Cover MLC MultiLook Complex MoLC Method of Log-Cumulants MoMLC Method of Matrix Log-Cumulants MPAC Modified Pappas Adaptive Clustering MRF Markov Random Field PAC Pappas Adaptive Clustering PDF Probability Density Function PolSAR Polarimetric SAR SAR Synthetic Aperture Radar SEM Stochastic EM SLC Single-Look Complex SVFMM Spatially Variant FMM SVM Support Vector Machine RCS Radar Cross Section


vii

LIST OF FIGURES Figure 1.1. Organization of the selected papers ............................................................................... 5

Figure 2.1. Example of the scattering profile of a simple flat-roof building model. ground

scattering a, double bounce b, scattering from vertical wall c, backscattering from roof d, shadow

area e. The gray values of the backscattering profile correspond to the relative amplitudes.

(Brunner et al. 2010) ....................................................................................................................... 19

Figure 3.1. Study area: rural-urban fringe of the GTA, ON, Canada with projections of the

RADARSAT-2 Ultra-Fine C-HH (blue frame) and PolSAR (red frame) data. ............................... 33

Figure 3.2. A demonstration of the RADARSAT-2 fine-quad polarimetric SAR data from June 11

and June 19 by processed Pauli parameters (refer to Table 4.1) R: (Log(T22))8bit G:

(Log(T33))8bit B: (Log(T11))8bit .................................................................................................. 34

Figure 3.3. Selected urban samples: Ground truths (first row), PolSAR Pauli images (second row)

and the corresponding polarimetric signatures: co-pol (third row), cross-pol (fourth row). ........... 35

Figure 4.1. Flow chart of multitemporal data fusion combining SVM and rule-based method ..... 39

Figure 4.2. SVM input vector formed by multitemporal features .................................................. 39

Figure 4.3. The structure of the proposed contextual SEM algorithm ........................................... 42

Figure 4.4. The flowchart of the proposed rule-based model selection. ......................................... 45

Figure 4.5. Candidate neighbourhoods for adaptive MRF ............................................................. 46

Figure 4.6. Multiscale MPAC average estimation strategy. The class-conditional scattering

averages are extracted from adaptive neighbors within windows Wj,s at spatial scale s, centered on

pixel j. ............................................................................................................................................. 48

Figure 4.7. Texture enhancement scheme. ..................................................................................... 50

Figure 5.1. Mapping results of the object-based approach: (a) Google map. (b) mapping results by

SVM and rules ................................................................................................................................ 53

Figure 5.2. Comparison of various polarimetric SAR parameters on the ascending (right) and the

descending (left) data stack ............................................................................................................. 54

Figure 5.3. The performances of selected date combinations ........................................................ 55

Figure 5.4. The trends of the total change rate in iterations using various PolSAR distribution

models. ............................................................................................................................................ 56

Figure 5.5. Pauli image samples (a) and corresponding results by G0p model with the proposed

adaptive MRF using beta 1 with learning chontrol (b) and without learning control (c). ............... 57

Figure 5.6. Mapping accuracies by the MRF Beta from 1 to 25 with scale 1 (a) and Beta from 5 to

75 with scale 5 (b) using G0p model. ............................................................................................. 58

Figure 5.7. Pauli image (a) and corresponding results by G0p model with proposed adaptive MRF

(b), adaptive neighbor but non-adaptive MRF energy fuction (c), and non-adaptive neighbor (d). 58

Figure 5.8. Overall Accuracy and Kappa for the different scale combinations in the proposed

algorithm by MPAC and adaptive MRF. ......................................................................................... 59

Figure 5.9. Google map (a), Pauli image (b) and corresponding results by Wishart model with

only adaptive MRF (c), original MPAC (d), proposed MPAC+adaptive MRF (e) and

SVM+adaptive MRF (f). ................................................................................................................. 60

Figure 5.10. The user (left) and producer (right) accuracies of PolSAR models ........................... 61

Figure 5.11. Google map Quick-bird image (a), descending Pauli image (b) and corresponding

classification results by G0p (c), Kp (d), KummerU (e) and Wishart (f) models. .......................... 62


viii

Figure 5.12. Overall Accuracy and Kappa for the different dictionaries. G=G0p, K=Kp,

U=KummerU, W=Wishart. Unit is in percent................................................................................. 63

Figure 5.13. Improvement by the HD-Road rule with the dictionary-based approach .................. 64

Figure 5.14. Google map (a), PolSAR Pauli image (b), and results by the PolSAR (c) and C-HH

SAR (d) data with (1) and without (2) texture analysis. ................................................................. 66

Figure 5.15. Changes of the Producer accuracies (red) and User accuracies (green) by only

contextual SEM from using 13 classes to using 10 classes on the PolSAR data. The solid bar is for

the increase, the hollow one is for the decrease. ............................................................................. 67

Figure 5.16. Changes of the Producer accuracies (red) and User accuracies (green) from using

only contextual SEM to the hybrid approach with texture analysis and rule-based approach on the

PolSAR data. The solid bar is for the increase, the hollow one is for the decrease. ....................... 68

Figure 5.17. The producer (left) and user (right) accuracy of the hybrid approach using PolSAR

data of A2, A3, A4 and C-HH SAR data, object-based approach using PolSAR data of A2, A3, A4

and using all the PolSAR data. ........................................................................................................ 71

Figure 5.18. Google map (a), Pauli image (b) and corresponding classification results by

contextual approach with only adaptive MRF (c), contextual approach and model selection (d),

hybrid approach (e) and object-based approach (f). ........................................................................ 72


ix

LIST OF TABLES

Table 2.1. Selected texture distribution and corresponding SAR intensity distributions ............... 11

Table 2.2. Selected texture distribution and corresponding PolSAR distributions ......................... 12

Table 2.3. Selected major spaceborne PolSAR sensors .................................................................. 18

Table 3.1. RADARSAT-2 Fine-Quad-Polarimetric SAR Imagery ................................................. 36

Table 3.2. RADARSAT-2 Ultra-Fine HH-SAR Imagery ................................................................ 36

Table 4.1. Selected feature sets for comparisons ............................................................................ 40

Table 5.1. Comparison of the C-HH SAR and PolSAR data and the improvement by texture

analysis (Unit: Percent) ................................................................................................................... 65

Table 5.2. Confusion matrix of the results using the PolSAR data A2, A3, A4 and three C-HH

SAR data by the hybrid approach (Unit: Percent) ........................................................................... 68

Table 5.3. Comparison of the pixel-based, object-based and hybrid approaches (Unit: percent) ... 70


x

LIST OF PAPERS This thesis is based on the following papers, which are referred to in the text by their Roman numerals. I. Xin Niu, Yifang Ban, 2013, Multitemporal RADARSAT-2 Polarimetric SAR Data for Urban Land Cover Classification using an Object-based Support Vector Machine and a Rule-based Approach. International Journal of Remote Sensing, Vol. 34, pp.1-26. II. Xin Niu, Yifang Ban, 2012, An Adaptive Contextual SEM Algorithm for Urban Land Cover Mapping using Multitemporal High-resolution Polarimetric SAR Data. IEEE Journal of Selected Topics In Applied Earth Observations And Remote Sensing, Vol. 5, pp. 1129-1139. III. Xin Niu, Yifang Ban, 2012, A Novel Contextual Classification Algorithm for Multitemporal Polarimetric SAR Data, Submitted to IEEE Geoscience and Remote Sensing Letters. IV. Xin Niu, Yifang Ban, 2012, Multitemporal Polarimetric RADARSAT-2 SAR Data for Urban Land Cover Mapping Through a Dictionary-based and a Rule-based Model Selection in a Contextual SEM Algorithm. Canadian Journal of Remote Sensing (Conditionally accepted). V. Xin Niu, Yifang Ban, 2012, RADARSAT-2 Fine-beam Polarimetric and Ultra-fine Beam SAR Data for Urban Land Cover Mapping: Comparison and Synergy. Submitted to IEEE Transaction on Geoscience and Remote Sensing.


1

1. INTRODUCTION 1.1 Background Urban areas are some of the most dynamic regions in the world. In 2008, more than half of the world’s population resided in urban areas. With the accelerating process of worldwide urbanization, the entire world’s population growth over the next 30 years will be further absorbed by urban areas (UN Department of Economic and Social Affairs 2009). Fast growing of urban sprawl with increasing numbers of city dwellers leads to various socioeconomic and environmental problems, such as resource allocation, environmental protection, ecological crisis, etc (Benfield et al. 1999; Cieslewicz 2002; Zhang et al. 2011). Most of the problems caused by urbanization actually have references to space. Therefore, to devise efficient management policies and to prevent negative consequences in the development, spatial knowledge is required for rational urban planning. Given that urbanization is a continuous process, not only the current spatial status is of importance, but also its development through time. As remote sensing can provide accurate and detailed observation data at various spatial and temporal scales, it can be used as an efficient tool to explore land use/land cover (LULC) information (Ward et al. 2000). Within many remote sensing systems, Synthetic Aperture Radar (SAR) has long been recognized as an effective tool for urban analysis (Henderson and Xia 1997, 1998; Ban et al. 2010; Soergel 2010), since it is less affected by weather conditions or solar illumination in contrast to optical or infrared sensors (Hayes and Gough 2009). SAR polarization has long been recognized as a useful tool to discover additional information about observed targets such as geometry, orientation and physical properties through polarimetric diversity (Dong et al. 1997; Ainsworth et al. 2009). Considerable progress for LULC applications using PolSAR techniques has been made recently, when high resolution PolSAR data has become routinely available due to the launch of many advanced spaceborne SAR sensors such as TerraSAR-X and RADARSAT-2. PolSAR data has become an attractive data source for various applications including agriculture, sea-ice, forestry, hydrology, oceans, etc. (Zebker and van Zyl 1991; McNairn and Brisco 2004; Lee et al. 2004b; Lee and Pottier 2009a, 2009d; Moon et al. 2010). Urban mapping using PolSAR data has also received increased interest from researchers in the remote sensing community. (Xia 1996; Boerner et al.


2

1998; Pellizzeri 2003; Alberga 2007; Alberga et al. 2008; Ainsworth et al. 2008, 2009; Gao and Ban 2009; Li et al. 2010; Zhang et al. 2010; Ban and Niu 2011; Qi et al. 2012; Zhu et al. 2012) Nevertheless, most studies about urban mapping using SAR or PolSAR data are limited in identifying the urban extents (Stasolla and Gamba 2008; Esch et al. 2010, 2011; Gamba et al. 2011; Qi et al. 2012) or mapping very few urban classes (Dell'Acqua and Gamba 2006; Alberga 2007; Alberga et al. 2008; Yin et al. 2011; Zhu et al. 2012). Very few studies have focused on detailed urban mapping using high resolution PolSAR data. The reason for the difficulty in detailed urban mapping using polarimetric SAR data is, for example, the complexity of the urban environment. The urban environment is comprised of various natural and man-made objects with several kinds of materials, different orientations, various shapes and sizes, etc., which complicates the SAR image interpretation (Franceschetti et al. 2002; Margarit et al. 2010). Problems can also originate from the nature of polarimetric SAR imaging such as inherent SAR speckle or geometry distortions such as shadow and layover. Although promising results have been reported in recent urban mapping studies (Ban et al. 2010; Hu and Ban 2012; Ban and Jacob 2012), detailed urban mapping using high resolution PolSAR data is still a challenging task. Regarding the methodology for urban mapping, methods can be divided into pixel-based or object-based classifications. As a traditional classification method, pixel-based methods are commonly employed for low resolution SAR data with reasonable results (Dean and Smith 2003). However, for high resolution SAR data, pixel-based approaches are usually insufficient (Scheiewe et al. 2001; Shaban and Dikshit 2001) due to the increased spectral variance within the classes. The pixel-based approach often results in “pepper-salt” effects due to the high variance of the pixel values (Solaiman et al. 1998). To cope with the problems of pixel-based approaches, some contextual analysis approaches (Stuckens et al. 2000; Hubert-Moy et al. 2001) such as Markov Random Field (MRF) (Geman and Geman 1984) have been developed. These have been shown to enhance classifications. However, traditional contextual approaches often lead to “over averaged” results, where urban details may be lost. Object-based methods, on the other hand, can explore the contextual information in a more convenient way. Shape characters and inner statistics of segmented regions can be directly obtained as object features. By using object-based analysis, problems caused by increased variance of the pixel values within a class can be better handled. Therefore, object-based methods have become increasingly popular for


3

LULC mapping using SAR data with ever increasing spatial resolution (Blaschke 2010). Nevertheless, preparation of meaningful objects by proper segmentation is essential to the object-based analysis. However, ideal segmentation on urban areas using SAR data is often hard to achieve. From the view of data modeling, classification methods can also fall into parametric or non-parametric approaches (Richards 1993). For the parametric approaches, modeling of the classified data is often required. For that purpose, many PolSAR data distribution models have been proposed such as Wishart (Goodman 1963), Kp (Lee et al. 1994b), G0p (Freitas et al. 2005) and KummerU (Bombrun and Beaulieu 2008). However, the efficiency of those models in detailed urban mapping have not been fully evaluated and compared in terms of accuracy and time efficiency. To estimate the parameters involved in these PolSAR models, the Expectation Maximum (EM) algorithm (Kersten et al. 2005) is often applied. The Finite Mixture Model (FMM) has been found to be a useful tool (McLachlan and Peel 2000) to represent the heterogeneity of the class model. However, effective incorporation of the contextual analysis into the EM algorithm and mitigation of the degenerative behaviors when dealing with many classes in the EM iterations are still required to be investigated in detailed urban mapping. Moreover, how to explore the advantages of different PolSAR models for further improvement in a multi-model approach by FMM or other method is still challenging in the parametric urban analysis. On the other hand, non-parametric approaches do not require modeling of the classified data. Therefore, they are more convenient and flexible to be employed in different situations such as multitemporal or multi-source data fusion (Jain et al. 2000), where the joint data is difficult to model. Methods such as Artificial Neural-Network (ANN) (Azimi-Sadjadi et al. 1993; Pellizzeri et al. 2003; Ban and Wu 2005; Alberga et al. 2008; Del Frate et al. 2008) and Support Vector Machine (SVM) (Waske and Benediktsson 2007; Griffiths et al. 2010; Yin et al. 2011) have been successfully applied in urban mapping in many previous studies. Especially, SVM has received increasing popularity in remote sensing classifications (Pal and Mather 2005). Nevertheless, such non-parametric approaches are usually time-consuming and the optimal classifier parameters (e.g. the parameters of the kernel function in SVM) are often hard to estimate. The efficiency of multitemporal high resolution PolSAR data has not yet been sufficiently evaluated for detailed urban mapping. Therefore, evaluation of various PolSAR features for urban analysis and the development of efficient approaches for detailed urban mapping using


4

multitemporal high resolution PolSAR data are of great importance for remote sensing applications. 1.2 Research Objectives This research investigates multitemporal high resolution PolSAR data for detailed urban mapping. Specific research objectives include the following: (1) Assess the efficiency of different polarimetric SAR features and combinations of multitemporal data for detailed urban mapping. (2) Develop effective contextual approaches exploring spatio-temporal information to improve classification results. (3) Compare the capabilities of various PolSAR distribution models for describing urban classes, and to explore efficient ways to integrate multi-models for further improvement. (4) Evaluate the significance of the SAR polarizations in mapping of urban details in comparison to single polarization SAR data. In order to achieve the above objectives, multitemporal C-band RADARSAT-2 fine-beam PolSAR and ultra-fine-beam HH SAR data collected over the same area within the same year were used for evaluation. 1.3 Thesis Organization This thesis has been organized as follows. Chapter 1 gives an introduction of the research background and objectives. Chapter 2 reviews the principles of PolSAR imaging and urban mapping approaches using PolSAR data. Chapter 3 specifies the study area and the data for this research. Chapter 4 describes the methodologies proposed for detailed urban mapping using polarimetric SAR data. Chapter 5 provides the mapping results with the discussions related to the suggested approaches. Chapter 6 presents conclusions and future work. This thesis is based on the following papers, which are referred to in the text by their Roman numerals.


5

I. Xin Niu, Yifang Ban, 2013, Multitemporal RADARSAT-2 Polarimetric SAR Data for Urban Land Cover Classification using an Object-based Support Vector Machine and a Rule-based Approach. International Journal of Remote Sensing, Vol. 34, pp. 1-26. II. Xin Niu, Yifang Ban, 2012, An Adaptive Contextual SEM Algorithm for Urban Land Cover Mapping using Multitemporal High-resolution Polarimetric SAR Data. IEEE Journal of Selected Topics In Applied Earth Observations And Remote Sensing, Vol. 5, pp. 1129-1139. III. Xin Niu, Yifang Ban, 2012, A Novel Contextual Classification Algorithm for Multitemporal Polarimetric SAR Data, Submitted to IEEE Geoscience and Remote Sensing Letters. IV. Xin Niu, Yifang Ban, 2012, Multitemporal Polarimetric RADARSAT-2 SAR Data for Urban Land Cover Mapping Through a Dictionary-based and a Rule-based Model Selection in a Contextual SEM Algorithm. Canadian Journal of Remote Sensing (Conditionally accepted). V. Xin Niu, Yifang Ban, 2012, RADARSAT-2 Fine-beam Polarimetric and Ultra-fine Beam SAR Data for Urban Land Cover Mapping: Comparison and Synergy. Submitted to IEEE Transaction on Geoscience and Remote Sensing. The relationship between the listed papers is illustrated in Figure 1.1.

Figure 1.1. Organization of the selected papers

Object-based Pixel-based Hybrid

Model

Comparison

and Selection

(Paper IV)

Multitemporal

PolSAR

Feature

Comparisons

using SVM and

Rules (Paper I)

Fusion of

fine-beam

PolSAR and

ultra-fine-beam

HH SAR data

(Paper V)

Object-features and rule-based approach

Contextual approaches

Adaptive MRF

(Paper II)

Modified MPAC

(Paper III)

Contextual SEM with single model

Contextual SEM with multi-model


6

Paper I explored an object-based SVM and a rule-based approach for detailed urban mapping. Various polarimetric SAR features with different processing steps were compared by such an object-based approach. Efficiencies of alternative multitemporal combinations considering the orbital complement and temporal relationships were assessed as well. Paper II proposed a pixel-based classification approach for detailed urban mapping. Using an adaptive MRF and the spatial variant FMM, a contextual SEM algorithm has been proposed. Moreover, a learning control scheme has been introduced to mitigate the degenerative problem in the SEM algorithm. Common PolSAR distributions models have also been compared by such an approach. Paper III put forward another contextual approach combining a modified MPAC and an adaptive MRF in the contextual SEM algorithm proposed in Paper II. The effect of the multiscale local feature averaging scheme has been evaluated. Comparisons have also been made between the adaptive MRF, the modified MPAC, the proposed approach and the contextual SVM. Paper IV further extended the PolSAR distribution model comparison conducted in Paper II on the whole data set. According to the different behaviors of those PolSAR models, a rule-based model selection has been introduced into the contextual SEM algorithm given by Paper II to improve mapping accuracy. Traditional dictionary-based model selection has also been compared with such a rule-based approach. Paper V conducted a multitemporal comparison and fusion of fine-beam PolSAR data and ultra-fine-beam HH SAR data using the contextual SEM algorithm introduced in Paper II. Texture analysis has been further explored to improve the identification of urban areas. To extract roads, streets and parks, a pixel- and object-based fusion approach using the object-based shape features and rules as employed in Paper I has been applied on the pixel-based mapping results.


7

2. LITERATURE REVIEW This chapter contains a review of urban mapping using PolSAR data. The principles of PolSAR imagery, a brief introduction to spaceborne PolSAR systems, the characteristics of observing urban areas using SAR data, enhancement approaches for SAR image processing and urban mapping methodologies using PolSAR data are reviewed in the following sections. 2.1 SAR Polarimetry This section gives an overview of the fundamental properties of PolSAR imagery, including the PolSAR statistical properties and decomposition theorem. First, the SAR Polarimetry foundation is presented. Then, the SAR imaging principles and the data format are introduced. 2.1.1 SAR Polarimetry Fundamentals An imaging PolSAR system detects targets by sending directional pulses of electromagnetic waves and receives the values of the backscatterings. A PolSAR image is formed from the backscattered energy and the response time. The measured backscattering depends on both the system configuration (such as wave frequency, polarization, incidence angle and looking direction) and the target characteristics (such as geometrical and dielectric properties). In a full polarimetric SAR system, the transmitting and receiving antennas can be freely configured with two orthogonal polarization states. Therefore, the scattering properties of the targets can be revealed in four polarimetric channels which provide more information about the observed targets. To model the interactions between the polarized waves and the targets, the Jones vector has been defined to represent the electromagnetic field (Boerner et al. 1998):

n

e

e

mjm m

mn jn n

E EE

E E

φ

φ

= = , (2-1)

where Em and En are the two electromagnetic components in general under the two orthogonal polarization basis {um, un} (e.g. the linear horizontal and vertical polarization basis: {H, V}). By the Jones vector, the transformation between the transmitted electromagnetic wave to the received wave can be connected by the scattering matrix S (van Zyl et al.


8

1987) which contains the information to interpret the scattering mechanisms. Under the linear horizontal and vertical polarization basis (which are the most commonly used polarization basis for modern PolSAR systems. Such basis is also assumed in the following discussions through this thesis), such transformation can be given as follow:

ikr ikrh h hh hv h

v v vh vv vrec trans trans

E E S S Ee e

E E S S Ekr kr

= =

S (2-2)

In the Monostatic Arrangement (MSA) case (receiver and transmitter are co-located) the scattering matrix is also called the Sinclair matrix. Further, with the reciprocal medium assumption, the Sinclair matrix is symmetric, i.e. vh hvS S= (Boerner et al. 1998). The coefficients in the scattering matrix are complex-valued, and they represent the scattering information in an individual look of PolSAR observation. Therefore, the data format which contains the scattering coefficients is also referred to as single-look complex (SLC) data. The PolSAR cross section can be related to the scattering coefficients by the following equation:

2

qp qp=4 Sσ π (2-3)

To simply represent the scattering matrix, the scattering vectors can be defined as the scattering coefficients namely the ‘Lexicographic Feature vector’

[ ]4

T

hh hv vh vvS S S S=s (2-4)

and the ‘Pauli Feature vector’

( )4

1

2

T

hh vv vv hh hv vh hv vhS S S S S S j S S= + − + − k (2-5)

In the monostatic case, wherevh hvS S= , the corresponding scattering vectors could be defined as: the ‘Lexicographic Feature vector’

3 2T

hh hv vvS S S = s (2-6)

and the ‘Pauli Feature vector’

[ ]3

12

2

T

hh vv vv hh hvS S S S S= + −k (2-7)

In the PolSAR imagery, an averaging operation known as multilooking (Cumming and Wong 2005; Massonnet and Souyris 2008) is often applied to generate a multilook image. Multilooking can be done either in


9

the frequency domain in the PolSAR imaging or in the spatial domain after the image has been produced. The multilooking process generates complex data which is known as multilook complex (MLC) data. The MLC data can be represented by the covariance matrix which is defined by averaging the lexicographic feature vector samples is :

1

1

nL Hn i ii

L −=

= ∑C s s (2-8)

and the coherency matrix which is defined by averaging the Pauli feature vector samples ik :

1

1

nL Hn i ii

L −=

= ∑T k k (2-9)

where nL indicates the nominal number of looks, and (.)H is the Hermitian transposition operator. Sample averaging is done in the multilook cell. The coherency matrix T and covariance matrix C are similar matrices. They can be transformed to each other through linear transformation. They all contain the complete polarimetric information of the observed targets. In terms of statistical analyses, the generated results are also equally valid for both data formats. Therefore, the choice of the MLC data is restricted to the covariance matrixC for the following discussions about PolSAR statistical analyses. 2.1.2 Statistical Properties of PolSAR Data The PolSAR statistics are normally based on the analysis of the SAR speckle. Speckle is an inherent property in SAR observations due to the interference phenomenon of the coherent imaging process (Goodman 1976). The scattering process can be explained by a random walk model (Goodman 2007; Lopes et al. 2008), which describes the measured electromagnetic wave as the vector sum of the wave components by the scatters in the same resolution cell. Such a process leads to a pixel-to-pixel variation which is represented as a granular speckle pattern on the SAR image. Based on the random walk model, the PolSAR statistic models can be derived by further assumptions about the speckle effect, which will be described as follows:

Product model and speckle statistics with single polarization

The relationship between speckle and other SAR properties can be


10

analyzed by the well-known product model (Oliver and Quegan 2004; Frery et al. 1995; Arsenault and April 1976). A single polarization observation S and the corresponding intensity I can be formed as the product of two independent contributions: speckle and texture. Accordingly, such a product model could be formulated as (Oliver and Quegan 2004):

2

1 2= = ( + ), = , =s s sS T X T X jX I T P P X . (2-10)

where the scalar random variable sT represents the texture. In the context of SAR statistics, the term “texture” means the variation of the mean SAR reflectivity in contrast to the conception in the image processing, where “texture” describes the spatial distribution patterns of the pixel values. The texture is also called radar cross section (RCS) in certain situations (Ulaby et al. 1986; Lopes and Sery 1997). Texture indicates the ideal scattering of the target. X is a complex random variable representing the independent speckle. Given a large amount of independent scatter in the resolution cell with no dominating type, the scatterings’ phases are normally assumed to follow a uniform distribution between 0 and 2π . For single-look data, by the central limit theorem, X will follow a circular complex Gaussian distribution with zero mean and variance 0.5 for each component:

1 2

2 21, 2, 1 2

1( ) = exp(- - ) X Xf x x x x

π (2-11)

Accordingly, the speckle intensity P follows a negative exponent distribution:

1( ) = exp( ), ( >0)Pf p p p

π− (2-12)

For the multi-look SAR data, the speckle intensity P can be demonstrated to follow a Gamma distribution:

LL-1L

( | L) = exp(-L ), (L, >0)(L)Pf p p p p

Γ (2-13)

where L is the equivalent number of looks (ENL), and ( )Γ ⋅ is the standard Euler gamma function. With the product model and speckle distribution, various multi-look SAR intensity distribution models can be derived by assuming different texture distributions for sT as


11

demonstrated in Table 2.1. Table 2.1. Selected texture distribution and corresponding SAR intensity distributions Intensity (Texture) distributions

Intensity Probability Density Function (PDF) Ref

Gamma (Constant)

1( | , ) exp( )( )Zf z z z

ααλλ α λ

α−= −

Γ

Goodman 1975

K (Gamma)

( +L)/2-1

L

2L L z Lz( | , ) (2 ), ( , , 0)

(L) ( )Zf z K zα

αα α αµ α α µα µ µ µ−

= > Γ Γ

Oliver 1984

G0I (Inverse Gamma)

( )( )

L L-1

L

L (L ) ( 1)( | , ) , ( 1, 0)

(L) ( ) L ( 1)Z

zf z

z

λ

λ

λ λ µµ λ λ µ

λ λ µ +

Γ + −= > >

Γ Γ + −

Frery 1997

Fisher (Fisher-Snedecor)

1

( )( | , , ) , ( 1, 0)

( ) ( )1

l

Z l m

lz

ml m lf z l m l m

l m m lz

m

µµµ

µ

−

+

Γ + = > >

Γ Γ +

Tison et al. 2004

The texture parameters for K, G0I and Fisher models areα , µ ,λ , l and m . No texture parameter for Gamma model.

Polarimetric SAR statistics

By using the product model, the SLC Lexicographic feature vector can also be decomposed as (Freitas et al. 2005):

sT=s X , (2-14)

where T is a positive random variable representing texture. X is a complex random vector following a zero mean multivariate complex Gaussian distribution. Therefore, the MLC covariance matrix can be written as:

1 1

1 1

n nL LH Hn i i n si i ii i

L s s L T− −= =

= =∑ ∑C X X (2-15)

For the covariance matrix, if the sample interval is short in the multilooking process or the multilook window size is small, a same RCS or texture value can be assumed for all the samples in the multilook cell. Therefore, siT can be replaced by a constant cT . And the covariance can be further written as:

1

1

nL Hn c i ii

L T T−=

= =∑C X X W (2-16)


12

whereW is a random matrix following a scaled multivariate complex Wishart distribution (Goodman 1963; Lee et al. 1994b; Srivastava 1965). T is a normalized scalar random variable with unit mean ( 0,E{ } 1T T> = ), thus all scale information is put into W . Therefore, alternative PolSAR MLC data distribution models can be derived by assuming different normalized texture distributions for T . Table 2.2 presents some common PolSAR MLC distribution models formed in this way.

Table 2.2. Selected texture distribution and corresponding PolSAR distributions PolSAR (Texture) distributions

Probability Density Function (PDF) of the covariance matrix Ref

Wishart (Constant)

LL 1

L

L exp( L ( ))( | ,L)

(L)

dd

W

d

C Tr Cf C

− −− ∑∑ =

Γ ∑

Goodman 1963

Kp (Gamma)

L LL 12 2

1LL

2 (L ) ( ( ))( | , L, ) (2 L ( )), ( 0)

(L) ( )

d dd

K d

d

C Tr Cf C K Tr C

α α

α

αα α α

α

+ −− −

−−

∑∑ = ⋅ ∑ >

Γ ∑ Γ

Lee et al. 1994b

G0p (Inverse Gamma)

LL1 L

L

L (L )( 1)( | , L, ) (L ( ) 1) , ( 0)

(L) ( )

ddd

G

d

C df C Tr C

λλλ λ

λ λ λλ

−− − −Γ + −

∑ = ⋅ ∑ + − >Γ ∑ Γ

Freitas et al. 2005

KummerU (Fisher- Snedecor)

LLL

L

1

L ( )( | ,L, , ) ( ) (L )

( ) ( ) 1(L)

(L ,L 1,L ( ) / ( 1)) ,( 0, 0)

ddd

U

d

Cf C d

U d d Tr C

ξ ζ ξξ ζ ζξ ζ ζ

ζ ξ ξ ζ ξ ζ

−

−

Γ +∑ = Γ +

Γ Γ −Γ ∑

⋅ + − + ∑ − > >

Bombrun and Beaulieu 2008

The texture parameters for Kp, G0p and KummerU models areα , λ , ξ and ζ . No texture parameter for Wishart model.

Parameter estimation

When statistical models are used to describe SAR data, correct modeling is essential for successful utilization. For a given model, the correct model form depends on the involved parameters such as ENL and the texture parameters. The ENL is a parameter which describes the averaging degree of the SAR image with certain data transformations or sample averaging operations such as multilooking processes. For single polarization SAR data, estimators of the ENL have been defined in many ways (Oliver and Quegan 2004; Cumming and Wong 2005). However, most of them need to define a homogenous region where the speckle is fully developed and RCS is assumed to be constant. For PolSAR data, the ENL has traditionally been obtained as the averaged value of the estimates from each polarization channel (Frery et al. 2007; Lee et al. 1994b). However, the PolSAR information is not fully explored by such channel averaging


13

approaches. To this end, Anfinsen et al. (2009) have proposed two superior PolSAR ENL estimators using the covariance matrix through an automatic estimation scheme. As demonstrated in the previous subsections, SAR distribution models are based on different texture assumptions. The texture forms are applicable to alternative LULC types according to their homogeneity. For example, the Wishart model fits the homogenous classes, while Kp is suitable to heterogeneous classes, G0p and KummerU can be used for extremely heterogeneous classes. The texture parameters involved in those distribution models describe the homogeneity of the class. For single polarization data, traditionally, those parameters were obtained by the moment estimators (Lee et al. 1994b; Frery et al. 1997, 2007; Freitas

et al. 2005). Tison et al. (2004) proposed a log-moment method by which superior parameter estimation could be acheived. For the PolSAR data, the parameters were conventionally obtained by averaging the mono-pol estimators (Frery et al. 1997, 2007; Freitas et al. 2005). To fully explore the polarimetric information, Anfinsen and Eltoft (2011) have proposed the method of matrix log-cumulants (MoMLC) as a superior estimation approach for the PolSAR distributions. Besides, the MoMLC approach has also been used for the goodness-of-fit test of the applied PolSAR models (Anfinsen et al. 2011). 2.1.3 Polarimetric Decompositions Polarimetric decomposition can provide physical interpretations of the PolSAR observations such as scattering mechanisms or polarimetric properties. Parameters obtained by the decomposition methods can be directly used as classification features in non-parametric classifiers such as SVM and ANN. Decomposition methods can be mainly divided into two major categories according to the assumptions of the target types. For the “pure target”, the coherent decompositions are applicable, while for the “distributed target”, the incoherent decompositions are appropriate. The pure target is the one which can be directly represented by the single scattering matrix. On the contrary, the distributed target, which needs to be described by statistical averaging, is represented by the second order Hermitian covariance matrix C or the coherency matrix T (Boerner et al. 1998). Most natural targets with varying properties are treated as distributed targets, while man-made objects with stationary properties are usually regarded as pure targets. In the following subsections, selected common polarimetric decompositions based on the two target types will be briefly reviewed.


14

Coherent decompositions

The coherent decompositions are carried on the scattering matrix S, which is used to represent the polarimetric properties of the “pure target”. In this case, the scattering matrix S is decomposed into many scattering components representing the modeled scattering of typical objects with the common expression:

1

k

i ii

c=

=∑S S , (2-17)

where iS stands for the scattering from a simple object i and ic serves as the weight. Therefore, the coherent decomposition interprets the physical properties through analyzing the simple object model iS . Brief descriptions of selected coherent decompositions are given as follows: Pauli decomposition (Cloude and Pottier 1996): The scattering matrix S is interpreted by four scattering mechanisms, namely sphere surface, dihedral, di-plane oriented at 45 degrees and helix related. They are respectively modeled on the following Pauli basis:

[ ] 1 01

0 12a

=

S , (2-18)

[ ] 1 01

0 12b

= −

S , (2-19)

[ ] 0 11

1 02c

=

S , (2-20)

[ ] 0 11

1 02d

− =

S . (2-21)

In the monostatic case with the reciprocal assumption, where vh hvS S= , the

Pauli basis can be reduced to [ ] [ ] [ ]{ }, ,a b c

S S S . And the scattering

matrix can be decomposed as follows:

[ ] [ ] [ ]hh hv

a b cvh vv

S S

S Sα β γ

= = + +

S S S S , (2-22)

where


15

2hh vvS Sα += , (2-23)

2hh vvS Sβ −= , (2-24)

2 hvSγ = . (2-25) The Pauli decomposition is usually used to illustrate the PolSAR data with an image by a false color scheme. Krogager decomposition (Krogager 1992): The scattering matrix S is represented as the combination of a sphere, a di-plane and a helix scattering model. The last two components are described with an angleθ . Cameron decomposition (Cameron and Leung 1990): The S matrix is firstly decomposed into the non-reciprocal part [Snr] and reciprocal part [Srec]. The latter one will be further factorized into the max symmetric component [Smax]sym and min symmetric one [Smin]sym. The max symmetric part will be used to compare the referenced measures to match the identical objects.

Incoherent decompositions

For the incoherent targets, the decompositions are applied on the second order descriptors: covariance and coherency matrices. Therefore, these kinds of decompositions can be expressed as:

1

k

i ii

p=

=∑C C (2-26)

or

1

k

i ii

q=

=∑T T , (2-27)

where the factors iC or iT represent certain physical scattering mechanisms. Some selected decomposition methods of this kind are concisely reviewed as follow: The model-based Freeman decomposition (Freeman and Durden 1998): The freeman decomposition is based on modeling which considers the covariance (or coherency) matrix as a superposition of three scattering mechanisms, namely “Volume scattering” (i.e. forest canopy), “Double-bounce scattering” (i.e. dihedral corner reflector) and “Surface or single-bounce scattering” (i.e. first-order Bragg surface scatterer). For example, in the reciprocal monostatic situation, the above mentioned


16

scattering models can be described by the following matrices (the initials of the model names are the subscripts):

[ ]3

1 0 1/ 3

0 2 / 3 0

1/ 3 0 1v

=

C , (2-28)

[ ]2

3

0

0 0 0

0 1d

α α

α ∗

=

C , (2-29)

[ ]2

3

0

0 0 0

0 1s

β α

α ∗

=

C . (2-30)

Therefore, the reciprocal covariance matrix 3C can be decomposed as:

[ ] [ ] [ ]3 3 3 3v d sv d sf f f= + +C C C C . (2-31)

Phenomenological Huynen decomposition (Huynen 1970): The basic idea of the Huynen decomposition is to divide the coherency matrix T into two parts which are the effective pure target T0 and the distributed N-target TN residue which represents nonsymmetric target parameters. The Huynen decomposition is used to explain the physical property and the geometry of the observed targets. Eigenvector based decomposition: The decomposition is based on the analysis of the eigenvectors and eigenvalues of the second order statistics. For the reciprocal coherency matrix, the eigen decomposition can be expressed as

3*

31

T Ti i i

i

λ µ µ=

=∑ , (2-32)

where iλ and iµ denote the i-th eigenvalue and eigenvector respectively. The eigenvector can be expressed as

[cos sin cos sin cos ]i ij j Ti i i i i ie eδ γµ α α β α β= . (2-33)

Based on the eigen decomposition of the reciprocal coherency matrix,


17

Cloude and Pottier (1997) have proposed a / /H A α decomposition with the following three parameters:

Entropy: 3

31

log ( )i ii

H p p=

= −∑ , (2-34)

where

3

1

ii

kk

pλ

λ=

=∑

, (2-35)

Anisotropy: 2 3

2 3

Aλ λλ λ

−=+

, (2-36)

Mean alpha angle: 3

1i i

i

pα α=

=∑ , (2-37)

where the mean alpha angle α indicates the main scattering mechanism (single-bounce scattering, volume scattering or double-bounce scattering, etc). The entropy H measures the degree of statistical disorder of each target. When H is low, one dominant scattering matrix corresponding to the largest eigenvalue can be extracted and the other eigenvalue components can be ignored. While the H is high, a mixture of different type of scatters must be considered. The anisotropy A, which is a parameter complementary to the polarimetric entropy H, gives the relative importance of the second and the third eigenvalues. However, it is most meaningful to use it while H >0.7 (Lee and Pottier 2009c). Biases of the Entropy/Alpha/Anisotropy parameters were evaluated by Lee et al. (2008) with a removal approach. 2.2 Spaceborne PolSAR Systems As an active remote sensor, the SAR system uses electromagnetic waves to illuminate the observed targets. In contrast to optical and infrared sensors, SAR is less affected by weather conditions or solar illumination (Hayes and Gough 2009). SAR has the capacity to penetrate a short distance into the ground. Among various SAR features, the SAR polarizations have long been recognized as a useful tool to study the backscattering behaviors of targets caused by their geometry, orientation and physical properties (Dong et al. 1997; Ainsworth et al. 2009). Polarization theory has been investigated since the middle of the nineteenth century. After World War II, classic radar polarization theory was established (Hubbert 1994; Zebker and van Zyl 1991; Touzi et al. 2004). In 1951, Carl Wiley of Goodyear discovered the Doppler


18

beam-sharpening, which can be considered as the birth of the SAR concept. Although SAR and radar polarization theory had been developed in the following years, rapid progress for the practical applications of the PolSAR systems was not made until Poelman (1981) proposed the concept of “virtual polarization”. This implies that, without changing the configuration of a PolSAR system, scattering information for other polarizations can be obtained. The complexity and cost of the PolSAR system has since significantly decreased. In 1985, the first airborne PolSAR sensor NASA CV990 PolSAR was created and tested for real use. This led to an upsurge of PolSAR applications (van Zyl et al. 1987; Zebker and van Zyl 1991). In one of these studies, the first terrain LULC classification using PolSAR was made by Kong (1988). The evolution of spaceborne SAR systems can be traced back to 1978, when the L-band SEASAT-A SAR sensor was launched by the USA. Since 1990, spaceborne PolSAR systems have been increasingly used for earth observation. With the launch of advanced spaceborne PolSAR sensors like RADARSAT-2 and TerraSAR-X, global observation using high-resolution PolSAR has become routinely available. Some spaceborne SAR sensors are listed in Table 2.3.

Table 2.3. Selected major spaceborne PolSAR sensors

Sensor Band Polarization Range x Azimuth Resolution (m x m)

Launch Year

Country

SIR-C/X-SAR L /C/X

Full polarization X: VV

13~26×30 1994 US/Germany

/Italy ENVISAT ASAR C HH/HV HH/VV

VH/VV

9~450×6~450 2002 ESA

ALOS PALSAR L Full polarization HH/HV VH/VV

Range: 7 ~89 2006 Japan

RADARSAT-2 SAR

C Full polarization 3~100×3~100 2007 Canada

TerraSAR-X X Full polarization Dual polarization

1~16×0.6~3.5 2007 Germany

COSMO-SkyMed X Dual polarization 1~100×1~100 2007 Italy

Although SAR systems have been used for civilian geoscience research since the late 1950s, less attention was paid to urban applications in the early years (Henderson and Xia 1998). Since SAR data was made commercially available in 1969, the potential of PolSAR systems for urban mapping has been reported in a growing number of studies. Rapid progress in urban applications using PolSAR has been made in recent years with increased amounts of high quality high resolution PolSAR data collected through spaceborne PolSAR systems. Typical urban mapping


19

using SAR data will be reviewed in the following sections. 2.3 SAR for Urban Analysis Urban area, which contains a mixture of man-made structures and natural objects, represents one of the most complicated areas to be studied through remote sensing. Particular characteristics of SAR observations in urban areas are reviewed in the following sections. 2.3.1 Interactions between the SAR Systems and Urban

Features The geometric distortions in SAR images are commonly introduced by man-made structures (Soergel et al. 2002; Brunner et al. 2010). Since SAR measures the distance between the target and the sensor, the top of the building will be detected first and recorded on the range profile, followed by the base of the building, which creates a reversed view of the facade as illustrated in Figure 2.1. This phenomenon, called “layover”, is more pronounced for high buildings. The ground obscured by the building from electromagnetic waves will leave a “shadow” in SAR images. The geometric distortions vary according to the looking direction and incidence angle which makes it difficult to fuse SAR data collected in different situations.

Figure 2.1. Example of the scattering profile of a simple flat-roof building model.

ground scattering a, double bounce b, scattering from vertical wall c, backscattering from roof d, shadow area e. The gray values of the backscattering profile correspond

to the relative amplitudes. (Brunner et al. 2010) The scattering mechanisms in urban areas are complex. As shown in Figure 2.1, the recorded observations are summed by the scatterings from the targets on the same wave front. For example, the layover part is mixed with the scatterings from the roof, wall and ground. Dong et al. (1997) concluded that the backscattering is dominated by single bounce

shadow

a a+c+d

b

d

e a

a b

c d

e a

ampl

itude

layover

double-bounce corner reflect


20

from roofs, and double bounce from ground-wall structures. Similarly, roads that go through dense built-up areas will contain both single and double bounce scattering. A mixture of volume scattering by vegetation and double-bounce scattering from buildings can also be observed in low density areas. Looking directions and the alignment of structures also influence SAR scattering. Man-made structures which are arranged perpendicularly to the illumination direction give rise to the power of echoes. This is called the cardinal effect (Raney, 1998) and is more evident in low frequency SAR data. 2.3.1.1 Frequency Frequency affects SAR observations in urban areas in many different ways (Xia 1996). Although the high frequency SAR system (short wavelength) is more sensitive to surface roughness, it is more influenced by atmospheric attenuation. Because of its low penetration ability, high frequency SAR systems are prone to observe more volume scattering from vegetation canopies, which can introduce confusion with man-made objects. Bryan (1975) reported that X-band data was better than the L-band data for interpreting urban features. The same observation has also been made by Haack (1984) when evaluating L- and X-band data for urban LULC identification. Henderson and Anuta (1980) used X-band, Ka-band and SEASAT HH-SAR data to study the effects of SAR frequency for urban area detection. Many other studies have also demonstrated the use of the Ka-band (Lewis 1968; Moore 1969; Peterson 1969), X-band (Dowman and Morris 1982; Esch et al. 2010, 2011), C-band (Dell'Acqua et al. 2003; Gao and Ban 2009; Hu and Ban 2012), L-band (Alberga 2007; Alberga et al. 2008; Zhu et al. 2012) and P-band (Qi and Jin 2007) for urban LULC mapping. Fusion of multifrequency SAR data has been found useful to improve the urban mapping accuracy (Lombardo et al. 2003; Pellizzeri et al. 2003; Sarker et al. 2008). 2.3.1.2 Polarization A study of SAR polarization for urban analysis was first reported by Lewis et al. (1969) who argued that cross-polarization could enhance the identification of commercial and industrial areas. Mapping of vegetation in urban areas was more successful with HH-polarized images. Moore (1969) found that the linear features of urban areas could be well identified by HV-polarized data. Wu and Sader (1987) demonstrated that the VH-polarized data and ratioed data (VH/HH, VH/VV) were better for separating urban and surrounding features than HH/VV-polarized data.


21

Dowman and Morris (1982) found that more built-up areas could be mapped by X-HH data than by X-HV data. Haack (1984) recorded the performance ranking for urban LULC delineations in descending order as X-HH, L-HH, X-HV and L-HV. Henderson (1979) noted that the cardinal effect in HV polarization is not as obvious as in HH polarization. Nevertheless, Wu (1984) demonstrated that the simultaneous use of multi-polarizations could better identify urban LULC than single polarization HH data. Henderson and Mogilski (1987) studied the ability of grey tone values from two polarizations to separate many man-made cultural features and natural terrain features. Ainsworth et al. (2009) compared the efficiency of dual-polarization data and quad-polarization data for urban classification. The polarimetric features such as Freeman decomposition parameters (Freeman and Durden 1998) and / /H A α decomposition parameters (Cloude and Pottier 1997) derived from the full-polarization data has been found useful for urban LULC analyses (Pellizzeri 2003; Lee et al. 2004a; Park and Moon 2007; Zhang et al. 2008; Gao and Ban 2009; Ban and Niu 2011; Bhattacharya and Touzi. 2011). The efficiency of those polarimetric features for LULC mapping was compared by Alberga (2007) and Alberga et al. (2008) using pixel-based approaches. Cloude and Pottier (1997) demonstrated that urban areas have medium polarimetric entropy, which indicates a medium degree of depolarization with multiple scattering mechanisms. Polarimetric characteristics of urban targets have also been studied by Margarit et al. (2010) from a geometric point of view. 2.3.1.3 Incidence Angle and Look Direction As discussed in the previous section, incidence angle and look direction greatly influence SAR images in urban areas. With regard to the optimal SAR incidence angle for urban applications, Lichtenegger et al. (1991) reported that SIR-A data with larger incidence angles are better than Seasat data for LULC mapping. Henderson (1990) studied three SIR-B images with different incidence angles and claimed that incidence angles less than 20-33 degrees were of minimum use for urban analysis. He found that interpretation accuracy increased at 41 degrees and decreased at 51 degrees. It was reported by Hussin (1995) that incidence angles around 40 degrees are suitable for urban analyses. However, the optimal angle varies with application. Although it poses problems for data fusion, multi-angle and multi-direction look SAR observations are valuable for urban analyses (Xia 1996). Multiangular fusion could balance the resolutions (which are higher with larger angles) and the man-made features like shadows and layovers (which are more prominent with smaller angles). With multi-angle or multi-direction look data fusion,


22

certain man-made features can be effectively separated from natural land-cover classes due to the varying representations of urban structures in SAR images. Dell'Acqua et al. (2003) showed that more streets could be extracted using simulated multi-angle data. Few studies on the fusion of multi-angle (Ban and Jacob 2012) or multi-direction look SAR exists for detailed urban mapping. 2.3.2 Spatial Resolution SAR spatial resolution is significant for urban applications. Due to limitations in spatial resolution, most of the previous urban LULC studies (Dell'Acqua et al. 2003; Lombardo et al. 2003; Pellizzeri et al. 2003; Lee et al. 2004a; Park and Moon 2007; Alberga et al. 2008) have focused on the detection of settlements or the mapping of very few urban features in SAR images. As one of the major improvements of modern spaceborne SAR sensors, increased resolution provides the opportunity to explore urban details. The advent of very high resolution SAR data with approximately one meter spatial resolution has the potential for more complex urban analysis (Soergel et al. 2005) such as the extraction of road networks (Lisini et al. 2006; Negri et al. 2006; Hedman et al. 2010) and reconstruction of man-made structures (Quartulli and Datcu 2004; Xu and Jin 2007). The high resolution or very high resolution SAR data also pose challenges for urban LULC mapping. The increased resolution also increases the scattering variance within classes. The computation time has greatly increased, especially for complicated algorithms. Most of the urban mapping using high resolution or very high resolution SAR data is concerned with the extraction of urban extents (Stasolla and Gamba 2008; Esch et al. 2010, 2011; Gamba et al. 2011) or the mapping of a few urban features (Chaabouni-Chouayakh and Datcu 2010a, 2010b; Yin et al. 2011). Few attempts have been made so far for detailed urban mapping (Ban and Niu 2011; Hu and Ban 2011) using multitemporal high resolution SAR data. 2.3.3 Multitemporal Observations Multitemporal data have been proven valuable to improve LULC classification (Lu and Weng 2007). Multitemporal data leads to improved urban mapping using SAR data, since most man-made structures are relatively stable across time in terms of scattering compared to natural classes. Moisture content which varies seasonally due to different weather conditions can also influence the SAR images collected at different dates. Increased moisture content will increase the SAR reflectivity (and image


23

brightness). For instance, soil surface and vegetation cover appear brighter in SAR images when they are wet (CCRS 2007). Dell'Acqua et al. (2003) stressed that an accurate characterization of stable features in urban areas is possible using sufficiently long temporal series of SAR data. Pellizzeri et al. (2003) suggested using long sequences of data to further improve urban feature characterization. Efficiency of multi-date PolSAR data combinations for urban mapping was first evaluated by Ban and Niu (2011). Ban and Jacob (2012) demonstrated fusion of 4-date SAR data and optical data with suitable temporal compositions can achieve similar urban mapping accuracy as fusion of 8-date SAR data and optical data. Zhu et al. (2012) have conducted a study on detailed urban mapping using nine dates of dual polarization data. The efficiency of the date combinations has not yet been extensively studied for multitemporal PolSAR data. 2.4 Enhancement Approaches SAR images have unique characteristics. To reduce the adverse effects and effectively explore the information contained in the SAR data for classification, certain SAR enhancement approaches should be considered for successful urban mapping. 2.4.1 Speckle Reduction Speckle is inherent in SAR images due to the coherent interference of the backscattering waves from the targets within a resolution cell. Strictly speaking, SAR speckle is not real noise but a kind of noise-like phenomenon. Speckle can for instance be used in SAR interferometry (Oliver and Quegan 2004). However, in most situations, speckle makes the SAR image segmentation and classification difficult. One conventional way to reduce SAR speckle is multilooking (Cumming and Wong 2005; Massonnet and Souyris 2008) which can be done either in the frequency domain, by averaging the sub-aperture images, or in the spatial domain, using an averaging window on the SAR image. However, the multilooking approach reduces the resolution of the SAR image. Many speckle filtering approaches have been proposed (Lee et al. 1994a; Touzi 2002; Lee and Pottier 2009b). Most of theses filters have been built based on the well known multiplicative or product speckle models using local statistics (Touzi 2002; Lee and Pottier 2009b). For single channel SAR filtering, classic filters such as the ones proposed by Lee (1980), Kuan et al. (1985), Frost et al. (1982) and Lopes et al. (1990a) have been


24

widely used. However, these conventional filters usually assume a fixed local analyzing structure which may lead to over-averaged results. In certain situations, the multiplicative model may not properly fit the SAR speckle. To this end, many advanced filters have been proposed such as the enhanced version of Lee, Frost and Kuan filters (Lopes et al. 1990b) and the adaptive gamma MAP filters (Baraldi and Parmiggiani 1995). Many wavelet-based filters (Aiazzi et al. 1998; Sveinsson and Atli Benediktsson 2003; Gleich and Datcu 2009) have also been proposed as they benefit from the properties of the wavelet transformation such as low entropy and multiresolution. Some non local approaches (Chen et al. 2011; Parrilli et al. 2012) such as the non-local means denoising algorithm (Buades et al. 2005; Mahmoudi and Sapiro 2005; Deledalle et al. 2009; Zhang and Zhang 2011; Zhong et al. 2011), which takes into account both local patches and similar nonlocal patches all over the image in the filtering process, has also been proposed with promising results. For PolSAR data filtering, early studies (Lee et al. 1991; Goze and Lopes 1993; Touzi and Lopes 1994; Lopes and Sery 1997) have mainly explored the correlations between the different polarization channels for despeckling such as the polarimetric whitening filter (Novak and Burl 1990; Novak et al. 1993). However, these filters introduce total correlation between polarization channels. The introduced cross talk between channels makes it difficult to preserve the polarimetric properties. To preserve the polarimetric information, cross talk should be avoided in the filtering. Therefore, the elements in the coherency or covariance matrix should be equally and independently filtered (Lee et al. 1999). Based on these PolSAR data filtering principles, advanced PolSAR filters (Lee et al. 1999, 2006; Lopez-Martinez and Fabregas 2003, 2008) have been proposed. They can reduce speckle and preserve the polarimetric information and spatial details. These filters can be used to further enhance segmentations or classifications. 2.4.2 Texture Analysis Image texture describes the pattern of the arrangement of color or intensities in space (Shapiro and Stockman 2001). Different LULC classes such as rocks, sea-ice, seawater, vegetation, urban areas, etc. can be characterized based on their distinct textural features (Kandaswamy et al. 2005). Since urbans have regular spatial patterns, unique urban textures can be extracted to, for instance, separate settlements from rural areas. Texture differences also exist among various urban types (Henderson and Xia, 1998). The texture analysis was first employed for urban applications by Fasler


25

(1980) using SEASAT L-band HH polarization data. Dekker (2003) demonstrated that texture analysis can be performed to identify the urban and industrial areas in medium resolution SAR data. Dell’Acqua and Gamba (2003) investigated the urban environments according to the built-up density using the Grey Level Co-occurrence Matrix (GLCM) (Haralick 1973). The GLCM texture has also been studied by Del Frate et al. (2008), Ban and Wu (2005) and Dell’Acqua and Gamba (2006) for various urban types such as low density and high density residential areas, industry and asphalt. Hu and Ban (2012) used the ultra-fine beam high resolution C-HH SAR GLCM as object features for detailed urban mapping. The efficiency of the GLCM for SAR data analysis has been compared with MRF by Clausi and Yue (2004). They found that the GLCM improved the discrimination ability of the classes with decreased window size compared to MRF, although the GLCM was found to be more sensitive to texture boundary confusion than MRF. In addition to the GLCM, many other texture measures have been studied in urban analysis (Dekker 2003). Although the semivariogram has been previously applied in remote sensing (Woodcock et al. 1988), its application in urban classification of SAR data is relatively new (Dekker 2003; Chamundeeswari et al. 2009). The efficiency of the semivariogram has also been evaluated and compared with GLCM by Carr and Miranda (1998) for SAR and optical data classification. Stasolla and Gamba (2008) employed spatial local indices (Sokal and Thomson 2006) for the extraction of human settlements from high resolution SAR data. Gamba et al. (2011) further extended this approach by combining spatial indices with GLCM for robust urban extraction from high resolution and very high resolution SAR data. A set of contextual descriptors based on 2-D Fourier spectra of SAR images has also been proposed by Popescu et al. (2012) for analyzing very high resolution SAR data. Other texture measures such as Gabor wavelets (Manjunath & Ma 1996) and wavelet packets (Pun & Lee 2003) are also popular in SAR image analysis. Nevertheless, challenges still remain on how to efficiently employ SAR textures in detailed urban mapping. 2.4.3 Contextual Analysis Contextual information has the potential to improve classifications (Flygare 1997; Stuckens et al. 2000; Hubert-Moy et al. 2001). Generally the concept of context covers both spatial and temporal relationships in the data. In the discussion below, we restrict context to spatial relationships only. Contextual analysis is used to improve the image classification by exploring spatial information available in the neighboring graphical elements (Lu and Weng 2007). The spatial


26

information can be explored using either object-based (Kartikeyan et al. 1998; Flanders et al. 2003; Benz et al. 2004) or pixel-based approaches (Kartikeyan et al. 1994; Stuckens et al. 2000). In object-based analysis, the contextual information is inherent as the object features. The local autocorrelations of neighboring pixels are implicitly represented by the result of the segmentation. However, as discussed in the previous paragraphs, the results of the object-based analysis strongly depend on the segmentations. Some object-based approaches will be reviewed in the following subsections. The pixel-based contextual analysis is reviewed below. Many problems in pixel-based approaches can be effectively reduced through the use of contextual analysis (Lu and Weng 2007). For example, confusion caused by intra-class variations can be mitigated by analyzing the spatial autocorrelation (Keuchel et al. 2003; Magnussen et al. 2004). Contextual analysis has also been considered as a useful way to cope with the small sample problem (Baraldi et al. 2006). The contextual classification approaches have a smoothing effect which can remove “Salt and Pepper” noise in the final classification maps (Keuchel et al. 2003; Magnussen et al. 2004). Contextual information is usually explored through model-based approaches. Models like Markov Random Field (MRF) (Geman and Geman 1984) based on the Bayesian framework with Markovian dependency assumption have been widely used in SAR image analysis. Methods based on MRF models such as the iterated conditional modes (ICM) (Cortijo and de la Blanca 1998; Magnussen et al. 2004) are some of the most frequently used in contextual analysis (Lu and Weng 2007). Urban mapping with SAR data through MRF analysis has previously been used (Solberg et al. 1996; Tison et al. 2004; Wu et al. 2008; Voisin et al. 2013; Kayabol and Zerubia 2013). However, traditional MRF analysis using stable neighborhood structures and fixed contextual parameters usually lead to over-averaged results, where detailed urban features disappear. To solve this problem, several adaptive MRF approaches (Smits and Dellepiane 1997; Garzelli 1999; Deng and Clausi 2005; Maillard et al. 2005; Yu and Clausi 2006; Zhong and Wang 2009; Lei et al. 2010) that define an adaptive neighborhood with adaptive parameters have been proposed for SAR data with encouraging results. Hierarchical MRFs have also been proposed for accurate and robust SAR data segmentations (Yang et al. 2006; Liu et al. 2011; Martinis et al. 2011; Voisin et al. 2013). Conditional Random Field (CRF) (Lafferty et al. 2001), a variant of MRF which impose directional dependency in Bayesian Network has recently been proposed with promising results (Su


27

et al. 2011; Yang et al. 2011; Wegner et al. 2011). However, due to computational complexity, it is difficult to apply CRF to high resolution SAR images at the pixel-level. Although great advances have been made for SAR applications by the model-based contextual approach, efficient methods for detailed urban mapping using PolSAR data need to be developed further. Instead of modeling the prior probability for the classified pixels through MRF analysis, the contextual information can be explored to represent the varying class characteristics using approaches like Pappas adaptive clustering (PAC) (Pappas 1992). Baraldi et al. (2000) have proposed a modified PAC (MPAC) approach with a multiscale analysis scheme, where estimations from neighborhoods within windows of different size and the global range are considered. Baraldi et al. (2006) have refined the MPAC with soft averaging techniques for local averaged class features. Nevertheless, the PAC approaches assume slow variation in the studied images, where the scattering texture is negligible. The PAC approaches have been applied to single date optical data or single-polarization SAR data assuming a Gaussian distribution. A test for detailed urban mapping using multitemporal PolSAR data is still required. 2.5 Methodology for Urban Classification using PolSAR In the following subsections, state-of-the-art urban classification methods will be reviewed with respect to the classified graphic elements and the modeling of the data. 2.5.1 Pixel- and Object-based Classifications Mapping methods can be defined as pixel-based or object-based according to whether the basic graphic elements are either individual pixels or segmented objects. In many previous LULC studies, pixel-based approaches have been used as traditional mapping methods with satisfactory accuracy especially for coarse resolution SAR data (Dean and Smith 2003). However, when dealing with high resolution SAR data, pixel-by-pixel classification is usually limited, mainly due to the increased spectral variance within the classes (Shaban and Dikshit 2001). Scheiewe et al. (2001) have argued that the pure pixel-based methods are not suitable for classification of very high resolution data, since the high spatial resolution increases the spectral variability, which can increase the confusion between the classes. Another problem is the so called “Salt and Pepper” effect (Solaiman et al. 1998) in pixel-based results. This effect is


28

more pronounced for applications using SAR data due to the inherent speckle. On the other hand, object-based approaches which use the contextual information from meaningful objects and the relationships between them have been increasingly used for high resolution data (Blaschke 2010). Beside spectral information, shape features of objects and their spatial relationships can be explored through object analysis (Flanders et al. 2003; Benz et al. 2004). Although some comparisons between pixel-based and object-based classifications have been made for urban LULC mapping using optical data (Cleve et al. 2008; Zhou et al. 2009; Lu et al. 2010), very few comparisons have been made using SAR data. Sarker et al. (2008) found that pixel-based approachs have advantages when used with coarse resolution SAR data. There are a few studies that have fused pixel-based and object-based methods (Shackelford and Davis 2003; Wang et al. 2004; Chaabouni-Chouayakh and Datcu 2010a). However, most of these studies have focused on optical data and have used pixel-based results to improve object-based mapping. There are few studies that have used object-based features to improve pixel-based results. The pixel-based classifications can be sorted into different groups based on different views. This will be further discussed in the next subsection. The review below mainly focuses on object-based classifications. The concept of object-based classification for remote sensing applications can be traced back to the 1970s, when Kettig and Landgrebe (1976) introduced the method called extraction and classification of homogeneous objects (ECHO). Segmentation has played an important role in the development of the object-based image analysis. Whether meaningful objects can be generated is essential for the successful employment of object-based classifications. The existing segmentation techniques used for SAR images can be divided into the following groups: 1) clustering-based (Smith 1996; Quan et al. 2008), 2) edge-based (Fjørtoft et al. 1998; Germain and Refregier 2001), 3) region-based (Li et al. 1999; Galland et al. 2003; Ayed et al. 2005), and 4) hybrid approaches (Hermes and Buhmann 2004; Carvalho et al. 2009; Yu et al. 2012). These segmentation techniques have been extensively discussed (Fu and Mui 1981; Haralick and Shapiro 1985; Pal and Pal 1993; Blaschke 2010). Of all existing segmentation algorithms, the region-based multi-resolution segmentation (Baatz and Schape 2000) is one of the more popular (Blaschke 2010) in remote sensing applications. Some studies have also found that better segmentation can be obtained in SAR images by integrating edge-based and region-based approaches (Hermes and Buhmann 2004; Yu et al. 2012). However, most of the previous segmentation techniques have been applied on coarse resolution SAR


29

data. Efficient segmentation approaches for analysis of high resolution or very high resolution SAR data over urban areas still needs to be developed. Although the object-based analysis has been employed in remote sensing for some time, most object-based urban applications were designed to process optical data. Only a few object-based studies over urban areas using SAR data exist. For instance, Benz and Pottier (2001) classified man-made structures, forest and vegetation types using polarimetric SAR data through analyzing the Cloude-Pottier decomposition parameters within objects. Corr et al. (2003) fused polarimetric and interferometric SAR data for object-based analysis in urban areas. Shape features and other object properties have been shown to be useful to reduce the confusion between different classes. Esch et al. (2005) developed a robust object-based approach to extract built-up areas using E-SAR data from three complete flight tracks. Esch et al. (2010) further explored the object-based approach to delineate the urban footprints from TerraSAR-X data. Thiel et al. (2008) applied an object-based approach to analyze speckle statistics detection of urban extents in TerraSAR-X data. Ban et al. (2010) proposed an object-based fusion approach of SAR and optical data for urban mapping. Qi et al. (2012) constructed a decision tree from PolSAR features for object-based classification of urban areas. Hu and Ban (2012) investigated the potential of the very high resolution SAR data for urban mapping using the object-based approach. So far, most studies have focused on mapping urban extents or a few urban classes. There is still a lack of studies on detailed urban mapping using high resolution PolSAR data. 2.5.2 Parametric and Non-parametric Classifications From the point of view of statistical modeling, classification methods can be defined as parametric approaches which usually assume a distribution model of the data. Non-parametric approaches can be defined as the opposite (Richards 1993). These concepts are mainly related to supervised classification, which is discussed below. Parametric classifiers such as maximum likelihood classifiers (MLC) (Ban 2003; Pellizzeri et al. 2003; Ban and Wu 2005; Li et al. 2010) and parametric minimum-distance classifiers (MDC) (Alberga 2007; Mishra et al. 2011) have long been used in remote sensing (Swain and Davis 1978). Normally, for the selected distribution models such as the SAR distributions introduced in the previous sections, the involved model parameters are determined from the training samples. In order to properly


30

estimate the parameters a sufficient number of training samples are required to ensure satisfactory performance (Hubert-Moy et al. 2001). However, in mapping using remote sensing images, it can be difficult to obtain enough training samples (especially for detailed LULC mapping). One reason is that the high feature dimensions due to the use of multitemporal and multisource data require more samples to accurately represent the class models (Jackson and Landgrebe 2001). Another reason is that spatial autocorrelation can reduce the informativeness of neighboring pixels by violating the assumption of sample independence (Congalton and Green 2008). To mitigate the small sample problem, some semi-supervised classifications such as the Expectation- Maximization (EM) algorithm (Kersten et al. 2005; Khan and et al. 2007; Jackson and Landgrebe 2001) have been proposed in remote sensing. The EM algorithm iteratively estimates the distribution model parameters from both the training data and the classified data. To avoid local extreme and complex pseudo-likelihood functions, the Stochastic EM (SEM) algorithm (Celeux and Diebolt 1985; Moser et al. 2006) has been proposed. It has been found that the SEM algorithm can provide more reliable convergence with less iteration than the EM algorithm (Dias and Wedel 2004). Finite Mixture Model (FMM) (McLachlan and Peel 2000) has been introduced into the parametric analysis of the SAR data (Moser et al. 2006; Krylov et al. 2011a, 2011b) to represent the heterogeneity of the observations expressed in a finite number of latent classes. To incorporate the contextual analysis, a Spatially Variant FMM (SVFMM) (Sanjay-Gopal and Hebert 1998; Celeux et al. 2003) has been proposed. Urban analysis using the EM algorithm and FMM can be found in very few studies (Moser et al. 2006; Krylov et al. 2011a, 2011b). These studies have dealt with very few urban classes. Since without any control scheme, the EM algorithm tends to converge to a degenerate solution (Ingrassia and Rocci 2011) especially in applications where many classes are involved. Non-parametric approaches such as Artificial Neural Network (ANN) (Jain et al. 2000; Ban 2003) and Support Vector Machine (SVM) (Mountrakis et al. 2011) have been widely used in remote sensing. In the case of multitemporal or multi-source data, the non-parametric approaches are preferable to the parametric ones, since the non-parametric approaches need no assumption regarding the distribution models of the unclassified data (Lu and Weng 2007). The non-parametric SVM approach has recently become increasingly popular (Pal and Mather 2005). SVM has also been introduced into object-based analyses (Tzotsos and Argialas 2008). The superiority of SVM over other classifiers has been reported by various studies in object-based and pixel-based


31

classifications (Buddhiraju and Rizvi 2010; Waske and van der Linden 2008; Mountrakis et al. 2011; Pal and Mather 2005). Pacifici et al. (2008) reported encouraging results using SAR in urban mapping with ANN. Usually, these non-parametric approaches are time-consuming (especially when dealing with noisy data) and the optimal values of the involved parameters are difficult to estimate (e.g. the kernel parameters in SVM) (Mountrakis et al. 2011). Very few examples of urban analysis with SAR data using ANN (Azimi-Sadjadi et al. 1993; Pellizzeri et al. 2003; Alberga et al. 2008; Del Frate et al. 2008) or SVM (Griffiths et al. 2010; Waske and Benediktsson 2007; Yin et al. 2011) exist. Most of these studies are limited in that they only focus on few urban classes. It has previously been reported that SVM used together with object- and rule-based approaches can map urban areas with detailed classes using high resolution SAR data (Hu and Ban 2012). So far, the potential of using multitemporal high resolution polarimetric SAR data for detailed urban mapping has not been fully evaluated. 2.6 Summary Spaceborne PolSAR data has been recognized as a useful remote sensing data source for urban analysis (Lee et al. 2004b; Lee and Pottier 2009a; Hayes and Gough 2009). When SAR data is used for mapping of urban areas, the higher frequency C-band and X-band have been found more efficient for LULC identification than the lower frequency bands (Bryan 1975; Haack 1984). Incidence angles of around 40 degrees have been noted suitable for urban observation (Lichtenegger et al. 1991; Henderson 1990; Hussin 1995). Fusion of multi-angle or multi-direction look SAR data has great potential for further improving urban mapping (Dell'Acqua et al. 2003; Ban and Jacob 2012). Increased spatial resolution has created both opportunities and challenges for detailed urban mapping (Soergel et al. 2005; Ban and Niu 2011; Hu and Ban 2011). Simultaneous use of multiple polarizations can significantly improve urban mapping accuracy (Wu 1984; Ainsworth et al. 2009). Polarimetric parameters generated from decomposition methods such as Freeman parameters (Freeman and Durden 1998) and / /H A α parameters (Cloude and Pottier 1997) are useful for interpreting urban features. Proper temporal relationship between the selected data is important in the multitemporal data fusion (Dell'Acqua et al. 2003; Ban and Niu 2011; Ban and Jacob 2012). To effectively utilize PolSAR data for urban mapping, proper enhancement approaches should be applied. With the reduction of SAR speckle, the performance of segmentation and classification can be further improved (Lee et al. 1999). Although the multilooking process


32

suppresses SAR speckle, the spatial resolution will also be reduced (Cumming and Wong 2005; Massonnet and Souyris 2008). Applying advanced PolSAR filters (Lee et al. 1999, 2006; Lopez-Martinez and Fabregas 2003, 2008), speckles in different PolSAR channels can be effectively reduced and shape details and polarimetric properties can be preserved. Urban mapping using SAR data can benefit from texture analysis (Stasolla and Gamba 2008; Gamba et al. 2011; Hu and Ban 2012). Applying the widely used GLCM measures, different urban classes may also be characterized by SAR textures (Henderson and Xia, 1998; Ban and Wu 2005; Dell’Acqua and Gamba 2006; Del Frate et al. 2008). Classification of SAR data can be greatly improved by exploring contextual information (Flygare 1997; Stuckens et al. 2000; Hubert-Moy et al. 2001). Out of the many contextual approaches for analyzing SAR data, MRF has been the most frequently employed in various applications (Lu and Weng 2007). Through adaptive MRF approaches (Smits and Dellepiane 1997; Garzelli 1999; Zhong and Wang 2009; Lei et al. 2010), homogenous mapping results can be obtained. In this case, shape details and linear features can be preserved. Such a development is encouraging for the detailed urban mapping using high resolution SAR data. Regarding the classification methodology, state-of-the-art urban mapping approaches using SAR data can be divided into pixel-based and object-based classifications using parametric or non-parametric methods. Because it has been found to be difficult to use pixel-based approaches on high resolution remote sensing data (Scheiewe et al. 2001; Shaban and Dikshit 2001), object-based approaches have been more frequently used (Blaschke 2010; Flanders et al. 2003; Benz et al. 2004). However, ideal segmentation considering all involved classes is usually hard to achieve on speckled SAR data in complex urban areas. In comparison with the parametric approach, the non-parametric approach is more suitable to process multitemporal or multi-source data (Lu and Weng 2007). Of the many non-parametric approaches, SVM has been increasingly employed in recent years (Pal and Mather 2005). However, non-parametric approaches are more time-consuming, and optimal system parameters are usually difficult to estimate (Mountrakis et al. 2011). Although many advanced approaches have been proposed for classifying SAR data, there is still a lack of studies concerning detailed urban mapping using high resolution multitemporal PolSAR data, which is the topic of this study.


33

3. STUDY AREA AND DATA DESCRIPTION

3.1 Study Area The study area is located in the Greater Toronto Area (GTA), Ontario, which is the most populated metropolitan area in Canada (Figure 3.1). The GTA is comprised of Metropolitan Toronto and four municipalities (Durham, Halton, Peel and York) and has a population of about six million. GTA is one of the fastest growing urban areas in North America over the past few decades. Rapid change of LULC has been taking place in the GTA, for instance, the urban encroachment towards the Oak Ridges Moraine (ORM), a significant and sensitive natural ecosystem north of Toronto. The ORM plays an important role for water quality in the region. In addition, most of the common biotopes including forests, wetlands with various plant and animal species are present there. However, the fast growing urban sprawl has put the ORM under the threat of water pollution, resource damage and loss of species diversity, which in turn affect the GTA’s environmental quality (Furberg and Ban 2012).

Figure 3.1. Study area: rural-urban fringe of the GTA, ON, Canada with projections of the RADARSAT-2 Ultra-Fine C-HH (blue frame) and PolSAR (red frame) data.


34

To evaluate the situation of the current urban expansion and the affected ecological context, multitemporal RADARSAT-2 SAR data were acquired. They were collected in 2008 from the area which covers the town of Vaughan, Richmond Hill, Aurora, Bolton, the western part of Markham, the southern part of Newmarket, and the eastern part of Brampton in the northern urban-rural fringe of the GTA, as illustrated in Figure 3.1 and Figure 3.2. This area represents one of the most dynamic regions in the GTA.

Figure 3.2. A demonstration of the RADARSAT-2 fine-quad polarimetric SAR data

from June 11 and June 19 by processed Pauli parameters (refer to Table 4.1) R: (Log(T22))8bit G: (Log(T33))8bit B: (Log(T11))8bit

The major LULC classes in the study area are: high-density residential areas (HD), low-density residential areas (LD), industrial and commercial areas (Ind.), construction sites (Cons.), major roads, streets, parks and golf courses. The studied LULCs also include a lot of natural classes such as forests, pasture, water and several types of agricultural crops. Within


35

those LULCs, the built-up classes such as HD, LD and Ind. have the most complicated polarimetric properties as discussed in the previous chapter. These classes can be refined into many sub-types according to their diverse polarimetric characteristics due to differences in geometry and orientation. Some examples are given in Figure 3.3 which show some selected urban samples and their corresponding polarimetric signatures (Boerner et al. 1998) calculated from the pixels in the yellow circles. This demonstrates that the three Ind. sub-types (a), (b), (c) have totally different polarimetric features. Although the Ind. sub-type (b) and the HD sub-type (d) appear similar in the PolSAR Pauli images, they are slightly different regarding their polarimetric signatures.

(a) (b) (c) (d)

Figure 3.3. Selected urban samples: Ground truths (first row), PolSAR Pauli images (second row) and the corresponding polarimetric signatures: co-pol (third row),

cross-pol (fourth row).


36

3.2 Data Description 3.2.1 RADARSAT-2 SAR Data The RADARSAT-2 data used in this study include two types: the fine-beam polarimetric SAR data and the ultra-fine beam high resolution HH-SAR data. The six-date fine-beam quad-polarimetric SAR (PolSAR) images contain the PolSAR information of the four linear polarization channels, namely: HH, HV, VH and VV. The centre frequency at this beam mode is 5.4GHz, i.e., C-band, and the pixel spacing is 4.7 meters in the range direction and 5.1 meters in the azimuth direction. Descriptions of these images are given in Table 3.1. Ascending and Descending data are marked as A and D.

Table 3.1. RADARSAT-2 Fine-Quad-Polarimetric SAR Imagery

Data Orbit Incidence angle range (degree)

Code

June 11 2008 Ascending 40.179~ 41.594 A1 June 19 2008 Descending 40.215~ 41.619 D1 July 05 2008 Ascending 40.182~ 41.597 A2 August 06 2008 Descending 40.197~ 41.612 D2 August 22 2008 Ascending 40.174~ 41.590 A3 September 15 2008 Ascending 40.173~ 41.588 A4

For comparison, three-date RADARSAT-2 ultra-fine beam HH-SAR images were also acquired. They are also C-band with the single polarization HH, and the pixel spacing is 1.56 by 1.56 meters. The descriptions of these images are given in Table 3.2.

Table 3.2. RADARSAT-2 Ultra-Fine HH-SAR Imagery

Data Orbit Incidence angle range (degree) June 25 2008 Ascending 30.622~ 32.038 August 12 2008 Ascending 30.611~ 32.043 September 05 2008 Ascending 30.606~ 32.023


37

3.2.2 Auxiliary Data In addition to RADARSAT-2 images, some auxiliary data were also used to support the experiments. To orthorectify the RADARSAT-2 images, one DEM issued by DMTI Spatial Inc with a resolution of 30 meters was used. To co-register the RADARSAT-2 images and verify the mapping results for some of the LULC classes, the National Topographic DataBase (NTDB) vector data with a scale of 1:50,000 were used as georeference. The NTDB includes vector features of roads, water bodies, forest, built-up areas, etc. To verify the results, some other references including Quickbird data from 2009 and Google Earth images from 2008 were employed. Some field data was also collected during the SAR observations. Soil conditions and vegetation types were gathered to a certain extent with the help of photographs.


38

4. METHODOLOGY In this research, multitemporal high resolution PolSAR data were investigated for detailed urban land cover mapping. To this end, both object-based and pixel-based classification schemes have been proposed. In this study, only supervised classification approaches were investigated. Using an object-based SVM and a rule-based approach, the efficiency of the multitemporal PolSAR features and the combination of different data sets were compared. In addition, pixel-based mapping approaches were proposed based on a contextual SEM framework. Common PolSAR data distribution models were compared using pixel-based methods. Adaptive schemes for distribution model selection were also investigated in this study. Possible enhancement of the mapping results by applying texture analysis was explored. The performance of detailed urban mapping using multitemporal ultra-fine beam high resolution C-HH SAR data was compared to the performance of multitemporal fine-beam high resolution PolSAR data using a pixel-based approach. For further improvement of detailed urban mapping, a scheme to fuse the pixel-based and object-based analysis was investigated. 4.1 Object-based Approach To conduct the object-based analysis, the raw multitemporal PolSAR data was first orthorectified using the same projection with a pixel-spacing of 5 meters. To explore the contextual information from different spatial scales, a multi-scale segmentation hierarchy was built up by the multiresolution segmentation algorithm (Baatz and Schape 2000) implemented in the software eCognition (Definiens 2011). By comparing the polarimetric parameters at different processing steps, the processed Pauli parameters (Cloude and Pottier 1996) were found to be the most suitable ones for segmentation since meaningful objects could be obtained in the speckled PolSAR images. The processed Pauli parameters were derived from raw Pauli parameters that had been processed by a refined Lee filter (Lee 1981), logarithm operation, linear scaling and truncation to 8-bit data. Detailed information on the efficiency of these PolSAR parameters used in the segmentation is given in Paper I. Using the segmented multitemporal PolSAR data set, an object-based classification scheme was developed as shown in Figure 4.1. The mapping procedure can be mainly divided into two parts. These contain the object-based SVM classification and the rule-based approaches for roads and streets extraction and the subsequent mapping improvement.


39

Figure 4.1. Flow chart of multitemporal data fusion combining SVM and rule-based method

SVM is especially suitable for classifying multitemporal or multi-source data (Pal and Mather 2005). For the object-based classification, the input feature vector to the SVM classifier is formed by the mean and standard deviation of the values from the multitemporal data layers of one specific polarimetric parameter set within the object. An example of using Cloude-Pottier parameters on the multitemporal data stack is shown in Figure 4.2.

Figure 4.2. SVM input vector formed by multitemporal features After the object-based SVM classification was performed, specific rules were developed to extract the major roads and streets. These rules took into account the polarimetric scattering values, the shape of the objects, and the spatial relationships between the neighboring classes. Finally, the


40

parks and the pastures were refined using their spatial relationships to the urban areas and the forests and crops were refined by removing isolated misclassified objects. This is one of the first studies to compare the performances of PolSAR parameters using object-based classification. The selected PolSAR parameters listed in Table 4.1 were compared using both the object-based SVM and the rule combined approaches. The efficiency of the different multitemporal data combinations and the fusion of ascending and descending data for urban mapping were investigated by using these object-based approaches. In this comparison, the effect of the segmentation of the selected data combination was considered. The Compressed Logarithmic Filtered Pauli parameter was selected for this multitemporal comparison over an area where all of the data overlapped. A more detailed description of this study can be found in Paper I.

Table 4.1. Selected feature sets for comparisons

Parameter Set (Code) Elements Intensity (I) |HH|2, |HV|2, |VV|2 Logarithmic Intensity (L I) Log (Intensity) Coherency Matrix (T) T Raw Pauli (R P) T11, T22, T33 Logarithmic Pauli (L P) Log (Raw Pauli)16-bit Compressed Logarithmic Pauli (CL P) (Log (Raw Pauli))8-bit Compressed Logarithmic Filtered Pauli (CLF P) (Log (Filtered Pauli))8-bit Freeman (Freeman) D, O, V Cloude-Pottier (Pottier) H, A, a


41

4.2 Pixel-based Approach 4.2.1 Parametric Classification by a Contextual SEM

Algorithm with FMM For the pixel-based analysis, a contextual SEM algorithm was proposed. The algorithm is based on modelling the probability of the observation by FMM (McLachlan and Peel 2000), which represents the heterogeneity of the observation by the mixture of finite latent classes. To cooperate with contextual analysis, the spatially variant FMM (SVFMM) (Celeux et al. 2003) was employed. Accordingly, the mixture model density function for the observation C at location s can be formulated as:

1 1

( ) ( ) ( | , ) ( | )g g

s si i s s r s i si i

f C f C p L i L r f Cρ η= =

= = = ∈ Θ∑ ∑ , (4-1)

where the ( | , )s r sp L i L r η= ∈ is the prior probability for the class of pixel s sL noted as i while the neighbouring pixels r in the neighbourhood

sη are marked as rL . Such prior probability can be analyzed by an adaptive MRF which is described below. ( | )i sf C Θ is the probability of the observation C at location s by the Probability Density Function (PDF)

(.)if with the parameter Θ . In this study, the multitemporal observation probability was simplified as the product of the PDFs from each date:

1( | ) ( | )

m

i s it st ittf C f C θ

=Θ = ∏ , (4-2)

where ( | )it st itf C θ is the PDF for class i of date t with the parameters itθ .

stC is the observation of the date t at location s. m is the number of the dates. Thereby, the mixture model density function (4-2) can be expressed as:

11 1

( ) ( | ) ( | , )( ( | ))g g

m

s si i s s r s it st itti i

f C f C p L i L r f Cρ η θ=

= =

= Θ = = ∈∑ ∑ ∏ . (4-3)

With the modelling of the observation by FMM, a contextual SEM algorithm was designed to estimate the applied SAR or PolSAR distribution models and iteratively classify the data. The flowchart of the algorithm is illustrated in Figure 4.3. This algorithm is based on a SEM framework with contextual analysis using an adaptive MRF (Paper II) and a modified MPAC (Paper III). To explore the different capabilities of


42

the alternative distribution models, the optional dictionary-based and knowledge-based model selection approaches (grey round frames in Figure 4.3) were introduced in the mapping process as optional steps, which will be described in the next subsection (Paper IV). To improve the mapping results, the texture analysis was also considered in the decision making. Improvement by the object-based analysis was introduced in the post processing (Paper V). Those optional steps (grey square frames in Figure 4.3) will also be described in the following subsection.

Figure 4.3. The structure of the proposed contextual SEM algorithm


43

The basic process of the SEM algorithm approaches could be described with the iteration index k as follows: Initialization: initializing the parameter 0itθ of class i on date t according to the training samples. Set the equal MRF prior probability for all the classes. Expectation-Step (E-step): for each pixel s, the local MRF probability

( | , )ks r sp L L r η∈ is calculated based on the classification result from the last

iteration and the weight of the class i is updated in the FMM as the probabilities according to (4-3),

1

1 1

( | , )( ( | ))

( | , )( ( | ))

mk ks r s it st itk t

si mg k ks r s lt st ltl t

p L i L r f C

p L l L r f C

η θτ

η θ=

= =

= ∈=

= ∈∏

∑ ∏. (4-4)

Stochastic-Step (S-step): randomly label each pixel s to a class i with the current probability k

siτ (therefore, a class with a higher probability is more likely to be selected). All the pixels on the map are finally classified into the g class groups1{ , , }k k

gQ QL .

Maximization-Step (M-step): update the distribution parameters k

itθ for all the distribution models on each date with the pixels belonging to the class group 1k

iQ − . For estimation of the texture parameters in the PolSAR distribution models, the method of matrix log-cumulants (MoMLC) (Anfinsen and Eltoft 2011) was applied. Moreover, to prevent the degenerate results such as unexpected merges of similar classes in PolSAR data mapping, a learning control was proposed by measuring the similarities between the class features through a symmetric similarity measure (Anfinsen et al. 2007) for two scale covariance matrices:

1 11( , ) ( )

2SRWd A B tr AB BA d− −= + − , (4-5)

where d is the polarization number. An updating criterion is given as:

0 0,1 1

( , ) min ( , )m mk k

SRW it jt i j G SRW it jtt td dα ∈= =

Σ Σ > ⋅ Σ Σ∏ ∏ , (4-6)

where G is the class set, kitΣ is the scale covariance matrix of class i on date t after iteration k. α is a custom relaxation parameter. Therefore,


44

before the Maximization-Step, class pairs which can’t satisfy the criterion (4-6) are marked as the closing pairs. The updating of their distribution models will accordingly stop. The SEM algorithm sequentially repeats the above E-, S- and M-steps until the convergence point is met, i.e. the percentage of pixels which are changing classes is under a pre-set value. Details of the contextual SEM algorithm are discussed in Paper II. 4.2.2 Model Selection by a Dictionary-based Approach and

a Knowledge-based Approach Since the different distribution models are derived based on alternative assumptions of the observed scenes, they have different capabilities to describe various LULC classes. To adaptively select the most suitable distribution model for the classified data, two model selection approaches were been proposed in our contextual SEM algorithm. Specifically, the dictionary-based and the knowledge-based model selections were studied for the PolSAR data classification. For that purpose, four common PolSAR distribution models namely the Wishart, Kp, G0p and KummerU were selected to form a finite distribution model dictionary 0{ , , , }M Wishart Kp G p KummerUD f f f f= as the candidates for the

distribution (.)itf in (4-3). Accordingly, in the M-step, the distribution models in such a dictionary will all be updated for each class. The aim of the dictionary-based model selection is to select the optimal model for each class. Therefore, in the calculation of the probability k

siτ in (4-4), for class i on date t, select the optimal model from

0{ , , , }M Wishart Kp G p KummerUD f f f f= as the distribution model for (.)itf . Such

optimal model selection is based on a Goodness-of-Fit (GoF) test. The model which best fits the distribution of one class will be assigned the optimal one for that class. A GoF test based on the Mellin Transform (Anfinsen et al. 2011) was employed for the PolSAR data. Another way to take advantage of multi-models is to explore the different separating capacities of alternative models for specific class pairs by using a knowledge-based approach. Comparing different PolSAR distribution models, it was found that the Wishart model was best at separating high-density built-up (HD) areas and adjacent roads. G0p and KummerU model were better at discriminating low-density built-up areas (LD) from forest. Kp, G0p and KummerU were better than Wishart to distinguish between low scattering classes like water, construction sites


45

and golf courses from each other. Based on these findings, a model selection scheme aimed to use the best model to separate the difficult class pairs was proposed. Specifically, this knowledge-based model selection was employed to ensure the changes of the pixel labels between the difficult class pairs in SEM iterations, as illustrated in Figure 4.4.

Figure 4.4. The flowchart of the proposed rule-based model selection. For example, since Wishart was better at separating HD and roads, whenever a pixel label was found to change from “HD” to “Road” in the S-Step, the process moved back to the E-Step to re-evaluate the probabilities using the Wishart model for all the distributions (.)itf in (4-3). The label of that pixel would then be assessed again in the S-Step using the revaluated Wishart probabilities. Three rules were therefore defined in the experiments 1. HD-Road rule: Use the Wishart model for separating the “HD” and “Road” pair. 2. LD-Forest rule: Use the G0p model for separating the “Forest” and “LD” pair. 3. Water-Golf-Cons. rule: Use the G0p model for separating the pairs among the “Golf”, “Water” and “Construction sites”.


46

Details of the above mentioned model approaches and the proposed rules can be found in Paper IV. 4.2.3 Contextual Analysis by an Adaptive MRF and

Modified MPAC 4.2.3.1 Adaptive MRF To explore contextual information, MRF is frequently used in many remote sensing applications. MRF assumes a local dependency for the spatial correlations. Under this assumption, the prior probability of a pixel s labelled as sL considering that its neighbouring pixel r in the neighbourhood sη are marked asrL can be modelled with Gibbs law:

1( | , ) exp{ ( | , )}s r s s r s

s

p L L r U L L rZ

η η∈ = − ∈ , (4-7)

where sZ is a normalizing factor and (.)U denotes an energy function. However, the traditional MRF which assumes a square neighbourhood structure and a fixed energy function usually leads to over averaged results. To preserve linear features and small shapes, an adaptive MRF was applied in this study. Details about the adaptive MRF are given in Paper II. Generally, the adaptivities of the proposed MRF are mainly demonstrated in the following two aspects: Adaptive neighbour structures: To preserve the shape details from over-averaging, a finite number of the neighbor templates were given to constrain the analysis region in specific shapes as illustrated in Figure 4.5. Specifically, the first square shape is used to analyze the homogenous area, while the linear templates are used for the boundaries and linear features. For each pixel, one candidate which has the lowest standard deviation on the span image will be selected from the five templates as the pixel’s neighborhood.

Figure 4.5. Candidate neighbourhoods for adaptive MRF


47

Adaptive energy function: The energy function in equation (4-7) was modified by adding an adaptive factor sb , which is used to represent the different impact of the contextual analysis according to the homogeneities of the area:

( | , ) ( (1 ) ( ))s

s r s s s rr

U L L r b L Lη

η β δ∈

∈ = − − −∑ , (4-8)

where β is a non-negative factor of the MRF model. It quantifies the basic impact of the prior probability. ( )δ ⋅ is the Kronecker delta function.

sb is a local homogeneity measure which is defined in Lee et al. (1999, 2003). For homogeneous areas, 0sb → , the MRF analysis has greater impact in the labelling decision. A uniform class type is most likely to be present in this area. For heterogeneous areas, 21/ (1 )sb σ→ + , (where σ indicates the speckle level (Lee et al. 1999)), the MRF has less effect. In this case different classes will be found. 4.2.3.2 Modified MPAC The multiscale modified Pappas adaptive clustering (MPAC) (Baraldi 2000) employs an adaptive class feature evaluation with a multiscale estimation scheme, where the class properties are simultaneously estimated from adaptive neighborhoods. By using such a strategy, the varying class features can be treated locally. However, the original MPAC has only been used for single date optical or single-polarization images. For the multitemporal PolSAR data classification, a modified MPAC was proposed. The proposed approach can be formulated as follows: For pixel j in a total of N classified pixels, the varying class feature value function is given as ˆ ( )W jst jlµ , where {1,..., }jl G∈ is the class label for pixel j selected from G classes, {1,..., }t M∈ is the date of M dates,

,j sW W⊂ indicates the local square window centered on pixel j of scale {1,..., }s S∈ , W represents the global scale and covers the whole mapping area. The width of window ,j sW is given as sWW , which increases with the scale s as 1 2 ... S WWW WW WW WW< < < < . One example of this multiscale MPAC average estimation strategy is given in Figure 4.6, where the averaged feature of class i for pixel j on date t at scale 1:

1ˆ ( )W j t jl iµ = will be estimated from the grey area inside the window ,1jW .


48

With the local averaged class feature, MPAC estimates the proper class

label jl for pixel j by maximizing the probability which is the product of

the prior probability ( | , )j r jp l l r η∈ as given in (4-1) and the observation

probability ( | )jl jf C Θ :

{1,.., }

ˆ arg max { ( | ) ( | , )}, 1,...,j

jj l j j r j

l Gl f C p l l r j Nη

∈= Θ ∈ = , (4-9)

where jC is the observed pixel value. The original MPAC employs a

fixed MRF limited to the 8-adjacency neighbourhood which is prone to over-averaging as discussed in the previous section. It is only performed on single date data assuming a Gaussian distribution. For the multitemporal PolSAR data, the MPAC is modified using the following strategies:

Figure 4.6. Multiscale MPAC average estimation strategy. The class-conditional scattering averages are extracted from adaptive neighbors within windows Wj,s at

spatial scale s, centered on pixel j. 1. The multitemporal observation probability( | )

jl jf C Θ in (4-9) is

estimated by the following scheme:

1

ˆ( | ) max{ ( | , ( )),...,ˆ ˆ( | , ( )), ( | , ( ))},

ˆif local statistic ( ) exists andis considered reliable, {1,..., }

ˆ( | ) ( | , ( ))

if local stati

j j

j j

j j

l j l j W j j

l j W jS j l j W j

W jst j

l j l j W t j

f C f C lf C l f C l

ls S

f C f C l

µµ µ

µ

µ

Θ = ΘΘ Θ

∀ ∈

Θ = Θ

ˆstic ( ) does not exists oris not considered reliable, {1,..., }

W jst jls S

µ

∀ ∈ , (4-10)

Wj,1

Wj,2

Pixel j

i

i

i

i


49

where 1ˆ ˆ( | , ( )) ( | , ( ))

j j

m

l j W js j l t jt it W jst jtf C l f C lµ θ µ

=Θ = ∏ is the product of the PDF

on each date with the locally averaged class feature ˆ ( )W jst jlµ estimated

from the corresponding date. This averaged class feature ˆ ( )W jst jlµ is

considered unreliable if the number of pixels of class jl within window ,j sW is less than the window width sWW .

2. Instead of the fixed MRF, the adaptive MRF which was proposed in the previous section is employed for the prior probability ( | , )j r jp l l r η∈ . The arbitrary labelling by the maximization process could be replaced by the proposed SEM algorithm to avoid the local extreme in the mapping process. Detailed information about the modified MPAC is given in Paper III.


50

4.3 Hybrid Approach Urban areas usually demonstrate a unique texture compared to natural objects. To explore the texture for further improvement, the Grey Level Co-occurrence Matrix (GLCM) (Haralick 1973) was analyzed. Especially the GLCM angular 2nd moment and dissimilarity features extracted from the ultra-fine beam high resolution C-HH SAR intensity data were found useful to identify the built-up areas from non-built-up. Moreover, the temporal characteristics of the GLCM textures were also helpful for distinguishing forest and crops. Based on the knowledge about the texture properties, an enhancement scheme was proposed and is illustrated in Figure 4.7.

Figure 4.7. Texture enhancement scheme. In the contextual SEM algorithm, the scheme in Figure 4.7 was implemented in the following way. First, each pixel was decided to be in the built-up or non-built-up areas by using the 2nd moment and dissimilarity textures. In the built-up urban area, the pixel would be classified as one of the “built-up classes” either as LD, HD or Ind. (The

Pixel

Averaged 2nd moment < th_2nd &

Averaged Dissimilarity>th_dis ?

Built-up area Non built-up area

LD HD or Ind..

Enhancement by the Averaged dissimilarity and 2nd moment

2nd moment of Jun. 25 enhancement

Ratio of the mean texture between Sep. 02 and Jun. 25 enhancement

Crop1 Forest Crop2

Most possible pairs

Crop1 & Forest Crop2 & Forest Others

yes no

Further classification


51

other classes are noted as “non built-up classes”.) This was implemented by confining the latent finite classes in FMM. Therefore, the class group in (4-1) is g={built-up classes} in built-up area and g={non built-up classes} in non-built-up area. Further, for the pixel has higher probability to be the classes in the four confused class pairs 1:(LD, Ind.), 2:(LD, HD), 3:(Crop1, Forest), 4:(Crop2, Forest), texture analysis have been applied for further identification. The textures were generated from C-HH SAR data. Specifically, the averaged dissimilarity and 2nd moment were used for pairs 1 and 2, 2nd moment of Jun. 25 were used for pair 3 and the ratio of the GLCM means between Sep. 02 and Jun. 25 were used for pair 4. For the confused class pairs (i, j) in which class i has the lower texture value, while class j has the higher, the probabilities in (4-4) were modified and the texture features for the two classes were considered from:

_ _ ( _ _ )

_ _ ( _ _ )

k ksi si

k ksj sj

with texture th ij val texture

with texture val texture th ij

τ ττ τ = ⋅ = ⋅

(4-11)

In urban areas, certain classes such as major roads or streets have complex scattering interactions with their surroundings. They usually have very high variance in the scattering domain and are easily confused with many other classes. For example, roads in an open area have low scattering values. While in dense urban areas, the road segments have high scattering values due to scattering interactions with buildings. However, the roads and streets have typical shape characteristics and are usually located near the other urban classes. To this end, object-based analysis has been performed, and the shape properties are described by some shape features like length, width, roundness and rectangular fit which are defined in the software eCognition (Definiens 2011). Using these shape features, streets are described as narrow curving segments with branches in built-up areas. A major road is noted as having long segments with moderate width and fewer curves. Major roads have low scattering values in the HV channel. Descriptions of roads and streets were defined using the mapping rules employed in the post processing according to the following: Major Roads: (a) Ratio of length to width > threshold_L/W, and Lowest_value <Width< Highest_value, and backscattering in HV< threshold_HV, or (b) Ratio of length to width is extremely high, or (c) Length is extremely long, and Lowest_value <Width< Highest_value


52

Streets: (a) Roundness*(1-Rectangular Fit) > threshold value, and Lowest_value <Width< Highest_value, or (b) Ratio of length to width > threshold_L/W, and Lowest_value <Width< Highest_value, With the rule-based approach, parks which had scattering features similar to pastures or other crops were marked by using spatial relationships. Crops and pastures mapped in the urban area were labelled as parks. The closeness of the pixel to the urban area is described by a built-up density ratio, which is measured as the percentage of the built-up pixels (including LD, HD, Ind.) to the total number of pixels within a square neighbourhood. Such a park mapping rule could be given as: Park: Pasture, crops with the built-up density ratio> Lowest_density_ratio. Paper V provides further the details on the texture improvement and the rule-based approach.


53

5. RESULTS AND DISCUSSION

5.1 Object-based Approach In Paper I, through the object-based SVM and the rule-based approach, thirteen LULC classes including seven urban classes were mapped with an overall accuracy of 90% and a Kappa coefficient 0.89. A mapping example using six date processed Pauli parameter is illustrated in Figure 5.1. With regard to the confusion between the urban classes, industrial and commercial areas have certain confusion with high density built-up areas due to the similar polarimetric properties. Whereas, some low density built-up areas were misclassified as high density built-up areas along borders. The omission errors of the streets and roads to the urban classes are mostly due to the segmentation in local areas where roads and streets were not fully extracted. Although results were also generated by using alternative polarimetric parameters, the misclassifications described above were still observed. More details on these comparisons and the analysis of the object-based approach can be found in Paper I.

(a) (b)

Figure 5.1. Mapping results of the object-based approach: (a) Google map. (b)

mapping results by SVM and rules 5.1.1 Object-based Polarimetric Feature Comparisons The performance of various polarimetric SAR parameters for object-based detailed urban mapping was compared twice, on the ascending and descending data stack, respectively. In both of these data


54

sets, the following was observed: 1. The best classification results were generated by the processed Pauli parameters. 2. The logarithmic data had better results than the raw data. 3. The compressed data produced results comparable to the uncompressed data. 4. The diagonal elements of the coherency matrix (T), the raw Pauli parameters, produced similar results, compared to using all of the matrix elements. 5. The Freeman parameters performed as well as the logarithmic intensity parameters. 6. The Pauli parameters yielded better results than intensity. 7. The Cloude-Pottier parameters yielded poor results compared to the others parameters. They are comparable only with the raw intensity data. Figure 5.2 compares the results of the polarimetric SAR parameters on the ascending and descending stack.

0.900.700.500.300.100.100.300.500.700.90

0.90 0.70 0.50 0.30 0.10 0.10 0.30 0.50 0.70 0.90

I

L–I

T

R–P

L–P

CL–P

CLF–P

Freeman

Pottier

SVM+Rules–Kappa SVM+Rules–OA SVM–Kappa SVM–OA Figure 5.2. Comparison of various polarimetric SAR parameters on the ascending

(right) and the descending (left) data stack 5.1.2 Multitemporal Combination Comparisons The efficiency of the multitemporal data combinations were evaluated by considering the following two aspects: 1. The capability of the ascending


55

and descending orbital complement. 2. The effect of the time span in the data selection. By studying the results of the selected date combinations illustrated in Figure 5.3 with the date code given in Table 3.1, the following was observed: 1. The improvement of using complementary information from ascending and descending orbital data was evident. For example, the 4 date ascending data combination A1234 was not as good as the 3 date combinations with 2 descending and 1 ascending images such as A3+D12. 2. Although the results were improved by continually adding data, the improvement rate decreased after combining 1 ascending and 2 descending images. The best results were obtained for the complete data stack A1234+D12. Some 3 date combinations such as A3+D12 also generated good results. 3. Using a wider time span improved the results. For example, the result of A4+D1 was found to be better than that of A1+D1, as the time span between A4 and D1 is wider than that of A1 and D1. This was also observed when using other combinations such as A1+D2 vs. A4+D2 and A2+D2 vs. A3+D2.

Figure 5.3. The performances of selected date combinations


56

5.2 Pixel-based Approach The proposed pixel-based approach was implemented using a contextual SEM algorithm. The properties of this approach, the performance of the SEM algorithm, the distribution model selection and the contextual analysis efficiency for detailed urban mapping are of importance to this study. 5.2.1 Convergence and Learning Control of the Contextual

SEM Algorithm By using the SEM algorithm, more reliable convergence could be met with less iteration than other iterative processes like EM (Dias and Wedel 2004). Due to the spatial correlation constraint in the MRF analysis and the learning control scheme, the application employing various PolSAR distribution models quickly converged. All the experiments demonstrated that the total change rate met at a convergence point of 0.5% after 7 or 8 iterations. An example of the running cases with the MRF β set to a value of 9 in (4-8) is given in Figure 5.4.

0

0.02

0.04

0.06

0.08

0.1

0.12

2 3 4 5 6 7

Wishart Kp G0p KummerU

Figure 5.4. The trends of the total change rate in iterations using various PolSAR distribution models.

In the experiment, it was found that the common degenerate problem of the SEM algorithm like merging of similar classes was mitigated by contextual analysis. When a more suitable higher β value was employed for the adaptive MRF, the unexpected class merging did not easily occur. There was no significant difference between the results with and without the proposed learning control scheme in a situation where twenty more classes were involved in the classification process. Nevertheless, given a smaller β value or without the MRF analysis, such degenerate behavior


57

became more obvious. The learning control scheme was effective in mitigating this problem. An example which demonstrates the efficiency of the learning control scheme is illustrated in Figure 5.5. The control scheme prevented the class of golf courses from being merged into the classes of pasture or water. Furthermore, the confusion between forest and urban classes was greatly reduced. The properties of the SEM algorithm are described in more detail in Paper II.

(a) (b) (c)

Figure 5.5. Pauli image samples (a) and corresponding results by G0p model with the proposed adaptive MRF using beta 1 with learning chontrol (b) and without learning

control (c). 5.2.2 Contextual Analysis by Adaptive MRF and MPAC Through using contextual analysis, “pepper-salt” effects in the conventional pixel based classification were avoided, and the mapping accuracy was improved. The effects of employing MRF and the modified MPAC are discussed below. A more extensive discussion is given in Paper II and III. 5.2.2.1 Beta and Adaptive Scheme in the MRF Analysis In the proposed adaptive MRF, the MRF β (4-8) denotes the impact of the MRF analysis. In general, a biggerβ value will lead to a more homogenous result, but some shape details may be lost in the averaging. An illustration of the effects of the MRF β on the six date data mapping accuracy using the G0p model is shown Figure 5.6. In this situation, aβ value greater than 5 and value smaller than 20 can generate better results. However, certain urban details such as streets in dense urban areas can dissapear when using a higherβ . An extremely high β value will make the contextual analysis more dominant in the decision making process,


58

and less details will be preserved. When β was set to a value bigger than 35, the overall accuracy started to decrease. The cause of this is the overflowed MRF probabilities in the calculation. A similar phenomenon can also be observed in other PolSAR models. The choice of the β value depends on the amount of data used in the mapping process. There also exists a tradeoff between accuracy and homogeneity.

0.6

0.65

0.7

0.75

0.8

0.85

1 3 5 7 9 11 13 15 17 19 21 23 25

OA Kappa

0.7

0.73

0.76

0.79

0.82

0.85

5 15 25 35 45 55 65 75

OA Kappa

(a) (b)

Figure 5.6. Mapping accuracies by the MRF Beta from 1 to 25 with scale 1 (a) and Beta from 5 to 75 with scale 5 (b) using G0p model.

Figure 5.7 shows a comparison between the contextual schemes in order to demonstrate the efficiency of the proposed adaptive MRF. The adaptive neighbor but non-adaptive MRF energy function (c) yielded similar results compared to the proposed adaptive MRF (b). However, some details such as streets within residential areas were lost. Without using any adaptive scheme (d) in the conventional MRF analysis an over-averaged result was obtained with unclear boundaries and less details.

(a) (b) (c) (d)

Figure 5.7. Pauli image (a) and corresponding results by G0p model with proposed adaptive MRF (b), adaptive neighbor but non-adaptive MRF energy fuction (c), and

non-adaptive neighbor (d).


59

5.2.2.2 Multiscale Effects in the Modified MPAC

To evaluate the multiscale analysis scheme in the modified MPAC, a comparison between different scale combinations with window sizes of 5x5, 7x7 and 11x11 was conducted. Results using the Wishart model on the six date data are shown in Figure 5.8 and the following was noted regarding the detailed mapping of urban areas using PolSAR data by MPAC: 1. A smaller analysis scale is more efficient. Generally, urban classes such as HD, LD and Ind. benefited the most from using a smaller scale. 2. The multiscale scheme did not provide further improvement and the accuracy is in middle of that by the other single scale schemes. Similar comparisons on the influence of scale have not been reported in previous studies using MPAC or PAC (Pappas 1992, Baraldi et al. 2000, Baraldi et al. 2006).

0.7

0.74

0.78

0.82

5 7 11 5_7 5_11 7_11 5_7_11

OA Kappa Figure 5.8. Overall Accuracy and Kappa for the different scale combinations in the

proposed algorithm by MPAC and adaptive MRF. 5.2.2.3 Comaprison of Adaptive MRF and MPAC Although the multiscale scheme has been found not always to be efficient, the local averaging approach for estimation of the adaptive class features has shown a certain effect. As demonstrated in the comparison with the adaptive MRF in Figure 5.9, the MPAC could bring more coherent local representation than the adaptive MRF in the HD and Ind. area. Moreover, classes like golf course were improved by using MPAC. Nevertheless, by combining the MPAC and the adaptive MRF, further improvement was obtained for LD, forest and pasture. Furthermore, the adaptive MRF was better at preserving the shape details than the fixed contextual analyzing scheme in the original MPAC. The SVM with the adaptive MRF was also compared with the proposed contextual approach. Although similar overall accuracy was achieved by the SVM algorithm and the adaptive MRF with the modified MPAC, most of the shape details were lost in the SVM. Paper II and III contain more detailed discussion on the use of the


60

modified MPAC.

(a) (b)

(c) (d)

(e) (f)

Figure 5.9. Google map (a), Pauli image (b) and corresponding results by Wishart model with only adaptive MRF (c), original MPAC (d), proposed MPAC+adaptive

MRF (e) and SVM+adaptive MRF (f).


61

5.2.3 Comparison and Selection of Various PolSAR Distribution Models

5.2.3.1 PolSAR Model Comparison Together with the parametric SEM algorithm, the PolSAR distribution models of Wishart, G0p, Kp and KummerU was employed. Using the G0p, Kp and KummerU models the proposed contextual SEM algorithm generated better results than the contextual SVM (Paper II). Generally, models with complex texture assumption such as G0p, Kp and KummerU overall produced better results than the Wishart model. For specific classes (Figure 5.10), G0p and KummerU were better at classifying LD. G0p and Kp were better at classifying Golf and Pasture, Crops and Forest. Although similar good results were produced by KummerU and G0p, KummerU was more time consuming than G0p. The Wishart model was best at separating roads from high density areas. G0p, Kp and KummerU all misclassified a large number of HD as Road. The Wishart model was poor at classifying low scattering classes such as Golf, Water, Construction sites and Road. That is why the Wishart model has higher user accuracy but lower producer accuracy for the Road class. Resulting classification maps using these models are shown over a small sample area in Figure 5.11. The model comparisons are discussed in more detail in Paper II and Paper IV.

Figure 5.10. The user (left) and producer (right) accuracies of PolSAR models


62

(a) (b)

(c) (d)

(e) (f)

Figure 5.11. Google map Quick-bird image (a), descending Pauli image (b) and

corresponding classification results by G0p (c), Kp (d), KummerU (e) and Wishart (f) models.


63

5.2.3.2 PolSAR Model Selection Since different PolSAR models have distinct abilities for the discrimination of alternative LULC classes, dictionary-based and rule-based model selection approaches were applied to further improve the results by integrating the advantages of different models. The results of the two model selections can be found in Paper IV. For the dictionary-based model selection, although certain effects have been noted by some previous researches (Moser et al., 2006; Krylov et al., 2011a, 2011b), the reported improvements were limited and the involved classes were few. As shown in Figure 5.12, a dictionary-based model selection does not necessarily produce better results compared to a single model approach. Although a more accurate distribution of classes was usually produced with the use of a dictionary-based approach, mapping accuracy was not always improved. Moreover, some model combinations even generated worse results. An explanation for this could be that some models are excluded. These models may be less accurate to describe the data distributions of some class types, but have better separating ability for discriminating certain class pair. For instance, the Wishart model was better at separating HD and Road. However, the chance of it being selected as the optimal model was less since the goodness of fit of Wishart was worse compared to the other models. Combining different models will complicate the calculation of the estimates due to the fact that different models have different estimation scales. Nevertheless, improvement was observed for some classes using the dictionary-based approach. For example, for G0p+Kp Road class producer accuracy increased from 86.15% to 90.14% and user accuracy increased from 62.03% to 64.45% compared to the the single G0p model.

77

78

79

80

81

82

83

84

85

86

G K U W GK GU KU WG WK WU GKU WGK WGU WKU WGKU

OA Kappa Figure 5.12. Overall Accuracy and Kappa for the different dictionaries. G=G0p,

K=Kp, U=KummerU, W=Wishart. Unit is in percent.


64

Since combining distribution models by using a dictionary-based approach provided limited improvement on the classification results, a rule-based approach, which adaptively selects one optimal model for separating pairs of similar classes, was applied. Specifically, the results from the rule-based model selection share the following properties: 1. The proposed rules mainly influence the involved classes. Side effects on other classes were very limited. For example the HD-Road rule only improves HD and Road classes. 2. The overall accuracy and Kappa were improved in all situations and the accuracy of classes targeted by the rules was significantly improved. 3. The results were further improved by adding more rules. For example, the HD-Road and LD-Forest rules improved the OA for the Kp model application by 3.4% and 0.7%. By combining the HD-Road and LD-Forest rules together in Kp, the OA was increased by 4.1%. The rule-based approach was also applied in combination with the dictionary-based approach. No further improvement was observed by using such a combination, since the dictionary-based approach is ineffective. The changes to the only dictionary-based approach were similar to the changes for the single model applications. An example of applying the HD-Road rule using the dictionary-based approach on various dictionaries is illustrated in Figure 5.13. Similar results were obtained using other rules.

Figure 5.13. Improvement by the HD-Road rule with the dictionary-based approach


65

5.2.4 Comparison and Fusion of PolSAR and C-HH SAR Data by Contextual SEM with Texture Analysis

To demonstrate the importance of the SAR polarizations in detailed urban mapping, multitemporal comparison of the PolSAR data and the C-HH SAR data was conducted using the contextual SEM with adaptive MRF. Results are shown in Table 5.1. Generally, classifications using the PolSAR data have higher accuracies, especially for the identification of the different urban classes. Moreover, without polarimetric information Cons. and LD were difficult to be discriminated. The Ind. and HD were generally difficult to distinguish from each other. Furthermore, the PolSAR data was effective in classifying low scattering classes such as water, golf and pasture. Crops and forest were also better classified with the PolSAR data. Table 5.1. Comparison of the C-HH SAR and PolSAR data and the improvement by

texture analysis (Unit: Percent)

Pro. User. Pro. Improvement User. Improvement

UH Pol UH Pol UH Pol UH Pol

Water 87.3 91.86 82.41 92.06 0 0.29 -0.54 1.92

Golf 64.53 88.63 63.08 85.19 0.14 1.93 -0.25 0.37

Pasture 70.58 86.22 59.88 89.47 -6.22 -8.25 0.33 1.61

Crop1 65.24 92.33 52.26 76.07 11.17 2.01 11.41 -0.11 Crop2 87.08 90.63 62.55 85.36 -0.7 -0.03 25.29 10.13

Forest 55.84 72.55 40.51 76.38 16.75 7.87 30.48 14.37 Cons. 20.08 88.68 33.69 88.28 -5.61 -1.04 -0.69 0.05

HD 36.22 65.85 36.96 58.98 -6.32 -3.13 10.46 5.81

Ind. 37.41 74.15 45.18 77.79 2.79 -1.14 6.91 -1.05

LD 8.4 62.06 39.33 80.97 77.41 23.66 11.64 -8.88

Although the PolSAR data is better for detailed urban mapping, the texture information generated from the C-HH SAR data has great potential for the identification of urban areas. The PolSAR textures are less efficient for this purpose. Using texture analysis in the SEM algorithm, significant improvement was observed for the LD, Forest and Crop classes in both the C-HH SAR and PolSAR data. A decrease in producer accuracy for HD and Pasture was observed. However, other classes were less affected by the texture analysis. The two data sets and the corresponding texture improved results are shown over a small sample area in Figure 5.14. A more extensive discussion on this analysis can be found in Paper V.


66

(a) (b)

(c1) (c2)

(d1) (d2)

Figure 5.14. Google map (a), PolSAR Pauli image (b), and results by the PolSAR (c)

and C-HH SAR (d) data with (1) and without (2) texture analysis.


67

5.3 Hybrid Approach Since the problematic classes, roads and streets, have characteristic shape features, a hybrid approach based on object-based analysis and pixel-based mapping was explored in order to further improve the classification results. Contrary to the other fusion approaches, this proposed method improves on pixel-based results by using object-based features and spatial contextual information. This hybrid fusion approach is discussed in detail in Paper V. 5.3.1 Comparison of Results with and without Texture

Analysis and Rules When the contextual SEM is employed without any rules, the major roads, streets and parks are difficult to classify together with the other classes. For example, using the PolSAR data A2, A3 and A4, the producer and user accuracies were only 15.2% and 28.7% for streets, 48.4% and 58.4% for major roads, and 32.4% and 59.2% for parks. The accuracies decreased sharply from classification of ten classes to thirteen classes including these three classes. Due to the confusions with these three classes, the accuracies of the other classes also decreased. The effect of introducing these three classes in the pure contextual SEM is demonstrated in Figure 5.15.

Figure 5.15. Changes of the Producer accuracies (red) and User accuracies (green) by only contextual SEM from using 13 classes to using 10 classes on the PolSAR data.

The solid bar is for the increase, the hollow one is for the decrease. By using a hybrid approach, including both object shape features and rules, major roads, streets and parks were well mapped. As discussed in the previous section, the classification of forest, LD and crops was significantly improved by the texture analysis with little impact on the other classes. Improvement by the hybrid approach on the use of the contextual SEM only is illustrated in Figure 5.16. A more thorough


68

discussion regarding these comparisons can be found in Paper V.

Figure 5.16. Changes of the Producer accuracies (red) and User accuracies (green) from using only contextual SEM to the hybrid approach with texture analysis and

rule-based approach on the PolSAR data. The solid bar is for the increase, the hollow one is for the decrease.

5.3.2 Fusion of Pixel-based Approach and Object-based

Analysis by a Rule-based Method A confusion matrix of the final results of the hybrid approach with texture analysis and rule-based feature extraction using the PolSAR data A2, A3, A4 and the three C-HH SAR data is shown in Table 5.2.

Table 5.2. Confusion matrix of the results using the PolSAR data A2, A3, A4 and three C-HH SAR data by the hybrid approach (Unit: Percent)

Ground truth

Water Golf Pastur Crop1 Crop2 Forest Cons. HD Ind. LD Street Road Park

Water 91.53 3.32 0.02 0 0 0.03 0.17 0.09 0.11 0.01 0.05 0.07 0.36

Golf 5 81.91 1.32 0.02 0.02 0.36 2.37 0.66 0.92 0.22 2.71 7.11 6.75

Pasture 0.2 2.1 77.87 0.17 0.15 1.53 1.3 0.28 0 0.44 0.3 0.93 0.55

Crop1 0.24 0.48 18.16 94.34 1.38 1.5 2.93 0.04 0 0.23 0.01 0.35 2.52

Crop2 0.04 0.24 0.63 1.94 90.6 1.11 3.17 0.01 0 0.04 0 0.21 0.01

Forest 0.04 0.7 1.19 2.3 3.55 80.39 0.14 0.38 0.17 0.52 0.16 0.49 2.23

Cons. 1.64 0.97 0.26 0.05 1.05 0.07 86.51 0.99 0.81 0.35 4 5.55 5.31

HD 0.02 0.21 0.13 0.14 0.43 4.79 0.39 57.77 17.36 7.26 3.24 0.58 0.67

Ind. 0.2 0.05 0.07 0.02 1.19 3.13 0.72 15.63 67.86 3.4 5.24 4.34 3.63

LD 0.02 0.06 0.07 1.02 1.64 5.72 0.15 13.8 6.24 79.78 9.4 1.12 5.45

Street 0 0 0 0 0 0.26 0 8.95 4.4 6.99 63.61 1.72 0.53

Road 1.06 9.19 0.1 0 0 0 1.43 0.3 1.91 0.06 9.72 78.92 1.19

Park 0 0.73 0 0 0 1.11 0.72 1.1 0.11 0.66 1.97 0.18 70.57

Pro. 91.53 81.91 77.87 94.34 90.6 80.39 86.51 57.77 67.86 79.78 63.61 78.92 70.57 User. 92.74 63.59 91.39 75.31 95.65 88.16 71.18 61.24 69.54 64.41 66.77 80.16 93.02

Generally, roads, streets and parks were mapped with reasonable


69

accuracy using a rule-based approach. By tracing the error source of the introduced classes in Table 5.2, the following confusions on these classes were noted. For parks, most of the omission was assigned to golf courses, Cons. and LD. However, very high user accuracy was observed for parks. Due to the strict urban constraint rules, few areas were incorrectly classified as parks. With the similarity shape features, certain straight and long streets were classified as major roads. Confusion of the roads caused by similar shape features could also be the found for golf courses and the Cons. and Ind. classes. Most of the confusion between streets (roads) and HD or LD was due to the poor segmentation locally, where the streets were not perfectly segmented out. However, due to the complex scattering interactions in dense urban areas, perfect segmentation is usually not easily obtained. 5.4 Evaluation of the Pixel-based, Object-based and Hybrid

Approaches In this study, the urban details have been mapped mainly by using pixel-based, object-based and hybrid approach. The performances of these approaches have been compared. Specifically, the pixel-based approach compared is the one which had better performance using adaptive MRF and G0p model with the road-HD rule for the rule-based model selection. As discussed in the previous section, roads, streets and parks are difficult to classify using the SAR scattering features only in a pure pixel-based approach. Although the classification of major roads could be improved by combining the ascending and descending data acquired on all of the dates, improvement was limited when using a pixel-based approach. Without considering streets and parks, comparisons between the three mapping approaches were made (see Table 5.3). The overall accuracy of the result by pixel-based approach using six dates PolSAR data is slightly better than that by hybrid approach using only three dates PolSAR data and three dates C-HH SAR data.The pixel-based approach could achieve higher accuracy on the crops, pasture, forest, golf, Cons. and HD, which is due to the use of more PolSAR information. But the hybrid approach gave better results on the Ind., LD and road than the pixel-based approach, which is due to the efficient C-HH SAR textures and the object-features. Moreover, the hybrid approach could further discriminate streets and parks with reasonable accuracy, which indicates a better performance than the use of the only scattering features in the pixel-based approach. Nevertheless, the best results were still obtained by the object-based


70

approach. Although using only three dates PolSAR data without the assistance of the C-HH SAR data, the results obtained by the object-based approach was generally better than those obtained by the hybrid or the pixel-based approach. The results were much more homogeneous. With the same PolSAR data the hybrid approach could better map the golf courses. The producer accuracy for LD and crop1 for the hybrid approach were obviously better than those obtained by the object-based approach. With all the PolSAR data, the object-based approach generated the best results. However, the producer accuracy of the forest and golf courses by the object-based approach was not as good as that acquired by the pixel-based approach. Table 5.3. Comparison of the pixel-based, object-based and hybrid approaches (Unit:

percent)

Object-based using

A1234D12

Object-based using

A234

Pixel-based using

A1234D12

Hybrid using

A234+3 dates C-HH

Pro. User Pro. User Pro. User Pro. User

Water 92.01– 100– 95.16– 93.26– 90.46– 97.25– 92.54– 93.92–Golf 79.58– 82.56– 76.61– 77.62– 96.23– 69.05– 85.03– 81.82–

Pasture 94.01– 95.89– 79.45– 90.45– 95.33– 94.61– 77.87– 92.72–Crop1 96.37– 97.54– 90.15– 83.25– 95.36– 92.63– 93.83– 76.21–Crop2 96.53– 99.12– 96.03– 94.85– 96.43– 97.64– 91.25– 95.56–Forest 94.02– 91.43– 92.65– 91.39– 97.24– 84.86– 81.11– 90.48–Cons. 89.72– 95.93– 87.92– 96.11– 88.89– 91.17– 85.82– 86.55–HD 79.91– 85.03– 68.91– 67.15– 70.25– 68.76– 62.01– 61.4–Ind. 91.3– 86.85– 82.79– 81.41– 58.12– 78.29– 70.48– 75.09–LD 94.21– 83.8– 78.41– 77.12– 65.97– 71.28– 82.03– 59.86–

Road 92.4– 89.03– 87.16– 86.8– 59.17– 58.27– 84– 88.62–OA 91.5 85.41 82.36 81.04

Kappa 90.5 83.69 80.36 78.9

Considering all thirteen classes, the hybrid and the object-based approach are compared in Figure 5.17. Generally, the object-based approach performs better. The overall accuracy and Kappa for the thirteen classes was 0.86 and 0.85 for the object-based approach using PolSAR data A1234D23, 0.78 and 0.75 for the object-based approach using PolSAR data A234, and 0.76 and 0.74 for the hybrid approach using PolSAR data A234 and three C-HH SAR data. Golf courses had higher producer accuracy for the hybrid approach. This is mainly due to the better segmentation of golf courses with the aid of the C-HH SAR data. Higher user accuracy for parks and streets could also be observed with the hybrid approach, which is probably due to the strict rules applied to the pixels.


71

1.000.800.600.400.200.000.200.400.600.801.00

1.00 0.80 0.60 0.40 0.20 0.00 0.20 0.40 0.60 0.80 1.00

Water

Golf

Pasture

Crop1

Crop2

Forest

Cons.

HD

Ind.

LD

Street

Road

Park

Hybrid A234+UHR Object A234 Object A1234D12 Figure 5.17. The producer (left) and user (right) accuracy of the hybrid approach

using PolSAR data of A2, A3, A4 and C-HH SAR data, object-based approach using PolSAR data of A2, A3, A4 and using all the PolSAR data.

The mapping results by the pixel-based approach using all the PolSAR data with G0p model and adaptive MRF (c), with further applied road-HD rule (d), by the hybrid approach using PolSAR data A234 and three C-HH SAR data (e) and by the object-based approach using PolSAR data A234 (f) are illustrated in Figure 5.18. It was found that the object-based approach was able to generate more homogeneous results. However, this approach strongly depends on the given segmentations. On the one hand, some classes could benefit from the proper segmentation with typical shape features for better identification such as roads and streets. On the other hand, different classes may need different segmentations (for example, the major roads need larger segments, while the golf courses and streets need smaller ones). It is hard to find a balance. A pure pixel-based approach is limited in mapping the high variance classes (such as roads, streets) and classes with similar scattering features but with different functional properties (such as parks). The hybrid approach can improve the results based on the pixel-based outcomes, but more efficient urban feature extraction methods need to be developed (e.g. roads, parks).


72

(a) (b)

(c) (d)

(e) (f)

Figure 5.18. Google map (a), Pauli image (b) and corresponding classification results by contextual approach with only adaptive MRF (c), contextual approach and model

selection (d), hybrid approach (e) and object-based approach (f).


73

6. CONCLUSIONS AND FUTURE RESEARCH

6.1 Conclusions The potential of multitemporal RADARSAT-2 fine-beam PolSAR data for urban LULC mapping has been evaluated in comparison with ultra-fine-beam HH SAR data at reduced resolution (10m). In general, the proposed object-based and pixel-based methods were able to effectively map detailed urban features with high accuracy. Considering the objectives of this research, specific conclusions can be drawn. By comparing various PolSAR features using the object-based SVM and rule-based approach, the efficiency of the Pauli parameters was revealed. Improvements by using certain processing steps such as logarithm and data stretching were observed in the segmentation and classification of the PolSAR data. The significance of using information from both the ascending and descending orbital modes has been demonstrated in the comparison of multitemporal data. A longer time span was found more efficient in the data fusion (Paper I). The importance of contextual analysis was revealed in the pixel-based classification using high resolution PolSAR data. Using the adaptive MRF and the SVFMM, the proposed contextual SEM was able to produce homogenous mapping results with high accuracy. Shape details were well preserved and no over-averaging was observed, which is usually the case in traditional fixed MRF. The proposed control scheme was effective in mitigating the degenerate problem like merging of similar classes in the SEM iteration, especially when the impact of the contextual analysis is decreased in the labeling decision. The modified MPAC was able to further improve the mapping accuracy through adaptive estimation of the varying class features by the local averaging scheme. The multiscale analysis scheme in the modified MPAC was demonstrated not to be effective. The best results were achieved using small scale analysis. The proposed contextual SEM with the modified MPAC and adaptive MRF generated better results than SVM. (Paper II, III) Regarding the PolSAR distribution models, G0p, Kp and KummerU were demonstrated to be more effective for detailed urban mapping than the Wishart distribution. Although relatively low overall accuracy was obtained, the Wishart model was better at separating built-up areas from surrounding roads. To combine the advantages of different distribution


74

models, the proposed rule-based model selection which adaptively uses the optimal model to discriminate the difficult class pairs was found more effective than the dictionary-based model selection. (Paper IV) Through comparison of the multitemporal fine-beam PolSAR data and the ultra-fine-beam HH SAR data, it was demonstrated that SAR polarizations significantly contributed to the identification of different urban patterns, such as low density residential areas, high density residential areas, industrial and commercial areas and construction sites. Textures generated from the ultra-fine-beam C-HH SAR data were found to be more efficient in discriminating urban area than the corresponding fine-beam high resolution PolSAR textures. Further improvement was achieved through the proposed texture analysis by fusion of two data sets with alternative resolution. The proposed rule-based fusion approach employing object features and spatial correlations could effectively extract the low backscatter classes such as roads, streets and parks with reasonable accuracy. (Paper V) 6.2 Future Research To further improve urban mapping with higher accuracy, the following topics will be investigated in the future. 1. Sophisticated contextual approaches need to be developed. Besides

MRF, methods like Conditional Random Field (CRF), watershed, wavelet transformation, etc. will be studied. Proper segmentation methods for high resolution SAR data and effective pixel-based and object-based fusion approaches will be further investigated.

2. Besides spatial analysis, decision synthesizing approaches for the multi-temporal, multi-source and multi-resolution data fusion will be further explored. Focus will be put on adaptive selection of reliable and representative information for different classes.

3. Specific pattern recognition approaches for extraction of typical urban features such as man-made structures and road networks will be studied. Especially for SAR or PolSAR data, practical and efficient methods to identify urban features and the way to use those features for urban mapping will be further explored.


75

REFERENCES Alberga, V., 2007. “A study of land cover classification using polarimetric

SAR parameters.” International Journal of Remote Sensing, 28(17): 3851-3870.

Alberga, V., G. Satalino, D.K. Staykova. 2008. “Comparison of polarimetric SAR observables in terms of classification performance.” International Journal of Remote Sensing, 29(14): 4129-4150.

Aiazzi, B., L. Alparone, S. Baronfi. 1998. “Multiresolution local-statistics speckle filtering based on a ratio Laplacian pyramid.” IEEE Transactions on Geoseience and Remote Sensing, 36(5):1466-1476.

Ainsworth, T.L., D.L. Schuler and J.-S. Lee. 2008. “Polarimetric SAR Characterization of Man-Made Structures in Urban Areas Using Normalized Circular-Pol Correlation Coefficients.” Remote Sensing of Environment, 112: 2876–85.

Ainsworth, T.L., J.P. Kelly and J.-S. Lee. 2009. “Classification comparisons between dual-pol, compact polarimetric and quad-pol SAR imagery.” ISPRS Journal of Photogrammetry and Remote Sensing, 64: 464-471.

Anfinsen, S.N., A.P. Doulgeris and T. Eltoft. 2009. “Estimation of the Equivalent Number of Looks in Polarimetric Synthetic Aperture Radar Imagery.” IEEE Transactions on Geoscience and Remote Sensing, 47(11): 3795–3809.

Anfinsen, S.N., A.P. Doulgeris and T. Eltoft. 2011. “Goodness-of-Fit Tests for Multilook Polarimetric Radar Data Based on the Mellin Transform.” IEEE Transactions on Geoscience and Remote Sensing, 49(7): 2764-2781.

Anfinsen, S.N. and T. Eltoft. 2011. “Application of the Matrix-Variate Mellin Transform to Analysis of Polarimetric Radar Images.” IEEE Transactions on Geoscience and Remote Sensing, 49(6):2281-2295.

Arsenault, H.H. and G. April. 1976. “Properties of speckle integrated with a finite aperture and logarithmically transformed.” Journal of the Optical Society of America, 66(11): 1160-1163.

Ayed, I.B., A. Mitiche and Z. Belhadj. 2005. “Multiregion level-set partition of synthetic aperture radar images.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(5): 793–800.

Azimi-Sadjadi, M.R., S. Ghaloum and R. Zoughi. 1993. “Terrain classification in SAR images using principal components analysis and neural networks.” IEEE Transactions on Geoscience and


76

Remote Sensing, 31(2): 511-515. Baatz, M. and A. Schape. 2000. “Multiresolution Segmentation – an

optimization approach for high quality multi-scale image segmentation.” In J. Strobl, T. Blaschke, and G. Griesebner. (Eds.) Angewandte Geographische Informationsverarbeitung XII, Beiträge Zum AGIT-Symposium Salzburg, 12-23, Salzburg, Austria, July 5-7. Heidelberg: Wichmann Verlag.

Ban, Y. 2003. “Synergy of multitemporal ERS-1 SAR and Landsat TM data for classification of agricultural crops.” Canadian Journal of Remote Sensing, 29(4): 518-526.

Ban, Y. and Q. Wu. 2005. “RADERSAT SAR data for landuse/land-cover classification in the rural-urban fringe of the greater Toronto area.” In: Toppen, F. and Painho, M. (Eds.). Proceedings of 8th AGILE Conference on Geographic Information Science, Estoril, Portugal, May 26-28. Lisboa :Universidade Nova de Lisboa.

Ban, Y., H. Hu and I.M. Rangel. 2010. “Fusion of QuickBird MS and RADARSAT SAR data for urban land-cover mapping: object-based and knowledge-based approach.” International Journal of Remote Sensing, 31:1391-1410.

Ban, Y. and X. Niu. 2011. “RADARSAT-2 Polarimetric SAR Data for Urban Land Cover Classification: A Multitemporal Dual-Orbit Approach.” In Lena Halounová (Ed.) 31th EARSeL Symposium, Paris, France, May 30-June 2. EARSeL.

Ban, Y. and A, Jacob. 2012. “Object-based Fusion of Multitemporal Multi-angle ENVISAT ASAR and HJ-1 Multispectral Data for Urban Land-Cover Mapping.” IEEE Transactions on Geoscience and Remote Sensing, in press.

Baraldi, A. and F. Parmiggiani. 1995. “A refined gamma MAP SAR speckle filter with improved geometrical adaptivity.” IEEE Transactions on Geoscience and Remote Sensing, 33(5): 1245-1257.

Baraldi, A., P. Blonda, F. Parmiggiani, and G. Satalino. 2000. “Contextual clustering for image segmentation.” Optical Engineering, 39(4): 1–17.

Baraldi, A., L. Bruzzone and P. Blonda. 2006. “A multiscale expectation-maximization semisupervised classifier suitable for badly posed image classification.” IEEE Transactions on Image Processing, 15(8): 2208-2225.

Benfield, F.K., M.D. Raimi and D.T. Chen. 1999. Once There Were Greenfields. How Urban Sprawl is Undermining America’s Environment, Economy and Social Fabric. New York: Natural Resources Defense Council.

Benz, U.C. and E. Pottier. 2001. Object Based Analysis of Polarimetric


77

SAR Data in Alpha-Entropy-Anisotropy Decomposition Using Fuzzy Classification by eCognition. In IEEE International Geoscience and Remote Sensing Symposium, 1427-1429, Sydney, NSW, July 9-18.

Benz, U.C., P. Hofmann, G. Willhauck, I. Lingenfelder and M. Heynen. 2004. “Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information.” ISPRS Journal of Photogrammetry and Remote Sensing, 58:239-258.

Bhattacharya, A. and R. Touzi. 2011. “Polarimetric SAR urban classification using the Touzi target scattering decomposition.” Canadian Journal of Remote Sensing, 37 (4): 323 – 332.

Blaschke, T.. 2010. “Object based image analysis for remote sensing.” ISPRS Journal of Photogrammetry and Remote Sensing, 65(1): 2-16.

Boerner, W-M., H. Mott, E. Lüneburg, C. Livingston, B. Brisco, R.J. Brown, and J.S. Paterson with contibutions by S.R. Cloude, E. Krogager, J.-S. Lee, D.L. Schuler, J.J. van Zyl, D. Randall, P. Budkewitsch and E. Pottier. 1998. “Polarimetry in radar remote sensing: Basic and applied concepts.” Chapter 5 in F.M. Henderson and A.J. Lewis (Eds.) Principles and Applications of Imaging Radar. Vol. 2 of Manual of Remote Sensing, 3rd ed. New York: John Wiley & Sons.

Bombrun, L. and J.-M. Beaulieu. 2008. “Fisher distribution for texture modeling of polarimetric SAR data.” IEEE Geoscience and Remote Letters, 5(3): 512–516.

Brunner, D., G. Lemoine, L. Bruzzone and H. Greidanus. 2010. “Building Height Retrieval From VHR SAR Imagery Based on an Iterative Simulation and Matching Technique.” IEEE Transactions on Geoscience and Remote Sensing, 48(3): 1487-1504.

Bryan, M.L. 1975. “Interpretation of an urban scene using multichannel radar imagery.” Remote Sensing of Environment, 4:49-66.

Buades, A., B. Coll and J.M. Morel. 2005. “A non-local algorithm for image denoising.” In IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2: 60 – 65, San Diego, CA, USA, June 20-25.

Buddhiraju, K.M. and I.A. Rizvi. 2010. “Comparison of CBF, ANN and SVM classifiers for object-based classification of high resolution satellite images.” In IEEE International Geoscience and Remote Sensing Symposium, 40-43, Honolulu, USA, July 25-30.

Cameron, W.L. and L.K. Leung. 1990. “Feature-motivated scattering matrix decomposition.” In IEEE International Radar Conference, 549-557, Arlington, VA, May 7-10.

Carr, J.R. and F.P. De Miranda. 1998. “Semivariogram in comparison to


78

the co-occurrence matrix for classification of image texture.” IEEE Transactions on Geoscience and Remote Sensing, 36(6): 1945-1952.

Carvalho, E., D. Ushizima, F. Medeiros, C. Martins, R. Marques and I. Oliveira. 2009. “SAR imagery segmentation by statistical region growing and hierarchical merging.” Digital Signal Processing, 20: 1365-1378.

CCRS (Canada Centre for Remote Sensing). 2007. Fundamentals of Remote Sensing. Available online at: http://www.nrcan.gc.ca/sites/www.nrcan.gc.ca.earth-sciences/files/pdf/resource/tutor/fundam/pdf/fundamentals_e.pdf (accessed 02 Nov, 2012).

Celeux, G. and J. Diebolt. 1985. “The SEM algorithm: A probabilistic teacher algorithm derived from the EM algorithm for the mixture problem.” Computational Statistics Quaterly, 2(1): 73–82.

Celeux, G., F. Forbes and N. Peyrard. 2003. “EM procedures using mean field-like approximations for Markov model-based image segmentation.” Pattern Recognition, 36(1): 131–144.

Chaabouni-Chouayakh, H. and M. Datcu. 2010a. “Coarse-to-Fine Approach for Urban Area Interpretation Using TerraSAR-X Data.” IEEE Geoscience and Remote Letters, 7(1):78-82.

Chaabouni-Chouayakh, H. and M. Datcu. 2010b. “Backscattering and Statistical Information Fusion for Urban Area Mapping Using TerraSAR-X Data.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 3(4):718-730.

Chamundeeswari, V.V., D. Singh and K. Singh. 2009. “An Analysis of Texture Measures in PCA-Based Unsupervised Classification of SAR Images.” IEEE Geoscience and Remote Sensing Letters, 6(2): pp.214-218.

Chen, J., Y. Chen, W. An, Y. Cui and J. Yang. 2011. “Nonlocal Filtering for Polarimetric SAR Data: A Pretest Approach.” IEEE Transaction on Geoscience and Remote Sensing, 49(5): 1744-1754.

Cieslewicz, D.J.. 2002. “The environmental impacts of sprawl.” In G.D. Squires (Ed.) Urban Sprawl, Causes, Consequences and Policy Responses, 23–38. Washington, DC: Urban Institute Press.

Clausi, D.A. and B. Yue. 2004. “Comparing cooccurrence probabilities and Markov random fields for texture analysis of SAR sea ice imagery.” IEEE Transaction on Geoscience and Remote Sensing, 42(1): 215- 228.

Cleve, C., M. Kelly, F. Kearns and M. Moritz. 2008. “Classification of the wild land-urban interface: A comparison of pixel and object-based classifications using high-resolution aerial photography.” Computers Environment and Urban Systems, 32:317-326.


79

Cloude, S.R. and E. Pottier. 1996. “A review of target decomposition theorems in radar polarimetry.” IEEE Transaction on Geoscience and Remote Sensing, 34(2): 498–518.

Cloude, S.R. and E. Pottier. 1997. “An entropy based classification scheme for land applications of polarimetric SAR.” IEEE Transaction on Geoscience and Remote Sensing, 35(1):549-557.

Congalton, R.G. and K. Green. 2008. Assessing the Accuracy of Remotely Sensed Data: Principles and Practices, 2nd ed. Boca Raton, FL: CRC Press/Taylor & Francis Group.

Corr, D.G., A. Walker, U. Benz, I. Lingenfelder and A. Rodrigues. 2003. “Classification of urban SAR imagery using object oriented techniques.” In IEEE International Geoscience and Remote Sensing Symposium, 188-190, Toulouse, France, July 21-25.

Cortijo, F.J. and N.P. De La Blanca. 1998. “Improving classical contextual classification.” International Journal of Remote Sensing, (19): 1591–1613.

Cumming, I.G. and F. H. Wong. 2005. Digital Processing of Synthetic Aperture Radar Data: Algorithms and Implementation. Norwood, USA: Artech House.

Dean, A.M. and G.M. Smith. 2003. “An evaluation of per-parcel land cover mapping using maximum likelihood class probabilities.” International Journal of Remote Sensing, 24: 2905–2920.

Definiens. 2011. Definiens Developer XD 1.5.2 Reference Book. Available online at: https://workspace.imperial.ac.uk/imagingfacility/public/DefiniensReferenceBook.pdf (accessed 02 Nov, 2012).

Dekker, R.J.. 2003. “Texture analysis and classification of ERS SAR images for map updating of urban areas in The Netherlands.” IEEE Transaction on Geoscience and Remote Sensing 41(9):1950-1958.

Del Frate, F., F. Pacifici and D. Solimini. 2008. “Monitoring Urban Land Cover in Rome, Italy, and Its Changes by Single-Polarization Multitemporal SAR Images.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 1(2): 87-97.

Deledalle, C.-A., L. Denis and F. Tupin. 2009. “Iterative Weighted Maximum Likelihood Denoising With Probabilistic Patch-Based Weights.” IEEE Transactions on Image Processing, 18(12): 2661-2672.

Dell’Acqua, F. and P. Gamba. 2003. “Texture-based characterization of urban environments on satellite SAR images.” IEEE Transaction on Geoscience and Remote Sensing, 41(1): 153-159.

Dell'Acqua, F., P. Gamba and G. Lisini. 2003. “Improvements to urban area characterization using multitemporal and multiangle SAR images.” IEEE Transaction on Geoscience and Remote Sensing,


80

41(9): 1996- 2004. Dell'Acqua, F. and P. Gamba. 2006. “Discriminating urban environments

using multiscale texture and multiple SAR images.” International Journal of Remote Sensing, 27(18): 3797-3812.

Deng, H and D.A. Clausi. 2005. “Unsupervised segmentation of synthetic aperture Radar sea ice imagery using a novel Markov random field model.” IEEE Transaction on Geoscience and Remote Sensing, 43(3): 528- 538.

Dias, J.G. and M. Wedel. 2004. “An empirical comparison of EM, SEM and MCMC performance for problematic Gaussian mixture.” Statistics and Computing, 14: 323-332.

Dong, Y., B. Forster and C. Ticehurst. 1997. “Radar backscatter analysis for urban environments.” International Journal of Remote Sensing, 18(6):1351–1364.

Dowman, I.J. and A.H. Morris. 1982. “The use of synthetic aperture radar for mapping.” Photogrammetric Record, 10:687-696.

Esch, T., A. Roth and S. Dech. 2005. “Robust approach towards an automated detection of built-up areas from high resolution radar imagery”. In M. Moeller, E. Wentz (Eds.) ISPRS WG VII/1 3rd International Symposium Remote Sensing and Data Fusion Over Urban Areas and 5th International Symposium Remote Sensing of Urban Areas, Tempe, AZ, USA, March 14-16.

Esch, T., M. Thiel, A. Schenk, A. Roth, A. Muller and S. Dech. 2010. “Delineation of Urban Footprints from TerraSAR-X Data by Analyzing Speckle Characteristics and Intensity Information.” IEEE Transactions on Geoscience and Remote Sensing, 48(2): 905-916.

Esch, T., A. Schenk, T. Ullmann, M. Thiel, A. Roth and S. Dech. 2011. “Characterization of Land Cover Types in TerraSAR-X Images by Combined Analysis of Speckle Statistics and Intensity Information.” IEEE Transactions on Geoscience and Remote Sensing, 49(6):1911-1925.

Fasler, F.. 1980. “Texture measurements from seasat-SAR images for urban land use interpretation.” In Fall Technical Meetings of the American Society of Photogrammetry, PS-2-B-1 to PS-2-B-10, New York, USA.

Fjørtoft, R., A. Lopes, P. Marthon and E. Cubero-Castan. 1998. “An optimum multiedge detector for SAR image segmentation.” IEEE Transactions on Geoscience and Remote Sensing, 36(3): 793–802.

Flanders, D., H. Mryka and P. Joan. 2003. “Preliminary evaluation of eCognition objectbased software for cut block delineation and feature extraction”, Canadian Journal of Remote Sensing, 20: 441-452.


81

Flygare, A.-M.. 1997. “A comparison of contextual classification methods using Landsat TM.” International Journal of Remote Sensing, 18:3835–3842.

Franceschetti, G., A. Iodice and D. Riccio. 2002. “A canonical problem in electromagnetic backscattering from buildings.” IEEE Transactions on Geoscience and Remote Sensing, 40(8): 1787-1801.

Freeman, A. and S.T. Durden. 1998. “A Three-Component Scattering Model for Polarimetric SAR Data.” IEEE Transaction on Geoscience and Remote Sensing, 36(3): 963-973.

Freitas, C.C., A.C. Frery and A.H. Correia. 2005. “The polarimetric G distribution for SAR data analysis.” Environmetrics, 16(1):13-31.

Frery, A.C., C.C.F. Yanasse and S.J.S. Sant'Anna. 1995. “Alternative distributions for the multiplicative model in SAR images.” In IEEE International Geoscience and Remote Sensing Symposium, 169-171, Florence, Italy, July 10-14.

Frery A.C., H.-J. Muller, C.C.F. Yanasse and S.J.S. Sant'Anna. 1997. “A model for extremely heterogeneous clutter.” IEEE Transactions on Geoscience and Remote Sensing, 35(3): 648-659.

Frery, A.C., A.H. Correia and C.d. da Freitas. 2007. “Classifying multifrequency fully polarimetric imagery with multiple sources of statistical evidence and contextual information.” IEEE Transactions on Geoscience and Remote Sensing, 45(10): 3098-3109.

Frost, V.S., J.A. Stiles, K.S. Shanmugan and J.C. Holtzman. 1982. “A Model for Radar Images and Its Application to Adaptive Digital Filtering of Multiplicative Noise.” IEEE Transactions on Pattern Analysis and Machine Intelligence , PAMI-4(2): 157-166.

Fu, M. and C. Mui. 1981. “A survey on image segmentation.” Pattern Recognition, 13: 3–16.

Furberg, D. and Y. Ban. 2012. “Satellite Monitoring of Urban Sprawl and Assessment of its Potential Environmental Impact in the Greater Toronto Area between 1985 and 2005.” Environment Management. (In press).

Gamba, P., M. Aldrighi and M. Stasolla. 2011. “Robust Extraction of Urban Area Extents in HR and VHR SAR Images.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 4(1) :27-34.

Gao, L. and Y. Ban. 2009. “Multitemporal RADARSAT-2 Polarimetric SAR Data for Urban Land-Cover Mapping.” In H. Guo and C. Wang (Eds.) 6th International Symposium of Digital Earth, 7841:78410N-78410N-9, Beijing, China. November 3. SPIE.

Galland, F., N. Bertaux and P. Refregier. 2003. “Minimum description length synthetic aperture radar image segmentation.” IEEE


82

Transactions on Image Processing, 12(9): 995–1006. Garzelli, A.. 1999. “Classification of polarimetric SAR images using

adaptive neighborhood structures.” International Journal of Remote Sensing, 20(8):1669-1675.

Geman, S. and D. Geman. 1984. “Stochastic relaxation, Gibbs distributions and the Bayesian restoration of images.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 6(6): 721-741.

Germain, O. and P. Refregier. 2001. “Edge location in SAR images: Performance of the likelihood ratio filter and accuracy improvement with an active contour approach.” IEEE Transactions on Image Processing, 10(1):72–78.

Gleich, D. and M. Datcu. 2009. “Wavelet-Based SAR Image Despeckling and Information Extraction, Using Particle Filter.” IEEE Transactions on Image Processing, 18(10):2167-2184.

Goodman, N.R.. 1963. “Statistical analysis based on a certain complex Gaussian distribution (an introduction).” Annals of Mathematical Statistics, 34:152–177.

Goodman, J.W.. 1976. “Some fundamental properties of speckle.” Journal of the Optical Society of America, 66(11):1145-1150.

Goodman, J.W. 2007. Speckle Phenomena in Optics: Theory and Applications. Greenwood Village, USA: Roberts and Company.

Goze, S. and A. Lopes. 1993. “A MMSE speckle filter for full resolution SAR polarimetric data.” Journal of Electromagnetic Waves and Applications, 7(5): 717–737.

Griffiths, P., P. Hostert, O. Gruebner and S. van der Linden. 2010. “Mapping megacity growth with multi-sensor data.” Remote Sensing of Environment, 114 (2):426-439.

Haack, B.N. 1984. “L- and X-band like- and cross-polarized synthetic aperture radar for investigating urban environments.” Photogrammetric Engineering and Remote Sensing, 50(3): 331-340.

Haralick, R.M., K. Shanmugam and I. Dinstein. 1973. “Textural features for image classification,” IEEE Transactions on Systems, Man and Cybernetics SMC-3(6): 610–621.

Haralick, R.M. and L.G. Shapiro. 1985. “Image segmentation techniques.” Computer Vision Graphics and Image Processing, 29:100–132.

Hayes, M.P. and P.T. Gough. 2009. “Synthetic Aperture Sonar: A Review of Current Status.” IEEE Journal of Oceanic Engineering, 34(3): 207-224.

Hedman, K., U. Stilla, G. Lisini and P. Gamba. 2010. “Road Network Extraction in VHR SAR Images of Urban and Suburban Areas by


83

Means of Class-Aided Feature-Level Fusion.” IEEE Transactions on Geoscience and Remote Sensing, 48(3): 1294 – 1296.

Henderson, F.M.. 1979. “Land use analysis of radar imagery.” Phorogrammerric Engineering and Remore Sensing, 41(3): 307-319.

Henderson, F.M. and M.A. Anuta. 1980. “Effects of radar systems parameters, population, and environmental modulation on settlement visibility.” International Journal of Remote Sensing, 1(2): 137–151.

Henderson, F.M. and K.A. Mogilski. 1987. “Urban land use separability as a function of radar polarization.” Internatiaonal Journal of Remote Sensing, 8(3): 441–448.

Henderson, F.M. 1990. “An analysis of settlement detectability in central Europe using SIR-B radar imagery." International Archives of Photogrammetry and Remote Sensing, 28(part 7-1): 202-214.

Henderson, F.M. and Z.-G. Xia. 1997. “SAR applications in human settlement detection, population estimation and urban land use pattern analysis: A status report.” IEEE Transactions on Geoscience and Remote Sensing, 35(1): 79– 85.

Henderson, F.M. and Z.-G. Xia. 1998. “Radar Applications in Urban Analysis, Settlement Detection and population Estimation.” Chapter 5 in F.M. Henderson and A.J. Lewis (Eds.) Principles and Applications of Imaging Radar. Vol. 2 of Manual of Remote Sensing, 3rd ed. New York: John Wiley & Sons.

Hermes, L. and J.M. Buhmann. 2004. “Boundary-constrained agglomerative segmentation.” IEEE Transactions on Geoscience and Remote Sensing, 42(9): 1984-1995.

Hu, H. and Y. Ban. 2012. “Multitemporal RADARSAT-2 ultra-fine beam SAR data for urban land cover classification.” Canadian Journal of Remote Sensing, 38(1): 1-11.

Hubbert, J.C. 1994. “A comparison of radar, optic, and specular null polarization theories.” IEEE Transaction on Geoscience and Remote Sensing, 32 (3):658-671.

Hubert-Moy, L., A. Cotonnec, L. Le Du, A. Chardin and P. Perez. 2001. “A Comparison of Parametric Classification Procedures of Remotely Sensed Data Applied on Different Landscape Units.” Remote Sensing of Environment, 75(2): 174-187.

Hussin,Y.A. 1995. “Effect of polarization and incidence angle on radar return from urban features using L-band aircraft radar data.” In IEEE International Geoscience and Remote Sensing Symposium, 1:178-180, Firenze, July 10-14.

Huynen, J.R. 1970. “Phenomenological theory of radar targets.” Ph.D. dissertation, Technical University Delft, Bronder Offset, Rotterdam,


84

The Netherlands. Ingrassia, S. and R. Rocci. 2011. “Degeneracy of the EM algorithm for

the MLE of multivariate Gaussian mixtures and dynamic constraints.” Computational Statistics and Data Analysis, 55(4): 1715-1725

Jackson, Q. and D. Landgrebe. 2001. “An adaptive classifier design for highdimensional data analysis with a limited training data set.” IEEE Transaction on Geoscience and Remote Sensing, 39(12): 2664–2679.

Jain, A.K., R.P.W. Duin and M. Jianchang. 2000. “Statistical Pattern Recognition: A Review.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(1): 4-37.

Kandaswamy, U., D.A. Adjeroh and M. Lee. 2005. “Efficient texture analysis of SAR imagery.” IEEE Transactions on Geoscience and Remote Sensing, 43(9): 2075-2083.

Kartikeyan, B., B. Gopalakrishna, M.H. Kalubarme and K.L. Majumder. 1994. “Contextual techniques for classification of high and low resolution remote sensing data.” International Journal of Remote Sensing, 15: 1037–1051.

Kartikeyan, B., A. Sarkar and K.L. Majumder. 1998. “A segmentation approach to classification of remote sensing imagery.” International Journal of Remote Sensing, 19: 1695–1709.

Kayabol, K. and J. Zerubia. 2013. “Unsupervised amplitude and texture classification of SAR images with multinomial latent model.” IEEE Transactions on Image Processing, in press.

Kersten, P.R., J.-S. Lee and T.L. Ainsworth. 2005. “Unsupervised classification of polarimetric synthetic aperture radar images using fuzzy clustering and EM clustering.” IEEE Transactions on Geoscience and Remote Sensing, 43(3): 519-527.

Kettig, R. and D. Landgrebe. 1976. “Classification of multispectral image data by extraction and classification of homogeneous objects.” IEEE Transactions on Geoscience Electronics: 14 (1):19-26.

Keuchel, J., S. Naumann, M. Heiler and A. Siegmund. 2003. “Automatic land cover analysis for Tenerife by supervised classification using remotely sensed data.” Remote Sensing of Environment, 86: 530–541.

Khan, K.U., J. Yang and W. Zhang. 2007. “Unsupervised classification of polarimetric SAR images by EM algorithm.” IEICE Transactions on Communications, 90(12): 3632-3642.

Kong, J.A., A.A. Swartz, H.A. Yueh, L.M. Novak and R.T. Shin. 1988. “Identification of terrain cover using the optimal polarimetric classifier.” Journal of Electromagnetic Waves and Applications, 2(2):171-194.


85

Krogager, E.. 1992. “Decomposition of the Sinclair Matrix into Fundamental Components with Application to High Resolution Radar Target Imaging.” Part 2 in W.-M. Boerner, et al. (Eds.) Direct and Inverse Methods in Radar Polarimetry, 1459-1478, Dordrecht, The Netherlands: Kluwer Academic Publishers.

Krylov, V.A., G. Moser, S.B. Serpico and J. Zerubia. 2011a. “Enhanced Dictionary-Based SAR Amplitude Distribution Estimation and Its Validation With Very High-Resolution Data.” IEEE Geoscience and Remote Sensing Letters, 8(1):148-152.

Krylov, V.A., G. Moser, S.B. Serpico, and J. Zerubia. 2011b. “Supervised High-Resolution Dual-Polarization SAR Image Classification by Finite Mixtures and Copulas.” IEEE Journal of Selected Topics in Signal Processing, 5(3): 554-566.

Kuan, D.T., A.A. Sawehuk, T.C. Strand and P. Chavel. 1985. “Adaptive Noise Smoothing Filter for Images with Signal-Dependent Noise.” IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-7(2): 165-177.

Lafferty, J., A. McCallum and F. Pereira. 2001. “Conditional random fields: Probabilistic models for segmenting and labeling sequence data.” in Carla E. Brodley, Andrea Pohoreckyj Danyluk (Eds.) International Conference on Machine Learning, 282-289, Williams College, Williamstown, MA, U.S., June 28 - July 1.

Lee, J.-S., 1980. “Digital image enhancement and noise filtering by use of local statistics.” IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-2: 165–168.

Lee, J.-S., 1981. “Refined filtering of images using local statistics.” Computer Graphics and Image Processing, 15:380-389.

Lee, J.-S., M.R. Grunes, and S.A. Mango. 1991. “Speckle reduction in multipolarization and multifrequency SAR imagery.” IEEE Transactions on Geoscience and Remote Sensing, 29: 535–544.

Lee, J.-S., L. Jurkevich, P. Dewaele, P. Wambacq and A. Oosterlinck. 1994a. “Speckle filtering of synthetic aperture radar images: A review.” Remote Sensing Reviews, 8(4): 313-340.

Lee, J.-S., D.L. Schuler, R.H. Lang and K.J. Ranson. 1994b. “K distribution for multi-look processed polarimetric SAR imagery.” In IEEE International Geoscience and Remote Sensing Symposium, 2179–2181, Pasadena, U.S., August 8-12.

Lee, J.-S., M.R. Grunes and G. de Grandi. 1999. “Polarimetric SAR speckle filtering and its implication for classification.” IEEE Transactions on Geoscience and Remote Sensing, 37(5): 2363-2373.

Lee, J.-S., S.R. Cloude, K.P. Papathanassiou, M.R. Grunes and I.H. Woodhouse. 2003. “Speckle filtering and coherence estimation of


86

polarimetric SAR interferometry data for forest applications.” IEEE Transactions on Geoscience and Remote Sensing, 41(10): 2254- 2263.

Lee, J.-S., M.R. Grunes, E. Pottier and L. Ferro-Famil. 2004a. “Unsupervised terrain classification preserving polarimetric scattering characteristics.” IEEE Transactions on Geoscience and Remote Sensing, 42(4): 722- 731.

Lee, J.-S., W.-M. Boerner, D.L. Schuler, T.L. Ainsworth, I. Hajnsek, K.P. Papathanassiou and E. Lünburg. 2004b. “A Review of Polarimetric SAR Algorithms and their applications.” Chinese (Taiwan) Journal of Photogrammetry and Remote Sensing, 9(3): 31-80.

Lee, J.-S., M.R. Grunes, D.L. Schuler, E. Pottier and L. Ferro-Famil. 2006. “Scattering-model-based speckle filtering of polarimetric SAR data.” IEEE Transactions on Geoscience and Remote Sensing, 44(1): 176- 187.

Lee, J.-S., T.L. Ainsworth, J.P. Kelly and C. Lopez-Martinez. 2008. “Evaluation and bias removal of multi-look effect on Entropy/Alpha/Anisotropy in polarimetric SAR Decomposition.” IEEE Transactions on Geoscience and Remote Sensing, 46(10): 3039 – 3052.

Lee, J.-S. and Pottier, E., 2009a. “Overview of Polarimetric Radar Imaging.” Charpter 1 in J.-S. Lee and E. Pottier (Eds.) Polarimetric Radar Imaging: From Basics to Applications. Boca Raton : CRC Press.

Lee, J.-S. and E. Pottier. 2009b. “Polarimetric SAR Speckle Filtering.” Charpter 5 in J.-S. Lee and E. Pottier (Eds.) Polarimetric Radar Imaging: From Basics to Applications. Boca Raton : CRC Press.

Lee, J.-S. and E. Pottier. 2009c. “H/A/a Polarimetric Decomposition Theorem.” Charpter 7 in J.-S. Lee and E. Pottier (Eds.) Polarimetric Radar Imaging: From Basics to Applications. Boca Raton : CRC Press.

Lee, J.-S. and E. Pottier. 2009d. “Selected Polarimetric SAR Applications.” Charpter 10 in J.-S. Lee and E. Pottier (Eds.) Polarimetric Radar Imaging: From Basics to Applications. Boca Raton : CRC Press.

Lei, X., Y. Li, N. Zhao and Y. Zhang. 2010. “Fast segmentation approach for SAR image based on simple Markov random field.” Journal of Systems Engineering and Electronics, 21(1):31-36.

Lewis, A.J. 1968. “Evaluation of Multiple Polarized Radar Imagery for the Detection of Selected Cultural Features.” The U. S. National Aeronautics and Space Administration Technical Letter, NASA-130, 56, October.

Lewis, A.J., H.C. MacDonald and D.S. Simonett. 1969. “Detection of


87

linear cultural features with multi-polarized radar imagery.” In 6th International Symposium on Remote Sensing of Environment, 879-893, Michigan, USA, October 13-16.

Li, W., G. B. Benie, D.C. He, S. Wang, D. Ziou, Q. Hugh and J. Gwyn. 1999. “Watershed-based hierarchical SAR image segmentation.” International Journal of Remote Sensing, 20(17): 3377–3390.

Li, X., E. Pottier, H. Guo and L. Ferro-Famil. 2010. “Urban land cover classification with high-resolution polarimetric SAR interferometric data.” Candaian Journal of Remote Sensing, 36: 236-247.

Lichtenegger, J., J.F. Dallemand, P. Reichart, P. Rehillad, and M. Buchroithner. 1991. “Multi-sensor analysis for land use mapping in Tunisia.” Earth Observation Quarterly, 33:1-6.

Lisini, G., C. Tison, F. Tupin, and P. Gamba. 2006. “Feature fusion to improve road network extraction in high-resolution SAR images.” IEEE Geoscience and Remote Sensing Letters, 3(2): 217- 221.

Liu, Y., A. Wong and P. Fieguth. 2011. “Synthesis of Remote Sensing Label Fields Using a Tree-Structured Hierarchical Model.” IEEE Transaction on Geoscience and Remote Sensing, 49(6): 2060-2070.

Lombardo, P., M. Sciotti, T.M. Pellizzeri and M. Meloni. 2003. “Optimum model based segmentation techniques for multifrequency polarimetric SAR images of urban areas.” IEEE Transaction on Geoscience and Remote Sensing, 41:1959-1975.

Lopes, A., E. Nezry, R. Touzi and H. Laur. 1990a. “Maximum a posteriori speckle filtering and first order texture models in SAR images.” In IEEE International Geoscience and Remote Sensing Symposium, 2409-2412, Washington, D.C., USA, May 20-24.

Lopes, A., B. Touzi and E. Nezry. 1990b. “Adaptive speckle filters and scene heterogeneity.” IEEE Transaction on Geoscience and Remote Sensing, 28(6): 992-1000.

Lopes, A., F. Sery. 1997. “Optimal speckle reduction for the product model in multilook polarimetric SAR imagery and the Wishart distribution.” IEEE Transactions on Geoscience and Remote Sensing, 35(3): 632-647.

Lopes, A., R. Garello and S. Le Hégarat-Mascle. 2008. “Speckle models.” Chapter 5 in H. Maître (Ed) Processing of Synthetic Aperture Radar Images, 87–142, London, UK :John Wiley & Sons.

Lopez-Martinez, C. and X. Fabregas. 2003. “Polarimetric SAR speckle noise model.” IEEE Transactions on Geoscience and Remote Sensing, 41(10): 2232- 2242.

Lopez-Martinez, C. and X. Fabregas. 2008. “Model-Based Polarimetric SAR Speckle Filter.” IEEE Transactions on Geoscience and Remote Sensing, 46(11): 3894-3907.


88

Lu, D. and Q. Weng. 2007. “A survey of image classification methods and techniques for improving classification performance.” International Journal of Remote Sensing, 28(5):823-870.

Lu, D., S. Hetrick and E. Moran. 2010. “Land cover classification in a complex urban-rural landscape with QuickBird imagery.” Photogrammetric Engineering and Remote Sensing, 76(10): 1159-1168.

Magnussen, S., P. Boudewyn and M. Wulder. 2004. “Contextual classification of Landsat TM images to forest inventory cover types.” International Journal of Remote Sensing, 25: 2421–2440.

Mahmoudi, M. and G. Sapiro. 2005. “Fast image and video denoising via nonlocal means of similar neighborhoods.” IEEE Signal Processing Letters, 2:839-842.

Maillard, P., D.A. Clausi and H. Deng. 2005. “Operational map-guided classification of SAR sea ice imagery.” IEEE Transactions on Geoscience and Remote Sensing, 43(12): 2940- 2951.

Manjunath, B. and W. Ma. 1996. “Texture features for browsing and retrieval of image data.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 18(8): 837-842.

Margarit, G., J.J. Mallorquí and L. Pipia. 2010. “Polarimetric Characterization and Temporal Stability Analysis of Urban Target Scattering.” IEEE Transactions on Geoscience and Remote Sensing, 48(4): 2038-2048.

Martinis, S., A. Twele and S. Voigt. 2011. “Unsupervised Extraction of Flood-Induced Backscatter Changes in SAR Data Using Markov Image Modeling on Irregular Graphs.” IEEE Transaction on Geoscience and Remote Sensing, 49(1): 251-263.

Massonnet, D. and J.-C. Souyris. 2008. Imaging with Synthetic Aperture Radar. Lausanne, Switzerland: EPFL Press.

McLachlan, G. and D. Peel. 2000. Finite Mixture Models. New York: John Wiley & Sons.

McNairn, H. and B. Brisco. 2004. “The Application of C-Band Polarimetric SAR for Agriculture: A Review.” Canadian Journal of Remote Sensing, 30(3): 525-542.

Mishra, P., D. Singh, and Y. Yamaguchi. 2011. “Land cover classification of palsar images by knowledge based decision tree classifier and supervised classifiers based on SAR observables.” Progress In Electromagnetics Research B, 30: 47-70.

Moore, E.G. 1969. “Side Looking Radar in Urban Research.” The U. S. Geological Survey Interagency Report, NASA-138, 24, January.

Moon, W.M., G. Staples, D.-J. Kim, S.-E. Park and K.-A. Park. 2010. “RADARSAT-2 and Coastal Applications: Surface Wind, Waterline, and Intertidal Flat Roughness.” Proceedings of the IEEE, 98(5):


89

800-815. Moser, G., J. Zerubia and S.B. Serpico. 2006. “Dictionary-based

stochastic expectation-maximization for SAR amplitude probability density function estimation.” IEEE Transactions on Geoscience and Remote Sensing, 44(1):188-200.

Mountrakis, G., J. Im and C. Ogole. 2011. “Support vector machines in remote sensing: A review.” ISPRS Journal of Photogrammetry and Remote Sensing, 66(3):247–259.

Nagao, M. and T. Matsuyama. 1980. A Structural Analysis of Complex Aerial Photographs. New York: Plenum Press.

Negri, M., P. Gamba, G. Lisini and F. Tupin. 2006. “Junction-aware extraction and regularization of urban road networks in high-resolution SAR images.” IEEE Transactions on Geoscience and Remote Sensing, 44(10): 2962-2971.

Novak, L.M. and M.C. Burl. 1990. “Optimal speckle reduction in polarimetric SAR imagery.” IEEE Transactions on Aerospace and Electronic Systems, 26: 293–305.

Novak, L.M., M.C. Burl and W.W. Irving. 1993. “Optimal polarimetric processing for enhanced target detection.” IEEE Transactions on Aerospace and Electronic Systems, 29: 234–243.

Oliver, C.J..1984. “A model for nonrayleigh scattering statistics,” Optica Acta, 31(6):701–722.

Oliver, C.J. and S. Quegan. 2004. Understanding synthetic aperture radar images, 2nd ed. Raleigh, USA: SciTech Publishing.

Pacifici, F., F. Del Frate, W.J. Emery, P. Gamba and J. Chanussot. 2008. “Urban Mapping Using Coarse SAR and Optical Data: Outcome of the 2007 GRSS Data Fusion Contest.” IEEE Geoscience and Remote Sensing Letters, 5(3): 331-335.

Pal, M. and P.M. Mather. 2005. “Support vector machines for classification in remote sensing.” International Journal of Remote Sensing, 26 (5):1007–1011.

Pal, R. and K. Pal. 1993. “A review on image segmentation techniques.” Pattern Recognition, 26 (9):1277-1294.

Pappas, T.N.. 1992. “An adaptive clustering algorithm for image segmentation.” IEEE Transactions on Signal Processing, 40(4): 901-914.

Park, S.-E. and W.M. Moon. 2007. “Unsupervised Classification of Scattering Mechanisms in Polarimetric SAR Data Using Fuzzy Logic in Entropy and Alpha Plane.” IEEE Transactions on Geoscience and Remote Sensing, 45(8): 2652-2664.

Parrilli, S., M. Poderico, C.V. Angelino and L. Verdoliva. 2012. “A Nonlocal SAR Image Denoising Algorithm Based on LLMMSE Wavelet Shrinkage.” IEEE Transactions on Geoscience and


90

Remote Sensing, 50(2): 606-616. Pellizzeri, T.M.. 2003. “Classification of polarimetric SAR images of

suburban areas using joint annealed segmentation and “H/A/alpha” polarimetric decomposition.” ISPRS Journal of Photogrammetry and Remote Sensing, 58(1-2): 55-70.

Pellizzeri, T.M., P. Gamba, P. Lombardo and F. Dell'Acqua. 2003. “Multitemporal/ multiband SAR classification of urban areas using spatial analysis: statistical versus neural kernel-based approach.” IEEE Transaction on Geoscience and Remote Sensing, 41(10):2338- 2353.

Peterson, F. 1969. “An urban land use study of Lawrence, Kansas using K-Band radar.” In The Utility of Radar and Other Remote Sensors in Thematic Land Use Mapping from Spacecraft: Annual Report, The U.S. Geological Survey Interagency Report, NASA-140, 3845 January.

Poelman, A.J.. 1981. “Virtual polarisation adaptation a method of increasing the detection capability of a radar system through polarisation-vector processing.” Communications, Radar and Signal Processing, IEE Proceedings F, 128(5): 261-270.

Popescu, A.A., I. Gavat and M. Datcu. 2012. “Contextual Descriptors for Scene Classes in Very High Resolution SAR Images.” IEEE Geoscience and Remote Sensing Letters, 9(1): 80-84.

Pun, C. and M. Lee. 2003. “Log-polar wavelet energy signatures for rotation and scale invariant texture classification.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(5): 590-603.

Qi, R.-Y. and Y.-Q. Jin. 2007. “Analysis of the Effects of Faraday Rotation on Spaceborne Polarimetric SAR Observations at P-Band.” IEEE Transaction on Geoscience and Remote Sensing, 45(5):1115-1122.

Qi, Z., A. Yeh, X. Li and Z. Lin. 2012. “A novel algorithm for land use and land cover classification using RADARSAT-2 polarimetric SAR data.” Remote Sensing of Environment, 118:21-39.

Quan, J.J., X.B. Wen, and X.Q. Xu. 2008. “Multiscale probabilistic neural network method for SAR image segmentation.” Applied Mathematics and Computation, 205(2):578–583.

Quartulli, M. and M. Datcu. 2004. “Stochastic geometrical modeling for builtup area understanding from a single SAR intensity image with meter resolution,” IEEE Transactions on Geoscience and Remote Sensing, 42(9): 1996–2003.

Raney, R.K., 1998. Radar Fundamentals: Technical Perspective, Chapter 2 in F.M. Henderson and A.J. Lewis (Eds.) Principles and Applications of Imaging Radar. Vol. 2 of Manual of Remote


91

Sensing, 3rd ed. New York: John Wiley & Sons. Richards, J.A., 1993. Remote Sensing Digital Image Analysis, 80, Berlin,

Germany: Springer-Verlag. Sarker, M.L.R., Y. Ban and J. Nichol. 2008. “Comparison of pixel-based,

object-based and sequential masking classification procedures for land use and land cover mapping using multiple sensor SAR in Sweden.” Asian Journal of Geoinformatics, 8(1):25-30.

Sanjay-Gopal, S. and T. Hebert. 1998. “Bayesian pixel classification using spatially variant finite mixtures and the generalized EM algorithm.” IEEE Transactions on Image Process, 7(7):1014–1028.

Scheiewe, J., L. Tufte and M. Ehlers. 2001. “Potential and problems of multi-scale segmentation methods in remote sensing.” Geo-Informations-Systeme, 6: 34–39.

Shaban, M.A., and O. Dikshit. 2001. “Improvement of classification in urban areas by the use of textural features: the case study of Lucknow City, Uttar Pradesh.” International Journal of Remote Sensing, 4:565–593.

Shackelford, A.K. and C.H. Davis. 2003. “A combined fuzzy pixel-based and object-based approach for classification of high-resolution multispectral data over urban areas.” IEEE Transaction on Geoscience and Remote Sensing, 41(10): 2354- 2363.

Shapiro, L.G. and G.C. Stockman. 2001. Computer Vision, Upper Saddle River, NJ, USA: Prentice–Hall.

Smith, D.M. 1996. “Speckle reduction and segmentation of synthetic aperture radar images,” International Journal of Remote Sensing, 17(11): pp. 2043–2057.

Smits, P.C. and S.G. Dellepiane. 1997. “Synthetic aperture radar image segmentation by a detail preserving Markov random field approach.” IEEE Transaction on Geoscience and Remote Sensing, 35(4): 844-857.

Soergel, U., U. Thoennessen and U. Stilla. 2002. “Utilization of LIDAR DEM for SAR Image Analysis in Dense Urban Areas.” In R. Kalliany, F. Leberl, F. Fraundorfer (Eds.) ISPRS Technical Commission III, Photogrammetric Computer Vision, ISPRS Archives, 34(Part 3B) :255-258, Graz, Austria, September 9-13.

Soergel, U., E. Michaelsen, U. Thoennessen and U. Stilla. 2005. “Potential of high-resolution SAR images for urban analysis.” In M. Moeller, E. Wentz (Eds.) ISPRS WG VII/1 3rd International Symposium Remote Sensing and Data Fusion Over Urban Areas and 5th International Symposium Remote Sensing of Urban Areas, Tempe, AZ, USA, March 14-16.

Soergel, U.. 2010. “Review of Radar Remote Sensing on Urban Areas.” Chapter 1 in U. Soergel (Ed.) Radar Remote Sensing of Urban


92

Areas, Dordrecht, Netherlands: Springer. Sokal, R.R. and B.A. Thomson. 2006. “Population structure inferred by

local spatial autocorrelation: an example from an Amerindian tribal population.” American Journal of Physical Anthropology, 129(1): 121-131.

Solaiman, B., R. Koffi, M.-C. Mouchot and A. Hillion. 1998. “An Information Fusion Method for Multispectral Image Classification Postprocessing.” IEEE Transaction on geoscience and remote sensing, 36(2): 395-406.

Solberg, A.H.S., T. Taxt and A.K. Jain. 1996. “A Markov random field model for classification of multisource satellite imagery.” IEEE Transactions on Geoscience and Remote Sensing, 34:100– 113.

Srivastava, M.S.. 1965. “On the complex Wishart distribution.” Annals of Mathematical Statistics, 36(1):313–315.

Stasolla, M. and P. Gamba 2008 “Spatial Indexes for the Extraction of Formal and Informal Human Settlements from High-Resolution SAR Images.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 1(2) 98-106.

Stuckens, J., P.R. Coppin and M.E. Bauer. 2000. “Integrating contextual information with per-pixel classification for improved land cover classification.” Remote Sensing of Environment, 71: 282–296.

Su, X., C. He, Q. Feng, X. Deng and H. Sun. 2011. “A Supervised Classification Method Based on Conditional Random Fields With Multiscale Region Connection Calculus Model for SAR Image.” IEEE Geoscience and Remote Sensing Letters, (8)3: 497-501.

Sveinsson, J.R. and J. Atli Benediktsson. 2003. “Almost translation invariant wavelet transformations for speckle reduction of SAR images.” IEEE Transactions on Geoscience and Remote Sensing, 41(10): 2404-2408.

Swain, P.H. and S.M. Davis. 1978. Remote Sensing; The Quantitative Approach. New York: McGraw-Hill.

Thiel, M., Esch, T. and Schenk, A., 2008. “Object-Oriented Detection of Urban Areas from TerraSAR-X Data.” In J. Chen, J. Jiang, A. Peled (Eds.) ISPRS 2008 Congress, Technical Commission VIII, ISPRS Archives, 37(Part 8): 23-26, Beijing, China, July 3-11.

Tison, C., J.-M. Nicolas, F. Tupin, and H. Maitre. 2004. “A new statistical model for Markovian classification of urban areas in high-resolution SAR images.” IEEE Transactions on Geoscience and Remote Sensing, 42(10): 2046–2057.

Touzi, R. and A. Lopes. 1994. “The principle of speckle filtering in polarimetric SAR imager.” IEEE Transactions on Geoscience and Remote Sensing, 32(5):1110-1114.

Touzi, R. 2002. “A review of speckle filtering in the context of estimation


93

theory.” IEEE Transaction on Geoscience and Remote Sensing, 40 (11): 2392-2404.

Touzi, R., W.M. Boerner, J.S. Lee and E. Lueneburg. 2004. “A review of polarimetry in the context of synthetic aperture radar: concepts and information extraction.” Canadian Journal of Remote Sensing, 30 (3): 380-407.

Tzotsos, A. and D. Argialas. 2008. “Support Vector Machine Classification for Object-Based Image Analysis.” In T. Blaschke, S. Lang and G. J. Hay (Eds.) Object-Based Image Analysis – Spatial Concepts for Knowledge-Driven Remote Sensing Applications, 663–667. Germany: Springer.

Ulaby, F.T., F. Kouyate, B. Brisco and T.H. Lee Williams. 1986. “Textural information in SAR images.” IEEE Transaction on Geoscience and Remote Sensing, 24(2):235-245.

UN Department of Economic and Social Affairs, 2009. World Urbanization Prospects: The 2009 Revision. Available at: http://esa.un.org/unpd/wup/Documents/WUP2009_Highlights_Final.pdf [Accessed January 19, 2010].

van Zyl, J.J., H.A. Zebker and C. Elachi. 1987. “Imaging radar polarization signatures: Theory and observation.” Radio Science, 22:529–543.

Voisin, A., V.A. Krylov, G. Moser, S.B. Serpico and J. Zerubia. 2013. “Classification of Very High Resolution SAR Images of Urban Areas Using Copulas and Texture in a Hierarchical Markov Random Field Model.” IEEE Geoscience and Remote Sensing Letters, 10(1): 96-100.

Wang, L., W.P. Sousa and P. Gong. 2004. “Integration of object-based and pixel-based classification for mapping mangroves with IKONOS imagery.” International Journal of Remote Sensing, 25(24):5655-5668.

Ward, D., S. Phinn and A. Murray. 2000. “Monitoring Growth in Rapidly Urbanizing Areas Using Remotely Sensed Data.” The Professional Geographer, 52(3): 371-386.

Waske, B. and J.A. Benediktsson. 2007. “Fusion of support vector machines for classification of multisensor data.” IEEE Transaction on Geoscience and Remote Sensing, 45:3858-3866.

Waske, B. and S. van der Linden. 2008. “Classifying multilevel imagery from SAR and optical sensors by decision fusion.” IEEE Transaction on Geoscience and Remote Sensing, 46:1457-1466.

Wegner, J.D., R. Hänsch, A. Thiele and U. Soergel. 2011. “Building Detection From One Orthophoto and High-Resolution InSAR Data Using Conditional Random Fields.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,


94

4(1):83-91. Woodcock, C., A. Strahler and D. Jupp, 1988. “The use of variograms in

remote sensing: I. scene models and simulated images; II. real digital images.” Remote Sensing of Environment, 25:323–348.

Wu, S.T. 1984. “Analysis of synthetic aperture radar data acquired over a variety of land cover.” IEEE Transactions on Geoscience and Remote Sensing, GE-22(6):550- 557.

Wu, S.T. and S.A. Sader. 1987. “Multipolarization SAR data for surface delineation and forest vegetation characterization.” IEEE Transactions on Geoscience and Remote Sensing, GE-25(1): 67-76.

Wu, Y., K. Ji, W. Yu and Y. Su. 2008. “Region-based classification of polarimetric SAR images using Wishart MRF.” IEEE Geoscience and Remote Sensing Letters, 5(4): 668-672.

Xia, Z.-G. 1996. “Applications of multi-frequency, multi-polarization and multi-incident angle SAR systems in urban land use and land cover mapping.” In IEEE International Geoscience and Remote Sensing Symposium, 4: 2310-2314, Lincoln, NE, USA, May 27-31.

Xu, F. and Y.-Q. Jin. 2007. “Automatic reconstruction of building objects from multi-aspect metric-resolution SAR images.” IEEE Transactions on Geoscience and Remote Sensing, 45(7): 2336-2353.

Yang, Y., H. Sun and C. He. 2006. “Supervised SAR Image MPM Segmentation Based on Region-Based Hierarchical Model.” IEEE Geoscience and Remote Sensing Letters, 3(4): 517-521.

Yang, W., L. Chen, D. Dai and G.-S. Xia. 2011. “Semantic labelling of SAR images with conditional random fields on region adjacency graph.” IET Radar, Sonar & Navigation, 5(8): 835-841.

Yin, H., Y. Cao and H. Sun. 2011. “Combining pyramid representation and AdaBoost for urban scene classification using high-resolution synthetic aperture radar images.” IET Radar, Sonar & Navigation, 5(1):58-64.

Yu, Q. and D.A. Clausi. 2006. “Filament Preserving Model (FPM) Segmentation Applied to SAR Sea-Ice Imagery.” IEEE Transaction on Geoscience and Remote Sensing, 44(12): 3687-3694.

Yu, P., A.K. Qin and D.A. Clausi. 2012. “Unsupervised Polarimetric SAR Image Segmentation and Classification Using Region Growing With Edge Penalty." IEEE Transactions on Geoscience and Remote Sensing, 50(4): 1302-1317.

Zebker, H.A. and J.J. van Zyl. 1991. “Imaging radar polarimetry: A review.” Proceedings of. IEEE, 79(11): 1583–1606.

Zhang, L., B. Zou, H. Cai and Y. Zhang. 2008. Multiple-Component Scattering Model for Polarimetric SAR Image Decomposition. IEEE Geoscience and Remote Sensing Letters, 5(4): 603-607.


95

Zhang, L., B. Zou, J. Zhang, and Y. Zhang. 2010. “Classification of Polarimetric SAR Image Based on Support Vector Machine Using Multiple-Component Scattering Model and Texture Features.” EURASIP Journal on Advances in Signal Processing, 2010: 1–9.

Zhang, Q., Y. Ban, J. Liu and Y. Hu. 2011. “Simulation and analysis of urban growth scenarios for the Greater Shanghai Area, China.” Computers, Environment and Urban Systems, 35(2): 126 139.

Zhang, W.G. and Q. Zhang. 2011. “SAR image despeckling combining target detection with improved nonlocal means.” Electronics Letters, 47(12): 724 -725.

Zhong, P. and R. Wang. 2009. “Image segmentation based on Markov random fields with adaptive neighborhood systems.” Optical Engineering, 45(9) 462-465.

Zhong, H., Y. Li and L.C. Jiao. 2011. “SAR Image Despeckling Using Bayesian Nonlocal Means Filter With Sigma Preselection.” IEEE Geoscience and Remote Sensing Letters, 8(4): 809-813.

Zhou, W., G. Huang, A. Troy and M. Cadenasso. 2009. “Object-based land cover classification of shaded areas in high spatial resolution imagery of urban areas: A comparison study.” Remote Sensing of Environment, 113(8): pp. 1769-1777.

Zhu, Z., C. Woodcock, J. Rogan and J. Kellndorfer. 2012. “Assessment of spectral, polarimetric, temporal, and spatial dimensions for urban and peri-urban land cover classification using Landsat and SAR data.” Remote Sensing of Environment, 117: 72-82.

Multitemporal Spaceborne Polarimetric SAR Data for Urban ...567158/FULLTEXT01.pdf · Multitemporal...

Documents

Transcript of Multitemporal Spaceborne Polarimetric SAR Data for Urban ...567158/FULLTEXT01.pdf · Multitemporal...