Urban Change Detection Using Multitemporal SAR Images814876/FULLTEXT01.pdf · unsupervised change...

Urban Change Detection Using Multitemporal SAR Images

Osama Yousif

June 2015

Doctoral Thesis in Geoinformatics

Royal Institute of Technology (KTH) School of Architecture and Built Environment (ABE)

Department of Urban Planning and Environment SE-100 44 Stockholm, Sweden

TRIA-SOM 2015-07 ISSN 1654-2754 ISNR KTH/SoM/2015-07/SE ISBN 978-91-7595-612-1 © Osama Yousif 2015 Doctoral Thesis Geoinformatics Division Department of Urban Planning and Environment Royal Institute of Technology (KTH) SE-100 44 Stockholm, Sweden

i

Abstract

Multitemporal SAR images have been increasingly used for the detection of different types of environmental changes. The detection of urban changes using SAR images is complicated due to the complex mixture of the urban environment and the special characteristics of SAR images, for example, the existence of speckle. This thesis investigates urban change detection using multitemporal SAR images with the following specific objectives: (1) to investigate unsupervised change detection, (2) to investigate effective methods for reduction of the speckle effect in change detection, (3) to investigate spatio-contextual change detection, (4) to investigate object-based unsupervised change detection, and (5) to investigate a new technique for object-based change image generation. Beijing and Shanghai, the largest cities in China, were selected as study areas. Multitemporal SAR images acquired by ERS-2 SAR and ENVISAT ASAR sensors were used for pixel-based change detection. For the object-based approaches, TerraSAR-X images were used.

In Paper I, the unsupervised detection of urban change was investigated using the Kittler-Illingworth algorithm. A modified ratio operator that combines positive and negative changes was used to construct the change image. Four density function models were tested and compared. Among them, the log-normal and Nakagami ratio models achieved the best results. Despite the good performance of the algorithm, the obtained results suffer from the loss of fine geometric detail in general. This was a consequence of the use of local adaptive filters for speckle suppression. Paper II addresses this problem using the nonlocal means (NLM) denoising algorithm for speckle suppression and detail preservation. In this algorithm, denoising was achieved through a moving weighted average. The weights are a function of the similarity of small image patches defined around each pixel in the image. To decrease the computational complexity, principle component analysis (PCA) was used to reduce the dimensionality of the neighbourhood feature vectors. Simple methods to estimate the number of significant PCA components to be retained for weights computation and the required noise variance were proposed. The experimental results showed that the NLM algorithm successfully suppressed speckle effects, while preserving fine

ii

geometric detail in the scene. The analysis also indicates that filtering the change image instead of the individual SAR images was effective in terms of the quality of the results and the time needed to carry out the computation.

The Markov random field (MRF) change detection algorithm showed limited capacity to simultaneously maintain fine geometric detail in urban areas and combat the effect of speckle. To overcome this problem, Paper III utilizes the NLM theory to define a nonlocal constraint on pixels class-labels. The iterated conditional mode (ICM) scheme for the optimization of the MRF criterion function is extended to include a new step that maximizes the nonlocal probability model. Compared with the traditional MRF algorithm, the experimental results showed that the proposed algorithm was superior in preserving fine structural detail, effective in reducing the effect of speckle, less sensitive to the value of the contextual parameter, and less affected by the quality of the initial change map.

Paper IV investigates object-based unsupervised change detection using very high resolution TerraSAR-X images over urban areas. Three algorithms, i.e., Kittler-Illingworth, Otsu, and outlier detection, were tested and compared. The multitemporal images were segmented using multidate segmentation strategy. The analysis reveals that the three algorithms achieved similar accuracies. The achieved accuracies were very close to the maximum possible, given the modified ratio image as an input. This maximum, however, was not very high. This was attributed, partially, to the low capacity of the modified ratio image to accentuate the difference between changed and unchanged areas. Consequently, Paper V proposes a new object-based change image generation technique. The strong intensity variations associated with high resolution and speckle effects render object mean intensity unreliable feature. The modified ratio image is, therefore, less efficient in emphasizing the contrast between the classes. An alternative representation of the change data was proposed. To measure the intensity of change at the object in isolation of disturbances caused by strong intensity variations and speckle effects, two techniques based on the Fourier transform and the Wavelet transform of the change signal were developed. Qualitative and quantitative analyses of the result show that improved change detection accuracies can be obtained by classifying the proposed change variables.

iii

Keywords: Change detection, High resolution, Image denoising, Kittler-Illingworth, MAP-MRF, Multitemporal SAR images, Nonlocal means, Object-based, Otsu, Outlier detection, Remote sensing, SAR speckle, Urban

iv

Sammanfattning

Multitemporala SAR bilder används allt oftare för att upptäcka olika typer av miljöförändringar. Upptäckten av ändringar i urban miljö med hjälp av SAR bilder är komplicerad på grund av den komplexa sammansättningen av stadsmiljön och SAR bildernas speciella egenskaper, till exempel förekomsten av speckle. I denna avhandling undersök hur förändringar i urban miljö kan upptäckas med hjälp multitemporala SAR bilder med följande särskilda mål: (1) att undersöka oövervakade metoder för upptåckt av förändringar, (2) att undersöka effektiva metoder för reduktion av speckle effekter vid detektering av förändringar, (3) att undersöka spatio-contextuella metoder för att upptäcka förändringar, (4) att undersöka objektbaserad oövervakad upptäckt av förändringar, och (5) att utveckla en ny teknik för objektbaserad generering av förändringsbilder. Peking och Shanghai, de största städerna i Kina, valdes som testområden. Multitemporal SAR bilder tagna av ERS-2 SAR och ENVISAT Asar sensorer användes för pixelbaserad upptäckt av förändringar. För objektbaserade metoder användes TerraSAR-X bilder.

I Paper I undersöktes oövervakad upptäckt av urban förändring med hjälp av Kittler-Illingworth algoritm. En modifierad ratio operator som kombinerar positiva och negativa förändringar användes för att konstruera förändringsbilden. Fyra täthetsfunktionsmodeller har prövats och jämförts. Bland dem, uppnådde log-normal och Nakagami ratio modeller de bästa resultaten. Trots de goda resultaten som erhölls med algoritmen, försämras de erhållna resultaten genom förlusten av fint geometriskt detalj i allmänhet. Detta orsakades av användningen av lokala adaptiva filter för minskning av speckle. I Papper II löses detta problem med hjälp av en algoritm som använder den icke-lokala medelvärden brusreduceringsalgoritm (NLM) för minskning av speckle och bevarande av detalj. I denna algoritm uppnåddes brusreducering med hjälp av ett rörligt viktad medelvärde där vikterna är en funktion av likheten mellan små bildfönster som definierats runt varje bildpunkt i bilden. För att minska beräkningskomplexiteten användes principalkomponentanalys (PCA)för att minska dimensionaliteten av grannskapens egenskapsvektorer. Enkla metoder för att uppskatta antalet signifikanta PCA komponenter som ska behållas för viktberäkningar och brusets varians föreslogs. De

v

experimentella resultaten visade att NLM algoritmen framgångsrikt minskar speckle effekter, samtidigt som fina geometrisk detaljer i scenen bevaras. Analysen visar också att det är effektivare att filtrera förändringsbilden istället för de enskilda SAR bilderna vad gäller kvalitet av resultaten och den tid som behövs för att utföra beräkningarna.

Algoritmen som detekterar förändringar baserad på Markov Random Field (MRF) modellen visade begränsad kapacitet att samtidigt behålla fina geometrisk detaljer i stadsområden och motverka effekten av speckles. För att övervinna detta problem, utnyttjar papper III NLM teorin för att definiera en icke-lokal begränsning för pixlar klassificering. Det itererade villkors mode (ICM) schema för optimering av MRF kriterium-funktionen utvidgas till att omfatta ett nytt steg som maximerar en icke-lokal sannolikhetsmodell. Jämfört med den traditionella MRF algoritmen visade de experimentella resultaten att den föreslagna algoritmen var överlägsen att bevara fina strukturella detaljer och effektivt minskade effekten av speckle. Den var mindre känslig för värdet av den kontextparametern och påverkades mindre av kvaliteten på den initiala förändringskartan.

Papper IV undersöker objektbaserad oövervakad förändringsupptäckt med hjälp av bilder från TerraSAR-X över tätorter med mycket hög upplösning. Tre algoritmer, dvs Kittler-Illingworth, Otsu och detektering av avvikande värden, testades och jämfördes. De multitemporala bilderna segmenterades med hjälp av multidate segmenteringsstrategi. Analysen visar att liknande noggrannhet uppnåddes med alla tre algoritmerna. De erhållna noggrannheterna var nära sitt störstmöjliga värde med tanke på den modifierade ratiobilden som var indata. Detta maximum hade dock inget särskilt högt värde. Detta anses bero på att den modifierade ratiobilden inte framhäver skillnaden mellan förändrade och oförändrade områden på ett bra sätt. Därför föreslås i Paper V en ny objektbaserad teknik för att generera förändringsbilden. De stora intensitetsvariationer som associeras med hög upplösning och effekterna av speckle gör objektens medelintensitet till en otillförlitlig egenskap. Den modifierade ratiobilden är därför mindre effektiv i att framhäva kontrasten mellan klasserna. En alternativ representation av förändringsdata föreslogs. För att mäta intensiteten av förändringen på objektet oberoende av störningar orsakade av starka intensitetsvariationer och speckle effekter, två tekniker utvecklades, en teknik baserad på Fourier-

vi

transformen och en baserad på Wavelet-transformen av ändringssignalen. Kvalitativa och kvantitativa analyser av resultatet visar att förbättrade noggrannheter i förändringsupptäckten kan erhållas genom att klassificera de föreslagna ändrings- variablerna.

Nyckelord: Förändringsupptäckt, hög upplösning, brusmonskning i bilder, Kittler-Illingworth, MRF, Multitemporala SAR bilder, icke-lokala medelvärden, objektbaserad, Otsu, upptäckt av avvikande värden, Fjärranalys, SAR fläck, Urban

vii

Acknowledgements

I would like to express my sincere gratitude to my supervisor Professor Yifang Ban, Head of Geoinformatics, Royal Institute of Technology (KTH), for her guidance, support, valuable discussions and thorough comments. I would also like to thank Associate prof Hans Hauska for his detailed comments and suggestions on the thesis.

I would also like to thank my colleagues in the Geodesy and Geoinformatics division for their help, support, and understanding. I have truly enjoyed our lunch-time conversations.

My deepest gratitude goes to my family for their unconditional love and support throughout my life, this thesis is simply impossible without them. Finally, I have to thank my Sudanese friends in Stockholm for making life easier during the last four years.

The research was funded by the Swedish National Space Board through project 'Change Detection in Urban Areas Using Multitemporal Spaceborne SAR and Fusion of SAR and Optical Data (Dnr 159/11, PI: Yifang Ban). This project is also part of the Urbanization project within the ESA/China MOST Dragon 2-3 program. The authors would like to thank the European Space Agency (ESA) for providing the ENVISAT ASAR and ERS-2 SAR data, and thank DLR for providing the TerraSAR-X data.

Osama Yousif

Stockholm, June 2015

ix

Table of Contents

ABSTRACT ......................................................................................... I

SAMMANFATTNING .................................................................... IV

ACKNOWLEDGEMENTS ........................................................... VII

TABLE OF CONTENTS ................................................................. IX

LIST OF FIGURES ........................................................................ XII

LIST OF TABLES ........................................................................... XV

LIST OF ABBREVIATIONS ....................................................... XVI

1. INTRODUCTION ....................................................................... 1

1.1. RESEARCH OBJECTIVES .......................................................... 3 1.2. THESIS STRUCTURE ................................................................ 3 1.3. DECLARATION OF CONTRIBUTIONS ......................................... 5

2. REVIEW OF CHANGE DETECTION TECHNIQUES ........ 7

2.1. CHANGE DETECTION PREPROCESSING .................................... 8 2.2. DERIVING CHANGE VARIABLES FROM MULTITEMPORAL

REMOTE-SENSING IMAGES ................................................................ 9 2.2.1. Driving Change Variables from Multitemporal Optical Images ........................................................................................ 9 2.2.2. Driving Change Variables from Multitemporal SAR Images ...................................................................................... 10

2.3. CHARACTERISTICS OF SAR SPECKLE .................................... 11 2.3.1. SAR Speckle Statistics .................................................. 12 2.3.2. Speckle Filtering ........................................................... 13 2.3.3. Speckle Filtering Strategies for Change Detection ...... 14

2.4. PIXEL-BASED CHANGE DETECTION ...................................... 15 2.4.1. Unsupervised Change Detection Algorithms ............... 15 2.4.2. Supervised Change Detection Algorithms .................... 17 2.4.3. Context-Based Change Detection Algorithms .............. 17

2.5. OBJECT-BASED CHANGE-DETECTION ALGORITHMS ............ 19

x

2.5.1. Segmentation Strategies for Change Detection ............ 19 2.5.2. Objects Comparison and Change Map Generation ..... 20

2.6. IMAGE FUSION FOR CHANGE DETECTION ............................. 22

3. STUDY AREAS AND DATA DESCRIPTION ...................... 24

3.1. STUDY AREAS ....................................................................... 24 3.2. DATA DESCRIPTION .............................................................. 25

4. METHODOLOGY .................................................................... 29

4.1. PIXEL-BASED CHANGE DETECTION ...................................... 29 4.1.1. Automatic Thresholding Algorithms ............................ 30

4.1.1.1. The Kittler-Illingworth Algorithm .................................... 31 4.1.1.2. Otsu Method ...................................................................... 33 4.1.1.3. Thresholding using outlier detection technique ................ 34

4.1.2. The Nonlocal Means Algorithm ................................... 35 4.1.3. A combined MRF and NLM Change Detection Algorithm .................................................................................... 37

4.2. OBJECT-BASED CHANGE DETECTION ................................... 40 4.2.1. Images Segmentation .................................................... 41

4.2.1.1. Multiresolution Segmentation ........................................... 42 4.2.2. Unsupervised Object-Based Change Detection ........... 42 4.2.3. A Novel Object-based Change Image Generation Approach ..................................................................................... 43

4.2.3.1. Supervised Classification of Change Variables ................ 47

5. RESULTS AND DISCUSSION ............................................... 49

5.1. PIXEL-BASED APPROACH ...................................................... 49 5.1.1. Results of the K&I, Otsu, Outlier Detection, MAP-MRF and MRF-NLM Algorithms ......................................................... 49 5.1.2. The Nonlocal Means Algorithm ................................... 55

5.2. OBJECT-BASED APPROACH ................................................... 59 5.2.1. Unsupervised Object-based Change Detection ............ 60 5.2.2. Change Intensity Quantification Technique Based on Change Signal Analysis .............................................................. 61

5.2.2.1. Wavelet-based Change Image Generation ........................ 62 5.2.2.2. Fourier-Based Change Image Generation ......................... 64

xi

5.2.2.3. Change Variables Comparison.......................................... 65 5.2.3. Accuracy Assessment .................................................... 67

6. CONCLUSIONS AND FUTURE RESEARCH ..................... 71

6.1. CONCLUSIONS ....................................................................... 71 6.2. FUTURE RESEARCH ............................................................... 72

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

xii

List of Figures

FIG. 1.1 ORGANIZATION OF THE SELECTED PAPERS ............................... 5 FIG. 3.1 BEIJING: (A) ERS-2 SAR 1998, AND (B) ENVISAT ASAR

2008 ............................................................................................. 26 FIG. 3.2 SHANGHAI: (A) ERS-2 SAR 1999, AND (B) ENVISAT ASAR

2009 ............................................................................................. 26 FIG. 3.3 BEIJING: (A) TERRASAR-X 2008, AND (B) TERRASAR-X 2013

..................................................................................................... 27 FIG. 3.4 SHANGHAI: (A) TERRASAR-X 2008, AND (B) TERRASAR-X

2011 ............................................................................................. 27 FIG. 4.1 PIXEL-BASED METHODOLOGICAL OVERVIEW ......................... 30 FIG. 4.2 FLOW CHART OF UNSUPERVISED CHANGE DETECTION USING

THE KITTLER-ILLINGWORTH ALGORITHM .................................... 32 FIG. 4.3 SCENARIOS TO ELIMINATE THE EFFECT OF SPECKLE NOISE IN

CHANGE DETECTION ANALYSIS .................................................... 36 FIG. 4.4 NONLOCAL MEANS DENOISING FLOW CHART ......................... 37 FIG. 4.5 (A) ICM FRAMEWORK FOR THE OPTIMIZATION OF THE MAP-

MRF CRITERIA, AND (B) FRAMEWORK FOR THE OPTIMIZATION OF

THE PROPOSED MRF-NLM ALGORITHM ...................................... 39 FIG. 4.6 OBJECT-BASED METHODOLOGICAL OVERVIEW ...................... 40 FIG. 4.7 SEGMENTATION STRATEGIES ................................................. 41 FIG. 4.8 AN EXAMPLE MULTITEMPORAL OBJECT (UNCHANGED)

EXTRACTED FROM SHANGHAI DATASET (A) OBJECT DATE I SERIES

( ), (B) OBJECT DATE II SERIES ( ), (C) CHANGE SIGNAL , AND

(D) FIRST DERIVATIVE OF .......................................................... 44 FIG. 4.9 WAVELET DECOMPOSITION OF A CHANGE SIGNAL AND CHANGE

INFORMATION ASSESSMENT ......................................................... 45 FIG. 4.10 FOURIER DECOMPOSITION OF THE CHANGE SIGNAL AND

CHANGE INFORMATION ASSESSMENT ............................................ 46 FIG. 5.1 BEIJING (A) DATE I IMAGE, (B) DATE II IMAGE, (C) MODIFIED

RATIO IMAGE, (D) RATIO MAGE, AND CHANGE MAPS GENERATED

USING (E) K&I: LN, (F) OTSU, (G) OUTLIER DETECTION, (H) MRF-NLM: LN, AND (I) MRF: LN ....................................................... 51

xiii

FIG. 5.2 SHANGHAI (A) DATE I IMAGE, (B) DATE II IMAGE, (C) MODIFIED

RATIO IMAGE, (D) RATIO MAGE, AND CHANGE MAPS GENERATED

USING (E) K&I: LN, (F) OTSU, (G) OUTLIER DETECTION, (H) MRF-NLM: LN, AND (I) MRF: LN ....................................................... 52

FIG. 5.3 BEIJING (A) DATE I, (B) DATE II, AND CHANGE VARIABLES

CONSTRUCTED BASED ON (C) UNFILTERED SAR IMAGES, (D) SAR

IMAGES FILTERED ONCE [5×5 GM], (E) SAR IMAGES FILTERED

TWICE [5×5 GM], (F) SAR IMAGES FILTERED USING PCA-NLM [N: 5×5], (G) CHANGE VARIABLE FILTERED USING PCA-NLM [N: 5×5]

AND (H) CHANGE VARIABLE FILTERED TWICE USING THE MEDIAN

FILTER [5×5] ................................................................................ 56 FIG. 5.4 SHANGHAI (A) DATE I, (B) DATE II, AND CHANGE VARIABLES

CONSTRUCTED BASED ON (C) UNFILTERED SAR IMAGES, (D) SAR

IMAGES FILTERED ONCE [5×5 EL], (E) SAR IMAGES FILTERED

TWICE [5×5 EL], (F) SAR IMAGES FILTERED USING PCA-NLM [N: 5×5], (G) CHANGE VARIABLE FILTERED USING PCA-NLM [N: 5×5]

AND (H) CHANGE VARIABLE FILTERED TWICE USING THE LEE FILTER

[5×5] ............................................................................................ 57 FIG. 5.5 ROCS OF THE ORIGINAL UNFILTERED CHANGE VARIABLE

(BLACK DASHED), CHANGE VARIABLE DERIVED FROM EL FILTERED

IMAGES (BLUE), CHANGE VARIABLE DERIVED FROM GM FILTERED

IMAGES (GREEN), CHANGE VARIABLE DERIVED FROM PCA-NLM

FILTERED IMAGES (BLACK), PCA-NLM FILTERED CHANGE

VARIABLE (RED), AND NLM FILTERED CHANGE VARIABLE

(DASHED CYAN) FOR (A) BEIJING WITH 5X5 NEIGHBORHOOD, AND

(B) SHANGHAI WITH 5X5 NEIGHBORHOOD ................................... 59 FIG. 5.6 BEIJING OBJECT-BASED (A) DATE I IMAGE, (B) DATE II IMAGE,

(C) MODIFIED RATIO IMAGE, AND CHANGE MAPS PRODUCED USING

(D) KITTLER-ILLINGWORTH ALGORITHM, (E) OTSU METHOD, AND

(F) OUTLIER DETECTION TECHNIQUE ............................................ 60 FIG. 5.7 SHANGHAI OBJECT-BASED (A) DATE I IMAGE, (B) DATE II

IMAGE, (C) MODIFIED RATIO IMAGE, AND CHANGE MAPS PRODUCED

USING (D) KITTLER-ILLINGWORTH ALGORITHM, (E) OTSU METHOD, AND (F) OUTLIER DETECTION TECHNIQUE .................................... 61

xiv

FIG. 5.8 BEIJING (A) DATE I IMAGE, (B) DATE II IMAGE, (C) MODIFIED

RATIO IMAGE (MR), CHANGE IMAGES BASED ON (D) TREND

SUBSIGNAL WT.A(3), (E) FIRST FLUCTUATION SUBSIGNAL WT.D(1), AND (F) SECOND FLUCTUATION SUBSIGNAL WT.D(2) .................... 62

FIG. 5.9 BEIJING: UNCHANGED AND CHANGED CLASSES BOXPLOTS

EVALUATED FOR THE (A) WT.A(3)

, (B) WT.D(1)

, (C) WT.D(2)

, AND

(D) WT.D(3)

CHANGE IMAGES ....................................................... 63 FIG. 5.10 BEIJING (A) DATE I IMAGE, (B) DATE II IMAGE, (C) MODIFIED

RATIO IMAGE (MR), CHANGE IMAGES BASED ON (D) FT.1, (E) FT.2, AND (F) FT.4 ................................................................................ 65

FIG. 5.11 BEIJING: UNCHANGED AND CHANGED CLASSES BOXPLOTS

EVALUATED FOR THE (A) FT.1, (B) FT.2, (C) FT.3, AND (D) FT.4

CHANGE IMAGES ........................................................................... 66 FIG. 5.12 BEIJING: CHANGE VARIABLES RECEIVER OPERATING

CHARACTERISTICS CURVE (ROC) ................................................. 67

xv

List of Tables

TABLE 2.1: SAR DATA MODELS .......................................................... 13 TABLE 3.1: MULTITEMPORAL SAR DATASETS ................................... 28 TABLE 5.2: ACCURACY ASSESSMENT OF SHANGHAI RESULTS (K&I,

OTSU, OUTLIER DETECTION, MAP-MRF, AND MRF-NLM) ........ 54 TABLE 5.1: ACCURACY ASSESSMENT OF BEIJING RESULTS (K&I, OTSU,

OUTLIER DETECTION, MAP-MRF, AND MRF-NLM) ................... 54 TABLE 5.3: ACCURACY ASSESSMENT OF BEIJING RESULTS ................. 68 TABLE 5.4: ACCURACY ASSESSMENT OF SHANGHAI RESULTS ............ 69

xvi

List of Abbreviations

ADB Adaptive Boosting

ANN Artificial Neural Network

CVA Change Vector Analysis

DEM Digital Elevation Model

EL Enhanced Lee Filter

EM Expectation Maximization Algorithm

FT Fourier Transform

GM Gamma MAP Filter

GG Generalized Gaussian model

ID Image Differencing

IR Image Ratioing

K&I Kittler-Illingworth Algorithm

LR Logistic Regression

LN Log-normal model

MRF Markov Random Field

MAP Maximum a Posteriori

NR Nakagami ratio Model

NLM Nonlocal means

ND Normal Distribution

PCA Principal Component Analysis

RCS Radar Cross Section

Radar Radio Detection and Ranging

xvii

SRTM Shuttle Radar Topography Mission

SVM Support Vector Machine

SAR Synthetic Aperture Radar

WT Wavelet Transform

WR Weibull ratio Model

1

1. Introduction

It is widely believed that global phenomena such as the earth’s increasing average temperature, deforestation and desertification, melting ice, lake drought, flooding, and so on, are triggered by or at least accelerated by human activity. Human lives and well-being are strongly affected by the occurrence of such types of phenomena; for example, the lives of over 22 million people residing in Niger, Nigeria, Cameroon, and Chad will be radically changed by the drought of Lake Chad. In Darfur, Sudan, people migration (mostly farmers with their families) to cities due to civil war, has reduced the area devoted for agricultural activity. This land use change has caused a significant impact in short-term production (Alix-Garcia et al., 2013). Natural phenomena and man-made conditions are in fact dynamic processes that interact with and affect each other in many different ways (Jensen, 2005). Measuring and documenting these dynamic processes can facilitate a deep understanding of the global changes occurring and, consequently, contribute to the development of better strategies for maintaining and sustaining the planet.

Owing to their fast rate of occurrence and strong negative impact, urban changes have received a great deal of attention recently. In 2008, and for the first time, the number of people living in urban areas exceeded the number of those living in rural areas (United Nations, 2008). In Africa for example, the percentage of people leaving in urban areas increased from 14% in 1950 to 35% in 2000, and expected to be 55% by year 2050 (United Nations, 2008). Two decades ago, fewer than 20% of China’s people lived in urban areas. Today it is approximately 50% and expected to be 60% by the year 2020. It is anticipated that China’s urban population will increase by 350 million by the year 2025 with 219 cities having a population over 1 million compared to 35 cities in Europe (Ban and Yousif, 2012). The high rate of human migration from rural to urban areas places a great deal of stress on large cities’ infrastructure and reduces the amount of land that dedicated to agricultural fields and forests. Quantifying urban change is therefore very important for city planners and decision makers.

Remotely sensed images can play an essential role in mapping and monitoring environmental changes in general and urban changes in

2

particular. Today, hundreds of governmental and commercial remote sensing satellites collect images of the earth surface on a regular basis. Examples include, but not limited to, Landsat, SPOT, WorldView, IKONOS, GeoEye, Hyperion, ERS2-SAR, Envisat-ASAR, ALOS POLSAR, Sentinel, Radarsat, TerraSAR-X, and COSMO-SkyMed. Remote sensing images provide a cost effective, comprehensive and easily accessible source of information about the earth. The extraction of change information from these images is normally carried out using change detection techniques. Change detection is the process of identifying differences in the state of an object or a phenomenon by observing it at different times (Lu et al., 2004). In remote sensing, the objective is to extract change information by comparing brightness values in images of the same geographic area acquired at different times. The basic assumption underlying this technique is that the variation in brightness values owing to change in the state of an object is greater than the variation caused by the sensor’s noise and by atmospheric conditions (Melgani et al., 2002).

Change detection can be carried out using either passive optical images or active radar images. Studies that combine both types of images constructively are also common. Radar images are independent of solar illumination and atmospheric conditions. Radar can penetrate haze, smog and smoke, conditions normally encountered in the urban environments. Change detection using multitemporal SAR images has been increasingly used in various areas of application, including the studies of flooding (e.g., Martinis et al., 2011; Giustarini et al., 2013) and deforestation (e.g., Phua et al., 2008; Servello et al., 2010), as well as damage assessment (e.g., Bovolo and Bruzzone, 2007a; Brunner et al., 2010). In addition, SAR images have also been used in the field of urban change detection (e.g., Grey et al., 2003; Gamba et al., 2006; Liu and Yamazaki, 2011; Du et al., 2012; Qin et al., 2013) .

The detection of urban changes using SAR images is complicated by many factors. Unlike other types of environmental changes (e.g., flooding), urban changes are characterized by the small size of the changed object (e.g., buildings, roads). Moreover, urban scenes are known to be heavily structured. Urban changes are also characterized by their diversity—that is, they involve several types of changes. The problem is further complicated by the existence of SAR speckle, which hampers the extraction of useful information from the

3

multitemporal SAR images. Developing a change detection strategy capable of dealing efficiently and simultaneously with all these issues turns out to be a difficult task. For example, the impact of speckle can be reduced through the application of any available local adaptive filter. However, filtering will also remove the fine geometric detail that normally characterizes urban changes. New developments in SAR sensor technology allow the production of images of better quality, with better spatial, spectral and radiometric resolution. The extraction of change information from these high-quality images requires the development of new suitable techniques.

1.1. Research Objectives

This thesis studies the detection of urban changes using multitemporal SAR images with five different approaches:

[1] To investigate the unsupervised detection of urban changes using multitemporal SAR images. The analysis focuses on the use of the Kittler-Illingworth algorithm.

[2] To investigate the possibility of improving the quality of the change image extracted from the multitemporal SAR images by considering various denoising algorithms and strategies.

[3] To investigate the potential of Markov random field in the context of urban change detection and study the possibility of improving its performance—that is, effective suppression of speckle and preservation of details in the scene.

[4] To investigate the object-based unsupervised detection of urban changes using high spatial resolution images and three automatic thresholding algorithms.

[5] To investigate a new technique for object-based change image generation that focuses on quantifying the dominant change behavior of an object in isolation of irrelevant disturbances.

1.2. Thesis Structure

Chapter 1 provides a short introduction to the thesis, along with the research objectives. A review of change detection techniques is given in Chapter 2. The study areas and the used datasets are described in

4

Chapter 3. The description of the proposed methods appears in Chapter 4. Chapter 5 analyses and discusses the results obtained. Conclusions and suggestions for future research are presented in Chapter 6.

This thesis is based on the following papers, referred to in the text by their Roman numerals:

[1] Y. Ban and Osama Yousif, 2012. “Multitemporal Spaceborne SAR Data for Urban Change Detection in China,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 5, no. 4, pp. 1087 –1094, Aug. 2012.

[2] Osama Yousif and Y. Ban, 2013. “Improving Urban Change Detection From Multitemporal SAR Images Using PCA-NLM,” IEEE Transactions on Geoscience and Remote Sensing, vol. 51, no. 4, pp. 2032–2041, 2013.

[3] O. Yousif and Y. Ban, 2014. “Improving SAR-Based Urban Change Detection by Combining MAP-MRF Classifier and Nonlocal Means Similarity Weights,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 7, no. 10, pp. 4288–4300, Oct. 2014.

[4] Osama Yousif and Y. Ban, 2015. “Object-based Unsupervised Detection of Urban Change Using High-Spatial Resolution SAR Images,” submitted to IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing.

[5] Osama Yousif and Y. Ban, 2015. “A Novel Approach for Object-based Change Image Generation Using Multitemporal High Resolution SAR Images,” submitted to IEEE Transactions on Geoscience and Remote Sensing.

The relationship between the listed papers is graphically illustrated in Fig. 1.1.

5

1.3. Declaration of Contributions

Paper I

Prof. Yifang Ban initiated the research on multitemporal SAR data for change detection, involved in the conception and analysis, and wrote part of the paper. Osama Yousif conceived the ideas of unsupervised change detection, performed the experiments and the analysis, and wrote part of the paper.

Paper II

Osama Yousif conceived and developed the ideas, performed the experiments and the analysis, and wrote the major part of the paper. Prof. Yifang Ban helped to improve the conception and analysis, and wrote part of the paper.

Paper III

Fig. 1.1 Organization of the selected papers

Pixel-based Analysis

Unsupervised Change Detection

Paper I

Object-based Analysis

Unsupervised Change Detection

Paper IV

Change Image Generation

Paper V

Change Variable Speckle Filtering

Paper II

Spatio-contextual Change Detection

Paper III

Change Detection Using Multitemporal SAR Images

6

Osama Yousif developed the ideas, performed the experiments and the analysis, and wrote the major part of the paper. Prof. Yifang Ban initiated the research on MRF for change detection, helped to improve the conception and analysis, and edited the paper.

Paper IV

Osama Yousif developed the ideas, performed the experiments and the analysis, and wrote the paper. Prof. Yifang Ban initiated the idea on object-based change detection using TerraSAR-X data, helped to improve the conception and analysis, and edited the paper.

Paper V

Osama Yousif developed the ideas, performed the experiments and the analysis, and wrote the paper. Prof. Yifang Ban helped to improve the conception and analysis, and edited the paper.

7

2. Review of Change Detection Techniques

The detection of ground changes using multitemporal remotely sensed images is complicated by uncertainties in the measured phenomenon, limitations in the ability of the imaging sensors to document (measure) the ground changes and the noise inherent in the imaging process. The specific characteristics of remotely sensed images including sensor characteristics (radar or optical), resolution (spatial, spectral and radiometric), noise, and types of distortions need to be carefully considered by the selected change detection algorithm.

Several types of change detection algorithms have been developed and tested over the years. These algorithms have focused on different aspects of the process—for example, extraction of detailed from-to change information using post-classification comparison algorithm (e.g., Alphan et al., 2008) , unsupervised change detection (e.g., Bovolo and Bruzzone, 2007b), change detection using multichannel SAR images (e.g., Moser and Serpico, 2009) speckle reduction in the context of change detection (e.g., Dekker, 1998; Yousif and Ban, 2013), change detection using polarimetric SAR images (e.g., Conradsen et al., 2003; Sabry, 2009), spatio-contextual change detection (e.g., Bruzzone and Prieto, 2002; Moser and Serpico, 2009; Yousif and Ban, 2014), the fusion of SAR and optical images for change detection (e.g., Poulain et al., 2011; Du et al., 2012), change detection by combining feature-based and pixel-based techniques (e.g., Gamba et al., 2006), object-based change detection (e.g., Desclée et al., 2006; Bontemps et al., 2008; Qin et al., 2013) and the comparison of multitemporal images (e.g., Bujor et al., 2004; Inglada and Mercier, 2007).

Remote sensing change detection techniques have been applied successfully in various fields of applications, including desertification studies (Dawelbait and Morari, 2012), forest dynamics (Bujor et al., 2001; Vollmar et al., 2013), flood mapping (Martinis et al., 2011), glacier change monitoring (Akbari et al., 2014), the detection of urban changes (Grey et al., 2003; DENG et al., 2009; Bouziani et al., 2010; Baselice et al., 2014) and disaster monitoring (Bovolo and Bruzzone, 2007a; Gamba et al., 2007). However, owing to the complexity of the problem, no single approach exists that can handle all types of change detection problems. Different applications call for different

8

approaches, and different types of data require special considerations. The development of new remote sensing systems that can produce higher quality images (i.e., with better spatial, spectral and radiometric resolution) will also require the development of new techniques.

2.1. Change Detection Preprocessing

For a successful change detection analysis, the multitemporal images should be preprocessed to ascertain that they are spatially and radiometrically comparable. The former is accomplished by image-to-image registration. It insures that corresponding pixels in the images refer to the same geographic location. Image-to-image registration is often carried out by manually selecting ground control points. Automatic techniques, with different levels of success, also exist. Image-to-image registration is particularly difficult when the analysis involves high spatial resolution images, or when the images contain high frequency components (e.g., edges and linear features). Inaccurate image-to-image registration is one of the main source of errors in change detection analysis, and can lead to a significant degradation in accuracy (Dai and Khorram, 1998). Compared to pixel-based, object-based change detection proved to be less sensitive to image-to-image registration errors (Chen et al., 2014, 2012).

The second preprocessing step aims at making the multitemporal images radiometrically comparable. Ideally, a ground object should show the same brightness values if no change has occurred. In reality, measured intensities are to a high degree sensitive to change in acquisition geometry. Moreover, atmospheric conditions have a serious impact on the measured intensity when using optical remote sensing images with their short wavelengths. Absolute atmospheric correction and image-to-image normalization are often used to eliminate the negative impact of the atmosphere and make the multitemporal optical images comparable. Absolute atmospheric correction converts digital numbers to scaled surface reflectance. It requires information about the atmospheric condition during image acquisition, which is not always easy to obtain. Image-to-image normalization consists of the linear transformation of the spectral characteristics of the image to be corrected to match those of a base (or reference) image (Gao et al., 2010). de Carvalho et al. (2006) developed a radiometric normalization technique that searches for

9

highest quality invariant features. Velloso et al. (2002) proposed to use artificial neural network to automatically perform a nonlinear image-to-image normalization. SAR has the advantage of being less affected by atmospheric condition. Nevertheless, SAR images should be calibrated if a proper change detection analysis is to be conducted. SAR image calibration calculates the radar cross section (or backscatter coefficient if area correction is applied). Calibration allows the comparison of radar images with different resolutions, transmitted power and processing gains (Oliver and Quegan, 2004).

2.2. Deriving Change Variables from Multitemporal Remote-Sensing Images

One of the essential steps in change detection analysis is the comparison of multitemporal remote sensing images. The aim of this step is to generate a change image that increases the contrast between changed and unchanged areas. A change map can then be generated by thresholding or classifying the change image using either supervised or unsupervised techniques.

2.2.1. Driving Change Variables from Multitemporal Optical Images

Lu et al. (2004) provided a comprehensive review of the mathematical operators that can be used to compare multitemporal optical remote sensing images—for example, image ratioing (IR), image differencing (ID), and image regression. Berberoglu and Akin (2009) found IR to be effective in reducing topographic effects like variation in illumination and shadowing. However, IR produces relatively poor results compared with ID. Change vector analysis (CVA) is an extension of the concept of image differencing that is particularly tailored to comparing multispectral multitemporal images. He et al. (2011) extended the CVA technique to include textural information layers. Bovolo & Bruzzone (2007b) found that the use of CVA magnitude does not in fact utilize all the information contained in the multispectral multitemporal difference image. They suggested transforming the spectral change vector from the Cartesian to the polar coordinate system, in which they developed rigorous statistical distributions for the magnitude and the direction random variables. Liu et al., (2015) used CVA to extract change information from

10

multitemporal hyperspectral images. In a similar line, Sicong Liu et al., (2015) proposed a sequential CVA technique that uses an iterative hierarchical scheme to extract different kinds of changes from multitemporal hyperspectral images.

Comparison of multitemporal optical remote sensing images can also be carried out in a new transformed space instead of in the raw data domain defined by the observed multitemporal images. A simple example is differencing multitemporal normalized difference vegetation index (NDVI) images, where the measured intensities in each image—that is, red and near infrared values—are first transformed to the NDVI space (Lyon et al., 1998). Similarly, Cakir et al. (2006) transformed each individual image in the multitemporal dataset to a component space using correspondence analysis. The difference image is then constructed in this new space.

2.2.2. Driving Change Variables from Multitemporal SAR Images

Imaging radar produces images using energy in the microwave part of the electromagnetic (EM) spectrum. As the name Radio Detection and Ranging indicates, a radar sensor transmits its own power and measures the distance to and the amount of power reflected from a target on the ground. The strength of the radar return from an object depends on the relative sensor-object geometry, in addition to the physical properties of the object (e.g., dielectric constant, size and roughness). Compared with optical, radar images are affected by different types of distortions (e.g., foreshortening, layover, and variations in the ground resolution along the range direction).

Comparison of multitemporal SAR images is commonly carried out using the ratio operator (Moser and Serpico, 2006; Rignot and van Zyl, 1993; Xiong et al., 2012). Ratio-related operators have also been used to compare SAR images, including the log-ratio (Bazi et al., 2005; Bovolo et al., 2013; Marin et al., 2015) and the normalized mean-ratio (Ma et al., 2012). The ratio image is known for robustness with respect to the SAR multiplicative radiometric errors (Bazi et al., 2005). The detection of changes in SAR ratio images can be performed equally well in regions with high and low levels of intensity (Rignot and van Zyl, 1993). Bujor et al. (2003) compared different types of parameters to quantify temporal changes based on SAR images and found the ratio operator to be especially suitable for the detection of

11

steplike changes. Hou et al. (2014) proposed a Gauss log-ratio operator for the comparison of multitemporal SAR images. This comparison operator applies a low-pass filtering to the logarithmically transformed SAR images. ID is then applied to the filtered images. Hachicha and Chaabane (2010) suggested two different types of change indicators that were developed based on the assumption that SAR amplitudes are Rayleigh distributed. The first indicator (Rayleigh ratio detector) works per pixel and uses first-order statistics, while the second one, the Rayleigh Kullback-Leibler divergence, utilizes higher order statistics for the comparison.

The comparison of multitemporal SAR images can also be carried out using similarity measures. These measures have been used extensively in the field of automatic image-to-image registration as a means of quantifying similarity in the spatial domain. In the context of change detection analysis, given two coregistered images, similarity measures can be used to quantify temporal rather than spatial similarity (Alberga, 2009). The strength of similarity measures lies in the fact that the estimation of the change indicator takes into account the pixel and its neighbourhood in contrast to traditional arithmetic operators, which work per pixel and normally ignore the contextual information (Inglada and Mercier, 2007).

In SAR-based change detection, it is common to transform the SAR change variable (e.g., the ratio image) to a new transformed space. The new space preferably allows for effective mitigation of speckle and preservation of geometric detail. In Bovolo and Bruzzone (2005) for example, the change variable was decomposed into many scale-dependent images using a wavelet transform. Each wavelet transformed image is a tradeoff between detail preservation and noise suppression. A suitable scale is selected to classify each pixel based on global and local statistics. Similarly, Celik (2010a) used a dual-tree complex wavelet transform to decompose the log-ratio image into different scales. The change maps generated at each scale using the expectation maximization algorithm are then combined using logical operators.

2.3. Characteristics of SAR Speckle

One of the most distinguishing characteristics of radar images is a phenomenon known as speckle. Speckle, which is common to all

12

coherent imaging systems, is the salt and pepper appearance of SAR image that occurs because of the existence of many scatterers within the resolution cell (Fletcher et al., 2007). Speckle is true electromagnetic measurement (not noise) that is utilized in many types of radar-image applications, including polarimetry and interferometry. Moreover, speckle is a repeatable phenomenon: the same speckle pattern can be obtained under the same acquisition conditions assuming an unchanged scene. The notion of speckle as noise comes into play only when analysing single-channel SAR images. There it is normally treated as a noise-like quantity multiplying the underlying radar cross section (RCS) (Oliver and Quegan, 2004).

2.3.1. SAR Speckle Statistics

Understanding the statistical characteristics of SAR speckle provides a solid foundation for the development of filtering, segmentation, classification, target detection, and change detection algorithms.

If there are a large number of statistically identical scatterers within the resolution cell, and the measured surface is rougher with respect to the radar wavelength, the speckle is described as fully developed (Lee and Pottier, 2009). For this type of speckle, the phase is uniformly distributed in the range (-π, π), and the measured in-phase and quadrature components are independent and identically distributed Gaussian random variables with zero mean and variance σ/2 (Oliver and Quegan, 2004). This model gives rise to Raleigh amplitude (A) and negative exponential intensity (I) models. Table 2.1 shows the parametric form of these statistical distributions.

Improving the radiometric accuracy through averaging L intensities or amplitudes (i.e., multilooking) gives rise to Gamma and square root Gamma (Nakagami model) distribution, respectively (Table 2.1). Although the abovementioned models have a sound physical basis, several empirical models were found to be useful in describing the statistics of SAR intensity or amplitude over homogenous land covers. Examples includes the log-normal distribution (defined by μz and σz

2 the logarithmic mean and variance, respectively) found to be effective for urban areas, and the Weibull model (defined by the parameters b and c) used in ocean measurements (Oliver and Quegan, 2004). These models are flexible in the sense that they can be used to describe intensity or amplitude in single or multilook images. Table 2.1 shows the parametric form of these models and the observations (i.e., random

13

variables) to which they apply.

2.3.2. Speckle Filtering

Although multilooking can improve the radiometric accuracy at the cost of reduced spatial resolution, it does not completely eliminate the effect of speckle. Several filters have been developed to reconstruct the radar cross section (RCS) from the speckled intensity. Gamma MAP, enhanced Lee, MMSE, Kuan and Frost are among the commonly used traditional speckle filters.

Despite the availability of many filters, speckle filtering is still an active area of research. For example, Xie et al. (2002) developed a SAR speckle filtering algorithm that combines the traditional wavelet denoising technique with Markov random field. Walessa and Datcu (2000) developed a despeckling algorithm that was specifically designed to preserve the textural quality of the image and maintain edges between uniform areas. The proposed approach models texture

Table 2.1: SAR data models

Model Parametric Form Observation

Rayleigh 22

expA

A AP A

Amplitude

Exponential expI

I IP I

Intensity

Gamma 1

expLL

I

I L LIP I

L

Averaged

intensity

Square root Gamma 2 1 22

expLL

A

A L LAP A

L

Averaged

amplitude

Lognormal 2

22

ln1exp

22

z

zz

zP z

z

any

Weibull 1

expcc

c

xz zP z

bb

any

14

areas using Gauss Markov random field. Le et al. (2014) developed an adaptive filtering approach specifically designed for denoising of a time series of SAR images.

Recently, the newly developed nonlocal means denoising algorithm specifically designed for additive noise filtering has been extended to SAR speckle filtering. For example, Deledalle et al. (2009) proposed an iterative weighted maximum likelihood denoising strategy, where a similarity criterion based on SAR specific noise distribution has been developed. Deledalle et al. (2012) investigated different types of similarity criteria and found the generalized likelihood ratio function to be more effective with respect to a high level of Gamma or Poison noise. Along a similar line, Parrilli et al. (2012) successfully modified the powerful block-matching 3D denoising algorithm to account for the non-additive nature of SAR speckle. Chen et al. (2011) extended the nonlocal means algorithm to the polarimetric SAR image case. In their work, the similarity of image patches is evaluated using the complex Wishart distribution, which is often used to model statistics of polarimetric covariance and coherence matrices.

According to Oliver and Quegan (2004), a good reconstruction filter should reduce the speckle in homogeneous regions, preserve image subtle features (e.g. edges), create no artifacts and avoid radiometric distortion. The large number of filters available today illustrates the complexity of the problem and the difficulty of fulfilling all the aforementioned requirements.

2.3.3. Speckle Filtering Strategies for Change Detection

Change detection based on multitemporal SAR images will be affected by the existence of speckle. Dekker (1998) has shown that speckle increases false and missed alarm rates when thresholding the SAR ratio image. In the context of unsupervised change detection Bazi et al. (2005) attributed the increase in the overlap between changed and unchanged classes in the 1D feature space to the existence of speckle.

In change detection using SAR images, the traditional strategy for eliminating the effect of speckle is to filter the individual SAR images before the comparison. Usually, an adaptive filter is iteratively applied to the SAR image until a satisfactory result is obtained. For example, Moser et al. (2006) found that the best change map is obtained by

15

filtering the individual SAR images twice using the Gamma MAP filter. Bazi et al. (2005) proposed an automatic method to estimate the optimum number of filtering iterations to be applied to SAR images. The method is based on analysis of the behavior of a criterion function that is related to the average classification error. Inversely, Dekker (1998) found that ratioing filtered SAR images causes the accumulation of individual filter errors and consequently the degradation of the quality of the change variable. The author suggested filtering the change image (e.g., the ratio image) rather than the individual SAR image. In Yousif and Ban (2013) the negative effects associated with using a local adaptive filters in change detection (e.g., the erosion of fine geometric detail) were addressed by using the nonlocal means denoising algorithm. There, the algorithm was applied directly to the change variable (i.e., the modified ratio image) in order to prevent the accumulation of individual filter errors and to reduce the computational burden.

The effects of speckle can be reduced during the classification phase (i.e., no speckle filtering is required). Moser and Serpico (2009) developed an unsupervised change detection algorithm, in which the use of the MRF model successfully helped in reducing the impact of speckle. In a similar way, Yousif and Ban (2014) extended the traditional MRF-based algorithm to include a nonlocal constraint that helps reducing the negative impact of speckle and maintain fine geometric detail in the image.

2.4. Pixel-Based Change Detection

2.4.1. Unsupervised Change Detection Algorithms

Unsupervised change detection algorithms have the important property that no prior knowledge about the study area (in the form of training data) is required. An unsupervised change detection algorithm normally accepts multitemporal images as input and outputs a binary (or ternary) change map that shows changed versus unchanged areas. The main disadvantage of this type of algorithm is the loss of detailed from-to change information.

There are several techniques for unsupervised change detection using remote-sensing images. Bruzzone and Prieto (2000) developed two unsupervised change detection methods. The first method

16

automatically selects the decision threshold that minimizes the probability of error. The second method analyses the difference image by taking into consideration the spatio-contextual information included in the neighbourhood of each pixel. Celik (2010b) considered unsupervised change detection to be an intensive search for a change mask that optimizes a minimum mean square (MSE) criterion function. The genetic algorithm (GA) is used to search for this optimum mask—that is, the change map. Celik (2009) used the PCA technique to map local neighbourhoods in the difference image to a higher dimensional space defined using nonoverlapping image blocks. The k-means algorithm was then used to automatically separate the changed from the unchanged areas. Moser et al. (2007) developed an unsupervised change detection algorithm based on Markov random field. In their work, the approximate change map necessary to run the MRF-based algorithm was obtained by applying the Kittler-Illingworth algorithm. Similarly, a context-based unsupervised change detection algorithm was proposed by Bruzzone and Prieto (2002), in which a Parzen estimate was used to model the likelihood function of the observations. Hao et al. (2014) proposed an unsupervised change detection approach based on the expectation-maximization-based level set method. Quin et al., (2014) develop an automatic change detection technique that was especially designed for the analysis of a SAR time series.

The minimum error thresholding algorithm proposed by Kittler and Illingworth (1986) has been used extensively in the field of change detection. This algorithm was developed based on Bayesian decision theory. It uses the histogram-fitting technique to estimate the unknown probabilities and an optimum threshold for separating the object from the background in the image. The Kittler-Illingworth algorithm is known to be a fast and effective thresholding tool. Melgani et al. (2002) used the algorithm successfully thresholding a change variable derived from multispectral multitemporal images. To deal with the non-Gaussian nature of the SAR amplitude and intensity images, the minimum-error thresholding algorithm was generalized by Bazi et al. (2005) and Moser and Serpico (2006). There, different types of density-function models suitable for describing the statistics of the changed and unchanged classes in a SAR change image were proposed. Essentially, the algorithm assumes the existence of one object (i.e., one typology of change) and one background. Bazi et al. (2006) successfully applied the algorithm to a case in which more than one typology of change existed in the study area—that is, to a case

17

with more than one threshold. The main drawback of the multithreshold version of the Kittler-Illingworth algorithm is its high computational complexity.

2.4.2. Supervised Change Detection Algorithms

Traditionally, supervised change detection is carried out using post-classification-comparison logic. It consists of classifying each image in the multitemporal dataset independently using the same classification scheme. The detailed from-to change information can then be extracted by comparing the classified images on a pixel-by-pixel basis. Yuan et al. (2005) applied the method to a series of Landsat images in order to study the dynamics of the land-cover change over the Twin Cities Metropolitan Area. Del Frate et al. (2008) used post-classification-comparison to extract change information from multitemporal SAR images. Instead of using the maximum likelihood classifier, the authors used an artificial neural network for the classification of each SAR image. In Castellana et al. (2007) the accuracy of the change detection process is improved by combining supervised post-classification logic with an unsupervised change detection algorithm.

Supervised change detection is not restricted to post-classification comparison logic. For example, Volpi et al. (2013) investigated supervised change detection using two techniques—namely, multidate classification and analysis of difference image. To address the problem of high intraclass variability the authors suggested using a support vector machine (SVM) classifier. Similarly, in Nemmour and Chibani (2006) urban growth in the Algerian capital was extracted from Landsat multitemporal images using SVM classifier.

The main problem with the supervised change detection method is the need for high-quality training data to classify each image in the multitemporal dataset. This turns out to be very difficult to achieve, especially for older images. Many semisupervised change detection algorithms have been developed that require only a limited amount of ground information or limited interaction from the analyst (Moser et al., 2002; Roy et al., 2012).

2.4.3. Context-Based Change Detection Algorithms

Contextual information plays an essential role in image understanding

18

and analysis (Li, 2009). In change detection, spatio-contextual information can be used at different levels of analysis. For example, spatio-contextual information may be utilized at the image-comparison stage; Im and Jensen (2005) used spatio-contextual information to extract change images (i.e., correlation, slope, and intercept images) from multitemporal optical remote sensing images. In an attempt to reduce the negative impact of speckle noise, Gong et al. (2012a) developed a comparison operator (i.e., neighbourhood-based ratio) that uses neighbouring pixels in the comparison process.

The most common use of spatio-contextual information in image analysis, however, is during classification rather than in the image comparison. Ghosh et al. (2007) used a Hopfield-type neural network that takes spatial correlation between neighbouring pixels into consideration to carry out an unsupervised change detection. Markov random field provides a rigorous statistical framework for modeling the contextual information in image analysis. It has been used extensively in the field of image classification (Garzelli, 1999; Tso and Olsen, 2005), in multisource image classification (Cossu et al., 2005; Solberg et al., 1996), and in change detection (Bruzzone and Prieto, 2002; Lorenzo Bruzzone and Prieto, 2000; Kasetkasem and Varshney, 2002).

Moser and Serpico (2009) developed an unsupervised change-detection algorithm specifically tailored for the analysis of multichannel-multitemporal SAR images. In that work the MRF model is used not only to model the spatio-contextual information but also to provide a framework for the fusion of the change information extracted from each SAR ratio channel. Similarly, Moser et al. (2007) developed an MRF-based change detection algorithm for the analysis of multichannel-multitemporal SAR images. In that study, the multichannel change image was transformed using Fisher transformation. Kasetkasem and Varshney (2002) pointed out that transformations such as image differencing, destroy the MRF properties of the image. Instead, they suggested a change detection algorithm in which the observed multitemporal images and the required change map are modeled using MRF. Wang et al. (2013) developed an unsupervised change detection algorithm for multitemporal SAR images that utilizes a triplet Markov field. Baselice et al. (2014) developed an MRF-based change detection algorithm for high resolution complex SAR data. Each complex SAR

19

image in the multitemporal dataset is modelled using Markov random field hyperparameters. The hyperparameters maps are compared using similarity measure.

2.5. Object-Based Change-Detection Algorithms

Although most of the available change detection algorithms work per pixel, the problem has also been approached using object-based logic (Zhong and Wang, 2006; Berberoglu and Akin, 2009; Jovanović et al., 2010; Bin et al., 2013; Rasi et al., 2013). Image segmentation techniques subdivide the image into meaningful homogeneous regions (i.e., objects) based not only on the spectral property, but possibly on the shape, texture, and size properties. In object-based techniques the contextual information is exploited in the segmentation phase rather than during classification or change detection analysis. For high spatial resolution imagery, object-based analysis could also help reducing the computational complexity. Moreover, object-based change detection proved to be less sensitive to images coregistration errors.

The resort to object-based analysis was triggered by the introduction of high spatial resolution imagery. With these images, the use of pixel-based classification or change detection algorithms often leads to low-accuracy and quite noisy results. Zhou et al. (2008) compared pixel-based and object-based post-classification change detection algorithms and found the latter to be more accurate in mapping changed and unchanged areas. Qin et al. (2013) performed supervised object-based change detection using the multidate composite logic.

2.5.1. Segmentation Strategies for Change Detection

Earlier, when only a few efficient segmentation algorithms existed, boundaries extracted from a per-parcel or per-field classification were used to identify the geometrical extent of objects (Blaschke, 2005). Essentially, the change detection analysis is carried out using an old map and a recently acquired image, where the intention is to update an existing geographic database (Bouziani et al., 2010; Walter, 2004). With this approach the demarcated segments/objects are generally not spectrally homogeneous, since the spectral information is not employed in the segmentation process.

20

With the development of several powerful segmentation techniques and software tools entirely devoted to image segmentation (e.g., eCognitionTM), several new segmentation possibilities, in the context of change detection, arose. Remote sensing change detection, by default, involves at least two images that are acquired over the same geographic area at different times. Consequently, there are two possible segmentation strategies. The first one consists of the independent segmentation of each image in the multitemporal dataset. A consequence of this technique is the creation of sliver polygons when comparing the segmented images. The main advantage of this approach is that it allows the inclusion of objects’ geometrical properties and topological relationships in the change detection analysis (Blaschke, 2005). Hazel (2001) developed a change detection framework, where a new image is segmented separately and then compared with an existing site model. The author proposed a segmentation algorithm that combines statistical spectral anomaly detector with competitive region growth extractor.

The second strategy, known as multidate segmentation, dominates the object-based change detection literature (Bontemps et al., 2008; Desclée et al., 2006; Hall and Hay, 2003; Im et al., 2008). It involves the segmentation of all images in the multitemporal dataset in one step. The result is objects that are spectrally, spatially, and temporally homogeneous (i.e., no sliver polygon). Because of the fixed geometrical extent of objects, there is a limited ability to utilize objects’ geometric relationships in the change detection analysis.

2.5.2. Objects Comparison and Change Map Generation

From certain perspective, the difference between pixel and segment representations of an image is merely of the used spatial unit/scale. In light of this view, most pixel-based comparison techniques are readily transferable to the object-based analysis. Instead of pixel intensity, object’s mean intensity can simply be used as a feature for image analysis tasks. Comparison of multitemporal objects can then be carried out using conventional mathematical operators usually used in pixel-based change detection (e.g. differencing, ratioing…etc.). This logic is especially simple to apply when multidate segmentation strategy is used. Hall and Hay (2003) for example, compared the multitemporal objects using the absolute value of differenced image known to cluster both types of changes (i.e., positive and negative

21

changes) together. Nielsen (2007) developed a pixel-based change detection algorithm for multi- and hyperspectral imagery that enhances the change information using multivariate alteration detection (MAD) transformation. Niemeyer et al. (2008) successfully extended this algorithm to the object-based case, where in addition to the spectral information the analysis also considered objects geometric features.

Similar to image comparison techniques, most of pixel-based thresholding/classification algorithms can directly be applied to segmented images. Post-classification comparison logic can be applied to extract change information using object-based paradigm. Walter (2004) and Zhou et al. (2008) successfully extracted detailed change information by comparing images classified using the object-based technique instead of per-pixel classification. Similarly, Anders et al. (2013) applied object-based post-classification comparison to obtain a detailed from-to geomorphological change information using multitemporal LiDAR-based digital terrain model (DTM). Im and Jensen (2005) developed a pixel-based change detection algorithm that extracts the detailed from-to change information using multidate composite image classification logic. Im et al. (2008) successfully extended this algorithm to the objet-based case.

Unsupervised change detection algorithms have also been used extensively in object-based change detection. Hall and Hay (2003) for example, applied the Otsu method to threshold an object-based change image constructed using image differencing technique. Desclée et al. (2006), proposed an automatic object-based change detection algorithm, in which objects are compared using image differencing technique. The final change map is generated using outlier detection technique, where the changed objects are conceived as outliers disturbing the statistical distribution of the unchanged objects. Similarly, Bontemps et al. (2008) developed an automatic algorithm to extract environmental changes from SPOT vegetation time series images. Their main target was to exploit the temporal dependence information in the time series to model intra- and inter-annual changes. To achieve this goal, a covariance matrix was used to model the temporal dependence between the observations. Object’s degree of change was then quantified using the Mahalanobis distance between an object feature vector and a reference that represents the unchanged ideal. To generate a binary change map, an outlier detection technique

22

was used.

2.6. Image Fusion for Change Detection

In remote sensing the ultimate goal of image fusion is to enhance the final product in light of the intended application: visual image interpretation, classification or change detection. Remote sensing image fusion typically involves different types of images—for instance, images with different spatial or spectral resolutions—and in many cases involves images with different modalities (as in the fusion of SAR and optical images).

In change detection analysis, the fusion of SAR and optical data is important from two points of view. First, on many occasions the limited availability of data forces the generation of a change indicator through the comparison of an image pair acquired over the same area but with different modalities. Although the images were acquired with sensors that have different modalities, they are two different representations of the same physical reality and consequently can be compared (Inglada and Giros, 2004). Recently, similarity measures have played an essential role in performing such complicated image comparison. Mercier et al. (2008), for example, successfully used Kullback-Leibler divergence to compare an ERS SAR image with a SPOT image.

Second, single-source multitemporal images (i.e., optical or radar) are known for their limited capacity to provide exhaustive documentation of changes that have occurred on the ground. Unlike optical, SAR images have a very high temporal resolution. The existence of speckle, however, impedes the accurate identification of shapes and edges. Optical images show more detail and allow the detection of sharp edges and region boundaries (Orsomando et al., 2007). Change detection analysis can benefit from the complementary nature of the change information represented by each type of data—that is, by SAR and by optical multitemporal datasets. In an attempt to improve the quality of the binary change map, Bruzzone and Prieto (2000) proposed an unsupervised change detection approach that uses consensus theory to integrate many change variables. Ban et al. (2014) developed a simple algorithm for the fusion of both optical and SAR data. In that work, a multidimensional change image was constructed by combining SAR and optical change variables. An iterative

23

classification strategy is then adopted to separate changed and unchanged classes.

Fusion in remote sensing analysis is not restricted to images with different resolution or modalities; it is also extended to include the fusion of different types of information extracted from the same source. Gong et al (2012b) improved the quality of the final change indicator through the fusion of different change variables extracted from the same multitemporal dataset using different comparison operators. The authors claimed that the mean-ratio image emphasizes changed areas in the scene, while the log-ratio image more reflects the background information, the unchanged areas. Consequently, two change variables were constructed using the mean-ratio and log-ratio operators. Each change variable can then be decomposed into low- and high-frequency components using a wavelet transform. The fusion of the change variables is performed in the wavelet domain, where different rules were developed for the low- and high-frequency components. It is worth mentioning that the same wavelet-based fusion approach was adopted by Ma et al. (2012), where different fusion rules were used. In a similar line, Hou et al. (2014) fused two change images produced using the Gauss log-ratio and log-ratio comparison operators.

24

3. Study Areas and Data Description

3.1. Study Areas

To evaluate the different change detection techniques presented in this thesis, two Chinese mega cities, Beijing and Shanghai, were selected as study sites. In 1978, only 20% of China population lived in urban areas. By the year 2010, the percentage jumped to 50% (Zhang et al., 2011). This population migration to urban areas causes great changes in cities infrastructures and negatively impacted the surrounding environment. As the biggest cities in China, Beijing and Shanghai experienced a rapid change in their urban structures. This high expansion rate makes them attractive candidates for the investigation of urban change detection using remote sensing images.

Beijing, the capital of China, is located in the Northern China plain, the main agriculture production region in the country. It is characterized by its active economy, large population, and its fast growing rate. In 2014 the population of Beijing was estimated to be 20 million, compared with 7 million in 1990 (United Nations, 2008). Corresponding to this rapid increase in population, Beijing has witnessed the swift expansion of its built-up area, mostly at the urban-rural fringe and in the city core. The expansion rate increased rapidly after the economic reform in 1978 (Zhang et al., 2011). The built-up area in Beijing increased from approximately 1100km2 in 1990, to almost 2500km2 in 2010 (Wang et al., 2012). This urban expansion caused a tremendous reduction in areas devoted for agricultural activity. Furthermore, the city suffers from soil contamination and high level of air pollution.

Shanghai is located in the Yangtze River Delta in Eastern China. This flat and fertile plain is a highly productive agricultural region, as well as an area in which rapid urban growth has taken place (Zhang et al., 2009). Shanghai has the largest population (23 million in 2010) among all Chinese cities. Its population was around 8 million in 1990 (United Nations, 2008). Zhang et al. (2009) found that the area of Shanghai has increased from 146.1km2 in 1979 to 1121.3km2 in the year 2007, with expansion rate of 34.8km2/year that was not spatially homogeneous. The development of Shanghai has occurred in two main phases. In the first phase, from 1949 to the mid-1990s, the

25

government policy was aimed at establishing Shanghai as a top Chinese industrial center. In the second phase, Shanghai started to transform gradually from a manufacturing center into an information- and service-oriented port (Zhang et al., 2009). Shanghai's significant economic expansion and corresponding high rates of urbanization have brought rapid changes to this megacity's urban spatial structure and greatly increased the amount of stress, in the form of waste and pollutants, on the ecosystem.

3.2. Data Description

A large collection of bi-temporal radar images has been used to study the temporal changes in Beijing and Shanghai. All the radar images used in the pixel-based approaches (i.e., Paper I, II, and III) were acquired with either an ERS-2 SAR or an ENVISAT ASAR sensor in C-band HH polarization. Fig. 3.1 shows two SAR image acquired in 1998 and 2008, respectively that cover a small area of Beijing. Fig. 3.2 shows two SAR image acquired in 1999 and 2009, respectively over a small area in Shanghai. Table 3.1 shows the acquisition date, sensor name, and the size of the images in each bi-temporal dataset.

For the object-based approaches used in Paper IV and V, two multitemporal datasets that consist of high resolution TerraSAR-X images in X-band HH polarization, were used. Fig. 3.3 shows two TerraSAR-X images acquired in 2008 and 2013, respectively, over small area in Beijing. Fig. 3.4 shows two TerraSAR-X SAR images acquired in 2008 and 2011, respectively, over small area of Shanghai. Between these two acquisitions a substantial amount of change occurred in Shanghai mostly related to new structures constructed in preparation for World Expo 2010. Table 3.1 shows the acquisition date, sensor name, and the size of the images in each bi-temporal dataset.

For Paper I, the images in each bi-temporal dataset were registered using a geocoded Landsat image that covers the corresponding area as a reference. The images were orthorectified using SRTM DEM with 90m resolution. All operations were carried out in PCI Geomatica. The images in Paper II and Paper III were geocoded, orthorectified and calibrated using the European space agency (ESA) software NEST. This software has shown greater capacity to automatically geocode and coregister the SAR images and to calibrate them using

26

the most up-to-date information. The datasets in Paper II, the Beijing and Shanghai datasets, cover smaller areas than those covered in Paper I. This is because the nonlocal means algorithm tested in Paper II is computationally intensive. Paper III focuses on the detection of positive and negative changes in urban areas. It achieves this goal by utilizing both Markov random field theory and ideas from the nonlocal means algorithm. Therefore, it requires even more computational power than the investigation in Paper II. Consequently, smaller datasets with many positive and negative changes have been chosen in Paper III.

(a) (b)

Fig. 3.1 Beijing: (a) ERS-2 SAR 1998, and (b) ENVISAT ASAR 2008

(a) (b)

Fig. 3.2 Shanghai: (a) ERS-2 SAR 1999, and (b) ENVISAT ASAR 2009

27

Paper IV and V focus on object-based change detection analysis. In order to evaluate the performance of the object-based approaches, two multitemporal datasets that cover the cities of Beijing and Shanghai were used. Each dataset consists of a pair of TerraSAR-X images 4000 × 4000 pixels in size. The images are enhanced ellipsoid corrected product that has a very high positional accuracy. Visual analysis of the multitemporal images reveals that image-to-image registration in not required.

(a) (b)

Fig. 3.3 Beijing: (a) TerraSAR-X 2008, and (b) TerraSAR-X 2013

(a) (b)

Fig. 3.4 Shanghai: (a) TerraSAR-X 2008, and (b) TerraSAR-X 2011

28

Table 3.1: Multitemporal SAR Datasets

Paper City Size (Pixels) Sensor Acquisition

Date

I

Beijing 1209 × 1755 ERS-2 SAR 1998-07-19

Envisat ASAR 2008-07-15

Shanghai 2031 × 1521 ERS-2 SAR 1999-09-07


II

Beijing 999 × 883 ERS-2 SAR 1998-09-27


Shanghai 666 × 718 ERS-2 SAR 1999-09-07


III

Beijing 600 × 600 ERS-2 SAR 1998-07-19


Shanghai 600 × 600 ERS-2 SAR 1999-06-29


IV & V

Beijing 4000 × 4000 TerraSAR- X 2008-05-09

TerraSAR- X 2013-05-09

Shanghai 4000 × 4000 TerraSAR- X 2008-08-20

TerraSAR- X 2011-09-16

29

4. Methodology

This chapter describes urban change detection methodologies using multitemporal SAR images. The chapter consists of two main parts. The first part focused on change detection using medium resolution SAR images (i.e., ERS2-SAR and ENVISAT ASAR at 30m resolution), where the analysis was carried out at the pixel level. In particular, three issues were addressed: unsupervised change detection, SAR change variable denoising, and change map generation through spatio-contextual classification.

The second part concentrates on change detection analysis based on very high resolution imagery (i.e., TerraSAR-X at 3m resolution). The analysis was carried out using the object-based paradigm. The use of the object-based approach is due to the need to avoid the creation of noisy change maps and to reduce computational complexity. Here, we investigate object-based unsupervised change detection and propose a new technique for object-based change image generation.

4.1. Pixel-Based Change Detection

Fig. 4.1 shows the general methodological structure followed in this section. As the figure indicates, the three different problems addressed within the context of change detection using SAR images are: automatic thresholding, denoising and spatio-contextual classification. The first research (Paper I), investigates the unsupervised generation of a binary change map using the Kittler-Illingworth algorithm. The existence of speckle complicates the detection of changed areas. Therefore, the SAR images were filtered using the enhanced Lee filter before applying the automatic thresholding algorithm. The second research (Paper II) focuses on the effective elimination of speckle noise, while preserving geometric detail in the scene. The capacity of the nonlocal means algorithm to fulfill these requirements is investigated. The third research (Paper III) concentrates on change map generation through a spatio-contextual classification of the SAR ratio image. To this end, the well-known Markov random field theory is used to model contextual information. To deal with the strong noise component of the SAR images, the MAP-MRF classifier is extended to include a nonlocal constraint motivated by the NLM theory.

30

4.1.1. Automatic Thresholding Algorithms

Thresholding involves the selection of an optimum threshold that can be applied to an image in order to separate an object from a background. It is an active research area in the digital image processing field, and several techniques, with varying degrees of success, have been developed (Huang and Wang, 1995; Kittler and Illingworth, 1986; Otsu, 1975). Thresholding algorithms differ in terms of the assumptions upon which they have been developed. Automatic thresholding techniques have the advantage that no training or prior knowledge about the properties of the object and background are required. The automatic algorithm often uses image content (e.g., its histogram) to estimate an optimum threshold.

The simplicity of automatic thresholding algorithms has encouraged their use in change detection analysis especially based on optical imagery. Single-threshold approaches have often been used for the generation of binary change maps (i.e., unchanged versus changed). However, few studies have considered multiple typology of change,

Thresholding

K&I

Speckle Filtering

Modified Ratio

Denoising

NLM

Orthorectification Coregistration

Modified Ratio

Spatio-contextual classification

MRF

Ratio

Multitemporal SAR images

Fig. 4.1 Pixel-based methodological overview

31

and consequently, adopted multiple-threshold approach (Bazi et al., 2006). A successful application of an unsupervised thresholding technique to change detection should consider the extent to which the characteristics of the submitted image obey the assumptions upon which the algorithm has been developed. For change detection using SAR image, speckle and its statistics are the most important issue that needs to be considered cautiously.

Next, three unsupervised thresholding algorithms—that is, Kittler-Illingworth, Otsu method, and outlier detection technique—are briefly introduced.

4.1.1.1. The Kittler-Illingworth Algorithm

Based on Bayesian decision theory, Kittler and Illingworth (1986) developed an unsupervised thresholding algorithm that uses the histogram-fitting technique to estimate an optimum threshold. In this algorithm, the probability distribution ( )P r of the image to be thresholded r is considered a combination of two different populations that correspond to the object and the background. The normalized histogram ( )ih r of the input image, which consists of L bins [1: ]i L , is considered a good approximation of ( )P r . Any histogram value ir can be used to divide the histogram into two classes. The Kittler-Illingworth algorithm specifies the optimum threshold r̂ as the one that minimizes a criterion function ( )J r that is related to the probability of the classification error. Fig. 4.2 shows the flow chart of unsupervised change detection using the Kittler-Illingworth algorithm. As indicated in Fig. 4.2 (the dashed square), for each possible value of the threshold r the algorithm estimates the unknown PDF’s parameters using the histogram-fitting technique and then evaluates the criterion function accordingly.

For unsupervised change detection, the input image is the change indicator that results from the comparison of the multitemporal SAR images. The comparison of SAR images is normally carried out using the ratio operator (Rignot and van Zyl, 1993). The application of the single-threshold Kittler-Illingworth algorithm to the ratio image implies the existence of a single typology of change in the study area. Changes in urban areas are characterized by the existence of both positive and negative changes. The identification of both types of change therefore requires the selection of two optimum thresholds.

32

Unfortunately, the double-threshold version of the Kittler-Illingworth algorithm is computationally intensive, since the search is carried out in 2D space instead of in the 1D feature space. The algorithm is also known to be sensitive to great variations in the relative sizes of the classes (e.g., the size of the object may be too small compared to the size of the background). To overcome these problems, the modified ratio operator is introduced. Given two SAR images 1X and 2X acquired over the same geographic area at two different times, the modified ratio image r is computed as follows:

1 2

1 2

max ,

min ,

i i

i

i i

X Xr

X X (4.1)

Fig. 4.2 Flow chart of unsupervised change detection using the Kittler-Illingworth algorithm

Change Image Histogram

Thresholding

PDF parameters Criterion function


Speckle filtering

Change Map

For each threshold

Input SAR images

ˆ argmin ( )i

ir

r J r

33

The effect of the modified ratio operator is conceptually similar to that of taking the absolute value of the difference image in the sense that it groups both positive (intensity increase over time) and negative (intensity decrease over time) changes together (Ban and Yousif, 2012). Grouping the positive and negative changes into a single changed class has two advantages. First, having only one changed class, versus the unchanged class, enables the use of the single-threshold Kittler-Illingworth algorithm to separate the classes. Second, the resulting new combined changed class is larger, preventing the mode of the changed class from disappearing inside the tail of the unchanged class. This also entails a larger sample size for the accurate estimation of the changed class statistics.

The application of the Kittler-Illingworth algorithm requires that the form of the density function be known in advance. Originally, the algorithm was developed assuming the classes to be Gaussian. In SAR ratio images, the statistics of the changed and unchanged class are known to be far from normal. In this research, four different density functions, usually used to describe the statistics of the classes in SAR ratio images, are considered: the generalized Gaussian, the log-normal, the Nakagami ratio, and the Weibull ratio models (Bazi et al., 2006, 2005; Moser and Serpico, 2006).

4.1.1.2. Otsu Method

Otsu method is a histogram-based automatic thresholding technique. Unlike Kittler-Illingworth algorithm, however, Otsu method is nonparametric in the sense that no probability density function assumption is required (Otsu, 1975). Otsu method is very popular among the digital image processing community. However, its application in remote sensing change detection based on SAR images is unusual.

Given a change image (i.e., the modified ratio image in 4.1), Otsu method searches for a threshold such that the output classes (i.e., the unchanged and changed classes) are maximally separated, and at the same time, having the smallest possible per class’s variance. This criterion can be accomplished by selecting the threshold r̂ that maximizes the difference between the classes’ means—that is, the between-class variance (Otsu, 1975). This criterion can equivalently be accomplished by searching for a threshold that minimizes the classes’ variances (i.e., within-class variance). The former is preferred

34

since it only requires the computation of the first-order moments (i.e., fewer parameters needs to be estimated).

Kittler-Illingworth and Otsu method differ mainly in terms of being parametric and nonparametric, respectively. Yan (1996) provided a unified mathematical formulation, in which their criterion functions were perceived as a weighted sum of the histogram grey levels. The weights are given by the corresponding cost function. As such, these algorithms are closely related except for the specific form of the cost function adopted by each. Kurita et al. (1992) show that the solution provided by Otsu method is equivalent to maximizing the likelihood of the conditional distribution assuming that the classes are normally distributed with common variance, while Kittler-Illingworth algorithm is equivalent to maximizing the likelihood of the joint distribution assuming normally distributed classes with different variances.

4.1.1.3. Thresholding using outlier detection technique

Outlier is defined “as an observation that deviates so much from other observations as to arouse suspicion that it was generated by a different mechanism (Ben-Gal, 2005)”. Outlier detection techniques can either be parametric or nonparametric. The former makes assumptions about the probability density function of the observations which can be used to identify possible outliers.

For change detection application, changed pixels can be envisaged as outliers that contaminate the probability distribution of the unchanged class. Change detection, therefore, boils down to the identification of region in the feature space, in which the null hypothesis (i.e., a pixel belongs to the unchanged class) will be rejected. Given the parametric form of the probability density function of the unchanged class p(r |u) , and a confidence level 1 - a , it is possible to estimate an optimum threshold ˆˆ ˆ: | (r r)r r p 1- a that separates changed pixels (i.e., outliers

from unchanged pixels.

In an unsupervised context, the density function parameters of the unchanged class are unknown. However, they can be estimated from the available data. Since the data under investigation includes outlying observations, the estimated PDF parameters will be biased. This will negatively affect the threshold determination and outliers detection and removal processes (i.e., masking and swamping effects). To overcome this problem outlier detection and removal are done often in

35

an iterative manner. Another solution suggested by Ben-Gal (2005) is to use statistics robust with respect to outliers (e.g., using the median instead of the mean).

SAR ratio image is known for being far from normal. Therefore, the unchanged class is assumed to be log-normally distributed. This PDF model proves to be suitable for urban change detection using multitemporal SAR images (Ban and Yousif, 2012).

4.1.2. The Nonlocal Means Algorithm

Like traditional neighbourhood filters, the nonlocal means denoising algorithm assumes an image affected by additive noise. Consequently, the NLM algorithm achieves denoising through a weighted averaging process. However, the NLM algorithm is unique in two ways. First, the averaging process extends beyond the local spatial extent to include all pixels in the image; hence the term nonlocal. Second, the averaging weights are a function of the similarity between small image patches, neighbourhood feature vectors, defined around each pixel in the image. As a result, the NLM algorithm tries to force pixels with similar local structures to have similar intensity values.

To compute the weights of the NLM algorithm, each pixel in the image is represented by a neighbourhood feature vector that consists of the neighbouring pixels’ intensities, thus creating a new multidimensional image. The feature vectors effectively capture the local structures in the image. The weight between any two pixels is a function of the similarity of their neighbourhood feature vectors. The local window used to extract the local neighbourhood is typically chosen to be of size 5×5 or 7×7 pixels. This in turn gives rise to a very high dimensional feature vector and, consequently, a high computational load. In Paper II the high dimensionality is reduced by projecting the neighbourhood feature vectors into a new space defined by the eigenvectors of the covariance matrix of the multidimensional image. The first PCA components corresponding to the first significant eigenvalues are then kept for the weight computations. The similarity weights are then evaluated in this low dimensional space. The remaining PCA components are noise-dominated and are therefore discarded from the weight computation. Paper II proposes an algorithm to estimate the number of significant components . In addition, a simple technique for estimating the variance of the noise based on the discarded PCA components is also given. The noise

36

variance is necessary in order to estimate the filter parameter.

To eliminate the effect of speckle noise in the context of change detection using multitemporal SAR images, Paper II investigates two denoising scenarios. The first scenario, shown in Fig. 4.3 (a), consists of constructing the change image using the noisy multitemporal SAR images. The filter is then applied so as to denoise the change variable. Fig. 4.4 shows a flow chart for denoising the SAR modified ratio image using the NLM algorithm. In the second scenario the filter is applied to the individual SAR images, after which the change variable is constructed using the comparison operator, as in Fig. 4.3 (b).

The first scenario has three advantages over the second one. First, it can be efficiently implemented because the filter is applied only once. Second, the quality of the change variable submitted to the filter is better since the use of the ratio/modified ratio reduces to a great extent the multiplicative radiometric errors in the SAR images. Finally, the first scenario, which involves running the filter once, avoids the accumulation of individual filter errors (Dekker, 1998).

In Paper II both scenarios are tested and compared using different filters. Regardless of the denoising strategy employed, the application of the NLM algorithm—either to the individual SAR images or to the

(a)

MR Filter

1I

2I

(b)

Filter

MR

Filter

1I

2I

Fig. 4.3 Scenarios to eliminate the effect of speckle noise in change detection analysis

37

SAR change image—requires logarithmic scaling of the input. This scaling is necessary to convert the multiplicative speckle noise to an additive noise. The algorithm has been developed based on the assumption of additive noise.

4.1.3. A combined MRF and NLM Change Detection Algorithm

Image analysis and understanding can be facilitated by the inclusion of spatio-contextual information (Li, 2009). In image analysis (e.g., classification and change detection), Markov random field (MRF) theory provides a well-established basis for modeling this type of information.

In urban change detection using multitemporal SAR images the ratio image can be considered a combination of three different classes: no

Fig. 4.4 Nonlocal means denoising flow chart

Modified Ratio Image

NLM Denoising

Neighbourhood Feature Vectors

Extraction


Denoised Change Variable

Input SAR images

Similarity Weights Computation

38

change, positive change and negative change. The change detection analysis can then be considered a simple classification problem with three classes. The maximum a posteriori probability (MAP) is one of the most commonly used statistical classification criteria. It simply states that, the best label configuration (i.e., change map) is the one that maximizes the joint posterior probability. Based on Bayes’ theorem, the posterior probability is represented in terms of the conditional distribution of the observations and the prior probability of the class labels.

MRF provides a means of modeling the prior probability component of the model using interaction between pixels’ labels. Hence, the criterion function is normally denoted by MAP-MRF to emphasize this fact. Given the MAP-MRF criterion function, several optimization algorithms exist that can be used to find the optimum classification/change map. Among them, the iterated conditional mode (ICM) algorithm is most frequently used. The ICM algorithm approximates the optimum solution by maximizing the posterior probability locally. Fig. 4.5 (a) shows the workflow for the implementation of the ICM algorithm to optimize the MAP-MRF criterion function. In Paper III, the initial change map required by the ICM algorithm was generated by assuming that percent of the scene has changed between the two acquisition dates. Given the form of the density function, initial values of its parameters could then be estimated. The iteration of the ICM algorithm continues until a certain convergence criterion is met—for example, until the percentage of pixels whose class labels have changed falls below a given threshold.

In change detection using SAR images, the MAP-MRF classifier shows a limited capacity to maintain the geometric detail in the image and to simultaneously eliminate the effect of speckle noise. This is especially true in an urban environment, with its highly complex structures. The problem arises because elimination of the speckle’s effect requires heavily weighting the contextual component of the model. In contrast, detail preservation requires alleviating the effect of this component. To overcome this limitation, a nonlocal probability model based on NLM theory is introduced in Paper III. The basic assumption of this nonlocal model is that pixels with similar local structures—that having similar neighbourhood feature vectors—are more likely to have similar class labels. Unlike the local MRF prior probability model that seeks intracontext homogeneity, the nonlocal

39

probability model strives for intercontext homogeneity. A detailed description of the nonlocal probability model and a method for estimating its value are given in Paper III.

To introduce the constraint imposed by the nonlocal probability model, the ICM implementation scheme is modified as shown in Fig. 4.5 (b). A new step that aimed at maximizing the nonlocal probability is introduced into the work flow just after the posterior probability maximization step. Unlike the traditional algorithm, the new scheme consists of three components: the likelihood function of the observations, the MRF prior probability model and the nonlocal

Fig. 4.5 (a) ICM framework for the optimization of the MAP-MRF criteria, and (b) framework for the optimization of the proposed MRF-NLM algorithm

For Each Iteration

For each pixel Find a class-label that maximizes the local posterior probability

For each pixel Find a class-label that maximizes the nonlocal probability (based on NLM theory)

Estimate the PDF parameters

End(b)

For Each Iteration

For each pixel Find a class-label that maximizes the local posterior probability

End

Estimate the PDF parameters

(a)

40

probability model. The proposed approach uses the same initialization logic as that used in the MAP-MRF algorithm and iterates until a certain convergence criterion is met.

The proposed algorithm will be referred to, in short, as the MRF-NLM algorithm. Note that the new implementation scheme violates the MRF model assumption that the class-labels are independent of the observations. As such the proposed algorithm is not Markovian in the rigorous sense, though MAP-MRF components are still part of the model.

4.2. Object-Based Change Detection

This section focuses on change detection using very high resolution images (i.e., TerraSAR-X). High resolution imagery is characterized by strong intensity variations within the same land-cover class. The

Fig. 4.6 Object-based methodological overview


Speckle Filtering

Multitemporal SAR Images

Segmentation

Change Image Construction

Automatic Thresholding

WT/FT Decomposition

Modified Ratio Change Signal (ƒ)

41

application of pixel-based change detection algorithms will invariably leads to the generation of a noisy change map, and hence, the resort to object-based techniques. Fig. 4.6 shows the general methodological structure followed in this section. As indicated by the figure, two different problems have been addressed within the context of object-based change detection.

The first research (Paper IV) investigates the unsupervised object-based change detection. Three automatic thresholding algorithms (i.e., Kittler-Illingworth algorithm, Otsu method, and outlier detection technique) are tested and compared. The second research (Paper V) proposed a new technique for change image generation. A new representation of the change data is provided. Wavelet and Fourier transforms are used to extract the dominant change behavior for each object.

4.2.1. Images Segmentation

Basically, there are two types of segmentation strategies for change detection applications. The first one consists of segmenting, separately, each image in the multitemporal dataset, before commencing to the

Segmentation

(b)

Segmentation Segmentation

(a)

Fig. 4.7 Segmentation Strategies

42

comparison phase as shown in Fig. 4.7 (a). A direct consequence of this technique is the creation of sliver polygons which might complicate the change detection analysis. With this approach, however, objects can be compared not just in terms of their spectral properties (e.g., object’s mean intensities), but also geometrically (e.g., objects areas and perimeters), and topologically (e.g., overlap).

In the second strategy, known as multidate segmentation (Desclée et al., 2006), all images in the multitemporal dataset are stacked and segmented together in one step as shown in Fig. 4.7 (b). With this technique, the geometric extent of the objects is not subject to change temporally (i.e., no sliver polygons). As a result, the difference between object- and pixel-based change detection is merely of the used spatial scale. Most existing pixel-based change detection techniques can simply be extended to the object-based case. Due to its simplicity and several advantages, this segmentation strategy is widely used in object-based change detection analysis (Desclée et al., 2006; Bontemps et al., 2008; Im et al., 2008).

4.2.1.1. Multiresolution Segmentation

Multiresolution is bottom-up region growing segmentation technique (Benz et al., 2004). It starts at the pixel level (the smallest possible object), and at each iteration neighbouring objects are merged as long as the heterogeneity of the newly created object does not exceed a certain threshold. The heterogeneity criterion involves two sub criterions: color and shape. The color heterogeneity assures that the created objects are spectrally homogeneous, while the shape heterogeneity guarantees that they are adequately compact and smooth (Baatz and Schäpe, 2000). Weights are used to balance the effects of color and shape heterogeneity criteria in the segmentation process.

4.2.2. Unsupervised Object-Based Change Detection

Multidate segmentation approach generates a single segments’ geometrical layout. Several object’s features can then be extracted (e.g., mean intensity, variance etc.). For change detection analysis, object mean intensity is often used as a feature. With this simple segmentation approach pixel- and object-based only differ in term of the used spatial unit, and numerous pixel-based algorithms can be used without more ado.

43

For unsupervised object-based change detection, the modified ratio image can readily be constructed by replacing pixel intensity in Eq. 4.1 with objects’ mean intensities. This approach was adopted in paper IV, where the unsupervised algorithms described previously—that are, Kittler-Illingworth, Otsu, and outlier detection—were used to threshold the resulted object-based change image and produce binary change maps.

4.2.3. A Novel Object-based Change Image Generation Approach

Given the segmented multitemporal images, the next step is to compare the objects and generate a change image. As mentioned above, object mean intensities are often used as a feature, and the comparison is often carried out using mathematical operators (e.g., the modified ratio in 4.1). The problem with the mean, as a measure of central tendency, is that it is quite sensitive to outlier observations. When the analysis is based on high spatial resolution SAR images, existence of unusual intensities within a segment is unavoidable. This can be attributed to the fact that high spatial resolution images (optical or radar) are characterized by increased intensity variations even within the same land cover class. This is more pronounced in urban areas (Schindler, 2012). With SAR images, the intensity variations exacerbates due to existence of speckle. As such, the ability of the mean intensity to accurately characterize the spectral property of an object is questionable. Change images constructed based on biased mean intensities, therefore, will not be precise in accentuating the contrast between unchanged and changed objects.

An alternative approach to assess the intensity of change is to descend to the pixel level within the object (Paper V). Let ir and is be two data series that consist of pixels intensities within object i at date I and date II images, respectively. Using ir and is multitemporal series, a change signal if can be constructed by inserting each intensity value from is immediately after the corresponding intensity from ir as in (4.2).

,0 ,0 ,1 ,1 , 1 , 1[ ]i i i i i i n i nf r ,s ,r ,s ....r ,s (4.2)

44

The new series is referred to as change signal because it contains all the available data that can be used to characterize the intensity of change at the corresponding object. The ratio of mean intensities provides one way to extract this information. From the definition of the change signal (4.2), unchanged/changed status of an object is strongly connected to the roughness of the change signal. That is, unchanged objects are characterized by similar multitemporal series, and consequently, will have a fairly smooth change signal. The reverse is true for changed objects which will have a very rough signal. Essentially, any measure that can evaluate the roughness of the change signal can be used as a change indicator. Texture measures such the variance and the contrast (Oliver and Quegan, 2004) are all possible candidates.

In reality, however, pixel intensities within an object are quite

0 10 20 30 40 500

500

1000

1500

(a) ri

0 10 20 30 40 500

500

1000

1500

(b) si

0 10 20 30 40 50 60 70 80 90 1000

500

1000

1500

(c) fi

0 10 20 30 40 50 60 70 80 90 100−1000

−500

0

500

1000

(d) ∇fi

Fig. 4.8 An example multitemporal object (unchanged) extracted from Shanghai dataset (a) Object date I series ( ), (b) Object date II series ( ), (c) Change signal , and (d) first derivative of

45

heterogeneous due to inherent intensity variations and speckle effect. Fig. 4.8 (a) and (b) show object multitemporal series ir and is , respectively, extracted from Shanghai multitemporal dataset. The figures, with no doubt, reflect the anticipated strong intensity variations within the object. Fig. 4.8 (c) show the change signal created based on the multitemporal series ir and is . From Fig. 4.8 (c), it is clear that not all the information available is actually relevant to change quantification. There are disturbances that preclude a clear conclusion about whether the object has changed or not. For example, some parts of change signal are smooth enough to conclude that the object is unchanged, while other parts show strong variations that might indicate change. This fact is more obvious in the first derivative of the change signal shown in Fig. 4.8 (d).

From the previous discussion, one can conclude that a measure is needed to assess the intensity of change of an object while being insensitive to fluctuations triggered by strong intensity variations and

Fig. 4.9 Wavelet decomposition of a change signal and change information assessment

Wavelet Transform

Change Signal (f) Extraction

f

A(1)

D(1)

D(2)

D(3)

A(3)

A(2)

WT.A(3)

WT.D(1)

WT.D(3)

WT.D(2)

46

speckle effects. For this purpose, two different techniques are proposed. The first approach uses Wavelet transform (WT) to decompose the change signal into detail and fluctuation subsignals. The fraction of energy captured by different subsignals emphasizes different types of information available in the change signal. Fig. 4.9 shows, schematically, a three scale wavelet decomposition of an object change signal. For each object, the fraction of energy captures by each subsignal is used as a change indicator and different change images (i.e., WT.D(1), WT.D(2), WT.D(3), and WT.A(3)) are constructed accordingly. Changed and unchanged objects are characterized by rough and smooth change signals, respectively. In the first fluctuation subsignal (i.e., WT.D(1)), corresponding to the highest frequency content, unchanged and changed objects will show low and high values, respectively. The situation is reversed for the trend subsignal (i.e., WT.A(3)), which emphasizes the lower frequency component of the change signal. Subsignals corresponding to intermediate frequencies (e.g., WT.D(2) and WT.D(3)) are mostly dominated by irrelevant disturbances that impede the accurate characterization of object’s change behavior.

Fig. 4.10 Fourier decomposition of the change signal and change information assessment

Change Signal (f) Extraction

Wavelet Transform

Power Spectrum

Frequency

(1) (2) (3) (4)

FT.1 FT.4 FT.3 FT.2

47

The second approach is based on Fourier transform (FT) of the change signal. FT bears a close resemblance to WT transform. It decomposes the input signal in terms of its frequency content, where temporal/spatial information is lost during transformation—that is, it does not provide outputs equivalent to WT detail and fluctuation subsignals. Nevertheless, by simple post processing step change information can be extracted in a similar way as in the WT case. Fig. 4.10 shows the Fourier power spectrum of a change signal. By dividing the power spectrum into regions of equal length along the frequency axis, the fraction of energy contained within each region can play the same role as that of WT subsignals. For example, the fraction of energy within the lower frequency region FT.1 provides similar information to that reflected by WT.A(3), while FT.4 is equivalent to WT.D(1). By analogy, FT.3 and FT.4 characterize intermediate frequencies content of the change signal and corresponds to unwanted disturbances. As such, using either FT or WT transform it is possible to isolate irrelevant disturbances and extract only the dominant meaningful change characteristics of the object.

4.2.3.1. Supervised Classification of Change Variables

To evaluate the potential of the proposed change variables, four supervised classification algorithms—logistic regression (LR), artificial neural network (ANN), support vector machine (SVM), and AdaBoost.M1 (ADB)—have been applied.

LR is one of the simplest classification algorithms. It is a linear classifier, in which the separating boundary is a linear function of the input observations. ANN is a generalization of LR in the sense that it comprises layers (e.g., input, hidden, and output layers), each of which consists of numerous activation functions (e.g., logistic sigmoid function). This structure allows for a solution with sophisticated decision boundaries. SVM is also a nonlinear classifier that uses kernel functions to map the input features to a higher dimensional space, where the classes can be separated using the largest margin hyperplane. Because of its merits, SVM has been used extensively in change detection (Bovolo et al., 2008; Volpi et al., 2013). ADB belongs to a family of classifiers (i.e., ensemble machine learning), in which the classification accuracy is improved by combining the outputs of many weak classifiers. A decision tree is normally used as a base classifier, although other classifiers can be used too. ADB uses a

48

sequential training strategy, where the training data submitted to the current classifier are weighted according to the performance of the previous classifier.

49

5. Results and Discussion

5.1. Pixel-based Approach

The approaches used in both Paper I (the Kittler-Illingworth algorithm) and Paper III (the MRF-NLM change detection algorithm) eventually generate a change map. Therefore, they are discussed and compared in section 5.1.1. Additionally, the K&I algorithm result is compared with results obtained using two other automatic thresholding techniques (i.e., Otsu method and outlier detection technique). The method investigated in Paper II focuses on using the nonlocal means denoising algorithm to reduce the impact of speckle noise in the context of change detection analysis; section 5.1.2 therefore discusses the NLM algorithm result and compares it with results obtained using different types of local adaptive filters.

5.1.1. Results of the K&I, Otsu, Outlier Detection, MAP-MRF and MRF-NLM Algorithms

Using the modified ratio image as input, the Kittler-Illingworth algorithm selects an optimum threshold that separates the changed class from the unchanged class. To decide whether the detected change in a pixel is positive or negative, the original intensities of the pixel in the multitemporal SAR images are compared, and a decision is made accordingly. Before the application of the K&I algorithm all original SAR images were filtered using the enhanced Lee filter with a 7×7 pixels window size. This was necessary because the existence of speckle impedes the algorithm’s capacity to separate the classes.

The performance of the K&I algorithm is compared with two other automatic thresholding algorithms (Otsu method and outlier detection technique). As in the Kittler-Illingworth case, existence of speckle will impede the capacity to discriminate between the classes. A despeckling step using enhanced Lee filter is therefore necessary before the application of these automatic thresholding algorithms.

To assess the performance of the proposed extended MRF change detection algorithm (MRF-NLM) and to compare it with the traditional MRF-based change detection algorithm (MRF), three different values of the parameter α necessary to construct an initial

50

change map have been tested (15, 20, and 25). This parameter represents a rough estimate of the percentage of the scene that has changed between the two acquisitions. It is used to estimate two thresholds to separate positive and negative changes from the unchanged class. A wide range of values of the contextual parameter β:[2,3,4…12] has been tested. This range extends between the two extreme case scenarios: when β is very small the contextual component has almost no effect in the solution, and the resulting change map is dominated by noise. When β is very large, the contextual component dominates, and an oversmooth change map is generated. In order to model the conditional distribution of the classes, three density functions (i.e., the log-normal, the generalized Gaussian, and the normal distribution) were used in Paper III.

In this section, all the abovementioned algorithms will be tested and compared using the two multitemproal datasets (Beijing and Shanghai multitemporal datasets) used in Paper III. For K&I, outlier detection, MAP-MRF, and MRF-NLM the analysis will only show the results obtained using the log-normal model. For K&I results obtained using other density functions, please refer to Paper I. For MAP-MRF and MRF-NLM evaluation using another density functions, see Paper III.

Fig. 5.1 (a) and (b) show Beijing date I and date II SAR images, respectively. Fig. 5.1 (c) shows the modified ratio image constructed using the Beijing multitemporal dataset. In this change image the brighter the pixel, the higher the probability of being changed, and vice versa. Fig. 5.1 (e) – (g) show the change maps generated using the Kittler-Illingworth algorithm (K&I) with the log-normal model, the Otsu algorithm, and the outlier detection technique, respectively. All these change maps were generated by thresholding the modified ratio image (Fig. 5.1c). In these change maps, blue and white colors indicate positive and negative changes, respectively. Unchanged areas are indicated by the black color. Because of the way the change image is constructed, these algorithms cannot identify positive and negative changes directly. To decide whether the change which has occurred at a pixel is positive or negative, the pixel’s multitemporal intensities are compared and a decision is made. Fig. 5.2 shows the same information for the Shanghai dataset. Fig. 5.1 indicates that the Otsu method identified more area as changed compared with the other two algorithms. Fig. 5.2 shows that both K&I and Otsu algorithms achieved comparable results, while with the outlier detection

51

technique less area has been assigned to the changed class. The first three rows in Tables 5.1 and 5.2 show the accuracy assessment for the Beijing and Shanghai change maps obtained using the three automatic algorithms (i.e., K&I, Otsu, and outlier detection). For the Beijing dataset, Table 5.1 shows that the Otsu method yielded a comparatively high false alarm rate. The best result in terms of kappa coefficient of agreement is achieved using the outlier detection technique, which

Fig. 5.1 Beijing (a) date I image, (b) date II image, (c) modified ratio image, (d) ratio mage, and change maps generated using (e) K&I: LN, (f) Otsu, (g) outlier detection, (h) MRF-NLM: LN, and (i) MRF: LN

(d) (e) (f)

(a) (b) (c)

(g) (h) (i)

52

shows a relatively small false alarm rate. This is also true for Shanghai dataset, where Table 5.2 shows that K&I and Otsu method achieved the highest false alarm rates.

The K&I algorithm is a parametric technique that allows the inclusion of existing knowledge we might have about the statistical distribution of the classes in the thresholding process. Paper I found that the best change maps were obtained using the log-normal and Nakagami ratio

Fig. 5.2 Shanghai (a) date I image, (b) date II image, (c) modified ratio image, (d) ratio mage, and change maps generated using (e) K&I: LN, (f) Otsu, (g) outlier detection, (h) MRF-NLM: LN, and (i) MRF: LN

(d) (e) (f)

(a) (b) (c)

(g) (h) (i)

53

models. The generalized Gaussian model overestimated the changed class; too much area of the scene has been identified as changed, while the Weibull ratio underestimated the changed class—that is, low detection accuracy. The quality of K&I change map will depend on the extent to which the assumed density model fits the statistics of the classes. In fact, studies have shown that very good results could be obtained using generalized Gaussian and Weibull ratio models (Bazi et al., 2005; Moser and Serpico, 2006). Otsu is a nonparametric technique that does not allow the inclusion of any type of knowledge in the thresholding process. The algorithm depends exclusively on the histogram of the input image. The quality of the result depends on how the input image fits the underlying assumptions of the algorithm. For example, the algorithm will give a biased threshold estimate if the classes variances are quite unequal.

Compared with K&I, the outlier detection technique is more flexible in the sense that only assumptions about the statistical distribution of the unchanged class are required. Several theoretical (Nakagami ratio; Moser and Serpico, 2006) or empirical (log-normal and Weibull ratio; Moser and Serpico, 2006, and the generalized Gaussian; Bazi et al., 2005) density models can be used to model the statistics of the unchanged class. Notice, in the K&I case, assumptions about both the unchanged and changed class density model are required. The latter is more difficult to model especially in urban environment where one can expect more than one typology of change. Another merit of the outlier detection technique is that the false alarm (or detection accuracy) can be controlled manipulating the confidence level. Discussions and comparison of these thresholding algorithms can be found in Paper IV.

A negative effect that can be observed from the K&I, Otsu, and outlier detection technique change maps is the bloblike appearance of the detected changed objects. This effect is due to the use of local adaptive filter (enhanced Lee filter) to eliminate speckle noise. Unfortunately, unfiltered speckle noise does not allow an accurate separation of the changed and unchanged classes in the 1D feature space. Despite filtering, it is possible to observe the noisy appearance of the change maps generated using the thresholding techniques. Another filtering iteration, before applying the algorithms, could help removing this problem. However, the cost will be the exacerbation of the bloblike appearance.

54

Table 5.2: Accuracy assessment of Shanghai results (K&I, Otsu, outlier detection, MAP-MRF, and MRF-NLM)

False Alarm

Positive Detection

Negative Detection

Overall Accuracy

Kappa

K&I 12.17 99.29 94.51 92.23 0.88

Otsu 12.17 99.29 94.51 92.23 0.88

Outlier 6.30 97.86 92.83 94.43 0.91

MAP-MRF 5.54 97.14 99.79 96.47 0.94

MRF-NLM 1.63 95.48 98.95 97.85 0.97

Table 5.1: Accuracy assessment of Beijing results (K&I, Otsu, outlier detection, MAP-MRF, and MRF-NLM)

False Alarm

Positive Detection

Negative Detection

Overall Accuracy

Kappa

K&I 1.58 82.87 84.50 91.49 0.86

Otsu 11.97 94.68 97.09 91.71 0.87

Outlier 03.15 85.42 90.07 92.54 0.88

MAP-MRF 9.56 99.07 95.16 93.60 0.90

MRF-NLM 3.36 97.69 95.64 96.66 0.95

Fig. 5.1 (d) shows the Beijing ratio image generated by dividing the date I by the date II image on a pixel by pixel basis. In this image, bright and dark pixels indicate possible negative and positive changes, respectively. Fig. 5.1 (h) shows the Beijing change map obtained using the approach proposed in Paper III (MRF-NLM) assuming the classes to be log-normally distributed. To generate this change map, the parameter α necessary to estimate the initial change map was set to 20, while β was set to 7. Fig. 5.2 (h) shows the same information for the Shanghai dataset. Compared with the corresponding K&I, Otsu, and outlier detection change maps, it is evident that the proposed approach gives better results in terms of geometric detail preservation. The detected changed areas obtained using the MRF-NLM approach do not suffer from the bloblike appearance noticed before. The

55

obtained change map is also less noisy. Note that the ratio image submitted to the MRF-NLM algorithm is constructed using the unfiltered SAR images.

Finally, Fig. 5.1 (i) shows Beijing change map generated using the traditional MAP-MRF change detection algorithm using the same set of parameters used in the MRF-NLM case. The corresponding Shanghai change map is shown in Fig. 5.2 (i). Comparing MRF and MRF-NLM change map, one can see that the latter performed much better in terms of detail preservation and noise suppression. In Paper III, the traditional MRF algorithm was found to be sensitive to the chosen density function model (e.g., a noisier change map was obtained when the GG mode was used). Similar behavior (i.e., result sensitivity to the chosen density model) was observed when the K&I results were evaluated. The MRF-NLM results look better regardless of the chosen density function. The good performance of the proposed approach, MRF-NLM, is attributable to the fact that MRF-NLM uses three components—conditional probability of the observation, MRF prior probability, and the nonlocal interaction between pixel labels—to determine the correct class of a pixel. In fact, quantitative analysis of the MRF-NLM results, compared with traditional MRF results, shows that the proposed algorithm is less sensitive to the value of the contextual parameter β, and to the initial change map. Tables 5.1 and 5.2 show that among all the techniques tested in this section, MRF-NLM achieved the best results. Further discussion and evaluation of the MRF and MRF-NLM algorithms can be found in Paper III.

5.1.2. The Nonlocal Means Algorithm

The NLM denoising algorithm was tested against different types of adaptive filters: enhanced Lee (EL), Gamma MAP (GM), Median and additive Lee. Two window sizes (i.e., 5×5 and 7×7 pixels) were tested to extract the neighbourhood feature vectors of the NLM algorithm. The local adaptive filters were also tested using two window sizes (i.e., 5×5 and 7×7 pixels). The search window size of the NLM algorithm was set to 21×21 pixels, which is the most commonly used size (Buades et al., 2005; Tasdizen, 2009). Two different despeckling strategies were tested and evaluated in the context of change detection: (1) comparing the images and then filtering the change variable and (2) filtering the SAR images then comparing them. Fig. 5.3 (a) and (b) show Beijing date I and date II SAR images, respectively. Fig. 5.3 (c)

56

– (h) show the Beijing change variables obtained from unfiltered SAR images, SAR images filtered once [5×5 GM], SAR images filtered twice [5×5 GM] and SAR images filtered using PCA-NLM [N: 5×5], respectively. In all these cases the modified ratio operator was applied after filtering the individual SAR images. The first four images in Fig.

Fig. 5.3 Beijing (a) date I, (b) date II, and change variables constructed based on (c) unfiltered SAR images, (d) SAR images filtered once [5×5 GM], (e) SAR images filtered twice [5×5 GM], (f) SAR images filtered using PCA-NLM [N: 5×5], (g) change variable filtered using PCA-NLM [N: 5×5] and (h) change variable filtered twice using the Median filter [5×5]

(c) (d) (e)

(a) (b)

(f) (g) (h)

57

5.4 show the same information for the Shanghai dataset, in which the EL filter was used instead of the GM filter.

Despite the noisy appearance, boundaries and linear features are clearly visible in the change variable in Fig. 5.3 (c) and Fig. 5.4 (c). As soon as the local adaptive filter is applied, whether it is the GM or

Fig. 5.4 Shanghai (a) date I, (b) date II, and change variables constructed based on (c) unfiltered SAR images, (d) SAR images filtered once [5×5 EL], (e) SAR images filtered twice [5×5 EL], (f) SAR images filtered using PCA-NLM [N: 5×5], (g) change variable filtered using PCA-NLM [N: 5×5] and (h) change variable filtered twice using the Lee filter [5×5]

(c) (d) (e)

(a) (b)

(f) (g) (h)

58

the EL filter, the noisy appearance of change variables decreases significantly, albeit with serious negative effects. Good examples of these negative effects are the existence of residual black and white flecks (see the yellow circle in Fig. 5.3 [d]), the lift or increase in intensity of unchanged areas surrounding a changed object, the fusion of nearby bright objects, and the blurring of narrow new linear features. Some of these negative effects are exacerbated by running additional adaptive filtering iterations. In contrast, the modified ratio image of the PCA-NLM filtered SAR images (Fig. 5.3 [f] and Fig. 5.4 [f]) is to a great extent free from these negative effects. However, one can still notices that narrow linear features are slightly blurred.

A superior result is obtained when the PCA-NLM algorithm is applied directly in order to filter the noisy change variable, as shown in Fig. 5.3 (g) and 5.4 (g). It is visually obvious that the result in this particular case is less affected by the shortcomings mentioned before. To summarize, the quality of the change variable improves significantly when the PCA-NLM algorithm is used. Filtering the change variable directly, as in the first scenario, using the PCA-NLM algorithm proved not only computationally efficient but also superior in terms of the quality of the results. Fig. 5.3 (h) shows the Beijing change variable filtered twice using the Median filter with a 5×5 pixel window size. Fig. 5.4 (h) shows the Shanghai change variable filtered twice using the additive Lee filter with a 5×5 pixel window size. Apparently, the performance under these generic filter types is very poor and incomparable with that obtained using the PCA-NLM algorithm. More details can be found in Paper II.

The receiver operating characteristic curve (ROC) plots the detection accuracies as function of the false alarm rates. It provides an effective tool to quantitatively evaluate the quality of the change variable independent of the classification/thresholding algorithm. Fig. 5.5 shows the ROC curve of the raw change variable constructed based on the unfiltered SAR images (black dashed), change variable obtained from images filtered once using EL (blue), change variable obtained from images filtered once using GM (green) and PCA-NLM filtered change variable (red) for Beijing dataset (a), and for Shanghai dataset (b). The figure clearly indicates that the PCA-NLM denoised change variable outperformed all change variables constructed based on images filtered using local adaptive filters (e.g., GM and EL). The figure also show that the first denoising scenario—that is, filtering the

59

0 10 20 30 40 5050

55

60

65

70

75

80

85

90

95

100

Fasle Alarm Rate

De

tect

ion

Acc

ura

cy

(a) (b)

0 10 20 30 40 5050

55

60

65

70

75

80

85

90

95

100

Fasle Alarm Rate

De

tect

ion

Acc

ura

cy

Fig. 5.5 ROCs of the original unfiltered change variable (black dashed), change variable derived from EL filtered images (blue), change variable derived from GM filtered images (green), change variable derived from PCA-NLM filtered images (black), PCA-NLM filtered change variable (red), and NLM filtered change variable (dashed cyan) for (a) Beijing with 5x5 neighborhood, and (b) Shanghai with 5x5 neighborhood

change variable using PCA-NLM—attains better result compared with the second denoising scenario—that is, filtering the individual SAR images followed by the comparison stage (black line). More details can be found in Paper II.

5.2. Object-based Approach

This section shows results obtained using multitemporal high spatial resolution SAR images, where the analysis is carried out using the object-based approach. Paper IV investigates the unsupervised object-based change detection, where three thresholding algorithms are considered. In Paper V, a new technique to generate the change image is proposed. Additionally, binary change maps are generated using four supervised classification algorithms.

In order to evaluate the performance of the object-based change detection, two bi-temporal datasets that cover the cities of Beijing and Shanghai were used. To reduce speckle negative impact on subsequent analysis, the images were filtered using enhance Lee filter with 7×7

60

pixels window size. The Multiresolution algorithm (eCognitionTM) was then used to segment the multitemporal images.

5.2.1. Unsupervised Object-based Change Detection

Fig. 5.6 (a) – (c) show Beijing date I, date II, and the corresponding modified ratio image, respectively. Fig. 5.6 (d) – (f) show Beijing change maps generated using the Kittler-Illingworth algorithm, Otsu method, and outlier detection technique, respectively. Blue and black colors in these change maps indicate changed and unchanged areas, respectively. Fig. 5.7 shows the same information as Fig. 5.6 but for the Shanghai dataset. From the figures, one can see the great similarity between the change maps produced using three techniques. This is interesting given the different assumptions upon which these algorithms have been developed. For example, unlike the Otsu method, the Kittler-Illingworth algorithm and the outlier detection technique are parametric—that is, they make assumption about the statistical

(a) (b) (c)

(d) (e) (f)

Fig. 5.6 Beijing object-based (a) date I image, (b) date II image, (c) modified ratio image, and change maps produced using (d) Kittler-Illingworth algorithm, (e) Otsu method, and (f) Outlier detection technique

61

distribution of the classes.

The comparison of these change maps with optical images (Google EarthTM historical imagery) reveals that many false alarms are due to small objects of nonpermanent nature (e.g., cars), which happened to exist in the area during image acquisitions. The visual analysis also shows that the use of multitemporal SAR images causes the detection of uninteresting changes from the urban point of view. For instance, the agricultural field, indicated by the yellow ellipse, appears dark in the first image (Fig. 5.6a) and bright in the second image (Fig. 5.6b) due to different moisture conditions. As a result, the field has been identified as changed, although from the urban change detection perspective it hasn’t.

5.2.2. Change Intensity Quantification Technique Based on Change Signal Analysis

Following segmentation, the change signal for each object is extracted.

(a) (b) (c)

(d) (e) (f)

Fig. 5.7 Shanghai object-based (a) date I image, (b) date II image, (c) modified ratio image, and change maps produced using (d) Kittler-Illingworth algorithm, (e) Otsu method, and (f) Outlier detection technique

62

The change information is then quantified using traditional measures of textures (i.e., the variance and the contrast of the change signal). Change information can also be extracted using wavelet or Fourier transforms. These last techniques allow the extraction of the dominant change behavior of an object in isolation of unrelated disturbances. Consequently, these two techniques will be the focus of discussion in this section.

5.2.2.1. Wavelet-based Change Image Generation

For each object, the change signal was decomposed using a three-level wavelet transform and the proportion of energy within each subsignal was used to construct a change image. Fig. 5.8 (a) – (c) show date I, date II, and corresponding modified ratio images, respectively, for a small area of Beijing. Fig. 5.8 (d) – (f) show the change images generated based on the proportion of energy contained in the trend subsignal (WT.A(3)), first fluctuation subsignal (WT.D(1)), and the

(a) (b) (c)

(d) (e) (f)

Fig. 5.8 Beijing (a) date I image, (b) date II image, (c) Modified ratio image (MR), change images based on (d) trend subsignal WT.A(3), (e) first fluctuation subsignal WT.D(1), and (f) second fluctuation subsignal WT.D(2)

63

second fluctuation subsignal (WT.D(2)), respectively. The WT.A(3) change image shown in Fig. 5.8 (d) gives a good overview of occurred change in the scene, though it looks like an inverted version of the MR change image in Fig. 5.8 (c). This is expected since unchanged objects are characterized by smooth change signal and consequently load heavily on the trend subsignal. The changed objects show exactly the opposite behavior. Fig. 5.8 (e) shows the WT.D(1) change image based on the first fluctuation subsignal. Compared with the multitemporal images, this image efficiently reflects the change that has occurred on the scene. Fig. 5.8 (f) shows the change image based on the second fluctuation subsignal (i.e., WT.D(2)). This change image emphasizes fluctuations that are not strong enough to be interpreted as change, or weak enough to be an indication of unchanged. This image represents disturbances caused by speckle and strong intensity variations, and does not reveal any useful change information. The WT.D(3) change

Fig. 5.9 Beijing: unchanged and changed classes boxplots evaluated for the

(a) WT.A(3)

, (b) WT.D(1)

, (c) WT.D(2)

, and (d) WT.D(3)

change images

Unchanged Changed

(a)

WT.A(3)

Unchanged Changed

(b)

WT.D(1)

Unchanged Changed

(c)

WT.D(2)

Unchanged Changed

(d)

WT.D(3)

64

image looks similar and is not shown for the sake of brevity.

Fig. 5.9 shows the boxplots of the change variables developed based on a WT decomposition of the change signal. Unchanged and changed training samples have been used to construct these boxplots. The boxplot is a visual tool that helps assessing the discrimination power of a given random variable. In the boxplot, the red line corresponds to the class’s median, while the upper and lower edges of the blue box correspond to the 25th and 75th percentiles, respectively. The whiskers (black horizontal lines) set the limits beyond which an observation is considered to be an outlier. From Fig. 5.9 (a) and (b), one can see that the unchanged and changed classes are quite separable under WT.A(3)

and WT.D (1) change variables. It is also possible to observe that with WT.D(1), the changed class shows no outlaying observations. As for WT.D(2) and WT.D(3) change variables, obtained from the second and third fluctuation subsignals, the classes experience strong overlap in feature space. These observations confirm the fact that useful change information exists either in the lowest frequency or the highest frequency subsignals. Mid-frequency subsignals are dominated by disturbances and do not provide useful information.

5.2.2.2. Fourier-Based Change Image Generation

Following the Fourier transformation of objects’ change signals, the frequency domain is divided into four equal segments and the normalized energy within each segment is calculated. The result is four different change images. Fig. 5.10 (a) – (c) show Beijing date I, date II, and the modified ratio image, respectively. Fig. 5.10 (d) shows the change image constructed based on the fraction of power in the first quadrant (i.e., FT.1). This change image is conceptually similar to WT.A(3)—that is, smooth signal (i.e., unchanged objects) will load heavily, while rough signal (i.e., changed objects) will have smaller contribution. Consequently, this change image looks like a negative of the MR image. Other than having inverted grey color shades, this change image seems to be informative and consistent with the expected change. The FT.2 change image (Fig. 5.10e) is based on proportion of power in the mid-frequency ranges, and therefore, emphasizes objects with moderate intensity of change that represent irrelevant disturbances. The FT.3 change image looks similar and is not shown here for brevity. Fig. 5.10 (f) shows the FT.4 change image constructed based on the proportion of power within the fourth power

65

spectrum section, which corresponds to the highest frequency region. In this frequency range, objects with rough signal (i.e., changed objects) appear bright (the brightness depends on the roughness of the signal) while smooth unchanged objects appear darker. Comparing this change image with Fig. 5.10 (a) – (c) one can see that this change image did a good job reflecting the changes that have occurred in the scene.

Fig. 5.11 shows the boxplots of the four change variables developed based on the distribution of power in the Fourier transform. The figure undoubtedly indicates the poor separability under FT.2 and FT.3, and the great discrimination power of FT.1 and FT.4. Compared with FT.1, one can also see that FT.4 achieved slightly better results (i.e., smaller classes overlap in the feature space).

5.2.2.3. Change Variables Comparison

In this section different types of change variables are compared quantitatively. The comparison includes the WT.D(1) change image

(a) (b) (c)

(d) (e) (f)

Fig. 5.10 Beijing (a) date I image, (b) date II image, (c) Modified ratio image (MR), change images based on (d) FT.1, (e) FT.2, and (f) FT.4

66

that results from the wavelet decomposition of the change signal and the FT.4 change image obtained from its Fourier transform. These change variables will be referred to as WT and FT, respectively. The comparison also considers the contrast (CN) of the change signal which uses all the information available at the change signal including unwanted disturbances. The traditional modified ratio (MR) image will be used as a benchmark.

Fig. 5.12 shows the ROC curves of the four change variable assessed using the training sample data for Beijing dataset. From the figure, the best performance (i.e., higher detection accuracy with lower false alarm rate) is achieved when using the FT (i.e., green line) and WT (i.e., blue line) change variables. The MR image shows inferior discrimination power. For example, at 10% false alarm rate, the detection accuracy under the WT and FT is 80%, while that of the MR variables is approximately 70%. CN achieved the poorest performance

Fig. 5.11 Beijing: unchanged and changed classes boxplots evaluated for the (a) FT.1, (b) FT.2, (c) FT.3, and (d) FT.4 change images

Unchanged Changed

(a)

FT.1

Unchanged Changed

(b)

FT.2

Unchanged Changed

(c)

FT.3

Unchanged Changed

(d)

FT.4

67

among all change variables. This was expected as no disturbances isolation is considered in the computation of this change variable.

5.2.3. Accuracy Assessment

Given the modified ratio image, three different binary change maps were generated automatically using Kittler-Illingworth, Otsu, and outlier detection techniques. Four different supervised classification algorithms—that are, the logistic regression (LR), artificial neural network (ANN), support vector machine (SVM), and AdaBoost.M1—were used to classify the modified ratio image, Wavelet-based change variable (i.e., WT), Fourier-based change variable (i.e., FT), and the contrast (i.e., CN) change variable.

Table 5.3 and 5.4 shows the accuracy assessment of Beijing and Shanghai change maps, respectively. From these tables, using the modified ratio image, supervised classifiers and thresholding algorithms achieved equivalent results in terms of kappa coefficient of

Fig. 5.12 Beijing: change variables receiver operating characteristics curve (ROC)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.70.3

0.4

0.5

0.6

0.7

0.8

0.9

1

False alarm

Det

ectio

n ac

cura

cy

MRFTWTCN

68

agreement. This is true except for the LR algorithm, which attained the lowest kappa coefficient in both datasets. For the Beijing dataset, the supervised classifiers achieved higher detection accuracies and understandably, higher false alarm rates than the unsupervised algorithms. This fact is less pronounced in the Shanghai dataset.

From Table 5.3, the WT achieved the highest kappa coefficient of

Table 5.3: Accuracy assessment of Beijing results

False Alarm

Detection Accuracy

Overall Accuracy

Kappa Coefficient

MR - K&I 13.74 73.81 80.06 0.60

MR - Otsu 18.32 79.32 80.50 0.61

MR - Outlier 13.74 72.32 79.32 0.59

MR - LR 11.08 65.33 77.17 0.54

MR - ANN 19.50 78.27 79.39 0.59

MR - SVM 18.02 77.23 79.61 0.59

MR - ADB 21.86 82.14 80.13 0.60

WT - LR 10.34 76.34 83.02 0.66

WT - ANN 13.59 81.70 84.06 0.68

WT - SVM 13.15 81.40 84.14 0.68

WT - ADB 12.41 80.65 84.14 0.68

FT - LR 10.19 74.85 82.36 0.65

FT - ANN 25.55 88.99 81.69 0.63

FT - SVM 10.19 75.00 82.43 0.65

FT - ADB 10.34 75.74 82.73 0.65

CN - LR 13.15 66.82 76.87 0.54

CN - ANN 25.55 84.38 79.39 0.59

CN - SVM 24.96 82.29 78.65 0.57

CN - ADB 29.10 88.24 79.54 0.59

69

among all the tested change variables. In comparison with MR, the WT variable shows 8% improvement in the kappa coefficient. The FT variable achieved slightly lower accuracies, although still better than those obtained using MR. As for the Shanghai dataset (Table 5.4), WT and FT achieved quite equivalent kappa coefficients. Compared with MR, the table shows that a 5% improvement in accuracy was achieved

Table 5.4: Accuracy assessment of Shanghai results

False Alarm

Detection Accuracy

Overall Accuracy

Kappa Coefficient

MR - K&I 11.14 85.58 87.27 0.74

MR - Otsu 7.17 82.04 87.60 0.75

MR - Outlier 8.07 82.72 87.47 0.75

MR - LR 2.82 67.89 82.98 0.66

MR - ANN 11.91 85.03 86.61 0.73

MR - SVM 11.14 84.49 86.74 0.73

MR - ADB 11.01 84.49 86.81 0.74

WT - LR 6.02 80.27 87.34 0.75

WT - ANN 15.24 94.01 89.25 0.79

WT - SVM 9.86 89.66 89.91 0.80

WT - ADB 10.76 89.93 89.58 0.79

FT - LR 4.87 76.05 85.88 0.72

FT - ANN 12.55 90.88 89.12 0.78

FT - SVM 9.73 88.98 89.64 0.79

FT - ADB 9.99 89.93 89.97 0.80

CN - LR 10.37 67.76 79.02 0.58

CN - ANN 16.39 79.05 81.40 0.63

CN - SVM 16.26 75.37 79.68 0.59

CN - ADB 23.69 84.08 80.08 0.60

70

when using either the WT or the FT change variable. The proposed change variables are considered to be better because of their ability to quantify the dominant change behavior of the object regardless of irrelevant disturbances. The worst results for both datasets are obtained when classifying the CN change variables. This change indicator uses all the information available at the change signal including outlying observations.

Regarding the supervised classifiers performance, one can observe that LR constantly achieved the worst result. SVM and ADB obtained results that are very similar in terms of the detection accuracies and false alarm rates. The latter, however, has a simple structure (i.e., the decision tree has been used as a base classifier). Moreover, it is very fast in learning and classification of new data instances. ANN achieved comparable results, although it shows an exceptionally high false alarm rate (25%) when classifying the FT change variable (Table 5.3).

Finally, one can observe that for the Beijing dataset, the obtained accuracies are lower compared to those obtained for the Shanghai dataset. This can be attributed to the fact that a large proportion of the change that has occurred in the Beijing study site is due to the replacement of old illegal settlements with new modern building. Unfortunately, the change intensity induced by this change was not strong enough to be detected.

71

6. Conclusions and Future Research

6.1. Conclusions

The research found that the detection of urban changes using multitemporal SAR images could be carried out efficiently using the Kittler-Illingworth algorithm (Paper I). The use of the modified ratio operator causes the clustering of positive and negative change together, allowing in turn the separation of the changed and unchanged classes using a single threshold. Moreover, it causes the prior probability of the changed class to increase which helped improving the performance of the algorithm. One of the algorithm’s shortcomings, however, is the difficulty of deciding in advance which density model to use. A density model that can be adjusted to suit different types of change detection problems might solve this problem. The result also indicated that the use of a local adaptive filter would cause the removal of fine geometric detail in the scene. This fact reveals the need to improve the quality of the change image submitted to the automatic thresholding algorithm.

The nonlocal means algorithm provided a good alternative to the traditional adaptive filters. It showed a great ability to reduce the impact of speckle, while maintaining the geometric structures in the scene (Paper II). The high computational complexity of the algorithm was reduced by using PCA transformation. To make the algorithm as user independent as possible, simple techniques to estimate the dimensionality of the new space and the unknown noise variance were developed. Quantitative and qualitative analyses of the results showed that the NLM algorithm outperformed the traditional local adaptive filters in terms of speckle elimination and fine-detail preservation. Furthermore, the experiment showed the direct denoising of the change image instead of the individual SAR images is not only computationally efficient but also superior in term of the quality of the result.

The theory behind the nonlocal means algorithm was further used to extend the MAP-MRF classifier in order to force pixels with similar local structures to have similar class labels (Paper III). To address the limitations associated with the MRF-based change detection technique when applied to SAR images, the optimization scheme of the MAP-

72

MRF criterion function using the ICM algorithm is modified. A new step based on the maximization of the nonlocal probability model was inserted. Compared with the Kittler-Illingworth and the traditional MRF algorithms, the proposed MRF-NLM algorithm produced a result that is better in terms of speckle suppression and fine geometric detail preservation. The main disadvantage of this algorithm is its high computational complexity. If details preservation is not required, then unsupervised algorithms can be used to obtain a good result in a reasonable time.

In change detection using high spatial resolution images, the use of object-based approach is dictated by the need to avoid the creation of a noisy change map and reduce the computational complexity. The unsupervised object-based change detection using three automatic thresholding algorithms (i.e., the Kittler-Illingworth, Otsu, and outlier detection technique) revealed that limited accuracies can be achieved (Paper IV). This was partially attributed to the way the change image was constructed (i.e., image comparison process). Consequently, Paper V focused on the development of a new intensity of change assessment technique. An alternative data representation (i.e., the change signal) is proposed. Intensity of change quantification boils down to the measurement of the roughness of the change signal. Two techniques are developed to measure the intensity of change at the object in isolation of disturbances caused mostly by the strong intensity variation and speckle effects—one based on Fourier transform and the other on Wavelet transform of the change signal. Analysis of the result shows that a satisfactory improvement has been achieved. Since change information is extracted by descending to the pixel level within each object and performing either Wavelet or Fourier transform, the proposed techniques are computationally demanding.

6.2. Future Research

Due to the lack of access to multitemporal polarimetric data, this research was focused on multitemporal single-frequency, single-polarization SAR images. Both medium (i.e., ENVISAT ASAR and ERS2-SAR) and high (i.e., TerraSAR-X) spatial resolution images were investigated. Although satisfactory results were obtained for mapping urban changes, future research is needed to address the

73

following issues:

[1] To overcome the limited information contained in single-frequency single-polarization images, future research will investigate the use of multitemporal polarimetric SAR images. Despite their many advantages, such data present a new type of challenge that requires the development of suitable change-detection algorithms.

[2] Instead of the conventional approach where each date is represented by a single image, future research will investigate change detection in which each date is represented by more than one image. Images acquired in different incidence angles, or at different seasons could help improving the quality of the change detection process.

[3] Finally, bearing in mind the limitations of multitemporal SAR or optical images in monitoring every types of change occurring on the ground, we will investigate the possibility of improving the quality of urban change detection through the fusion of SAR and optical data.

75

References

Akbari, V., Doulgeris, A., Eltoft, T., 2014. Monitoring Glacier Changes Using Multitemporal Multipolarization SAR Images. IEEE Transactions on Geoscience and Remote Sensing 52, 3729–3741. doi:10.1109/TGRS.2013.2275203

Alberga, V., 2009. Similarity Measures of Remotely Sensed Multi-Sensor Images for Change Detection Applications. Remote Sensing 1, 122–143. doi:10.3390/rs1030122

Alix-Garcia, J., Bartlett, A., Saah, D., 2013. The landscape of conflict: IDPs, aid and land-use change in Darfur. Journal of Economic Geography 13, 589–617.

Alphan, H., Doygun, H., Unlukaplan, Y.I., 2008. Post-classification comparison of land cover using multitemporal Landsat and ASTER imagery: the case of Kahramanmaraş, Turkey. Environ Monit Assess 151, 327–336. doi:10.1007/s10661-008-0274-x

Anders, N.S., Seijmonsbergen, A., Bouten, W., 2013. Geomorphological Change Detection Using Object-Based Feature Extraction From Multi-Temporal LiDAR Data. IEEE Geoscience and Remote Sensing Letters 10, 1587–1591. doi:10.1109/LGRS.2013.2262317

Baatz, M., Schäpe, A., 2000. Multiresolution Segmentation: an optimization approach for high quality multi-scale image segmentation, in: Proc. Angew. Geog. Inf. XII, Beiträge Zum AGIT-Symp. Salzburg, pp. 12–23.

Ban, Y., Yousif, O.A., 2012. Multitemporal Spaceborne SAR Data for Urban Change Detection in China. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 5, 1087 –1094. doi:10.1109/JSTARS.2012.2201135

Ban, Y., Yousif, O., Hu, H., 2014. 18 Fusion of SAR and Optical Data for Urban Land Cover Mapping and Change Detection. Global Urban Monitoring and Assessment through Earth Observation 353.

Baselice, F., Ferraioli, G., Pascazio, V., 2014. Markovian Change Detection of Urban Areas Using Very High Resolution Complex SAR Images. IEEE Geoscience and Remote Sensing Letters 11, 995–999. doi:10.1109/LGRS.2013.2284297

Bazi, Y., Bruzzone, L., Melgani, F., 2006. Automatic Identification of the Number and Values of Decision Thresholds in the Log-Ratio Image for Change Detection in SAR Images. IEEE Geoscience and Remote Sensing Letters 3, 349–353. doi:10.1109/LGRS.2006.869973

76

Bazi, Y., Bruzzone, L., Melgani, F., 2005. An unsupervised approach based on the generalized Gaussian model to automatic change detection in multitemporal SAR images. IEEE Transactions on Geoscience and Remote Sensing 43, 874–887. doi:10.1109/TGRS.2004.842441

Ben-Gal, I., 2005. Outlier Detection, in: Maimon, O., Rokach, L. (Eds.), Data Mining and Knowledge Discovery Handbook. Springer US, pp. 131–146.

Benz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I., Heynen, M., 2004. Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS Journal of Photogrammetry and Remote Sensing, Integration of Geodata and Imagery for Automated Refinement and Update of Spatial Databases 58, 239–258. doi:10.1016/j.isprsjprs.2003.10.002

Berberoglu, S., Akin, A., 2009. Assessing different remote sensing techniques to detect land use/cover changes in the eastern Mediterranean. International Journal of Applied Earth Observation and Geoinformation 11, 46–53. doi:10.1016/j.jag.2008.06.002

Bin, W., Jian, Y., Zhongming, Z., Yu, M., Anzhi, Y., Jingbo, C., Dongxu, H., Xingchun, L., Shunxi, L., 2013. Parcel-Based Change Detection in Land-Use Maps by Adopting the Holistic Feature. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing Early Access Online. doi:10.1109/JSTARS.2013.2291773

Blaschke, T., 2005. Towards a framework for change detection based on image objects. Göttinger Geographische Abhandlungen 113, 1–9.

Bontemps, S., Bogaert, P., Titeux, N., Defourny, P., 2008. An object-based change detection method accounting for temporal dependences in time series with medium to coarse spatial resolution. Remote Sensing of Environment 112, 3181–3191. doi:10.1016/j.rse.2008.03.013

Bouziani, M., Goïta, K., He, D.-C., 2010. Automatic change detection of buildings in urban environment from very high spatial resolution images using existing geodatabase and prior knowledge. ISPRS Journal of Photogrammetry and Remote Sensing 65, 143–153. doi:10.1016/j.isprsjprs.2009.10.002

Bovolo, F., Bruzzone, L., 2007a. A Split-Based Approach to Unsupervised Change Detection in Large-Size Multitemporal Images: Application to Tsunami-Damage Assessment. IEEE Transactions on Geoscience and Remote Sensing 45, 1658–1670. doi:10.1109/TGRS.2007.895835

Bovolo, F., Bruzzone, L., 2007b. A Theoretical Framework for Unsupervised Change Detection Based on Change Vector Analysis in

77

the Polar Domain. IEEE Transactions on Geoscience and Remote Sensing 45, 218–236. doi:10.1109/TGRS.2006.885408

Bovolo, F., Bruzzone, L., 2005. A detail-preserving scale-driven approach to change detection in multitemporal SAR images. IEEE Transactions on Geoscience and Remote Sensing 43, 2963–2972. doi:10.1109/TGRS.2005.857987

Bovolo, F., Bruzzone, L., Marconcini, M., 2008. A Novel Approach to Unsupervised Change Detection Based on a Semisupervised SVM and a Similarity Measure. IEEE Transactions on Geoscience and Remote Sensing 46, 2070–2082. doi:10.1109/TGRS.2008.916643

Bovolo, F., Marin, C., Bruzzone, L., 2013. A Hierarchical Approach to Change Detection in Very High Resolution SAR Images for Surveillance Applications. IEEE Transactions on Geoscience and Remote Sensing 51, 2042–2054. doi:10.1109/TGRS.2012.2223219

Brunner, D., Bruzzone, L., Lemoine, G., 2010. Change detection for earthquake damage assessment in built-up areas using very high resolution optical and SAR imagery, in: Geoscience and Remote Sensing Symposium (IGARSS), 2010 IEEE International. Presented at the Geoscience and Remote Sensing Symposium (IGARSS), 2010 IEEE International, pp. 3210–3213. doi:10.1109/IGARSS.2010.5651416

Bruzzone, L., Prieto, D.F., 2002. An adaptive semiparametric and context-based approach to unsupervised change detection in multitemporal remote-sensing images. IEEE Transactions on Image Processing 11, 452–466.

Bruzzone, L., Prieto, D.F., 2000. Automatic analysis of the difference image for unsupervised change detection. IEEE Transactions on Geoscience and Remote Sensing 38, 1171–1182.

Bruzzone, L., Prieto, D.F., 2000. Unsupervised change detection in multisource and multisensor remote sensing images, in: Proceedings. IEEE 2000 International Geoscience and Remote Sensing Symposium. Presented at the IGARSS 2000, pp. 2441–2443 vol.6. doi:10.1109/IGARSS.2000.859602

Buades, A., Coll, B., Morel, J.-M., 2005. A non-local algorithm for image denoising, in: IEEE Computer Society Conference on Computer Vision and Pattern Recognition, CVPR 2005. pp. 60–65.

Bujor, F.T., Nicolas, J.-M., Trouve, E., Rudant, J.-P., 2003. Application of log-cumulants to change detection on multi-temporal SAR images, in: IEEE International Geoscience and Remote Sensing Symposium. Presented at the IGARSS 2003, pp. 1386–1388 vol.2. doi:10.1109/IGARSS.2003.1294117

78

Bujor, F., Trouve, E., Valet, L., Nicolas, J.-M., Rudant, J.-P., 2004. Application of log-cumulants to the detection of spatiotemporal discontinuities in multitemporal SAR images. IEEE Transactions on Geoscience and Remote Sensing 42, 2073–2084. doi:10.1109/TGRS.2004.835304

Bujor, F.T., Valet, L., Trouvw, E., Mauris, G., Classeau, N., Rudant, J.-P., 2001. Data fusion approach for change detection in multi-temporal ERS-SAR images, in: IEEE International Geoscience and Remote Sensing Symposium. Presented at the IGARSS 2001, pp. 2590–2592 vol.6. doi:10.1109/IGARSS.2001.978098

Cakir, H.I., Khorram, S., Nelson, S.A.C., 2006. Correspondence analysis for detecting land cover change. Remote Sensing of Environment 102, 306–317. doi:10.1016/j.rse.2006.02.023

Castellana, L., D’Addabbo, A., Pasquariello, G., 2007. A composed supervised/unsupervised approach to improve change detection from remote sensing. Pattern Recognition Letters 28, 405–413. doi:10.1016/j.patrec.2006.08.010

Celik, T., 2010a. A Bayesian approach to unsupervised multiscale change detection in synthetic aperture radar images. Signal Processing 90, 1471–1485. doi:10.1016/j.sigpro.2009.10.018

Celik, T., 2010b. Change Detection in Satellite Images Using a Genetic Algorithm Approach. IEEE Geoscience and Remote Sensing Letters 7, 386–390. doi:10.1109/LGRS.2009.2037024

Celik, T., 2009. Unsupervised Change Detection in Satellite Images Using Principal Component Analysis and k-Means Clustering. IEEE Geoscience and Remote Sensing Letters 6, 772–776. doi:10.1109/LGRS.2009.2025059

Chen, G., Hay, G.J., Carvalho, L.M.T., Wulder, M.A., 2012. Object-based change detection. International Journal of Remote Sensing 33, 4434–4457. doi:10.1080/01431161.2011.648285

Chen, G., Zhao, K., Powers, R., 2014. Assessment of the image misregistration effects on object-based change detection. ISPRS Journal of Photogrammetry and Remote Sensing 87, 19–27. doi:10.1016/j.isprsjprs.2013.10.007

Chen, J., Chen, Y., An, W., Cui, Y., Yang, J., 2011. Nonlocal Filtering for Polarimetric SAR Data: A Pretest Approach. IEEE Transactions on Geoscience and Remote Sensing 49, 1744–1754. doi:10.1109/TGRS.2010.2087763

Conradsen, K., Nielsen, A.A., Schou, J., Skriver, H., 2003. A test statistic in the complex wishart distribution and its application to change detection in polarimetric SAR data. IEEE Transactions on

79

Geoscience and Remote Sensing 41, 4–19. doi:10.1109/TGRS.2002.808066

Cossu, R., Chaudhuri, S., Bruzzone, L., 2005. A Context-Sensitive Bayesian Technique for the Partially Supervised Classification of Multitemporal Images. IEEE Geoscience and Remote Sensing Letters 2, 352–356. doi:10.1109/LGRS.2005.851541

Dai, X., Khorram, S., 1998. The effects of image misregistration on the accuracy of remotely sensed change detection. IEEE Transactions on Geoscience and Remote Sensing 36, 1566–1577. doi:10.1109/36.718860

Dawelbait, M., Morari, F., 2012. Monitoring desertification in a Savannah region in Sudan using Landsat images and spectral mixture analysis. Journal of Arid Environments 80, 45–55. doi:10.1016/j.jaridenv.2011.12.011

De Carvalho, O.A., Guimaraes, R.F., Gomes, R., de Carvalho, A.P.F., da Silva, N.C., 2006. Normalization of Multi-Temporal Images Using a New Change Detection Method Based on the Spectral Classifier, in: IEEE International Conference on Geoscience and Remote Sensing Symposium, 2006. IGARSS 2006. Presented at the IEEE International Conference on Geoscience and Remote Sensing Symposium, 2006. IGARSS 2006, pp. 771–774. doi:10.1109/IGARSS.2006.198

Dekker, R.J., 1998. Speckle filtering in satellite SAR change detection imagery. International Journal of Remote Sensing 19, 1133–1146.

Deledalle, C.-A., Denis, L., Tupin, F., 2012. How to Compare Noisy Patches? Patch Similarity Beyond Gaussian Noise. Int J Comput Vis 99, 86–102. doi:10.1007/s11263-012-0519-6

Deledalle, C.-A., Denis, L., Tupin, F., 2009. Iterative Weighted Maximum Likelihood Denoising With Probabilistic Patch-Based Weights. IEEE Transactions on Image Processing 18, 2661–2672. doi:10.1109/TIP.2009.2029593

Del Frate, F., Pacifici, F., Solimini, D., 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, 87–97. doi:10.1109/JSTARS.2008.2002221

DENG, J.-S., WANG, K., LI, J., DENG, Y.-H., 2009. Urban Land Use Change Detection Using Multisensor Satellite Images. Pedosphere 19, 96–103. doi:10.1016/S1002-0160(08)60088-0

Desclée, B., Bogaert, P., Defourny, P., 2006. Forest change detection by statistical object-based method. Remote Sensing of Environment 102, 1–11. doi:10.1016/j.rse.2006.01.013

80

Du, P., Liu, S., Gamba, P., Tan, K., Xia, J., 2012. Fusion of Difference Images for Change Detection Over Urban Areas. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 5, 1076–1086. doi:10.1109/JSTARS.2012.2200879

Fletcher, K., European Space Agency, European Space Research and Technology Centre, ESTEC, 2007. InSAR principles: guidelines for SAR interferometry processing and interpretation. ESA Publications Division, ESTEC, Noordwijk, the Netherlands.

Gamba, P., Dell’Acqua, F., Lisini, G., 2006. Change Detection of Multitemporal SAR Data in Urban Areas Combining Feature-Based and Pixel-Based Techniques. IEEE Transactions on Geoscience and Remote Sensing 44, 2820–2827. doi:10.1109/TGRS.2006.879498

Gamba, P., Dell’Acqua, F., Trianni, G., 2007. Rapid Damage Detection in the Bam Area Using Multitemporal SAR and Exploiting Ancillary Data. IEEE Transactions on Geoscience and Remote Sensing 45, 1582–1589. doi:10.1109/TGRS.2006.885392

Gao, F., Masek, J.G., Wolfe, R.E., Tan, B., 2010. Normalizing Landsat and ASTER data using MODIS data products for forest change detection, in: Geoscience and Remote Sensing Symposium (IGARSS), 2010 IEEE International. Presented at the Geoscience and Remote Sensing Symposium (IGARSS), 2010 IEEE International, pp. 3206–3209. doi:10.1109/IGARSS.2010.5650978

Garzelli, A., 1999. Classification of polarimetric SAR images using adaptive neighbourhood structures. International Journal of Remote Sensing 20, 1669–1675. doi:10.1080/014311699212678

Ghosh, S., Bruzzone, L., Patra, S., Bovolo, F., Ghosh, A., 2007. A Context-Sensitive Technique for Unsupervised Change Detection Based on Hopfield-Type Neural Networks. IEEE Transactions on Geoscience and Remote Sensing 45, 778–789. doi:10.1109/TGRS.2006.888861

Giustarini, L., Hostache, R., Matgen, P., Schumann, G.J.-P., Bates, P.D., Mason, D.C., 2013. A Change Detection Approach to Flood Mapping in Urban Areas Using TerraSAR-X. IEEE Transactions on Geoscience and Remote Sensing 51, 2417–2430. doi:10.1109/TGRS.2012.2210901

Gong, M., Cao, Y., Wu, Q., 2012a. A Neighborhood-Based Ratio Approach for Change Detection in SAR Images. IEEE Geoscience and Remote Sensing Letters 9, 307–311. doi:10.1109/LGRS.2011.2167211

Gong, M., Zhou, Z., Ma, J., 2012b. Change Detection in Synthetic Aperture Radar Images based on Image Fusion and Fuzzy Clustering.

81

IEEE Transactions on Image Processing 21, 2141–2151. doi:10.1109/TIP.2011.2170702

Grey, W., Luckman, A., Holland, D., 2003. Mapping urban change in the UK using satellite radar interferometry. Remote Sensing of Environment 87, 16–22. doi:10.1016/S0034-4257(03)00142-1

Hachicha, S., Chaabane, F., 2010. Comparison of Change Detection Indicators in SAR Images, in: 8th European Conference on Synthetic Aperture Radar. Presented at the EUSAR 2010, pp. 1–4.

Hall, O., Hay, G.J., 2003. A Multiscale Object-Specific Approach to Digital Change Detection. International Journal of Applied Earth Observation and Geoinformation 4, 311–327. doi:10.1016/S0303-2434(03)00010-2

Hao, M., Shi, W., Zhang, H., Li, C., 2014. Unsupervised Change Detection With Expectation-Maximization-Based Level Set. IEEE Geoscience and Remote Sensing Letters 11, 210–214. doi:10.1109/LGRS.2013.2252879

Hazel, G.G., 2001. Object-level change detection in spectral imagery. IEEE Transactions on Geoscience and Remote Sensing 39, 553–561. doi:10.1109/36.911113

He, C., Wei, A., Shi, P., Zhang, Q., Zhao, Y., 2011. Detecting land-use/land-cover change in rural–urban fringe areas using extended change-vector analysis. International Journal of Applied Earth Observation and Geoinformation 13, 572–585. doi:10.1016/j.jag.2011.03.002

Hou, B., Wei, Q., Zheng, Y., Wang, S., 2014. Unsupervised Change Detection in SAR Image Based on Gauss-Log Ratio Image Fusion and Compressed Projection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing Early Access Online. doi:10.1109/JSTARS.2014.2328344

Huang, L.-K., Wang, M.-J.J., 1995. Image thresholding by minimizing the measures of fuzziness. Pattern Recognition 28, 41–51. doi:10.1016/0031-3203(94)E0043-K

Im, J., Jensen, J.R., 2005. A change detection model based on neighborhood correlation image analysis and decision tree classification. Remote Sensing of Environment 99, 326–340. doi:10.1016/j.rse.2005.09.008

Im, J., Jensen, J.R., Tullis, J.A., 2008. Object-based change detection using correlation image analysis and image segmentation. International Journal of Remote Sensing 29, 399–423. doi:10.1080/01431160601075582

82

Inglada, J., Giros, A., 2004. On the possibility of automatic multisensor image registration. IEEE Transactions on Geoscience and Remote Sensing 42, 2104–2120. doi:10.1109/TGRS.2004.835294

Inglada, J., Mercier, G., 2007. A New Statistical Similarity Measure for Change Detection in Multitemporal SAR Images and Its Extension to Multiscale Change Analysis. IEEE Transactions on Geoscience and Remote Sensing 45, 1432–1445. doi:10.1109/TGRS.2007.893568

Jensen, J.R., 2005. Introductory digital image processing: a remote sensing perspective. Pearson Prentice Hall.

Jovanović, D., Govedarica, M., Đorđević, I., Pajić, V., 2010. Object based image analysis in forestry change detection, in: International Symposium on Intelligent Systems and Informatics. Presented at the SISY 2010, pp. 231–236. doi:10.1109/SISY.2010.5647487

Kasetkasem, T., Varshney, P.K., 2002. An image change detection algorithm based on Markov random field models. IEEE Transactions on Geoscience and Remote Sensing 40, 1815–1823. doi:10.1109/TGRS.2002.802498

Kittler, J., Illingworth, J., 1986. Minimum error thresholding. Pattern recognition 19, 41–47.

Kurita, T., Otsu, N., Abdelmalek, N., 1992. Maximum likelihood thresholding based on population mixture models. Pattern Recognition 25, 1231–1240. doi:10.1016/0031-3203(92)90024-D

Lee, J.-S., Pottier, E., 2009. Polarimetric Radar Imaging: From Basics to Applications. CRC Press.

Le, T.T., Atto, A., Trouvé, E., Nicolas, J.-M., 2014. Adaptive Multitemporal SAR Image Filtering Based on the Change Detection Matrix. IEEE Geoscience and Remote Sensing Letters 11, 1826–1830. doi:10.1109/LGRS.2014.2311663

Li, S.Z., 2009. Markov Random Field Modeling in Image Analysis, 3rd ed. Springer.

Liu, S., Bruzzone, L., Bovolo, F., Du, P., 2015. Hierarchical Unsupervised Change Detection in Multitemporal Hyperspectral Images. IEEE Transactions on Geoscience and Remote Sensing 53, 244–260. doi:10.1109/TGRS.2014.2321277

Liu, S., Bruzzone, L., Bovolo, F., Zanetti, M., Du, P., 2015. Sequential Spectral Change Vector Analysis for Iteratively Discovering and Detecting Multiple Changes in Hyperspectral Images. IEEE Transactions on Geoscience and Remote Sensing 53, 4363–4378. doi:10.1109/TGRS.2015.2396686

83

Liu, W., Yamazaki, F., 2011. Urban monitoring and change detection of central Tokyo using high-resolution X-band SAR images, in: IEEE International Geoscience and Remote Sensing Symposium. Presented at the IGARSS 2011, pp. 2133–2136. doi:10.1109/IGARSS.2011.6049587

Lu, D., Mausel, P., Brondízio, E., Moran, E., 2004. Change detection techniques. International Journal of Remote Sensing 25, 2365–2401. doi:10.1080/0143116031000139863

Lyon, J.G., Yuan, D., Lunetta, R.S., Elvidge, C.D., 1998. A change detection experiment using vegetation indices. Photogrammetric Engineering and Remote Sensing 64, 143–150.

Ma, J., Gong, M., Zhou, Z., 2012. Wavelet Fusion on Ratio Images for Change Detection in SAR Images. IEEE Geoscience and Remote Sensing Letters 9, 1122–1126. doi:10.1109/LGRS.2012.2191387

Marin, C., Bovolo, F., Bruzzone, L., 2015. Building Change Detection in Multitemporal Very High Resolution SAR Images. IEEE Transactions on Geoscience and Remote Sensing 53, 2664–2682. doi:10.1109/TGRS.2014.2363548

Martinis, S., Twele, A., Voigt, S., 2011. Unsupervised Extraction of Flood-Induced Backscatter Changes in SAR Data Using Markov Image Modeling on Irregular Graphs. IEEE Transactions on Geoscience and Remote Sensing 49, 251–263. doi:10.1109/TGRS.2010.2052816

Melgani, F., Moser, G., Serpico, S.B., 2002. Unsupervised change-detection methods for remote-sensing images. Optical Engineering 41, 3288–3297.

Mercier, G., Moser, G., Serpico, S.B., 2008. Conditional Copulas for Change Detection in Heterogeneous Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing 46, 1428–1441. doi:10.1109/TGRS.2008.916476

Moser, G., Melgani, F., Serpico, S.B., Caruso, A., 2002. Partially supervised detection of changes from remote sensing images, in: IEEE International Geoscience and Remote Sensing Symposium. Presented at the IGARSS 2002, pp. 299–301 vol.1. doi:10.1109/IGARSS.2002.1025019

Moser, G., Serpico, S.B., 2009. Unsupervised Change Detection From Multichannel SAR Data by Markovian Data Fusion. IEEE Transactions on Geoscience and Remote Sensing 47, 2114–2128. doi:10.1109/TGRS.2009.2012407

Moser, G., Serpico, S.B., 2006. Generalized minimum-error thresholding for unsupervised change detection from SAR amplitude

84

imagery. IEEE Transactions on Geoscience and Remote Sensing 44, 2972–2982. doi:10.1109/TGRS.2006.876288

Moser, G., Serpico, S., Vernazza, G., 2007. Unsupervised Change Detection From Multichannel SAR Images. IEEE Geoscience and Remote Sensing Letters 4, 278–282. doi:10.1109/LGRS.2007.890549

Nemmour, H., Chibani, Y., 2006. Multiple support vector machines for land cover change detection: An application for mapping urban extensions. ISPRS Journal of Photogrammetry and Remote Sensing 61, 125–133. doi:10.1016/j.isprsjprs.2006.09.004

Nielsen, A., 2007. The Regularized Iteratively Reweighted MAD Method for Change Detection in Multi- and Hyperspectral Data. IEEE Transactions on Image Processing 16, 463–478. doi:10.1109/TIP.2006.888195

Niemeyer, I., Marpu, P.R., Nussbaum, S., 2008. Change detection using object features, in: Blaschke, T., Lang, S., Hay, G.J. (Eds.), Object-Based Image Analysis, Lecture Notes in Geoinformation and Cartography. Springer Berlin Heidelberg, pp. 185–201.

Oliver, C., Quegan, S., 2004. Understanding Synthetic Aperture Radar Images. SciTech Publishing.

Orsomando, F., Lombardo, P., Zavagli, M., Costantini, M., 2007. SAR and optical data fusion for change detection, in: Urban Remote Sensing Joint Event, 2007. pp. 1–9.

Otsu, N., 1975. A threshold selection method from gray-level histograms. Automatica 11, 23–27.

Parrilli, S., Poderico, M., Angelino, C.V., Verdoliva, L., 2012. A Nonlocal SAR Image Denoising Algorithm Based on LLMMSE Wavelet Shrinkage. IEEE Transactions on Geoscience and Remote Sensing 50, 606–616. doi:10.1109/TGRS.2011.2161586

Phua, M.-H., Tsuyuki, S., Furuya, N., Lee, J.S., 2008. Detecting deforestation with a spectral change detection approach using multitemporal Landsat data: A case study of Kinabalu Park, Sabah, Malaysia. Journal of Environmental Management 88, 784–795. doi:10.1016/j.jenvman.2007.04.011

Poulain, V., Inglada, J., Spigai, M., Tourneret, J.-Y., Marthon, P., 2011. High-Resolution Optical and SAR Image Fusion for Building Database Updating. IEEE Transactions on Geoscience and Remote Sensing 49, 2900–2910. doi:10.1109/TGRS.2011.2113351

Qin, Y., Niu, Z., Chen, F., Li, B., Ban, Y., 2013. Object-based land cover change detection for cross-sensor images. International Journal of Remote Sensing 34, 6723–6737. doi:10.1080/01431161.2013.805282

85

Quin, G., Pinel-Puyssegur, B., Nicolas, J.-M., Loreaux, P., 2014. MIMOSA: An Automatic Change Detection Method for SAR Time Series. IEEE Transactions on Geoscience and Remote Sensing 52, 5349–5363. doi:10.1109/TGRS.2013.2288271

Rasi, R., Beuchle, R., Bodart, C., Vollmar, M., Seliger, R., Achard, F., 2013. Automatic Updating of an Object-Based Tropical Forest Cover Classification and Change Assessment. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 6, 66–73. doi:10.1109/JSTARS.2012.2217733

Rignot, E.J., van Zyl, J.J., 1993. Change detection techniques for ERS-1 SAR data. IEEE Transactions on Geoscience and Remote Sensing 31, 896–906.

Roy, M., Ghosh, S., Ghosh, A., 2012. Search-based semi-supervised clustering algorithms for change detection in remotely sensed images, in: Annual IEEE India Conference. Presented at the INDICON 2012, pp. 503–507. doi:10.1109/INDCON.2012.6420670

Sabry, R., 2009. A New Coherency Formalism for Change Detection and Phenomenology in SAR Imagery: A Field Approach. IEEE Geoscience and Remote Sensing Letters 6, 458–462. doi:10.1109/LGRS.2009.2016359

Schindler, K., 2012. An Overview and Comparison of Smooth Labeling Methods for Land-Cover Classification. IEEE Transactions on Geoscience and Remote Sensing 50, 4534–4545. doi:10.1109/TGRS.2012.2192741

Servello, E.L., Kuplich, T.H., Shimabukuro, Y.E., 2010. Tropical land cover change detection with polarimetric SAR data, in: IEEE International Geoscience and Remote Sensing Symposium. Presented at the IGARSS 2010, pp. 1477–1480. doi:10.1109/IGARSS.2010.5653215

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

Tasdizen, T., 2009. Principal Neighborhood Dictionaries for Nonlocal Means Image Denoising. IEEE Transactions on Image Processing 18, 2649–2660. doi:10.1109/TIP.2009.2028259

Tso, B., Olsen, R.C., 2005. A contextual classification scheme based on MRF model with improved parameter estimation and multiscale fuzzy line process. Remote Sensing of Environment 97, 127–136. doi:10.1016/j.rse.2005.04.021

United Nations, 2008. World urbanization prospects: the 2014 revision.

86

Velloso, M.L.F., de Souzal, F.J., Simoes, M., 2002. Improved radiometric normalization for land cover change detection: an automated relative correction with artificial neural network, in: Geoscience and Remote Sensing Symposium, 2002. IGARSS ’02. 2002 IEEE International. Presented at the Geoscience and Remote Sensing Symposium, 2002. IGARSS ’02. 2002 IEEE International, pp. 3435–3437 vol.6. doi:10.1109/IGARSS.2002.1027207

Vollmar, M., Rasi, R., Beuchle, R., Simonetti, D., Stibig, H.-J., Achard, F., 2013. Combining Landsat TM/ETM+ and ALOS AVNIR-2 Satellite Data for Tropical Forest Cover Change Detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 6, 102–109. doi:10.1109/JSTARS.2013.2241017

Volpi, M., Tuia, D., Bovolo, F., Kanevski, M., Bruzzone, L., 2013. Supervised change detection in VHR images using contextual information and support vector machines. International Journal of Applied Earth Observation and Geoinformation 20, 77–85. doi:10.1016/j.jag.2011.10.013

Walessa, M., Datcu, M., 2000. Model-based despeckling and information extraction from SAR images. IEEE Transactions on Geoscience and Remote Sensing 38, 2258–2269. doi:10.1109/36.868883

Walter, V., 2004. Object-based classification of remote sensing data for change detection. ISPRS Journal of Photogrammetry and Remote Sensing 58, 225–238. doi:10.1016/j.isprsjprs.2003.09.007

Wang, F., Wu, Y., Zhang, Q., Zhang, P., Li, M., Lu, Y., 2013. Unsupervised Change Detection on SAR Images Using Triplet Markov Field Model. IEEE Geoscience and Remote Sensing Letters 10, 697–701. doi:10.1109/LGRS.2012.2219494

Wang, L., Li, C., Ying, Q., Cheng, X., Wang, X., Li, X., Hu, L., Liang, L., Yu, L., Huang, H., Gong, P., 2012. China’s urban expansion from 1990 to 2010 determined with satellite remote sensing. Chin. Sci. Bull. 57, 2802–2812. doi:10.1007/s11434-012-5235-7

Xie, H., Pierce, L.E., Ulaby, F.T., 2002. SAR speckle reduction using wavelet denoising and Markov random field modeling. IEEE Transactions on Geoscience and Remote Sensing 40, 2196–2212. doi:10.1109/TGRS.2002.802473

Xiong, B., Chen, Q., Jiang, Y., Kuang, G., 2012. A Threshold Selection Method Using Two SAR Change Detection Measures Based on the Markov Random Field Model. IEEE Geoscience and Remote Sensing Letters 9, 287–291. doi:10.1109/LGRS.2011.2166149

87

Yan, H., 1996. Unified formulation of a class of image thresholding techniques. Pattern Recognition 29, 2025–2032. doi:10.1016/S0031-3203(96)00050-7

Yousif, O., Ban, Y., 2014. Improving SAR-Based Urban Change Detection by Combining MAP-MRF Classifier and Nonlocal Means Similarity Weights. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 7, 4288–4300. doi:10.1109/JSTARS.2014.2347171

Yousif, O., Ban, Y., 2013. Improving Urban Change Detection From Multitemporal SAR Images Using PCA-NLM. IEEE Transactions on Geoscience and Remote Sensing 51, 2032–2041. doi:10.1109/TGRS.2013.2245900

Yuan, F., Sawaya, K.E., Loeffelholz, B.C., Bauer, M.E., 2005. Land cover classification and change analysis of the Twin Cities (Minnesota) Metropolitan Area by multitemporal Landsat remote sensing. Remote Sensing of Environment 98, 317–328. doi:10.1016/j.rse.2005.08.006

Zhang, Q., Ban, Y., Hu, Y., Liu, J., 2009. The trajectories of urban land and industrial land in Shanghai over the past 30 years, in: Urban Remote Sensing Event, 2009 Joint. Presented at the Urban Remote Sensing Event, 2009 Joint, pp. 1–7. doi:10.1109/URS.2009.5137640

Zhang, Q., Ban, Y., Liu, J., Hu, Y., 2011. Simulation and analysis of urban growth scenarios for the Greater Shanghai Area, China. Computers, Environment and Urban Systems 35, 126–139. doi:10.1016/j.compenvurbsys.2010.12.002

Zhong, P., Wang, R., 2006. Object Detection Based on Combination of Conditional Random Field and Markov Random Field, in: 18th International Conference on Pattern Recognition, 2006. ICPR 2006. Presented at the ICPR 2006, pp. 160 –163. doi:10.1109/ICPR.2006.876

Zhou, W., Troy, A., Grove, M., 2008. A Comparison of Object-Based with Pixel-Based Land Cover Change Detection in the Baltimore Metropolitan Area using Multitemporal High Resolution Remote Sensing Data, in: IEEE International Geoscience and Remote Sensing Symposium. Presented at the IGARSS 2008, pp. IV – 683–IV – 686. doi:10.1109/IGARSS.2008.4779814

Urban Change Detection Using Multitemporal SAR Images814876/FULLTEXT01.pdf · unsupervised change...

Documents

Transcript of Urban Change Detection Using Multitemporal SAR Images814876/FULLTEXT01.pdf · unsupervised change...