A segment-based approach to classify agricultural lands by using multi-temporal optical and...

This article was downloaded by: [The UC Irvine Libraries]On: 02 November 2014, At: 16:22Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of RemoteSensingPublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/tres20

A segment-based approach to classifyagricultural lands by using multi-temporal optical and microwave dataAsli Ozdarici Ok a & Zuhal Akyurek ba Graduate School of Natural and Applied Sciences, Geodeticand Geographic Information Technologies , Middle East TechnicalUniversity , 06800 , Ankara , Turkeyb Department of Civil Engineering , Middle East TechnicalUniversity , 06800 , Ankara , TurkeyPublished online: 28 Jun 2012.

To cite this article: Asli Ozdarici Ok & Zuhal Akyurek (2012) A segment-based approach to classifyagricultural lands by using multi-temporal optical and microwave data, International Journal ofRemote Sensing, 33:22, 7184-7204, DOI: 10.1080/01431161.2012.700423

To link to this article: http://dx.doi.org/10.1080/01431161.2012.700423

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the“Content”) contained in the publications on our platform. However, Taylor & Francis,our agents, and our licensors make no representations or warranties whatsoever as tothe accuracy, completeness, or suitability for any purpose of the Content. Any opinionsand views expressed in this publication are the opinions and views of the authors,and are not the views of or endorsed by Taylor & Francis. The accuracy of the Contentshould not be relied upon and should be independently verified with primary sourcesof information. Taylor and Francis shall not be liable for any losses, actions, claims,proceedings, demands, costs, expenses, damages, and other liabilities whatsoever orhowsoever caused arising directly or indirectly in connection with, in relation to or arisingout of the use of the Content.

This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &

http://www.tandfonline.com/loi/tres20

http://www.tandfonline.com/action/showCitFormats?doi=10.1080/01431161.2012.700423

http://dx.doi.org/10.1080/01431161.2012.700423

Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14

http://www.tandfonline.com/page/terms-and-conditions

http://www.tandfonline.com/page/terms-and-conditions

International Journal of Remote SensingVol. 33, No. 22, 20 November 2012, 7184–7204

A segment-based approach to classify agricultural lands by usingmulti-temporal optical and microwave data

ASLI OZDARICI OK*† and ZUHAL AKYUREK‡†Graduate School of Natural and Applied Sciences, Geodetic and Geographic

Information Technologies, Middle East Technical University, 06800 Ankara, Turkey‡Department of Civil Engineering, Middle East Technical University, 06800 Ankara,

Turkey

(Received 18 October 2011; in final form 22 May 2012)

This research study aims to classify crop diversity in agricultural land with asegment-based approach using multi-temporal Kompsat-2 and EnvironmentalSatellite (Envisat) advanced synthetic aperture radar (ASAR) data acquired inJune, July and August on Karacabey Plain, Turkey. Analyses start with the imagesegmentation process applied to the fused optical images to search homogenousobjects. The segmentation outputs are evaluated using multiple goodness measures,which take into consideration area and location similarities. Image classificationsare performed on each multispectral (MS) single date image. In order to com-bine the most probable classes of the thematic maps, distance maps are generated.Evaluations of the thematic maps are performed through confusion matrices basedon pixel-based and segment-based approaches. The results indicate that the high-est overall accuracy of 88.71% and a kappa result of 0.86 are provided for thesegment-based approach of the combined thematic map along with the microwavedata, which is around 10% higher than the related pixel-based results.

1. Introduction

Due to the rapid increase in urban populations, agricultural lands have gone througha dramatic decrease. As a result, the demand for agricultural products has gainedgrowing importance all over the world. In order to make accurate yield estimation,automated methods based on the development stages of agricultural crops are nec-essary. In traditional agricultural applications, up-to-date information on crops isacquired by farmer declarations and/or ground visits to the fields. As Penã-Barragánet al. (2011) suggest, this procedure generates not only some errors and discrepan-cies but is also a time-consuming and expensive process. In this regard, the analysisof satellite images and/or aerial photographs can be a more reliable and cost-effectiveway to monitor agricultural areas. With the development of satellite sensor technology,the availability of high spatial resolution images has increased (Geoeye, QuickBird,Kompsat-2, and IKONOS, etc.).

While this improvement is helpful in detecting distinct small objects more preciselyin agricultural practices, it may negatively affect the final accuracies of the thematicmaps due to the within-field spectral variability (e.g. Gong and Howarth 1990, Smith

*Corresponding author. Email: [email protected]

International Journal of Remote SensingISSN 0143-1161 print/ISSN 1366-5901 online © 2012 Taylor & Francis

http://www.tandfonline.comhttp://dx.doi.org/10.1080/01431161.2012.700423

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14

A segment-based approach to classify agricultural lands 7185

and Fuller 2001, De Wit and Clevers 2004). In order to overcome this problem, thepixels need to be considered as groups based on their texture and context informationto delineate more meaningful objects, which is the major interest of the object-basedimage analysis (OBIA) community (e.g. Gong et al. 1992, Gong and Howarth 1992,Yu et al. 2006). Although a wide variety of results can be obtained through differentparameter combinations and different software (e.g. Schoenmakers et al. 1994, Cheng1995, Rydberg and Borgefors 2001, Mueller et al. 2003, Zhan et al. 2005, Chen et al.2006, Lee and Warner 2006, Li and Xiao 2007, Lu and Weng 2007, Corcoran et al.2010, Wang et al. 2010b, Xiao et al. 2010), additional steps are required to find themost appropriate segmentation results. The main problem is the lack of uniformcriteria to evaluate the results (Liu and Yang 1994, Zhang 1996, 2001, Martin et al.2004, Chabrier et al. 2006, Ge et al. 2006, Li and Xiao 2007), because each evaluationmethod indicates only one aspect of algorithm performance (Weidner 2008). In thisstudy, we define the final segmentation results through the use of multiple goodnessmeasures to find the optimum match between the well-defined reference fields andthe segments generated. The defined segments for each image are then employedin the proposed multi-temporal classification approach. The practicality of themulti-temporal classification on agricultural applications has been examined in manystudies so far (e.g. Parmuchi et al. 2002, Ban 2003, Blaes et al. 2005, Turker andArikan 2005, Liu et al. 2006, Stankiewicz 2006, Wang et al. 2010a, Penã-Barragánet al. 2011, Skriver et al. 2011) and reliable thematic maps have been produced withhigh accuracy (e.g. Parmuchi et al. 2002, Ban 2003, Blaes et al. 2005, Turker andArikan 2005, Stankiewicz 2006, Liu et al. 2006, Wang et al. 2010a, Penã-Barragánet al. 2011, Skriver et al. 2011). However, most of them are based on a rule-basedapproach, which requires considerable knowledge regarding the data (e.g. Turker andArikan 2005, Ban et al. 2010). Therefore, much more effective methods are necessaryto combine such multi-temporal data. This study proposes a new multi-temporalclassification strategy that combines different individual classification results in a jointprobabilistic approach. Owing to the different response characteristics, microwavedata (Environmental Satellite (Envisat) advanced synthetic aperture radar (ASAR))are utilized along with the Kompsat-2 multispectral (MS) data in this study (Liu et al.2006, Blaes et al. 2007, Ban et al. 2010).

To summarize, a segment-based crop identification methodology is proposed for theassessment of eight crop types in an agricultural region of Turkey. The objectives are:(i) to perform the evaluation of the segmentation results through multiple goodnessmeasures; (ii) to explore the effect of a multi-temporal and multi-sensor data set ofKompsat-2 and Envisat ASAR in identifying eight crop types cultivated in the studyarea; and (iii) to determine the influence of phenological characteristics of the cropsin crop discrimination.

2. Study area and data description

The study area is situated in Karacabey, an agricultural area in Bursa, northwestTurkey (figure 1(a)). It covers approximately 100 km2 and is centred at 28

◦14′ 12′′

and 40◦

11′ 09′′ (figure 1(b)). It is one of the most productive and valuable agriculturalregions of Turkey, with its rich and loamy soils and good weather conditions. Theregion has a temperate and semi-humid climate with a mean annual temperature of14.4

◦C and mean annual precipitation of 706 mm. The area has a flat terrain and the

mean elevation above sea level is 10 m. Most of the fields have a regular shape in the

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14

7186 A. O. Ok and Z. Akyurek

(a)

(b)

0 1000 m 2000 m

Ankara

50 0 50 150 km

N 40° 13′ 43″N 40° 07′ 44″

E 28° 20′ 28″E 28° 10′ 31″

Black Sea

MarmaraSea

AegeanSea

Bursa

Istanbul

Study Area

N

Figure 1. (a) Whole study area and (b) true colour composite of the MS Kompsat-2 image(July).

area based on a land consolidation project performed between 1988 and 1992 (Turkerand Ozdarici 2011).

The major crop types cultivated in the test site are corn, tomato, wheat, rice, sugarbeet and pea. The area also contains several grassland fields to provide feed to ani-mals. The phenological characteristics of the crops are given in figure 2. In the figure,

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


Janu

ary

Febr

uary

Mar

h

Apr

il

May

June

July

Aug

ust

Sept

embe

r

Oct

ober

Now

embe

r

Dec

embe

r

CornLate cornWheatTomatoSugar beetRiceGrasslandPea

Bare soil Open canopy closure Dense canopy closure

Figure 2. Phenological characteristics of the crops.

black boxes represent dense canopy closure, grey boxes indicate open canopy closureand white boxes imply the areas having no vegetation, bare soil. Based on figure 2,the first planting phase of corn starts at the end of April and the fields are coveredwith small leaves of the corn in May. Leaves grow gradually from June to September,when dense vegetation covers the fields. In November, the harvesting period of thecorn fields starts and it continues until the end of November. There are also late cornfields cultivated in the area between July and December. These fields are generallyplanted after harvesting some fields of pea, tomato and sugar beet. Wheat is anothermajor crop type in the area. Its planting date lies between November and the next July.Except for the wheat, each field is irrigated in the area. It is observed that tomato andrice fields have similar growing dates between May and October. The rice is an inter-esting crop type in the region. This is because the water requirement of this crop ismore than other crop types. The rice fields are filled with water at the beginning of thedevelopment stages, April, May, June and July. A root drying process is applied to therice fields after 15 or 30 days, and then the rice plots are filled with water again untilmid-July. The planting period of the sugar beet starts in March and ends in November.The class sugar beet has the third longest planting period following the grassland andwheat in the area. Another interesting crop type in the region is the grassland, whichgenerally looks green all the year provided that there is no snow on it. The crop peahas a short planting cycle when compared with the other crop types. Its planting datesare between April and July (Turker and Ozdarici 2011).

Three dates of early-, mid- and late-season Kompsat-2 panchromatic (PAN) andMS optical data and Envisat ASAR data were acquired in June, July and Augustin 2008. The specific acquisition dates of the optical Kompsat-2 images are 13 June(early season), 11 July (mid season) and 18 August (late season). Each of the opti-cal data is obtained in a clear-sky condition. Kompsat-2 satellite has two products:PAN (1 m) and MS (4 m). The spectral range of a PAN image is between 0.5 and 0.9µm. Kompsat-2 MS data have four spectral bands, namely blue, green, red and near-infrared (NIR). The spectral ranges of these bands lie between 0.45–0.52, 0.52–0.60,0.63–0.69 and 0.76–0.90 µm, respectively (NIK System 2008). The preprocessing levelof the Kompsat-2 image is level 2A, in which radiometric correction is applied to theimages to minimize the sensor-based radiometric errors. The geometric correctionshave also been applied to the images in a standard cartographic projection (UTMWGS 84) without using any ground control points (GCPs) (table 1). The EnvisatASAR images were acquired in 28 June, 18 July and 3 August 2008 in image mode

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


Table 1. Technical characteristics of optical and microwave data.

Kompsat-2 Envisat ASAR

Products PAN MS –Spectral bands (µm) Blue: 0.45–0.52 C-band

0.50–0.90 Green: 0.52–0.60Red: 0.63–0.69NIR: 0.76–0.90

Frequency range – – 5.331 GHzSpatial resolution 1 m 4 m 15 m

Radiometric resolution 16 bit unsignedFootprint 15 km × 15 km 56 km × 105 kmAcquisition

configuration– IS6 and IS7

Viewing angle Revisit rate of 3 days with a rollangle of 30◦

39.1◦ and 42.8◦–42.5◦ and 45.2◦

Preprocessing level Level 2A Level 1bPolarization – VVDatum WGS 84Map projection UTMZone number 35

Precision Image format. This mode provides HH and VV polarization images withspatial resolutions between 15 and 150 m and coverage of 56 × 105 km2. Due to tech-nical problems which occurred during image acquisition, only VV polarization imagesare utilized in the study. Envisat ASAR operates in the C-band and it provides variousincidence angles between 15◦ and 45.2◦. Seven acquisition configurations (IS1–IS7) areavailable for the Envisat ASAR data; however, due to the limitations which occurredon data acquisition, only the configurations of IS2, IS6 and IS7 could be employed(ASAR Product Handbook 2009). The radiometric resolution of each data type is16 bit unsigned. To provide computational efficiency and facilitate interpretation, theradiometric resolution of the images is converted to the 8 bit unsigned level prior to theanalyses. Table 1 provides a detailed description of the optical and microwave images.

3. Methodology

The overall methodology of the study is presented in figure 3. As shown in figure 3,the study is composed of five main parts: (i) image fusion, (ii) image segmentation, (iii)image classification, (iv) likelihood estimation and (v) accuracy assessment.

3.1 Image fusion of Kompsat-2 data

In order to produce colour composite images with higher spatial resolution (1 m),Kompsat-2 PAN (1 m) and MS (4 m) data are fused in this study. Nine different imagefusion methods, frequently used in the literature, are tested to determine a propermethod for the study. The fused results are evaluated visually, statistically and in termsof classification performance. In statistical evaluations, multiple evaluation indicators,namely relative mean difference, relative variation difference, correlation, peak signal-to-noise ratio, universal image quality index and erreur relative globale adimensionnelledesynthèse (ERGAS) are utilized. Based on the analyses, it is observed that the least

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


Figure 3. Methodology of the study.

squares fusion (LSF) method provides the best performance (Ozdarici and Akyurek2009, Ozdarici Ok and Akyurek 2011); thus, it is employed to fuse the Kompsat-2 PANand MS images.

3.2 Atmospheric correction of Kompsat-2 data

Following the image fusion process, atmospheric correction is applied to the fusedKompsat-2 data to make reliable measurements of the images taken on different dates(Jensen 2005). In this regard, the optical data are atmospherically corrected using theAtmospheric Correction of Flat Areas (ATCOR-2) module of PCI Geomatica (PCIGeomatics, Richmond Hill, ON, Canada) (Richter 1990).

3.3 Image filtering of Envisat ASAR data

The presence of speckle is mostly encountered in microwave data, which seriouslyaffects the interpretability of the images, although it provides valuable information onthe imaging system itself (Henderson and Lewis 1998). Hence, prior to the analyses,the speckle should be reduced in order to interpret the images properly. One tradi-tional way of reducing the speckle is image filtering, which is a local operation that

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


modifies the original pixels of an image with its neighbours (Lillesand et al. 2004).So far, multiple image filtering methods have been tested in a number of studies (e.g.Lee 1980, Frost et al. 1982, Kuan et al. 1985, Serkan et al. 2008). Besides the filter-ing methods, filter size is an important factor affecting the image quality. In orderto determine the method and filter size of the study, seven image filtering methodsare examined on the Envisat ASAR data (Ozdarici and Akyurek 2010). The filteringresults are evaluated by multiple statistical evaluation indicators: difference of mean,standard deviation between the filtered results and the original image, correlation andquality factor. Analyses suggest that a Lee filter with 5 × 5 windows is the most suit-able method to minimize the speckle effect of the Envisat ASAR data. A small part ofthe original data and the Lee filtered image with a filter size of 5 × 5 can be seen infigure 4.

3.4 Orthorectification of the images

Geometric integrity among data is essential to make accurate analyses in a study.Both optical and microwave images are orthorectified in order to provide propergeometric conditions. A digital elevation model (DEM) is generated using the 1:25000 scale digital maps obtained from General Command of Mapping of Turkey. Well-distributed GCPs are gathered from the fieldwork by differential global positioningsystem (DGPS) measurements in the subpixel level (Mini MAX 2004). As a geomet-ric model, a rigorous ‘Satellite Orbital Modelling’ (PCI Geomatica 2009) is employedfor all of the Kompsat-2 and Envisat ASAR data. In the orthorectification process,at least 6 and 14 evenly distributed GCPs are selected for the Envisat ASAR andKompsat-2 data sets, respectively, and all root mean square error (RMSE) values ofthe geometric model are computed to be less than one pixel size. Table 2 provides thenumber of GCPs, RMSEs and the methods used for the resampling of each image.

3.5 Image segmentation of Kompsat-2 data

One of the most widely used information extraction methods in the remote-sensingcommunity is image segmentation, which is utilized to search homogenous regions

0 1000 m 2000 m

(b)(a)

Figure 4. A small part of (a) the original Envisat ASAR data and (b) the Lee filtered imagewith 5 × 5 windows.

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


Table 2. Number of GCPs, the geometric correction and resampling method and the RMSEfor each data.

Data Acquisition date Number of GCPs Resampling RMSE (pixel)

Envisat ASAR 28 June 2008 6 NN 0.5518 July 2008 8 NN 0.453 August 2008 10 NN 0.50

The fused 13 June 2008 14 NN 0.80Kompsat-2 11 July 2008 15 NN 0.85

18 August 2008 20 NN 0.75

Note: NN, nearest neighbour.

in an image by defining some homogeneity criteria (Cheng et al. 2001). To over-come the problems caused by the heterogeneous pixels and crop variability within thefield, object-based image analyses have been increasingly employed in remotely senseddata (Blaschke 2010). In this study, an old pattern recognition procedure, mean-shiftsegmentation method, is applied to the 1 m spatial resolution-fused Kompsat-2 images(green, red and NIR) to determine homogenous objects. The mean-shift is used in thisstudy for many reasons. First of all, the mean-shift procedure does not require anyassumption about the data distribution and shape of the clusters, which facilitateshandling of the complex objects in a real feature space. Moreover, it is a simple itera-tive mode seeking procedure that smoothes the images prior to the clustering processwhile preserving and sharpening its discontinuities. Additionally, the mean-shift pro-cedure can be applied not only to grey images but also to colour images by a jointspatial-range domain. Furthermore, it requires only a few parameters to apply themean-shift procedure to the images (Comaniciu and Meer 2002). It is based on kerneldensity estimation, which applies an iterative shifting procedure to data values in orderto produce a new centre of mass. The new mean point of x in mk(x) is computed bythe following formula (equation (1)):

mK (x) ≡∑n

i=1 xiK( x−xi

h

)∑n

i=1 K( x−xi

h

) − x, (1)

where K(x) is the kernel defined for the mean-shift process, x denotes the centre of thekernel used, h is the size of the kernel and n is the number of data values (Comaniciuand Meer 2002).

In respect of the images, the method works on two domains: (i) spatial (position inthe image, hs) and (ii) range (grey level, hr). A minimum region (MR) parameter is thethird parameter that needs to be defined. More detailed explanation of the mean-shiftcan be found in Comaniciu and Meer (1997, 2002).

As for the optical images fused, in total 972 combinations of the three parameters –spatial (hs), range (hr) domains and minimum region (MR) – are tested on {3,4,5,. . . ,20} × {3,4,5, . . . ,20} × {1000}, respectively. In other words, 324 parameter com-binations are tested on each single date Kompsat-2 image fused separately. The MRparameter (1000 pixels) is defined based on the total number of pixels that belongs tothe smallest parcel in the area.

In the evaluation of the segmentation results and searching for the optimum param-eters, the segmentation outputs are evaluated in a wide perspective through multiple

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


0 100 m 200 m

(b)(a)

Figure 5. Segmentation outputs of a small part of the image (taken in July) with theparameters of (a) hs = 2, hr = 2, MR = 200 and (b) hs = 5, hr = 10, MR = 1000, respectively.

goodness measures. In total, 13 measures – over-segmentation, under-segmentation,area fit index (AFI), count over, count under, relative area metric, similar size index,quality rate, under merge, over merge indices, qLoc, relative position indices andsome weighted products of theirs – are utilized to compute both area and locationsimilarities between the segments and the well-defined reference fields selected out ofthe vector data. More detailed information about the goodness measures can be foundin Clinton et al. (2010). To make a comparison, a subset of representative 10% of allthe polygons, human delineated agricultural fields, is selected. Having computed thegoodness measures, the optimum parameters of the segments are determined basedon a ranking process of the results in terms of the goodness measures followed bymaking an optimization process. The evaluations indicate that the optimum param-eter combinations of the Kompsat-2 image (green, red and NIR) taken in June werefound as hs = 14, hr = 3 for the spatial and range parameters, respectively. The bestparameter combinations of the Kompsat-2 images taken in July and August wereestablished as hs = 5, hr = 10 and hs = 12, hr = 3, respectively. After the segmen-tation, two postprocessing operations – (i) buffering and (ii) line simplification – wereapplied to the segments to generate more representative objects. A small portion of thesegmented Kompsat-2 image taken in July is presented with two different parametercombinations in figure 5.

3.6 Image classification

In traditional remote sensing, image classification operations are performed based onthree categories: (i) unsupervised, (ii) supervised and (iii) hybrid (Lillesand et al. 2004).Although a significant number of different studies are available in the literature, mostof them use the supervised approach to obtain more reliable results (e.g. Lillesandand Kiefer 2000, Lu and Weng 2007, Turker and Ozdarici 2011). In the supervisedapproach, a sufficient number of samples is required as prior information to producerepresentative parameters for each class. On the other hand, defining these samplesis a very critical process, because the quality of the samples directly affects the per-formance of the classification (Chen and Stow 2002, Lu and Weng 2007). In orderto overcome those problems and minimize possible bias in the classifications, the

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


segmentation results are utilized to define homogenous training samples in this study.In this regard, first, the corresponding segmentation result is overlaid with each band(blue, green, red and NIR) of the single date MS Kompsat-2 image to find the bestrepresentative regions among all the available segments. Next, standard deviations ofthe pixels within the segments are computed for each image channel and stored ina database. Then, means of the standard deviations for the segments are computed.Finally, the segments having a standard deviation smaller than 2 are extracted as train-ing samples and labelled automatically with the help of reference information. Thisprocess is repeated for each single date Kompsat-2 image taken on different dates.In this way, a representative 10% of all the pixels are determined as training samplesfor each image (figure 6). The ranges of the training sizes lie between 0.1 and 1.2 ha,0.1 and 2.3 ha and 0.1 and 2.2 ha for the images taken in June, July and August,respectively. Spectral signatures of the crops computed for each date by the proposedapproach are provided in table 3. The selected training objects are then utilized asprior information to carry out the state-of-the-art maximum likelihood classification(MLC) of the 4 m MS Kompsat-2 images (blue, green, red and NIR) and 15 m EnvisatASAR data. We also test the classification results of the fused Kompsat-2 products.However, the overall accuracies do not exceed 0.2% of the classification results of theMS data. Therefore, we utilize the MS Kompsat-2 data of the classification processesto minimize the effect of spectral within-field variability, increase the computationalefficiency and save time. In order to make use of different characteristics of the EnvisatASAR data, backscattering coefficients (σ 0) are computed using the information ofincidence angle and calibration constants provided in the header file of the microwaveimages (Liu et al. 2006) (equation (2)):

σij = 10 log10

(DN2

ij

Ksin(αij))

, (2)

N 40° 13′ 43″N 40° 07′ 44″

E 28° 20′ 28″E 28° 10′ 31″

Figure 6. Distribution of training data overlaid with the green band of July image.

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


Tab

le3.

Spec

tral

sign

atur

esof

the

six

crop

type

sco

mpu

ted

for

each

date

byth

epr

opos

edtr

aini

ngsi

tese

lect

ion

stra

tegy

.

Kom

psat

-2M

S-bl

ueba

ndK

omps

at-2

MS-

gree

nba

ndK

omps

at-2

MS-

red

band

Kom

psat

-2M

S-N

IRba

ndE

nvis

atA

SAR

Cla

sses

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

June C

orn

48.0

38.

7640

.79

8.31

48.2

99.

4413

4.89

15.5

17.

931.

63To

mat

o10

2.84

13.4

096

.32

15.4

610

3.67

16.8

510

1.64

7.96

10.8

72.

02R

ice

45.1

121

.88

55.2

919

.84

61.3

719

.34

33.5

714

.78

7.98

3.60

Suga

rbe

et10

8.17

19.4

311

2.12

16.8

184

.19

18.6

513

2.37

17.3

914

.32

2.30

Whe

at79

.33

20.3

784

.47

18.8

112

1.82

22.0

976

.94

17.7

910

.35

2.73

Gra

ssla

nd70

.67

21.6

370

.58

19.5

480

.19

16.8

711

5.73

21.5

57.

771.

95Ju

ly Cor

n30

.15

7.43

43.1

811

.02

23.4

211

.46

98.3

730

.85

123.

5029

.76

Tom

ato

35.1

94.

1848

.04

5.77

23.5

46.

7813

2.15

30.6

416

0.04

23.5

0R

ice

23.9

93.

2844

.91

3.29

21.2

33.

4763

.82

12.6

416

5.48

31.1

5Su

gar

beet

39.7

83.

6554

.45

4.28

18.9

43.

0317

2.78

9.20

166.

0318

.81

Whe

at48

.56

3.14

73.1

04.

1262

.17

5.29

93.2

94.

1365

.80

18.8

3G

rass

land

51.6

85.

3072

.64

6.31

53.4

35.

2899

.61

10.9

981

.99

20.1

3A

ugus

tC

orn

11.2

23.

1021

.69

4.46

19.4

52.

9710

0.34

9.90

133.

8026

.43

Tom

ato

21.9

67.

6736

.57

5.68

28.2

23.

9610

3.90

17.0

416

7.80

24.4

1R

ice

9.82

2.33

31.0

75.

7825

.27

3.80

132.

3012

.33

124.

0938

.67

Suga

rbe

et14

.45

2.66

35.3

83.

0227

.96

2.48

119.

79.

5716

1.74

24.9

0W

heat

54.7

17.

3364

.36

7.07

48.4

05.

2078

.65

7.47

76.5

930

.81

Gra

ssla

nd40

.40

4.91

50.9

36.

2539

.01

4.81

72.9

07.

7576

.35

20.4

7

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


Figure 7. Mean backscatter values of the crops for each date.

where σ ij is in dB, DNij is the digital number of the (i, j) pixel, αij is the angle of the(i, j) pixel and K is the calibration constant.

In this way, characteristics of the crops are observed and this information is uti-lized in the classification operations (figure 7). During the classification processes,speckle-filtered Envisat ASAR data are included in the classifications as an additionalband.

In the first part of the classifications, six different crop types – corn, grassland, rice,sugar beet, tomato and wheat – that have a common planting period are classified.When performing the classifications, villages, roads and water canals are excludedmanually from the study area. First, the pixel-based classifications of the crops areperformed for each date (June, July and August). Next, the thematic maps producedby the pixel-based MLC are overlaid with the corresponding segment file. The fre-quencies of the pixels in each segment are then computed and the majority of the pixelvalue is assigned as a class label to the segments. In this way, segment-based classifica-tion results are acquired for the same crop types, independently, for each image. A newapproach is proposed to combine the information of all of the thematic maps producedin this study. The proposed approach is based on a hard classification strategy, in whichdistance maps are generated for each image and the maximum membership value isthen assigned as a class label to the pixels. Mahalanobis distance is utilized to computethe values of distance maps (equation (3)). In a distance map, the distance value andthe membership function of a class are inversely proportional. If the distance betweenthe pixel and the class mean is high, the membership function of the pixel is small andthe pixel is most likely to be incorrectly classified, which means the Mahalanobis dis-tance between the pixel and the signature mean is comparatively large (Pouncey andSwanson 1999, Jensen 2005). The distances computed for each single date image are

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


N 40° 13′ 43″ N 40° 07′ 44″ N 40° 13′ 43″

E 28° 20′ 28″E 28° 10′ 31″E 28° 20′ 28″E 28° 10′ 31″

N 40° 07′ 44″

(a) (b)

Figure 8. (a) The combined distance map and (b) the thematic output.

compared to each other, and then the smallest distances are calculated and the corre-sponding labels are combined in one thematic map by implementing a simple Matlabcode (figure 8):

D2 = (x − m)T C−1 (x − m) , (3)

where D2 is the Mahalanobis distance, x is the vector of data, m is the vector of meanvalues of independent variables, C–1 is the inverse covariance matrix of independentvariables and T indicates vector should be transformed.

According to the fieldwork, it is observed that a crop rotation occurs for the classpea and late corn in the study site. Therefore, in the second part of the classifications,the class pea and late corn are classified by using the single date images taken in Juneand July, respectively. Based on the phenological characteristics of the crops, only theJune image includes one part of the planting cycle of the class pea. After June, thiscrop type is replaced with the other crops (such as tomato, corn and sugar beet) in theregion. In order to find the fields cultivated with pea, a distance threshold is applied onthe distance map of June based on chi-square statistics at 95% confidence level. Thatthreshold is utilized to find out which pixels are most likely to be wrongly classified(figure 9). After that, the incorrectly classified pixels are extracted from the thematicmap and those areas are then reclassified. The same process is applied to the July imageto classify the late corn.

N 40° 07′ 44″

E 28° 10′ 31″ E 28° 20′ 28″

N 40° 13′ 43″N 40° 07′ 44″

E 28° 10′ 31″ E 28° 20′ 28″

N 40° 13′ 43″

(a) (b)

Figure 9. (a) Correctly classified pixels and (b) the reclassified map of June with the class pea.

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


3.7 Evaluation of the thematic maps

The accuracy of the thematic maps produced is computed with the help of a mostlyused evaluation method, the confusion matrix. The reference data are preparedby updating the database of the existing vector data (agricultural field boundariesproduced by digitizing cadastral maps) gathered out of the fieldwork performed con-currently with the image acquisitions. Three visits are made to the test site, and cropinformation such as canopy developments, irrigation and fertilization activities isrecorded into a database of the vector data. After transferring the vector data intoa raster format based on the crop information, around 30% of all the reference pixelsare utilized for validation in the accuracy assessment process. During the computa-tions, the calibration data are excluded from the validation part to prevent possiblebias on the assessment.

Evaluations are performed based on a simple random sampling method, in whichrandom points are distributed to the reference fields based on its class percentages.A set of 567 samples are determined based on equation (4) and used to perform theaccuracy assessment process (Jensen 2005). At the end of the analyses, overall accu-racy, kappa statistic and individual class accuracies are computed for the thematicmaps:

N = B∏

i

(1 −∏

i

)b2

i

, (4)

where N refers to the sample size,∏

i is the proportion of a population in the ith classout of k classes that has the proportion closest to 50%, bi is the desired precession forthis class (e.g. 5%), B explains the upper (α/k) × 100th percentile of the chi-square(χ2) distribution with one degree of freedom and k is the number of classes.

4. Results and discussion

Based on the pixel-based and segment-based approaches, 14 different confusion matri-ces are computed for separate sets of thematic maps to find the crop types (corn,tomato, rice, sugar beet, wheat and grassland) having a common planting cycle(table 4).

It is observed that the segment-based approach using the combined June–July–August images improves the classification accuracies compared to the pixel-basedapproach (table 5). The highest user’s and producer’s accuracies computed for thesegment-based approach are obtained for the class rice. This is due to the use ofEnvisat ASAR data in addition to optical data where the dielectric property of watercan be used in the classification. In Envisat ASAR images, rice fields appear dark dur-ing the early vegetative phase when the fields are flooded. Rice fields are retrieved fromthe multi-temporal radar imagery making use of the unique backscattering signatureof rice fields, which is significantly different from that of the other land cover. Thecrops corn, grassland, sugar beet, tomato and wheat yielded producer’s accuraciesof over 85%, which denotes a good identification performance. Nevertheless, user’saccuracies of corn and tomato are 78% and 77%, respectively, which demonstratesthat the segment-based approach using the combined images slightly overestimatesthe classification performance of these crops. The major confusion is between corn andtomato with a very similar spectral response and similar phenological characteristics.

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


Table 4. Pixel-based and segment-based results of the images having a common set of (six) croptypes.

Kompsat-2 MS Kompsat-2 MS and Envisat ASAR

Pixel-basedresults (%)

Segment-basedresults (%)

Pixel-basedresults (%)

Segment-basedresults (%)

Overallaccuracy

Overallkappa

Overallaccuracy

Overallkappa

Overallaccuracy

Overallkappa

Overallaccuracy

Overallkappa

June 45.67 0.34 51.85 0.41 52.20 0.42 61.19 0.53July 74.25 0.69 76.36 0.71 75.13 0.70 80.42 0.76August 71.95 0.66 78.13 0.73 79.18 0.75 82.71 0.79June–July 75.66 0.70 84.30 0.81 77.07 0.72 87.47 0.84June–August 72.66 0.67 84.48 0.81 78.83 0.74 86.59 0.83July–August 76.01 0.71 82.71 0.79 78.66 0.74 85.36 0.82Combined

map75.66 0.70 85.18 0.82 79.18 0.75 88.71 0.86

Table 5. Pixel-based and segment-based results of the combined image.

Corn Grassland Rice Sugar beet Tomato Wheat Row T UA (%)

Pixel-based results

Corn 67 6 6 1 19 4 103 65.04Grassland 1 77 0 0 1 19 98 78.57Rice 0 0 84 2 0 0 86 97.67Sugar beet 1 0 3 63 6 0 73 86.30Tomato 2 3 2 17 64 1 89 71.91Wheat 1 23 0 0 0 94 118 79.66Column T 72 109 95 83 90 118 567PA (%) 93.05 70.64 88.42 75.90 71.11 79.66

Overall A: 79.18 (%); kappa: 0.75

Segment-based results

Corn 63 0 2 1 9 6 81 77.77Grassland 0 93 0 0 0 2 95 97.89Rice 0 0 90 0 0 0 90 100Sugar beet 2 0 0 71 3 0 76 93.42Tomato 7 0 3 11 78 2 101 77.22Wheat 0 16 0 0 0 108 124 87.09Column T 72 109 95 83 90 118 567PA (%) 87.50 85.32 94.73 85.54 86.66 91.52

Overall A: 88.71 (%); kappa: 0.86

Note: PA, producer’s accuracy; UA, user’s accuracy. The bold numbers refer to diagonalelements of the confusion matrix. These numbers are employed to compute the overall accuracy.

It is observed that most of the individual accuracies computed for the pixel-basedapproach of the combined thematic map (June–July–August) are improved to someextent when the Envisat ASAR data are included in the classification analysis. Themaximum improvement of 11.11% is obtained for the class tomato in producer’saccuracy level and it is computed as 71.11%. User’s accuracy of the same crop type

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


is computed as 71.91%, which is around 4% higher than the user’s accuracy of thecorresponding result computed only for the thematic map of single date Kompsat-2image. This improvement is followed by the class corn with the producer’s anduser’s accuracies of around 6% and 8%, respectively. The individual accuracies of thegrassland and rice are also improved when the Envisat ASAR data are included inthe classification. While a slight decrease is observed around 2% for the producer’saccuracy of the grassland (68.80%), a reasonable increase is computed for the user’saccuracy level around 6% for the same crop type. The class rice provides an improve-ment of around 4% in the producer’s accuracy level, while this improvement is 1.29%in the user’s accuracy level of this crop. No significant change is observed for the sugarbeet and wheat when the Envisat ASAR data are included in the classification analysis.

The individual class accuracies of the segment-based approach performed with theMS Kompsat-2 images and Envisat ASAR data indicate that a dramatic increase ofabout 18% is observed for the user’s accuracy of the class corn (77.77%) while a rea-sonable drop of around 7% is obtained in the producer’s accuracy level of the samecrop type. Another dramatic improvement is computed for the class tomato around25.55% in the producer’s accuracy level (86.66%). A slight increase around 2% isobserved for the user’s accuracy of the tomato and computed as 77.22%, while theother classes exhibit a stable trend.

In table 6, the segment-based classification results of the single date MS Kompsat-2 and Envisat ASAR data taken in June are computed as 76.72% (overall accuracy)and 0.72 (kappa), respectively. The class pea exhibits around 80% producer’s accu-racy. It means 63 pixels out of 79 are correctly classified as pea and 16 pixels areomitted. Those pixels are incorrectly classified as corn (2), rice (5), sugar beet (2) andwheat (7) on the producer’s accuracy column. The user’s accuracy of the pea is com-puted as 64.28%. Only 63 pixels out of 98 classified as pea actually represent thatcategory on the ground. The highest confusion is observed between the corn and peawith 13 wrongly classified pixels. This can be explained by similar spectral responsecharacteristics of the crops. Accuracies of the other crops are not examined in detaildue to relatively low overall accuracies with respect to the results of the combinedmulti-temporal images.

Table 6. Segment-based results of the reclassified MS Kompsat-2 and Envisat ASAR data takenin June.

Corn Grassland Rice Sugar beet Tomato Wheat PeaRowtotal UA (%)

Corn 22 1 0 1 16 0 2 42 52.38Grassland 11 70 1 0 0 27 0 109 64.22Rice 1 0 70 1 0 3 5 80 87.50Sugar beet 5 0 0 79 1 0 2 87 90.80Tomato 0 0 0 0 58 0 0 58 100Wheat 5 7 1 0 0 73 7 93 78.49Pea 13 1 1 3 7 10 63 98 64.28Column total 57 79 73 84 82 113 79 567PA (%) 38.59 88.60 95.89 94.04 70.73 64.60 79.74

Overall A: 76.72%; kappa: 0.72


Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


Table 7. Segment-based results of the reclassified MS Kompsat-2 and Envisat ASAR data takenin July.

Corn Grassland RiceSugarbeet Tomato Wheat

Latecorn

Rowtotal UA (%)

Corn 67 1 1 1 29 1 11 111 60.36Grassland 0 71 0 0 0 3 0 74 95.94Rice 0 0 84 0 0 0 0 84 100Sugar beet 0 0 0 70 0 0 0 70 100Tomato 1 0 2 10 37 0 0 50 74Wheat 0 5 0 0 0 91 10 106 85.84Late corn 0 1 3 0 9 3 56 72 77.77Column total 68 78 90 81 75 98 77 567PA (%) 98.52 91.02 93.33 86.42 49.33 92.85 72.72

Overall A: 83.95%; kappa: 0.81


The overall accuracies and kappa values of the reclassified single date opticaland microwave images taken in July are computed as 83.95% and 0.81, respectively(table 7). Individual class accuracies of the late corn indicate that 56 pixels out of 77 arecorrectly classified as late corn and the producer’s accuracy of that crop is computedas 72.72%. Confusions are observed between the late corn and corn (as 11 pixels) andwheat (as 10 pixels). User’s accuracy of that class is obtained around 78%. Based onthe matrix, 56 pixels are correctly classified as late corn on the ground. The highestconfusion of 9 pixels is observed between tomato and late corn.

5. Conclusion

In this study, a segment-based classification approach is proposed to classify eight croptypes on a multi-temporal data set of Kompsat-2 and Envisat ASAR data acquiredover Karacabey Plain, Turkey. The classified thematic maps are evaluated by confusionmatrices based on pixel-based and segment-based approaches. Due to the inconsis-tent segments compared with the related reference objects, evaluation based on thesegment-based approach is still a challenging process. Therefore, the accuracies ofthe thematic maps are tested based on sampling random points on the study area.The results indicate that the highest overall accuracy of 88.71% and a kappa valueof 0.86 are achieved for the combined thematic map of June–July–August. When theEnvisat ASAR data are included in the segment-based classifications, the overall accu-racy is improved around 10% for the single date Kompsat-2 and Envisat ASAR imagestaken in June and around 4% for the single date optical and microwave images taken inJuly and August. Except for the class rice, the optical and microwave images taken inJune generally provide lower results than the others. This is due to the fact that most ofthe crops are planted in June and bare soil affects the spectral reflectance of the crops.Due to the water content of the rice fields, the class rice exhibits more different charac-teristics than the other crop types do. That characteristic makes the class rice distinctin the region and provides relatively high accuracies. In terms of the combined the-matic map of June–July–August, microwave data increase the accuracies around 3%.It is worth noting that the cumulative effect of the microwave data is incontrovertible

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


with respect to the accuracies of the combined thematic maps. The microwave dataimprove the individual class accuracies of the class corn, grassland, rice and tomatofor the pixel-based classification results. In the segment-based approach, the individualaccuracies of the class corn and tomato are improved when the Envisat ASAR dataare added to the classification. The class pea provides relatively low user’s accuracy(64.28%) when the classification of the single date June image is performed. In June,the pea fields are in the last planting period with less canopy development and baresoil effect, which may cause some spectral confusion between the pea and the othercrop types (especially for the corn and wheat). The low accuracy of the pea can beimproved by analysing additional images acquired in April and May. The same case isalso valid for the late corn. An image taken in September can increase the accuracy ofthis class.

Through the study, the performance of multi-temporal classification strategytogether with the segment-based approach is examined to classify multiple crop typesin agricultural land by using optical and microwave data. In the proposed methodol-ogy, without any complex rules, the crop yield estimation is performed reliably withhigh accuracies over 85% for the test site. It is observed that multi-temporal andmulti-sensor data sets improve the crop identification through a traditional classifi-cation method, MLC. The training site selection, based on objects obtained from thesegmentation results, is helpful in the case of lack of expert knowledge.

The final products produced by the study can be utilized effectively to control waterrequirements and prepare irrigation plans of agricultural lands, which positively affectthe management strategies. In order to strengthen our results, further research maybe conducted to evaluate the method on other agricultural lands having different croptypes and topology with multiple image classification methods.

AcknowledgementsThe Kompsat-2 data were provided by a project called DAP-2008-07-02-07, fundedby the Geodetic and Geographic Information Technologies (GGIT) Department ofthe Middle East Technical University (METU) in Turkey. The Envisat ASAR datawere supplied by a Category-1 ESA project (Project No: 4825). This study wasalso supported for six months in 2010 by the Scientific and Technological ResearchCouncil of Turkey. We thank Assoc. Prof. Dr Lutfi Suzen (Department of GeologicalEngineering at METU) and Assistant Prof. Dr Ilkay Ulusoy (Department of Electricaland Electronics Engineering at METU) for their valuable comments during the studyand our colleagues Ali Ozgun Ok and Resat Gecen who contributed to the fieldworkof the study. We thank the farmers and staff working in the irrigation department ofthe region for contributing to our study with respect to the field observations. Finally,we thank two anonymous reviewers for their valuable contributions to improve thestudy.

ReferencesASAR PRODUCT HANDBOOK, 2009, Available online at: http://envisat.esa.int/handbooks/asar

(accessed 15 November 2009).BAN, Y., 2003, Synergy of multitemporal ERS-1 SAR and Landsat TM data for classification

of agricultural crops. Canadian Journal of Remote Sensing, 29, pp. 518–526.BAN, Y., HU, H. and RANGEL, I.M., 2010, Fusion of QuickBird MS and RADARSAT SAR

data for urban land-cover mapping: object-based and knowledge-based approach.International Journal of Remote Sensing, 31, pp. 1391–1410.

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14

http://envisat.esa.int/handbooks/asar


BLAES, X., HOLECZ, F., VAN LEEUWEN, H.J.C. and DEFOURNY, P., 2007, Regional crop mon-itoring and discrimination based on simulated ENVISAT ASAR wide swath modeimages. International Journal of Remote Sensing, 28, pp. 371–393.

BLAES, X., VANHALLE, L. and DEFOURNY, P., 2005, Efficiency of crop identification based onoptical and SAR image time series. Remote Sensing of Environment, 96, pp. 352–365.

BLASCHKE, T., 2010, Object based image analysis for remote sensing. ISPRS Journal ofPhotogrammetry and Remote Sensing, 65, pp. 2−16.

CHABRIER, S., EMILE, B., ROSENBERG, C. and LAURENT, H., 2006, Unsupervised performanceevaluation of image segmentation. EURASIP Journal on Applied Signal Processing,2006, 96306, doi:10.1155/ASP/2006/96306.

CHEN, D. and STOW, D., 2002, The effect of training strategies on supervised classification atdifferent spatial resolutions. Photogrammetric Engineering and Remote Sensing, 68, pp.1155–1161.

CHEN, Z., ZHAO, Z., GONG, P. and ZENG, B., 2006, A new process for the segmentation ofhigh resolution remote sensing imagery. International Journal of Remote Sensing, 27,pp. 4991–5001.

CHENG, H.D., JIANG, X.H., SUN, Y. and WANG, J., 2001, Color image segmentation: advancesand prospects. Journal of Pattern Recognition, 34, pp. 2259–2281.

CHENG, Y., 1995, Mean shift, mode seeking, and clustering. IEEE Transactions on PatternAnalysis and Machine Intelligence, 17, pp. 790–799.

CLINTON, N., HOLT, A., SCARBOROUGH, J., YAN, L.I. and GONG, P., 2010, Accuracy assessmentmeasures for object-based image segmentation goodness. Photogrammetric Engineeringand Remote Sensing, 76, pp. 289–299.

COMANICIU, D. and MEER, P., 1997, Robust analysis of feature spaces: color image segmenta-tion. In IEEE Conference on Computer Vision and Pattern Recognition, 17–19 June 1997,San Juan, Puerto Rico, pp. 750–755.

COMANICIU, D. and MEER, P., 2002, Mean shift: a robust approach toward feature spaceanalysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24, pp.603–619.

CORCORAN, P., WINSTANLEY, A. and MOONEY, P., 2010, Segmentation performance evalua-tion for object-based remotely sensed image analysis. International Journal of RemoteSensing, 31, pp. 617–645.

DE WIT, A.J.W. and CLEVERS, J.G.P.W., 2004, Efficiency and accuracy of per-field classifica-tion for operational crop mapping. International Journal of Remote Sensing, 25, pp.4091–4112.

FROST, V.S., STILES, J.A., SHANMUGAN, K.S. and HOLTZMAN, J.C., 1982, A model for radarimages and its application to adaptive digital filtering of multiplicative noise. IEEETransactions on Geosciences and Remote Sensing, 4, pp. 157–166.

GE, F., WANG, S. and LIU, T., 2006, Image-segmentation evaluation from the perspectiveof salient object extraction. In Proceedings of the Computer Society Conference onComputer Vision and Pattern Recognition (CVPR’06) 17–22 June 2006, New York, NY,pp. 1146–1153.

GONG, P. and HOWARTH, P.J., 1990, The use of structural information for improving land-cover classification accuracies at the rural-urban fringe. Photogrammetric Engineeringand Remote Sensing, 56, pp. 67–73.

GONG, P. and HOWARTH, P.J., 1992, Frequency-based contextual classification and grey-levelvector reduction for land-use identification. Photogrammetric Engineering and RemoteSensing, 58, pp. 423–437.

GONG, P., MARCEAU, D. and HOWARTH, P.J., 1992, A comparison of spatial feature extrac-tion algorithms for land-use mapping with SPOT HRV data. Remote Sensing ofEnvironment, 40, pp. 137–151.

HENDERSON, F.M. and LEWIS, A.J., 1998, Principles & Applications of Imaging Radar, Manualof Remote Sensing, 3rd ed., pp. 67–68 (New York, NY: John Wiley & Sons).

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14


JENSEN, J.R., 2005, Introductory Digital Image Processing, 3rd ed., pp. 175, 198–202 (New York,NY: John Wiley & Sons).

KUAN, D.T., SAWCHUCK, A.A., STRAND, T.C. and CHAVEL, P., 1985, Adaptive noise smoothingfilter for images with signal-dependent noise. IEEE Transactions on Pattern Analysis andMachine Intelligence, 7, pp. 165–177.

LEE, J.S., 1980, Digital image enhancement and noise filtering by use of local statistics. IEEETransactions on Pattern Analysis and Machine Intelligence, 2, pp. 65–168.

LEE, J.Y. and WARNER, T.A., 2006, Segment based image classification. International Journal ofRemote Sensing, 27, pp. 3403–3412.

LI, P. and XIAO, X., 2007, Multispectral image segmentation by a multichannel watershed-basedapproach. International Journal of Remote Sensing, 28, pp. 4429–4452.

LILLESAND, T.M. and KIEFER, R.W., 2000. Remote Sensing and Image Interpretation (NewYork: John Wiley and Sons).

LILLESAND, M., KIEFER, R.W. and CHIPMAN, J.W., 2004, Remote Sensing and ImageInterpretation, 5th ed., p. 638 (Hoboken, NJ: Wiley).

LIU, J. and YANG, Y.-H., 1994, Multi-resolution color image segmentation. IEEE Transactionson Pattern Analysis and Machine Intelligence, 16, pp. 689–700.

LIU, L., WANG, J., BAO, Y., HUANG, W., MA, Z. and ZHAO, C., 2006, Predicting winter wheatcondition, grain yield and protein content using multi-temporal Envisat-ASAR andLandsat TM satellite images. International Journal of Remote Sensing, 27, pp. 737–753.

LU, D. and WENG, Q., 2007, A survey of image classification methods and techniques forimproving classification performance. International Journal of Remote Sensing, 28, pp.823–870.

MARTIN, D., FOWLKES, C., TAL, D. and MALIK, J., 2004, Learning to detect natural imageboundaries using local brightness, color and texture. IEEE Transactions on PatternAnalysis and Machine Intelligence, 26, pp. 530–549.

MINIMAX, 2004, Reference Manual (Calgary: CSI wireless).MUELLER, M., SEGL, K. and KAUFMANN, H., 2003, Extracting characteristic segments in high-

resolution panchromatic imagery as basic information for objects-driven image analysis.Canadian Journal of Remote Sensing, 29, pp. 453–457.

NIK SYSTEM, 2008, Kompsat-2 technical notes. Available online at: http://nik.com.tr/2008/tr/sistem/uydu_goruntuleri/kompsat2.html.

OZDARICI, A. and AKYUREK, Z., 2009, Evaluating the contribution of image fusion meth-ods into the classification accuracies of agricultural lands. In Proceedings of TürkiyeUlusal Fotogrammetri ve Uzaktan Algılama Birligi V. Teknik Sempozyumu (TUFUAB),cd, 23–25 February 2009, Turkey.

OZDARICI, A. and AKYUREK, Z., 2010, A comparison of SAR filtering techniques on agri-cultural area identification. In Proceedings of ASPRS 2010 Annual Conference, cd,26–30 April, San Diego, CA.

OZDARICI OK, A. and AKYUREK, Z., 2011, Evaluation of image fusion methods on agriculturallands. Journal of Earth Science and Engineering 1, 107–113.

PARMUCHI, M.G., KARSZENBAUM, H. and KANDUS, P., 2002, Mapping wetlands using multi-temporal RADARSAT-1 data and a decision-based classifier. Canadian Journal ofRemote Sensing, 28, pp. 175–186.

PCI GEOMATICA, 2009, Geomatica OrthoEngine Course Guide (Richmond Hill, ON: PCIGeomatics Enterprises Inc.).

PENÃ-BARRAGÁN, M.J., NGUGI, M.K., PLANT, R.E. and SIX, J., 2011, Object-based cropidentification using multiple vegetation indices, textural features and crop phenology.Remote Sensing of Environment, 115, pp. 1301–1316.

POUNCEY, R. and SWANSON, K., 1999, ERDAS Manual, 5th ed., p. 540 (Norcross, GA:ERDAS).

RICHTER, R., 1990, A fast atmospheric correction algorithm applied to LANDSAT TM images.International Journal of Remote Sensing, 11, pp. 159–166.

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14

http://nik.com.tr/2008/tr/sistem/uydu_goruntuleri/kompsat2.html.

http://nik.com.tr/2008/tr/sistem/uydu_goruntuleri/kompsat2.html.


RYDBERG, A. and BORGEFORS, G., 2001, Integrated method for boundary delineation of agri-cultural fields in multispectral satellite images. IEEE Transactions on Geoscience andRemote Sensing, 39, pp. 2514–2520.

SCHOENMAKERS, R.P.H.M., VAN LEEUWEN, H.J.C., LEMOINE, G.G. and NEZRY, E., 1994,Segmentation of combined high resolution optical and radar imagery for the determina-tion of field inhomogenities. Proceedings of the IGARSS, 8–12 August 1994, Pasadena,CA, pp. 2137–2139.

SERKAN, M., MUSAOGLU, N., KIRKICI, H. and ORMECI, C., 2008, Edge and fine detailpreservation in SAR images through speckle reduction with an adaptive mean filter.International Journal of Remote Sensing, 29, pp. 6727–6738.

SKRIVER, H., MATTIA, F., SATALINO, G., BALENZANO, A., PAUWELS, V.R.N., VERHOEST,N.E.C. and DAVIDSON, M., 2011, Crop classification using short-revisit multitempo-ral SAR data. IEEE Journal of Selected Topics in Applied Earth Remote Sensing, 4, pp.423–431.

SMITH, G.M. and FULLER, R.M., 2001, An integrated approach to land cover classification:an example in the Island of Jersey. International Journal of Remote Sensing, 22, pp.3123–3142.

STANKIEWICZ, K.A., 2006, The efficiency of crop recognition on ENVISAT ASAR imagesin two growing season. IEEE Transactions on Geoscience and Remote Sensing, 44, pp.806–814.

TURKER, M. and ARIKAN, M., 2005, Sequential masking classification of multi-temporalLandsat7 ETM+ images for field-based crop mapping in Karacabey, Turkey.International Journal of Remote Sensing, 26, pp. 3813–3830.

TURKER, M. and OZDARICI, A., 2011, Field-based crop classification using SPOT4, SPOT5,IKONOS, and QuickBird imagery for agricultural areas: a comparison study.International Journal of Remote Sensing, 32, pp. 9735–9768.

WANG, D., LIN, H., CHEN, J., ZHANG, Y. and ZENG, Q., 2010a, Application of multi-temporal ENVISAT ASAR data to agricultural area mapping in the Pearl River Delta.International Journal of Remote Sensing, 31, pp. 1555–1572.

WANG, L., GONG, P., YING, Q., YANG, Z., CHENG, X. and RAN, Q., 2010b, Settlement extrac-tion in the north china plain using Landsat and Beijing-1 multispectral data with animproved watershed segmentation algorithm. International Journal of Remote Sensing,31, pp. 1411–1426.

WEIDNER, U., 2008, Contribution to the assessment of segmentation quality for remote sens-ing applications. In Proceedings of the 21st Congress for the International Society forPhotogrammetry and Remote Sensing, 3–11 July, Beijing, China.

XIAO, P., FENG, X., AN, R. and ZHAO, S., 2010, Segmentation of multispectral high-resolutionsatellite imagery using log gabor filters. International Journal of Remote Sensing, 31, pp.1427–1439.

YU, Q., GONG, P., CLINTON, N., BIGING, G. and SCHIROKAUER, D., 2006, Object-baseddetailed vegetation mapping using high spatial resolution imagery. PhotogrammetricEngineering and Remote Sensing, 72, pp. 799–811.

ZHAN, Q., MOLENAAR, M., TEMPFLI, K. and SHI, W., 2005, Quality assessment for geo-spatialobjects derived from remotely sensed data. International Journal of Remote Sensing, 26,pp. 2953–2974.

ZHANG, Y.J., 1996, A survey on evaluation methods for image segmentation. PatternRecognition, 29, pp. 1335–1346.

ZHANG, Y.J., 2001, A review of recent evaluation methods for image segmentation.In International Symposium on Signal Processing and its Applications (ISSPA),13–16 August 2001, Kuala Lumpur, Malaysia, pp. 148–151.

Dow

nloa

ded

by [

The

UC

Irv

ine

Lib

rari

es]

at 1

6:22

02

Nov

embe

r 20

14

A segment-based approach to classify agricultural lands by using multi-temporal optical and...

Documents

Transcript of A segment-based approach to classify agricultural lands by using multi-temporal optical and...