A new approach for land cover classification and ...

Remote Sensing of Environment 177 (2016) 37–47

Contents lists available at ScienceDirect

Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse

A new approach for land cover classification and change analysis: Integrating backdating and an object-based method Wenjuan Yu a,b, Weiqi Zhou a,⁎, Yuguo Qian a, Jingli Yan a,b a b

State Key Laboratory of Urban and Regional Ecology, Research Center for Eco-Environmental Sciences, Chinese Academy of Sciences, No. 18 Shuangqing Road, Beijing 100085, China University of Chinese Academy of Sciences, No. 19A Yuquan Road, Beijing 100049, China

a r t i c l e

i n f o

Article history: Received 18 July 2015 Received in revised form 2 February 2016 Accepted 12 February 2016 Available online xxxx Keywords: Land cover classification Change analysis Backdating Object-based method Beijing

a b s t r a c t Accurate information on land use and land cover (LULC) change is crucial for ecosystem monitoring and environmental change studies. Updating/backdating approaches have been increasingly used for LULC classification and change analysis, but mostly based on pixels. Here, we presented a new approach, an object-based backdating approach which integrates the backdating approach with an object-based method, and further compared it with the pixel-based backdating approach. We tested the new approach by using Landsat TM data collected in 2001 and 2009 at the Beijing metropolitan region. We found that: 1) an object-based backdating approach achieved higher accuracy for change detection, LULC classification and change analysis than the pixel-based backdating approach. With the object-based approach, the overall accuracies for the classification and change analysis were 84.33% (versus 69.33% for a pixel-based approach), and 80.00% (versus 70.50% for a pixel-based approach), respectively. 2) The object-based backdating approach greatly increases the efficiency because classification and change analysis are only conducted for locations with changes. The increase in efficiency is particularly important for LULC classification and change analysis conducted at a large area, for example, at the national or global scale. © 2016 Elsevier Inc. All rights reserved.

1. Introduction Land use/land cover (LULC) change affects local and regional climate, carbon, water, and biodiversity, and therefore has been recognized as one of the major components of environmental change (Grimm et al., 2008; Turner, Lambin, & Reenberg, 2007). Accurate information on LULC and change is crucial for ecosystem monitoring, environmental change studies, and land management and planning (Turner et al., 2007). Remote sensing data have been widely used for LULC classification and change analysis, as these data explicitly reveal spatial patterns of LULC and change over a large geographic area in a recurrent and consistent way (De Fries, Hansen, & Townshend, 1998; Homer et al., 2007; Peng, Liu, Shen, Han, & Pan, 2012; Vogelmann, Howard, & Yang, 2001; Zhang et al., 2014). Various methods have been developed for land cover change analysis using remotely sensed data (Coppin, Jonckheere, Nackaerts, Muys, & Lambin, 2004; Hussain, Chen, Cheng, Wei, & Stanley, 2013; Lu, Mausel, Brondizio, & Moran, 2004; Tewkesbury, Comber, Tate, Lamb, & Fisher, 2015). These methods may be broadly classified into two categories: post-classification comparison and pre-classification change detection (Lu et al., 2004; Singh, 1989; Zhou, Troy, & Grove, 2008). The first approach generates the multi-temporal LULC maps independently, and then identifies and quantifies the changes by comparing the ⁎ Corresponding author. E-mail address: [email protected] (W. Zhou).

http://dx.doi.org/10.1016/j.rse.2016.02.030 0034-4257/© 2016 Elsevier Inc. All rights reserved.

classification maps (Deng, Wang, Hong, & Qi, 2009; Ellis & Porter-Bolland, 2008; Yuan, Sawaya, Loeffelholz, & Bauer, 2005). Preclassification change detection techniques typically identify changes by comparing multi-temporal imagery directly, without classification. With high temporal resolution data, such as AVHRR, SPOTVEGETATION and MODIS data, change detection can be conducted based on characterizing spectral trajectories of land cover by dense time series data (Bontemps, Bogaert, Titeux, & Defourny, 2008; Eklundh, Johansson, & Solberg, 2009; Hansen & DeFries, 2004). With the opening of Landsat data archive on the long-term data accumulation, there has been increasing interest in applying dense time series Landsat data on change detection (Wulder, Masek, Cohen, Loveland, & Woodcock, 2012; Zhu & Woodcock, 2014). Because of the advantage of high temporal frequency, many quite subtle disturbance events of forest, such as defoliation, diseases, insect pests and regeneration, can be captured based on the change of vegetation spectral attribution (Goodwin et al., 2008; Hermosilla, Wulder, White, Coops, & Hobart, 2015; Zhu, Woodcock, & Olofsson, 2012). In addition to its wide applications in forest ecosystems, such method has also been applied to quantify changes of impervious surfaces in urban environments (Powell, Cohen, Yang, Pierce, & Alberti, 2008; Schneider, 2012), coral reef health (Palandro et al., 2008) and fire events (Röder, Hill, Duguy, Alloza, & Vallejo, 2008). Pre-classification change detection techniques, whether using dense time series image data or not, generally only generates “change” vs. “no-change” maps, but do not specify the type of change (Berberoglu & Akin, 2009; Lu et al., 2004; Singh, 1989).

38

W. Yu et al. / Remote Sensing of Environment 177 (2016) 37–47

Updating/backdating, an approach that has long been used in visual interpretation (Linke et al., 2009; McDermid et al., 2008; Zhou, Huang, & Cadenasso, 2011), has been increasingly applied in automatic change analysis and classifications (e.g., Xian & Homer, 2010, Jin et al., 2013). An updating/backdating approach can be considered as a synthesis of the post-classification comparison and pre-classification change detection (Xian & Homer, 2010). This approach typically started with an existing map, frequently referred as the reference map, based on which the classification and change analysis are conducted. One of the most remarkable examples is the generation of the 2006 National Land Cover Database (NLCD 2006) by updating the NLCD 2001 (Xian & Homer, 2010), which is now used as the reference dataset to create the 2011 NLCD, also using an updating approach (Jin et al., 2013). Previous research has shown that an updating/backdating approach has several advantages in terms of both efficiency and accuracy (Linke et al., 2009; Xian, Homer, & Fry, 2009). For example, with this approach, classification is only conducted at locations with changes, which greatly reduce the time for classifications of the entire area of interest, compared with the post-classification comparison approach (Xian et al., 2009). An updating/backdating approach also helps maintain the consistency of the features with on changes, and largely reduce “false changes” (Xian et al., 2009). In addition, this approach provides an opportunity to correct the errors of reference map in the change analysis process, which greatly improving the reliability of change analysis (Perdigao & Annoni, 1997). An updating/backdating approach can be implemented on pixels, or image objects (Linke et al., 2009; Xian et al., 2009). Object-based image analysis is quickly gaining acceptance among remote sensors, and has demonstrated great potential for classification and change detection, compared to pixel-based approach (Blaschke, 2010; Myint, Gober, Brazel, Grossman-Clarke, & Weng, 2011; Zhou et al., 2008). A considerable amount of research have shown that an object-based approach is superior to a pixel-based approach because it can greatly reduce the “salt and pepper” effect, provide an effective way in incorporating spatial, textural, and neighborhood relation in classification and change analysis (Blaschke, 2010; Hansen & Loveland, 2012; Zhou & Troy, 2008). However, previous studies using an automatic updating/ backdating approach, have been largely applying a pixel-based method (e.g., Xian & Homer, 2010, Jin et al., 2013). Relatively few studies have investigated how an object-based updating/backdating approach performs. This study aims to fill this gap. Here, we present a new approach that integrates the backdating approach with an object-based method for LULC classification and change analysis. We tested this approach using the Beijing metropolitan region as a case study, where great changes occurred during the time period (i.e., 2001–2009) we chose. We further compared this new approach with the typically used pixel-based backdating approach. 2. Study area and data acquisition Our research focused on the Beijing metropolitan region (Fig. 1). This study area contains an urban–suburban–rural gradient which presents the land use intensity diminishes from the central Beijing city to the rural fringe. During the period of 2001 to 2009, Beijing metropolitan has experienced a dramatic LULC change from agriculture land to developed land in the suburban area. In addition, a mix of farmland and forest in the rural fringe was also highly dynamic. Therefore, this study area is well suited for the goals of this research. We used two scenes of Landsat TM data collected in 2001 (2001/08/ 31) and 2009 (2009/09/22), and a land cover thematic map of 2009 (Hereinafter referred to as Map2009) (Fig. 1). These two images were obtained from United States Geological Survey (USGS), with primary process through Level 1 Product Generation System (LPGS), which included systematic radiometric and geometric corrections (Chander, Markham, & Helder, 2009). We further normalized the 2001 TM data using the 2009 TM data as the reference (Yang & Lo, 2000). Map2009

has six land cover types including forest, grass, water, farmland, developed land and barren. Developed land includes urban residential, commercial, industrial and transportation lands, and rural residential lands (Homer et al., 2007). It generated from the 2009 TM data, using an object-based classification approach, and thus had the spatial resolution of 30 m. We did extensive manual editing to refine the classification by referring to higher resolution SPOT image data (2.5 m). Therefore, Map2009 had an overall accuracy of 96%. 3. Methods For comparison purpose, we implemented the backdating approach with two different methods: 1) the one integrating backdating and an object-based method (BOB, hereafter); and 2) the other integrating backdating and pixel-based method (BPB, hereafter). For both methods, we first used change vector analysis (CVA) to identify areas with changes (image objects in BOB, and pixels in BPB) from 2001 to 2009. We then classified these areas with changes, and backdated the areas with no changes based on Map2009 (Fig. 2). We did not use any ancillary data to aid in classification, and not conduct manual editing. 3.1. Integrating backdating and an object-based method 3.1.1. Image segmentation With the BOB approach, we first segmented the 2001 image into objects. We used the multi-resolution segmentation algorithm that was embedded in eCognition software (Baatz & Schäpe, 2000). When implementing the segmentation, the classification map, Map2009, was used as the thematic layer. Consequently, the generated objects were not allow to across any of the borders separating different thematic classes of Map2009, and thus fell within or shared the boundaries of different land cover class of the thematic layer (Zhou et al., 2008) (Fig. 3). The multi-resolution segmentation algorithm uses a bottom-up region merging technique, with each pixel initialized as a single segment (Baatz & Schäpe, 2000). Spatially adjacent segments are then merged based on the degree of heterogeneity that is largely defined by the parameter - scale. The process stops when there are no more possible merges given the defined scale parameter (Zhou & Troy, 2008). The greater the scale parameter, more heterogeneity allowed within each object, and the larger the average size of the objects. Two other parameters, color and shape, can also be set to determine the relative weighting of reflectance and shape in defining segments. The total weighted value of color and shape equals to one (Trimble, 2012). Previous studies showed that a higher weight, typically up to 0.9, should be given to color for better segmentation results (Mathieu, Aryal, & Chong, 2007; Pu, Landry, & Yu, 2011). Therefore, we set the weights as 0.9 and 0.1, respectively. As the average size of land cover patches and their changes varied by different types, there was no one scale fitting for all the land cover types. For example, the forest patches were larger than patches of water and grass. Therefore, we created a 3-level hierarchy of image objects with the scale parameters setting as 10 (Level 1), 30 (Level 2), and 50 (Level 3) (Benz, Hofmann, Willhauck, Lingenfelder, & Heynen, 2004). The three values of the scale parameter were determined by testing different parameter values and visually interpreting the image segmentation results (Mallinis, Koutsias, Tsakiri-Strati, & Karteris, 2008; Zhou & Troy, 2008). Specifically, Objects created at Level 1 were used to detect changes for the classes of water, grass and barren, the size of whose changes tended to be small. Objects at Level 2 were created to identify changes for farmlands and developed lands, and those at Level 3 for forested land (Fig. 3). 3.1.2. Change detection Change vector analysis has been widely used for land cover change detection (Chen, Gong, He, Pu, & Shi, 2003; Johnson & Kasischke, 1998; Nackaerts, Vaesen, Muys, & Coppin, 2005; Xian et al., 2009).


39

Fig. 1. Study area and datasets that were used in this research. Panel A: Landsat TM data acquired on August 31, 2001; panel B: Landsat TM data acquired on September 22, 2009; panel C: land cover classification map of 2009 (Map2009).

With the CVA method, the first step is to calculate the change vector (CV) based on normalized images acquired at two dates (Eq. (1)). 3 r1 −s1 6 r2 −s2 7 7 ΔV ¼ R−S6 4 ⋮ 5 rn −sn 2

3

2

CV j ðx; yÞ ¼

3

s1 r1 6 r2 7 6 s2 7 7 6 6 R ¼ 4 5; S ¼ 4 7 ⋮ ⋮ 5 rn sn

ð2Þ

the change magnitude is calculated with h i1=2 jΔVj ¼ ðr 1 −s1 Þ2 þ ðr 2 −s2 Þ2 þ … þ ðrn −sn Þ2 :

(

ð1Þ

Here, R and S are the vectors of two images in date t1 and t2, respectively, and n is the number of bands. 2

a multi-threshold method, that is, we identified the threshold values of change for different land cover classes (Xian et al., 2009). The threshold for each land cover type was defined by the following constrains

ð3Þ

One of the key steps of CVA is to determine the threshold of change magnitude for changes in land cover types. This has been typically done empirically. In general, it is inappropriate to use a single threshold to identify changes for all the land cover classes, which leads to either over extraction or under extraction (Chen et al., 2003). Here, we used

change if ΔV j ðx; yÞ ≥ V J þ a j σ j nochange f ΔV j ðx; yÞbV J þ a j σ j

ð4Þ

where j is a type of land cover, jV J j is the mean of change vector CVj for the land cover type j, σj is the standard deviation of the CVj, and aj is an adjustable parameter. The value of aj typically ranges from 0.0 to 1.5. Here, the value of aj was set as 1.5, which was determined followed the approach used in Morisette and Khorram (2000). To determine the values of jV J j and σj for each land cover class, we need a classification map to identify the land cover type for each object (Xian & Homer, 2010; Xian et al., 2009). Map2009 was used to serve this need. 3.1.3. Change analysis and classification Following the change detection, we created a class hierarchy for change analysis and classification (Zhou et al., 2008). Using this class hierarchy, we generated a series of rules (so-called process tree in eCognition). With the backdating approach, we obtained not only the resultant layer of change analysis, but also land cover classification

40


Fig. 2. Flowchart of classification and change analysis: integrating backdating and an object-based method (BOB, left); integrating backdating and a pixel-based method (BPB, right). We highlighted the different processes between the two approaches in gray.

map for 2001. Below we briefly discuss the class hierarchy, change analysis and classification process to classify these objects with changes. Objects with no changes from 2001 to 2009 were first classified as “Nochange”, which were later backdated, that is, assigned as the class types of 2009, based on Map2009. Objects with changes were assigned to the class of “Possible change”. The image objects of “Possible change” were then classified into 6 classes: ToForest, ToGrass, ToWater, ToFarmland, ToDevelopedLand and ToBarren, based on the classification map of Map2009. For example, if an object with change was forest in 2009, then this object was classified as ToForest, which meant that the object changed to forest from 2001 to 2009. In order to obtain the final change analysis results, each object had been assigned into one of the 6 classes was further classified into subclasses, by recognizing the land cover class of the object in 2001. For example, if an object assigned as “ToDevelopedLand” was forested in 2001, then this object would be classified as “01Fo-09De” meant that the object was forest in 2001 and changed to developed land in 2009. Theoretically, each of these 6 classes could have 5 subclasses, resulting in 30 subclasses in total. However, after change analysis, we found there were only 15 types of changes occurred in our study area, as shown in Fig. 4 in white. Some types of the other 15 classes of change (Fig. 4 in gray) such as conversion of grass to developed land that were not found in our study area, may occur in other regions (Yuan et al., 2005). Meanwhile, some types of change, such as the land cover changes from developed land to farmland, are also unlikely to occur, if not impossible, in other region. Consequently, we had only 15 subclasses in the resultant layer of change analysis (Fig. 4; Table 1). Following the class hierarchy, we created rules to classify each object into a subclass. The features and the value of features were listed in Table 1. Here, we briefly described the classified process and feature parameter for the final existing subclasses. For the objects of ToForest, if the value of Normalized Difference Water Index (NDWI), which was calculated based on green and near-infrared wavebands, was greater than 0, the objects were classified into the subclass of 01Wa-09Fo. The rest objects of ToForest, whose brightness (defined as the mean of the

DN values of 6 TM bands) was equal or greater than 45 and the Normalized Difference of Vegetation Index (NDVI) was equal or greater than 0.5, were assigned as 01Fa-09Fo. Then the 01De-09Fo was identified when the value of NDVI was less than 0.15 (Table 1). For the objects assigned as the other 5 classes, we used similar method for the subclasses classification based on the class hierarchy. We found that some subclasses, such as 01Wa-09Gr, 01Wa-09Fa and 01Wa-09De, were extracted by a same parameter with same value (Table 1). Some subclasses were extracted by combining the spatial and spectrum information. For instance, when NDVI was greater than 0.3 the objects assigned as ToDevelopLand were further classified as the class of 01Fa-09De. Further, the objects of ToDevelopLand, whose shape index value greater than 1.1 and lesser than 1.4 and the NDVI was greater than 0.25, were also assign as 01Fa-09De (Table 1). After obtaining the change layer of LULC from 2001 to 2009, Map0109_OB, we further produced the final LULC classification map of Map2001_OB by assigning the attribute of Nochange and merging the subclasses. The land cover attribute would be retained the same as Map2009 for the class of Nochange and the 15 change subclasses were merged if the attribute of LULC was similar in 2001. For instance, the class of water in Map2001_OB contained 4 change subclasses: 01Wa09Fo, 01Wa-09Gr, 01Wa-09Fa and 01Wa-09De. We then classified these 4 subclasses into the class of Water in Map2001_OB. The final LULC map of Map2001_OB only contained six typical land cover classes. 3.2. Integrating backdating and a pixel-based method Similar to the BOB approach, the first step of the BPB approach is to conduct change detection using the change vector analysis. The reference map, Map2009, was also used to calculate the thresholds for change. However, change detection was conduct at the pixel level, rather than image objects. Following the change detection, we used a maximum likelihood algorithm to classify the pixels with changes from 2001 to 2009. Map2009 was not used in this classification process, but used to backdate the class types of pixels with no change from 2001


Fig. 3. Image objects generated from segmentation at Level 1, Level 2 and Level3, overlaid on the image of 2001 and Map_2009.

Fig. 4. The class hierarchy for object-based change detection and classification. The change classes that did not exist in our study were marked in gray.

41

42


Table 1 Class name, code and description of Nochange class and the 15 subclasses which represent the land cover change from 2001 to 2009. The features used for classifications, and the thresholds of the features were also listed. Class name

Class description

Features

Value of features

Nochange (I) 01Wa-09Fo (II-1-1) 01Fa-09Fo (II-1-2) 01De-09Fo (II-1-3) 01Fo-09Gr (II-2-1) 01Wa-09Gr (II-2-2) 01Fa-09Gr (II-2-3) 01De-09Gr (II-2-4) 01Ba-09Gr (II-2-5) 01Gr-09Wa (II-3-1) 01De-09Wa (II-3-2) 01Wa-09Fa (II-4-1) 01Fo-09De (II-5-1) 01Wa-09De (II-5-2) 01Fa-09De (II-5-3) 01Fa-09Ba (II-6-1)

Land cover with no change Change from water to forest Change from farmland to forest Change from developed land to forest Change from forest to grass Change from water to grass Change from farmland to grass Change from developed land to grass Change from barren to grass Change from grass to water Change from developed to water Change from water to farmland Change from forest to developed land Change from water to developed land Change from farmland to developed land Change from farmland to barren

– NDWI NDVI, brightness NDVI NDVI NDWI NDVI NDVI NDVI, Brightness NDVI NDVI, brightness NDWI NDVI, RBFO NDWI NDVI, shape index NDVI

– NDWI N 0 Brightness ≥ 45 & NDVI N 0.5 NDVI b 0.15 NDVI N 0.5 NDWI N 0 NDVI N 0.3 NDVI b 0.15 NDVI b 0.1 & Brightness > 54 NDVI N 0.15 NDVI b 0.15 & 52 b brightness b 70 NDWI N 0 RBFO N 0.3 & NDVI N 0.5 NDWI N 0 NDVI N 0.3 and then (NDVI N 0.25 & 1.1 b shape index b 1.4) NDVI N 0.3

RBFO: Relative border to "forest with no change", with the value ranging from 0 to 1. RBFO means the ratio of an object's border shared with neighboring fine textured vegetation objects to the total border length. NDWI = (Xgreen − Xnir) / (Xgreen + Xnir), where Xgreen refers to the green band and Xnir refers to the near infrared band (McFeeters, 1996).

to 2009. Following the classification, result of change layer from 2001 to 2009, Map01-09_PB, was also generated. 3.3. Accuracy assessment We conducted the accuracy assessment for 6 layers. These included the two classification maps, Map2001_OB from BOB and Map2001_OB from BPB, the two “from-to” change layers, and the two layers from change detection. For all the layers, we used a stratified random sampling scheme to generate the checking points. For the classification maps and the change detection layers, we generated a total number of 300 sample points for each classification map, with at least 30 samples for each category (Fuller, Smith, & Devereux, 2003; Yuan et al., 2005). As the change layers had more class types, we used 400 sample points for the accuracy assessment. We used reference data created from visual interpretation of very high spatial resolution image data. Specifically, reference data for 2001 were generated from a combination of 1 m IKONOS and 0.6 m QuickBird image data, and that for 2009 were created from 2.5 m SPOT imagery. Error matrices were generated to calculate the overall accuracies, user's and producer's accuracy, and the Kappa statistics. 4. Results 4.1. Land cover classification accuracy and methods comparison The classification map generated from the BOB approach, Map2001_OB had overall accuracy and Kappa statistic of 84.33% and 0.80, respectively, higher than that of 69.33% and 0.62 for Map2001_PB derived from the BPB approach (Appendix A, Table A1 and Table A2). The User's Accuracy (User's Acc.) of Map2001_OB ranged from 50% to 91.43% (Appendix A, Table A1), whereas the User's Acc. of Map2001_PB ranged from 36.67% to 79.75% (Appendix A, Table A2). The User's Acc. for some classes, such as forest and developed land, in Map2001_OB was over 90%, while those in Map2001_PB were under 80%. Both the User's Acc. of grass in Map2001_OB and Map2001_PB were low (50% and 40.54%). The producer's accuracy for Farmlands increased from 54.55% to 91.67%, when using the object-based approach. This is because farmlands with no vegetation cover were frequently misclassified into developed lands in Map2001_PB, but not in the Map2001_OB (Fig. 5 D3). In addition, farmlands covered by vegetation were frequently misclassified into grass in Map2001_PB (Fig. 5 C3). These misclassification, however, were greatly reduced by using the object-based approach (Appendix A, Table A1 and Table A2). In

addition, the landscape in Map2001_PB was more fragmented than that in Map2001_OB (Fig. 5A, B, C3 and D3). The change layer of land cover classification from 2001 to 2009 by the BOB approach, Map01-09_OB (Fig. 6A), had overall accuracy and Kappa statistic of 80% and 0.78, respectively. The User's Acc. ranged from 56.67% to 94.34% (Appendix B, Table B1). The class of Nochange had the highest User's Acc. of 94.34%, whereas the class of 01Wa-09De had the lowest accuracy of 56.67%. Many of the subclasses had the User's Acc. greater than 70% except for the 01Wa-09De, 01Fa-09Gr and 01Fo-09Gr (Appendix B, Table B1). The Producer's Acc. for Nochange, however, was only 51.54% because the misclassifications of classes with changes into the class of Nochange. It should be noted that the accuracy values for the 4 classes, 01Wa-09Fo, 01Wa-09Gr, 01De-09Wa and 01Wa-09Fa, were absent because no random points could be generated for them due to the very small proportion of cover of these four classes (less than 0.01%). In order to compare the change analysis result, we selected same classes from the changed layer by BPB approach, Map01-09_PB (Fig. 6B), for the accuracy assessment. The Map01-09_PB had lower overall accuracy and Kappa statistic of 70.50% and 0.67, respectively (Appendix B, Table B2). The User's Acc. ranged from 43.33% to 90%. Comparatively, there only three subclasses had the User's Acc. over 80%, whereas there are seven by BOB approach. The class of Nochange had the User's Acc. of 79.71%, and the lowest Producer's Acc. of 46.67%, which was lower than derived from BOB approach. Additionally, except for the same four subclasses had no accuracy by BPB, the 01Ba-09Gr absent for the same reason. We found, with the BPB approach, there were more changes compared with the BOB approach. This is because the BPB approach identified some classes of the change that did not occur in reality in the Beijing metropolitan, such as the land cover change from developed land to farmland (01De-09Fa). 4.2. Change detection accuracy and method comparisons The portion of areas with changes detected by the BOB approach was slightly higher than that by BPB (16.76% vs. 13.77%). The accuracy assessment of change detection showed that the overall accuracy of BOB was 86.67%, greater than 79.33% of BPB. For the class of Possible change, the User's Acc. of the BOB was 77.92%, slightly higher than that of the BPB approach, 75.61% (Table 2). However, the User's Acc. of 89.69% for Nochange by BOB was significantly higher than that of 79.92% by BPB. The comparisons between the two change detection maps showed that the pixel-based method produced much more “salt and pepper” noises. The “salt and pepper” noises refer to sparsely occurring pixels/ objects that were classified differently from their neighboring/


43

Fig. 5. Land cover classification and change detection results. Panel A: Map2001_OB, classification map for 2001 generated from the object-based backdating approach; panel B: Map2001_PB, classification map for 2001 generated from the pixel-based backdating approach; panel C1: subset area of image 2009 (left) and image 2001 (right); panel C2: the change detection results generated by BOB (left), and BPB (right); panel C3: classification results generated by BOB (left), and BPB (right). The meanings of panels D1, D2, D3, E1, E2 and E3 are the same as C1, C2 and C3.

surrounding pixels/objects, but in fact belonged to the same class of the neighboring pixels/objects (Blaschke, Lang, Lorup, Strobl, & Zeil, 2000). Especially in locations with dramatic changes, most isolated pixels were detected as Possible change (right panel of Fig. 5C2 and D2). By contrast, the map derived from the object-based method had much less noises (left panel of Fig. 5C2 and D2). The BPB approach failed to detect changes of land cover with similar spectral information. For example, pixels that were developed land in 2009 converted from farmlands with no crop in 2001 were commonly considered as no change (left

panel of Fig. 5C2). Such changes, however, frequently could be detected using the BOB approach. 5. Discussion The comparisons on the two approaches indicated that the objectbased approach was superior to the pixel-based one. This is largely because the object-based approach provides an effective way in incorporating the spatial and textural features, and neighboring relations in

Fig. 6. The change analysis result presented as change layer from 2001 to 2009, where panel A was generated by BOB and panel B was generated by BPB.

44


Table 2 Error matrices and accuracy assessment for change detection of the BOB and BPB. Change detection assessment of BOB Classified data

Reference data Nochange

Possible change

Nochange 200 23 Possible change 17 60 Total 217 83 Producer's Acc. (%) 92.17% 72.29% Overall classification accuracy = 86.67% Overall Kappa statistics = 0.66

User's Acc. (%) Total 223 77 300

89.69 77.92

Change detection assessment of BPB Classified data

Reference data Nochange

Possible change

Nochange 207 52 Possible change 10 31 Column total 217 83 Producer's acc. (%) 95.39% 37.35% Overall classification accuracy = 79.33% Overall Kappa statistics = 0.39

User's Acc. (%) Total 259 41 300

79.92 75.61

the backdating process of classification and change analysis. More importantly, the object-based method provides an effective way to incorporate expert knowledge in change detection and the later classification. Rules can be developed to reduce the errors that were propagated from the attribute errors associated with the reference map, and position errors due to spatial misregistration (Zhou et al., 2008). For example, with the object-based approach, the false changes identified in the change detection process (Table 2) have been largely eliminated in the latter classification process by using the reference map and the expert knowledge, as suggested by the User's Acc. of 94.34% for the type of “Nochange” in the final change layer. Taking changes in farmland as an example (Fig. 5), during the change detection process, lots of farmlands with no real change from 2001 to 2009 were identified as changes, due to the vegetation cover changes in farmlands (Fig. 5, E2). These types of false changes were latterly reduced during the classifications and change analysis process because conversions from developed lands in 2001 to farmlands in 2009 were very unlikely to occur (Fig. 5, E3). Additionally, the BOB approach can largely reduce the errors caused by spatial misregistration, leading to greater accuracies of LULC classification and change analysis. Precise spatial registration is difficult, if not impossible, to achieve. Therefore, the misregistration error presented as the boundary inconsistent of land cover in change analysis is difficult to eliminate completely (Blaschke, 2005; Dai & Khorram, 1998; McDermid et al., 2008). Our results clearly showed the inconsistent boundary issue

Fig. 7. The misregistration issue presented under the two approaches. Panel A shows the result in thematic map, Map2009, panel B shows the results from the approach that integrates backdating and an object-based method, while panel C shows the one that integrates backdating and a pixel-based approach.


caused by spatial misregistration when applying the BPB approach (Fig. 7C). This problem, however, was largely eliminated with the BOB approach (Fig. 7B). With the object-based method, the reference map (i.e., Map2009) was used as the thematic layer in the segmentation procedure. Consequently, the generated objects were not allowed to cross any of the borders of different land cover classes defined in the reference map, resulting in consistent and reconcilable borders between data layers of 2001 and 2009. A backdating approach can greatly improve the efficiency for classification and change analysis, since it only conducts classifications in areas with changes, which may only cover a small proportion of the area of interest. Our results showed that, even in a rapidly urbanizing region such as the Beijing metropolitan area, less than 20% (16.76% from BOB and 13.77% from BPB) of the study region had LULC change. Therefore, compared to the post-classification comparison approach, a backdating/updating approach could be much more efficient (Xian & Homer, 2010). Because of this, a backdating/updating approach will be particularly useful if the study area expands to a national or global extent, or the observed period contains multi time series. In this study, we tested the new approach with the most commonly used medium resolution satellite imagery, Landsat TM data. It is one of the most important data sources for LULC monitoring at the global, national and regional scale. However, because of the limitation of spatial resolution of 30 m, relatively small but abundant changes in LULC, such as changes in urban greenspace in the inner cities, cannot be detected by TM data (Qian, Zhou, Yu, & Pickett, 2015). Imagery with higher spatial resolutions is needed to quantify such changes at very fine scales. Change detection based on very high spatial resolution data, however, could be very challenging, particularly when using pixel-based approaches (Hussain et al., 2013). For example, it is difficult to address the abundant false changes due to spatial misregistration (Zhou et al., 2008). With the increasing applications of the object-based approach on widely available high-spatial resolution imagery, therefore, it would be interesting to test the object-based updating/backdating approach for change analysis using high-spatial resolution image data. The BOB approach had similar accuracies for land cover classification and change analysis with those from previous studies (Aguirre-Gutiérrez, Seijmonsbergen, & Duivenvoorden, 2012; Chen,

45

Chen, Shi, & Yamaguchi, 2012; Desclee, Bogaert, & Defourny, 2006). But we did not conduct any manual editing for refinement, and did not use any ancillary data to aid in the change detection and classification. As ancillary data could be a significant contributor for the enhancement of land cover classification and change analysis (Rogan et al., 2003; Treitz & Howarth, 2000), we would expect better accuracies for LULC classification and change analysis with the new approach proposed in this study, with ancillary data. In fact, an object-based approach provides an effective means for integrating ancillary data in classification and change analysis (Zhou & Troy, 2008; Zhou et al., 2008). 6. Conclusion Accurate information on LULC change is crucial for ecosystem monitoring, environmental change studies, and land management and planning. Updating/backdating approaches have been increasingly used for LULC classification and change analysis, but mostly based on pixels. Here, we presented a new approach, an object-based backdating approach, and further compared it with pixel-based backdating approach. We found that: 1) an object-based backdating approach achieved higher accuracy for change detection, LULC classification and change analysis than the pixel-based backdating approach. Without any ancillary data, the BOB approach achieved acceptable overall accuracies for LULC classifications and change analysis. 2) The BOB approach greatly increases the efficiency for LULC classification and change analysis. Even in the rapidly urbanizing Beijing metropolitan region, less than 20% of the LULC changed. That means, we only need to classify less than 20% of study region, which provides distinct advantage for LULC classification and change analysis for a large area, for example, at the national or global scale. Acknowledgments The support of the National Natural Science Foundation of China (Grant nos. 41371197 and 41422104) and the One Hundred Talents program of Chinese Academy of Sciences is gratefully acknowledged. The authors would like to thank the editor in chief and anonymous reviewers for their helpful comments and suggestions.

Appendix A. Error matrix and accuracy assessment for the land cover classification Table A1 Error matrix and accuracy assessment of the six classes for Map2001_OB by using the approach of integrating backdating and an object-based method. Classified data

Reference data Forest

Forest 64 Grass 7 Water 1 Farmland 3 Developed land 2 Barren 0 Total 77 Producer's Acc. 83.12 Overall classification accuracy = 84.33% Overall Kappa statistics = 0.80

Grass

Water

Farmland

Developed land

Barren

Total

User's Acc. (%)

0 16 0 1 0 0 17 94.12

0 0 27 1 0 0 28 96.43

3 2 0 55 0 0 60 91.67

2 6 4 3 70 9 94 74.47

1 1 0 0 1 21 24 87.50

70 32 32 63 73 30 300

91.43 50.00 84.38 87.30 95.89 70.00

Table A2 Error matrix and accuracy assessment of the six classes for Map2001_PB by using the approach of integrating backdating and a pixel-based method. Classified data

Forest Grass Water Farmland

Reference data

User's Acc. (%)

Forest

Grass

Water

Farmland

Developed land

Barren

Total

53 3 1 9

0 15 1 0

0 1 24 0

9 7 4 42

5 11 1 2

2 0 1 0

69 37 32 53

76.81 40.54 75.00 79.25 (continued on next page)

46


Table A2 (continued) Classified data

Reference data Forest

Developed land 7 Barren 6 Total 79 Producer's Acc.(%) 67.09 Overall classification accuracy = 69.33% Overall Kappa statistics = 0.62

User's Acc. (%) Grass

Water

Farmland

Developed land

Barren

Total

0 1 17 88.24

1 1 27 88.89

7 8 77 54.55

63 3 85 74.12

1 11 15 73.33

79 30 300

79.75 36.67

Appendix B. Error matrix and accuracy assessment for land cover change layer from 2001 to 2009 Table B1 Error matrix and accuracy assessment for land cover change layer from 2001 to 2009 by the BOB approach. Classified data

Reference data I

II-1-1

I 50 0 II-1-1 0 0 II-1-2 4 0 II-1-3 2 0 II-2-1 4 0 II-2-2 0 0 II-2-3 6 0 II-2-4 0 0 II-2-5 1 0 II-3-1 3 0 II-3-2 0 0 II-4-1 0 0 II-5-1 7 0 II-5-2 13 0 II-5-3 5 0 II-6-1 2 0 Total 97 0 Producer's Acc. (%) 51.54 – Overall classification accuracy (%) Overall Kappa statistics

II-1-2

II-1-3

II-2-1

II-2-2

II-2-3

II-2-4

II-2-5

II-3-1

II-3-2

II-4-1

II-5-1

II-5-2

II-5-3

II-6-1

0 0 29 1 1 0 1 0 0 1 0 0 0 0 0 0 33 87.88

0 0 0 27 0 0 0 3 0 0 0 0 0 0 0 0 30 90.00

0 0 0 0 20 0 0 0 0 0 0 0 0 0 0 0 20 100.00 80.00 0.78

0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 –

1 0 0 0 1 0 19 0 0 1 0 0 0 0 0 0 22 86.36

0 0 0 1 0 0 0 27 3 0 0 0 0 0 0 0 31 87.10

0 0 0 0 1 0 0 0 26 0 0 0 0 0 0 0 27 96.30

0 0 0 0 1 0 0 0 0 23 0 0 0 0 0 0 24 95.83

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 –

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 –

0 0 0 0 2 0 0 0 0 0 0 0 22 0 2 0 26 84.62

0 0 0 0 0 0 0 0 0 0 0 0 0 17 0 0 17 100.00

2 0 0 0 0 0 4 0 0 2 0 0 1 0 35 3 47 74.47

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 25 25 100.00

Total

User's Acc. (%)

53 0 33 31 30 0 31 30 30 30 0 0 30 30 42 30 400

94.34 – 87.88 87.10 66.67 – 61.29 90.00 86.67 76.67 – – 73.33 56.67 83.33 83.33

Total

User's Acc. (%)

69 0 35 35 30 0 35 35 0 30 0 0 30 35 36 30 400 400

79.71 – 74.29 45.71 43.33 – 51.43 85.71 – 90 – – 83.33 77.14 69.44 63.33

Table B2 Error matrix and accuracy assessment for land cover change layer from 2001 to 2009 by the BPB approach. Classified data

Reference data I

II-1-1

I 56 0 II-1-1 0 0 II-1-2 8 0 II-1-3 7 3 II-2-1 11 0 II-2-2 0 0 II-2-3 11 0 II-2-4 2 0 II-2-5 0 0 II-3-1 3 0 II-3-2 0 0 II-4-1 0 0 1 0 II-5-1 II-5-2 2 0 II-5-3 11 0 II-6-1 8 0 Total 120 3 Producer's Acc. (%) 46.67 – Overall classification accuracy (%) Overall Kappa statistics

II-1-2

II-1-3

II-2-1

II-2-2

II-2-3

II-2-4

II-2-5

II-3-1

II-3-2

II-4-1

II-5-1

II-5-2

II-5-3

II-6-1

0 0 26 4 0 0 6 0 0 0 0 0 0 0 0 0 36 72.22

2 0 0 16 0 0 0 2 0 0 0 0 0 0 0 0 20 80

1 0 0 0 13 0 0 0 0 0 0 0 1 0 0 0 15 86.67 70.50 0.67

1 0 0 2 0 0 0 0 0 0 0 0 0 3 0 0 6 –

2 0 0 0 1 0 18 1 0 0 0 0 0 0 0 0 22 81.82

0 0 0 3 0 0 0 30 0 0 0 0 0 0 0 0 33 90.91

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 –

0 0 0 0 0 0 0 0 0 27 0 0 0 0 0 0 27 100

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 –

0 0 0 0 0 0 0 0 0 0 0 0 0 3 0 0 3 –

3 0 0 0 4 0 0 0 0 0 0 0 25 0 0 0 32 78.13

1 0 0 0 0 0 0 0 0 0 0 0 0 27 0 0 28 96.43

3 0 1 0 1 0 0 0 0 0 0 0 3 0 25 3 36 69.44

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 19 19 100

References Aguirre-Gutiérrez, J., Seijmonsbergen, A. C., & Duivenvoorden, J. F. (2012). Optimizing land cover classification accuracy for change detection, a combined pixel-based and object-based approach in a mountainous area in Mexico. Applied Geography, 34, 29–37.

Baatz, M., & Schäpe, A. (2000). Multiresolution segmentation: An optimizating approach for high quality multi-scale image segmentation. Beitra¨ge zum AGIT-Symposium (pp. 12–23). Salzburg: Angewandte Geographische Informationsverarbeitung XII. 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, 58, 239–258.

W. Yu et al. / Remote Sensing of Environment 177 (2016) 37–47 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. Blaschke, T. (2005). Towards a framework for change detection based on image objects. Göttinger Geographische Abhandlungen, 113, 1–9. Blaschke, T. (2010). Object based image analysis for remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing, 65, 2–16. Blaschke, T., Lang, S., Lorup, E., Strobl, J., & Zeil, P. (2000). Object-oriented image processing in an integrated GIS/remote sensing environment and perspectives for environmental applications. Environmental Information for Planning, Politics and the Public, 2, 555–570. 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. Chander, G., Markham, B. L., & Helder, D. L. (2009). Summary of current radiometric calibration coefficients for Landsat MSS, TM, ETM +, and EO-1 ALI sensors. Remote Sensing of Environment, 113, 893–903. Chen, J., Gong, P., He, C. Y., Pu, R. L., & Shi, P. J. (2003). Land-use/land-cover change detection using improved change-vector analysis. Photogrammetric Engineering and Remote Sensing, 69, 369–379. Chen, X. H., Chen, J., Shi, Y. S., & Yamaguchi, Y. (2012). An automated approach for updating land cover maps based on integrated change detection and classification methods. ISPRS Journal of Photogrammetry and Remote Sensing, 71, 86–95. Coppin, P., Jonckheere, I., Nackaerts, K., Muys, B., & Lambin, E. (2004). Digital change detection methods in ecosystem monitoring: A review. International Journal of Remote Sensing, 25, 1565–1596. Dai, X., & Khorram, S. (1998). The effects of image misregistration on the accuracy of remotely sensed change detection. Geoscience and Remote Sensing, IEEE Transactions, 36, 12. De Fries, R. S., Hansen, M., & Townshend, J. R. G. (1998). Global land cover classifications at 8 km spatial resolution: The use of training data derived from Landsat imagery in decision tree classifiers. International Journal of Remote Sensing, 19, 28. Deng, J. S., Wang, K., Hong, Y., & Qi, J. G. (2009). Spatio-temporal dynamics and evolution of land use change and landscape pattern in response to rapid urbanization. Landscape and Urban Planning, 92, 187–198. Desclee, B., Bogaert, P., & Defourny, P. (2006). Forest change detection by statistical object-based method. Remote Sensing of Environment, 102, 1–11. Eklundh, L., Johansson, T., & Solberg, S. (2009). Mapping insect defoliation in scots pine with MODIS time-series data. Remote Sensing of Environment, 113, 1566–1573. Ellis, E. A., & Porter-Bolland, L. (2008). Is community-based forest management more effective than protected areas? A comparison of land use/land cover change in two neighboring study areas of the Central Yucatan Peninsula, Mexico. Forest Ecology and Management, 256, 1971–1983. Fuller, R. M., Smith, G. M., & Devereux, B. J. (2003). The characterisation and measurement of land cover change through remote sensing: Problems in operational applications? International Journal of Applied Earth Observation and Geoinformation, 4, 11. Goodwin, N. R., Coops, N. C., Wulder, M. A., Gillanders, S., Schroeder, T. A., & Nelson, T. (2008). Estimation of insect infestation dynamics using a temporal sequence of Landsat data. Remote Sensing of Environment, 112, 3680–3689. Grimm, N. B., Faeth, S. H., Golubiewski, N. E., Redman, C. L., Wu, J. G., Bai, X. M., & Briggs, J. M. (2008). Global change and the ecology of cities. Science, 319, 756–760. Hansen, M. C., & DeFries, R. S. (2004). Detecting long-term global forest change using continuous fields of tree-cover maps from 8-km advanced very high resolution radiometer (AVHRR) data for the years 1982–99. Ecosystems, 7, 695–716. Hansen, M. C., & Loveland, T. R. (2012). A review of large area monitoring of land cover change using Landsat data. Remote Sensing of Environment, 122, 66–74. Hermosilla, T., Wulder, M. A., White, J. C., Coops, N. C., & Hobart, G. W. (2015). Regional detection, characterization, and attribution of annual forest change from 1984 to 2012 using Landsat-derived time-series metrics. Remote Sensing of Environment, 121–132. Homer, C., Dewitz, J., Fry, J., Coan, M., Hossain, N., Larson, C., ... Wickham, J. (2007). Completion of the 2001 National Land Cover Database for the conterminous United States. Photogrammetric Engineering and Remote Sensing, 73, 337–341. Hussain, M., Chen, D. M., Cheng, A., Wei, H., & Stanley, D. (2013). Change detection from remotely sensed images: From pixel-based to object-based approaches. ISPRS Journal of Photogrammetry and Remote Sensing, 80, 91–106. Jin, S. M., Yang, L. M., Danielson, P., Homer, C., Fry, J., & Xian, G. (2013). A comprehensive change detection method for updating the National Land Cover Database to circa 2011. Remote Sensing of Environment, 132, 159–175. Johnson, R. D., & Kasischke, E. S. (1998). Change vector analysis: A technique for the multispectral monitoring of land cover and condition. International Journal of Remote Sensing, 19, 411–426. Linke, J., McDermid, G. J., Pape, A. D., McLane, A. J., Laskin, D. N., Hall-Beyer, M., & Franklin, S. E. (2009). The influence of patch-delineation mismatches on multi-temporal landscape pattern analysis. Landscape Ecology, 24, 157–170. Lu, D., Mausel, P., Brondizio, E., & Moran, E. (2004). Change detection techniques. International Journal of Remote Sensing, 25, 2365–2407. Mallinis, G., Koutsias, N., Tsakiri-Strati, M., & Karteris, M. (2008). Object-based classification using Quickbird imagery for delineating forest vegetation polygons in a Mediterranean test site. ISPRS Journal of Photogrammetry and Remote Sensing, 63, 237–250. Mathieu, R., Aryal, J., & Chong, A. K. (2007). Object-based classification of ikonos imagery for mapping large-scale vegetation communities in urban areas. Sensors, 7, 2860–2880. McDermid, G. J., Linke, J., Pape, A. D., Laskin, D. N., McLane, A. J., & Franklin, S. E. (2008). Object-based approaches to change analysis and thematic map update: Challenges and limitations. Canadian Journal of Remote Sensing, 34, 462–466.

47

McFeeters, S. K. (1996). The use of the normalized difference water index (NDWI) in the delineation of open water features. International Journal of Remote Sensing, 17, 1425–1432. Morisette, J. T., & Khorram, S. (2000). Accuracy assessment curves for satellite-based change detection. Photogrammetric Engineering and Remote Sensing, 66, 875–880. Myint, S. W., Gober, P., Brazel, A., Grossman-Clarke, S., & Weng, Q. H. (2011). Per-pixel vs. object-based classification of urban land cover extraction using high spatial resolution imagery. Remote Sensing of Environment, 115, 1145–1161. Nackaerts, K., Vaesen, K., Muys, B., & Coppin, P. (2005). Comparative performance of a modified change vector analysis in forest change detection. International Journal of Remote Sensing, 26, 839–852. Palandro, D. A., Andrefouet, S., Hu, C., Hallock, P., Muller-Karger, F. E., Dustan, P., ... Beaver, C. R. (2008). Quantification of two decades of shallow-water coral reef habitat decline in the Florida Keys National Marine Sanctuary using Landsat data (1984–2002). Remote Sensing of Environment, 112, 3388–3399. Peng, J., Liu, Y. H., Shen, H., Han, Y., & Pan, Y. J. (2012). Vegetation coverage change and associated driving forces in mountain areas of Northwestern Yunnan, China using RS and GIS. Environmental Monitoring and Assessment, 184, 4787–4798. Perdigao, V., & Annoni, A. (1997). Technical and methodological guide for updating CORINE land cover data base. Brussels.: E.C. Joint Research Centre. Powell, S. L., Cohen, W. B., Yang, Z., Pierce, J. D., & Alberti, M. (2008). Quantification of impervious surface in the Snohomish water resources inventory area of Western Washington from 1972–2006. Remote Sensing of Environment, 112, 1895–1908. Pu, R. L., Landry, S., & Yu, Q. (2011). Object-based urban detailed land cover classification with high spatial resolution IKONOS imagery. International Journal of Remote Sensing, 32, 3285–3308. Qian, Y., Zhou, W., Yu, W., & Pickett, S. T. A. (2015). Quantifying spatiotemporal pattern of urban greenspace: New insights from high resolution data. Landscape Ecology, 30, 1165–1173. Röder, A., Hill, J., Duguy, B., Alloza, J. A., & Vallejo, R. (2008). Using long time series of Landsat data to monitor fire events and post-fire dynamics and identify driving factors. A case study in the Ayora region (eastern Spain). Remote Sensing of Environment, 112, 259–273. Rogan, J., Miller, J., Stow, D., Franklin, J., Levien, L., & Fischer, C. (2003). Land-cover change monitoring with classification trees using Landsat TM and ancillary data. Photogrammetric Engineering and Remote Sensing, 69, 793–804. Schneider, A. (2012). Monitoring land cover change in urban and peri-urban areas using dense time stacks of Landsat satellite data and a data mining approach. Remote Sensing of Environment, 124, 689–704. Singh, A. (1989). Digital change detection techniques using remotely-sensed data. International Journal of Remote Sensing, 10, 989–1003. Tewkesbury, A. P., Comber, A. J., Tate, N. J., Lamb, A., & Fisher, P. F. (2015). A critical synthesis of remotely sensed optical image change detection techniques. Remote Sensing of Environment, 160, 1–14. Treitz, P., & Howarth, P. (2000). Integrating spectral, spatial, and terrain variables for forest ecosystem classification. Photogrammetric Engineering and Remote Sensing, 66, 305–317. Trimble (2012). eCognition Developed Reference Book. Trappentreustr. 1D-80339 München, Germany: München, Germany GmbH. Turner, B. L., Lambin, E. F., & Reenberg, A. (2007). The emergence of land change science for global environmental change and sustainability. Proceedings of the National Academy of Sciences of the United States of America, 104, 20666–20671. Vogelmann, J. E., Howard, S. M., & Yang, L. (2001). Completion of the 1990s National Land Cover Data Set for the conterminous United States from Landsat Thematic Mapper data and ancillary data sources. Photogrammetric Engineering and Remote Sensing, 67. Wulder, M. A., Masek, J. G., Cohen, W. B., Loveland, T. R., & Woodcock, C. E. (2012). Opening the archive: How free data has enabled the science and monitoring promise of Landsat. Remote Sensing of Environment, 122, 2–10. Xian, G., & Homer, C. (2010). Updating the 2001 National Land Cover Database Impervious Surface Products to 2006 using Landsat imagery change detection methods. Remote Sensing of Environment, 114, 1676–1686. Xian, G., Homer, C., & Fry, J. (2009). Updating the 2001 National Land Cover Database land cover classification to 2006 by using Landsat imagery change detection methods. Remote Sensing of Environment, 113, 1133–1147. Yang, X. J., & Lo, C. P. (2000). Relative radiometric normalization performance for change detection from multi-date satellite images. Photogrammetric Engineering and Remote Sensing, 66, 967–980. 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. Zhang, Z. X., Wang, X., Zhao, X. L., Liu, B., Yi, L., Zuo, L. J., ... Hu, S. G. (2014). A 2010 update of National Land Use/Cover Database of China at 1:100000 scale using medium spatial resolution satellite images. Remote Sensing of Environment, 149, 142–154. Zhou, W. Q., Huang, G. L., & Cadenasso, M. L. (2011). 90 years of forest cover change in an urbanizing watershed: Spatial and temporal dynamics. Landscape Ecology, 26, 645–659. Zhou, W. Q., & Troy, A. (2008). An object-oriented approach for analysing and characterizing urban landscape at the parcel level. International Journal of Remote Sensing, 29, 3119–3135. Zhou, W. Q., Troy, A., & Grove, J. M. (2008). Object-based land cover classification and change analysis in the Baltimore metropolitan area using multitemporal high resolution remote sensing data. Sensors, 8, 1613–1636. Zhu, Z., & Woodcock, C. E. (2014). Continuous change detection and classification of land cover using all available Landsat data. Remote Sensing of Environment, 144, 152–171. Zhu, Z., Woodcock, C. E., & Olofsson, P. (2012). Continuous monitoring of forest disturbance using all available Landsat imagery. Remote Sensing of Environment, 122, 75–91.