Effects of landscape characteristics on accuracy of land cover change detection

Transcription

1 Effects of landscape characteristics on accuracy of land cover change detection Yingying Mei, Jingxiong Zhang School of Reote Sensing and Inforation Engineering, Wuhan University, 129 Luoyu Road, Wuhan , China Abstract The effects of landscape characteristics on land cover change detection accuracy were evaluated. Logistic regression odels with change/no-change classes were extended to include bi-teporal explanatory variables: patch density, largest patch index and landscape shape index at both fro and to ties. It was found that, these biteporal variables had significant effects on change detection accuracy. To validate the odels, a leave-one-out procedure was applied in which the absolute difference between the actual and the odel-estiated was suarized over 20 saple area units in Wuhan. The su of differences reduced fro 1.68 to 1.47 after adding such biteporal variables as patch density, largest patch index and landscape shape index to the odels concerning binary change/ no-change classes. Therefore, these landscape characteristics led to a substantial iproveent in the estiation of change detection accuracy. Keywords: classification, change detection accuracy, landscape characteristics, logistic, bi-teporal. 1. Introduction Change detection, which seeks to identify differences in the states of spatially distributed phenoena, has been widely used in resources anage and environental onitoring (Macleod, 1994). Quantifying the extent and rate of land-cover change, and developing odels to relate the processes driving changes in land use to observed changes in land-cover are iportant challenges in the field of global change science. The liited knowledge concerning the accuracy of land-cover change products directly affects our ability to ake accurate predictions of change (Foody, 2002). Error atrix, an extension of the single-date classification error atrix, has been widely used in accuracy assessent in change detection (Congalton and Macleod, 1994; Conglaton and Green, 1999), and is iportant in an error propagation analysis or estiation of the fitness for use of the data for a specific region. However, accuracy assessent is hapered by sorts of difficulties (Achard et al., 2002; De Zeeuw and Hazeu, 2001) iplying that uch work is needed for quantitative estiation of change detection accuracy. Up to now, any ethods for the estiation of change detection accuracy have been proposed. Although siple ethods ay be adopted to derive accuracy easures of change detection based on individual classification aps (Congalton and Green,1999; Carel et al.,2001; Sohl et al.,2004), ore elaborated approaches should be explored for iproved reliability in accuracy assessent, when dealing with spatial data which is

2 characterized by data dependency and heterogeneity (Van Oort, 2005; Burnicki et al., 2007). More recent developents applying transition rules to land cover change trajectories to assess the accuracy of changes predicted across ore than two classified iages (Liu and Zhou, 2004). Logistic regression odels have been used to analyse the relationship between land cover classification accuracy at single tie and landscape pattern indices (Sith et al., 2002; Van Oort et al., 2002). A odel was established to illustrate the ipact of locations on the uncertainty of land-cover aps (Carnel and Dean, 2004). Also, landscape pattern analysis (LPA) is an iportant tool to quantify changes in land cover aps (Turner et al.2001; Farina, 2006). This paper seeks to extend landscape characteristics as potential explanatory variables for uncertainty evaluation fro single-date data sets to change categorization. We established the logistic regression odels with landscape characteristics as the explanatory variables with change detection accuracy. In order to validate the odels, a leave-one-out procedure was applied. A substantial iproveent in the estiation of change detection accuracy was achieved. 2. Methodology 2.1 Data Two sub-scene ETM+ iages of Wuhan flown on May 2012 (tie1) and August 2013 (tie2) with bands 1-5 and 7 at 30 resolution were used to calculate landscape variables. Twenty area units, each of which contains 500*697 pixels, were sapled fro the bi-teporal iages at the sae location. Post-classification was used to deterine the changes in land cover inforation. Two high resolution land cover aps of the sae region at corresponding tie were used as reference data to estiate change detection accuracy. Figure1 shows the distribution of the 20 units. The black unit was fitted to validate the percentage correctly classified of the other 19 units (white). Figure1. Wuhan with the saple area units (rectangles) used.

3 2.2 Explanatory variables Table1 shows the explanatory variables used. Table 1: Explanatory variables. Variable Ab. Key aspects of variable No-change/Change Class Class two binary variables are used to indicate change or no-change Patch Density PD the extent to which the landscape is fragented Largest Patch Index LPI a standardized easure of the plaque concentration and landscape doinance Landscape Shape Index LSI a standardized easure of total edge or edge density that adjusts for the size of the landscape 2.3 Statistical analysis Logistic Regression The logistic regression (Sith et al., 2002; Lin et al., 2008) was used to calculate the distribution of cells correct classification p as a function of the landscape variables introduced above. The logistic regression odel with intercept 0 and with x equals: k1k explanatory variables k exp( 0 k xk) k 1 p K 1 exp( x ) Forula (1) can be expressed as follows: 0 K k 1 K p logit(p) ln x 1 k k (1) 0 k k (2) p k1 Regression coefficients are calculated by iniizing the -2 log likelihood of the odel. In order to evaluate the ipact of these variables on land-cover change accuracy, a set of odels including different variables were built (Table2), where PD1 LPI1 and LSI1 express the landscape variables on tie1, PD2 LPI2 and LSI2 on tie2.

4 Model nuber() 0 0 Table 2: Description of odels. Model Interception 1 0 Class 1 2 Interception, Class 2a 0 12 Class+ 3X 3+ 4X4 Variables Interception, Class,PD1,PD2 2b 0 12 Class+ 5X 5+ 6X6 Interception, Class,LPI1,LPI2 2c 0 12 Class + 7 X7 + 8 X8 Interception,Class,LSI1,LSI2 3a 0 12 Class+ 3X 3+ 4X 4+ 5X 5+ 6X6 Interception, Class,PD1,PD2, LPI1,LPI2 3b 0 12 Class + 3X 3+ 4X 4+ 7 X7 + 8 X8 Interception, Class,PD1,PD2, LSI1,LSI2 3c 0 12 Class + 5X 5+ 6X 6+ 7 X7 + 8 X8 Interception, Class, LPI1,LPI2, LSI1,LSI Class + 3X 3+ 4X 4+ 5X 5+ 6X 6+ 7 X7+ 8 X8 Interception, Class, PD1,PD2, LPI1,LPI2, LSI1,LSI Validation A leave-one-out procedure (Van Oort et al., 2004) was used to find the odel containing the highest nuber of significant explanatory variables. Each one of the 20 area units was used to test the odel based on the other 19 units. The resulting 20 paraeter sets fored versions of a odel. The easure SM (Equation (3)) gives the residuals of each odel, which is better with a sall value. R0 (Equation (4)) and R1 (Equation (5)) describe the relative iproveent of odel to the odel 0 assuing the sae probability of correct classification for all cells and odel 1 containing no-change/change class, respectively. 20 n(r) SM p (c) y(c) (3) r1 c1 R0 100% (SM SM ) / SM (4) 0 0 R1 100% (SM SM ) / SM (5) 1 1 where p (c) is the correct classification probability of odel, n(r) is the nuber of cells in unit r, the binary variable y(c) indicts if the cells c is correctly classified or not. 3. Results and Discussion 3.1 Change/no change error atrix Table 3 shows accuracy easures at two dates. Table 3: Accuracy easures. Users accuracy(%) PCC(%) water urban forest farland tie tie

5 The change/no change error atrix is shown in Table 4. Table 4: Change/no change error atrix. No Change Change Total No Change Change Total Change detection accuracy= ( )/34173=91.05%. 3.2 Model selection Table 5 gives the estiated regression coefficients for a selection of odels. Model1 shows the effect of Class on change detection accuracy, odel 2a-2c show the effect of the landscape characteristics variables PD LPI and LSI when Class is accounted for, odel 3a-3c show the interaction of two different landscape indices accopanied with Class. Model 4 is the odel that shows the coprehensive effects of the four explanatory variables. Table 5: Estiated regression coefficients. Model nuber() Regression coefficients a b c a b c Model validation Table 6 gives the validation results of odels established. It indicates that all odel estiates were better than odel 0( R 0 >0 for all ). Addition of a landscape variable to a odel containing Class (odels 2a-2c) iproved odel 1 estiates by 2.69 to 10.67%. The interaction of two of the three landscape indices to the odel containing variable Class (odels 3a-3c) was significant: R 13 a =12.06%, R 13 b =5.12%. Relatively to odels 0 and 1, the result of odel 4 is closest to the actual value: R =38.39% and R 14 =12.53%. 0 4

6 Model nuber() Model description Table 6: Validation. Relative to odels 0 and 1 SM R 0 (%) 1 R (%) 0 sae PCC in all units Class(=error atrix) a Class&PD1&PD b Class&LPI1&LPI c Class&LSI1&LSI a Class&PD1&PD2& LPI1&LPI b Class& PD1&PD2& LSI1&LSI c Class& LPI1&LPI2& LSI1&LSI Class&PD1&PD2& LPI1&LPI2& LSI1&LSI Conclusion It is difficult for reote sensing classification ethods only based on spectral response to distinguish objects with siilar spectral characteristics but of totally different classes. This eans that accuracy in classification and change detection will be affected by class inseparability inherent to reotely sensed data and other factors in labelling class and detecting change. To estiate accuracy in classification and change detection, it is iportant to investigate the ethods by which accuracy can be inferred fro the spatial patterns in the aps being exained. Existing literature indicates that landscape variables are significant predictors of classification accuracy, and can be used to iprove the prediction reliability of change detection accuracy. This paper evaluated landscape indices as potential explanatory variables of variability in change detection accuracy considering the class is change or not. Logistic regression odels were developed to include three landscape indices: patch density, largest patch index and landscape shape index as the explanatory variables for uncertainty evaluation in change categorization. In order to quantify the ipact of landscape indices on accuracy of change detection, a leave-one-out procedure was applied to evaluate the estiated residuals of each odel. Several rearks can be ade as follows: 1) Logistic regression odel can be used effectively for analyzing land-cover change and its environental effects. Quantitative analysis of the relationship between the change detection accuracy and binary landscape index was used to siulate the land cover change tendency. The principle of the land-cover evolution can be used in change detection prediction. 2) Addition of landscape characteristics variables PD LPI and LSI to spectral classification led to a substantial iproveent in the estiation of change detection accuracy. The three landscape indices were all iproved the reliability of estiation results in different degrees, where LSI had the weakest ipact aong the three landscape indices. 3) It was found that change detection accuracy was significantly higher for Logistic odels with ore explanatory variables. Synergistic effect of the three landscape indices is ore than single variables above. The highest estiates of change detection

7 accuracy were obtained in odels containing the four variables above, where the estiation results are closest to the actual values. Future research will focus on how the ipacts of landscape characteristics ay vary across different land cover and change types and over space. Regression-based ethods for analyzing accuracy of change detection can be usefully cobined with other analytical approaches to accuracy estiation to ake better use of teporal correlation and spatial patterns of isclassification errors. Acknowledgents The research reported in this paper is supported by the National Natural Science Foundation of China (grant no ). References Macleod, R. D., Congalton, R. G. (1998). A quantitative coparison of change-detection algoriths for onitoring eelgrass fro reotely sensed data. Photograetric engineering and reote sensing, 64(3), Foody G M(2002). Status of land cover classification accuracy assessent[j]. Reote sensing of environent, 80(1): Congalton R G, Macleod R D(1994). Change detection accuracy assessent on the NOAA Chesapeake Bay Pilot Study[C]. Aerican Society of Photograetry and Reote Sensing, Williasburg: Congalton R G, Green K(2008). Assessing the accuracy of reotely sensed data: principles and practices[m]. CRC press. Achard, F., Eva, H. D., Stibig, H. J., Mayaux, P., Gallego, J., Richards, T., & Malingreau, J. P. (2002). Deterination of deforestation rates of the world's huid tropical forests. Science, 297(5583), Zeeuw, C. D., & Hazeu, G. W. (2001). Monitoring land use changes using geo-inforation: possibilities, ethods and adapted techniques. Alterra-rapport. Carel Y, Dean D J, Flather C H(2001). Cobining location and classification error sources for estiating ulti-teporal database accuracy[j]. Photograetric Engineering and Reote Sensing, 67(7): Sohl T L, Gallant A L, Loveland T R(2004). The characteristics and interpretability of land surface change and iplications for project design[j]. Photograetric Engineering and Reote Sensing, 70(4): Van Oort P A J(2005). Iproving land cover change estiates by accounting for classification errors[j]. International Journal of Reote Sensing, 26(14): Burnicki A C, Brown D G, Goovaerts P(2007). Siulating error propagation in land-cover change analysis: the iplications of teporal dependence[j]. Coputers, Environent and Urban Systes, 31(3): Liu H, Zhou Q(2004). Accuracy analysis of reote sensing change detection by rule-based rationality evaluation with post-classification coparison[j]. International Journal of Reote Sensing, 25(5):

8 Carel Y, Dean D J(2004). Perforance of a spatio-teporal error odel for raster datasets under coplex error patterns[j]. International Journal of Reote Sensing, 25(23): Sith, J. H., Wickha, J. D., Stehan, S. V., & Yang, L. (2002). Ipacts of patch size and landcover heterogeneity on theatic iage classification accuracy. Photograetric Engineering and Reote Sensing, 68(1), Van Oort, P. A., Bregt, A. K., de Bruin, S., de Wit, A. J., & Stein, A. (2004). Spatial variability in classification accuracy of agricultural crops in the Dutch national land-cover database. International Journal of Geographical Inforation Science, 18(6), Turner M G(2001). Landscape ecology in theory and practice: pattern and process[m]. Springer. Farina A(2006). Principles and ethods in landscape ecology: towards a science of the landscape[m]. Springer. Lin, H., Wang, J., Bo, Y., & Bian, J. (2008). The Ipacts of Landscape Patterns on the Accuracy of Reotely Sensed Data Classification. InProceedings of 8 th International Syposiu on Special Accuracy Assessent in Natural Resources and Environental Sciences (Shanghai, China (pp ).