Spatial and Temporal Patterns of Land Cover Change

Spatial and Temporal Patterns of Land Cover Change In Chengdu, China, 1978-2002Annemarie Schneider a, Karen C. Seto b, and Curtis E. Woodcock aa Department of Geography and Center for Remote Sensing, Boston University, Boston, Massachusetts 02215, USAb Center for Environmental Science and Policy, Stanford University, Stanford, CaliforniaCorresponding author: anmarie@, /urbanAbstract –The goal of this work is to monitor urban growth and the drivers of urban expansion in Chengdu, China. Between 1978 and 2002, the municipal population increased from 8 to 10.2 million and GDP grew seven-fold, changes which have had profound impacts on land cover. To address these issues, this research involves two tasks. First, we monitor land cover change using a supervised, multi-date decision tree method that exploits spectral as well as seasonal differences in Landsat imagery. Results indicate that urban areas have increased by more than 350% between 1978 and 2002. Second, we estimate changing urban density by applying spatial landscape metrics and gradient analysis to the map of land cover change. Using a moving window technique, changes in urban pattern are estimated for seven urban-rural corridors. Results show that western corridors exhibit more disaggregate growth, possibly as the result of policies attracting foreign investment. These results have important implications for urban sustainability, resource use, and potential degradation of the environment.I. I NTRODUCTION AND B ACKGROUNDLand use/cover change in China is becoming extremely important due to the close correlation between land use change and global biogeochemical cycles, which impact climate [1]. Rapid change in China’s cities has attracted attention because of accelerated conversion of agriculture to urban uses during the last two decades [2]. This transition, unprecedented in size and scale, has resulted in part from a series of economic reforms initiated in 1978, when China established 18 economic regions and development zones in major coastal cities. Combined with land tenure reforms, these changes have fostered a massive shift of the population from rural to urban areas. Inland cities such as Chengdu, Chongqing, and Wuhan have been targeted for similar development during the 1990s.As a result, China’s cities have redefined functions, including reorganization of production activities to outside the city, restructuring of residential settlements, and substantial expansion of smaller cities and towns near large cities [3]. In turn, it is becoming increasingly clear that the physical form of urban development affects economic potential, social and political stability, and ecological outcomes [4].Quantifying the spatial and temporal patterns of urban growth dynamics is key to understanding urban change. Satellite remote sensing has revolutionized the process of measuring urban land use during the last three decades [5]. Many studies have used remote sensing to provide accurate measurement of the urban form [6], [7], [8]. Only recently have investigators begun to use satellite information on spatial pattern to assess, manage, and predict growth in cities [9].The objective of this paper is to develop a better understanding of urban growth and the factors driving land conversion in Chengdu, capital of Sichuan province in Central China. Three tasks were involved in this study. First, a supervised decision tree method was used to map land use change in greater Chengdu using a multi-date series of Landsat data. In the second task, density and pattern of land use was quantified for the study area over time and across space, using a combination of landscape metrics [10] and gradient analysis [11]. A third task linking spatial information to socioeconomic data is currently in progress.II. M ETHODOLOGYA. Change DetectionThe method involved three main steps: (1) image preprocessing, (2) change detection using supervised methods, and (3) accuracy assessment. Eight images (WRS path/row 129/39) of the study area were acquired between 1978 and 2002 (Table 1). Preprocessing included registra-tion of all images to a master using a second order transformation with <0.5 RMS error and nearest neighbor resampling to 30 by 30 m pixels.In the second step, the quantity and type of land use/land cover changes was estimated. One common problem when mapping urban change is the confusion between bare agricultural plots and new urban areas. This problem is further compounded by the multi-cropping practices of the Chengdu region. A reliable means to distinguish agriculture from urban change is use of seasonal images: spring green up (April), peak greenness of primary crop (August), and peak greenness of secondary crop (December) [12]. A supervised decision tree classification algorithm (C4.5) was chosen, since this algorithm has been used and tested rigorously in the machine learning community [13]. Decision trees are especially effective for remote sensing problems because they require no a priori assumptions regarding the distribution of the input data, and they can handle complex, nonlinear relations between features and classes. A total of 1052 training sites (15,720 pixels) were selected during field work in Chengdu, ranging in size from one to ten pixels and classified as stable urban,stable vegetation, stable agriculture, stable water, change to urban 1988, change tourban 1991, change to urban 1995, change to urban 2000,TABLE 1I MAGERY U SED IN THE A NALYSISSatellite Acquisition dateLandsat MSS 3 21 August 1978Landsat TM 4 1 May 1988Landsat TM 4 24 April 1991Landsat TM 5 16 August 1992Landsat TM 5 5 May 1995Landsat ETM 7 2 November 2000Landsat ETM 7 23 December 2001Landsat ETM 7 7 October 2002and change to urban 2002. Following convention to train the classifier with the majority of the sample and test classifier accuracy with the remaining sites unseen by the algorithm, 152 sites (~15%) were set aside for validation.A relatively new technique known as “boosting” has been widely shown to increase classification accuracies [14. Boosting improves accuracy by estimating multiple classifiers while systematically varying the training sample. Once the classifier iterations are complete, the boosting algorithm computes a weighted vote for each pixel, assigning probabilities of class membership for each class at every pixel [15. C4.5 with boosting thus produces two outputs: (1) probabilities of each class (0 - 100%), and (2) a predicted class label based on the class with the highest probability.Initial results from the predicted class map (output 2) revealed class confusion, and overall map accuracy was below 50%. Change areas were instead labeled stable urban or stable agriculture. Inspection of the class probabilities (output 1) revealed that the incorrectly labeled pixels consistently had a second most probable class that was the correct class. With this in mind, a density slicing technique was used to compare the probability maps for the five change classes directly to the Landsat images. Once the values corresponding to change pixels were determined, a binary mask of change was created and overlaid on the map of stable areas. By combining the two, a final map was created for each date. Finally, these maps were validated using 152 ground truth sites held out from the classification process. B. Spatial AnalysisUsing the change maps, the density and pattern of land use was quantified for the study area over time and across space by merging landscape metrics [11] and gradient analysis [12]. The study area was partitioned into seven transects or corridors based on major transportation networks (Fig. 1). The roads were buffered by 2 km, and a series of analyses was conducted along each corridor using spatial pattern metrics in a 4 x 4 km moving window. Two simple metrics were chosen: mean patch size (MPS) and landscape shape index (LSI). MPS conveys the size of contiguous area of a given land cover type within each corridor. If urban patches are small and scattered, for instance, MPS will be low, while increasing MPS will result as “densification” occurs. LSI is a standardized measure of perimeter length of all patches of a given type. Increasing LSI indicates that urban areas are becoming more fragmented, while decreasing LSI results as the urban fabric becomes more continuous.III. R ESULTSFig. 2 presents the results of the change detection algorithm, revealing explicit patterns of growth in the southeast, east and northeast regions of the city. The extent of urban, built up land increased more than 350% from 1978 to 2002. Field assessments confirm a high overall accuracy of the land-use change map (88.0%).The transect analysis provided information additional to trends seen in the land use maps. Results of MPS from 1988 to 2002 (Fig. 3) for the seven corridors show an initial peak indicating the edge of the city proper and secondary peaks corresponding directly to satellite cities and clusters of growth along the corridors. Six of the seven corridors (Fig. 3a, 3c-3g) exhibit dramatic growth outside the city, suggesting urban clusters as far as 40 km from the city. Construction of a new airport and nearby land development is apparent in Fig. 3e, where MPS increases at a much greater rate within 10 km of the city. Despite increasing amounts of urban land in the north, northwest and west (Fig. 3a, 3f, 3g), these areas remain patchy in comparison to the southwest corridor (Fig. 3e). Results of LSI (not shown) provided additional evidence corroborating these findings.IV. C ONCLUSIONSThe main objective of this research was to improve understanding of the growth and pattern of urban areas in Chengdu, China. Change and stable land use was mapped using a decision tree algorithm, and overall map accuracy was greater than 87%. Patchiness and contiguity of urban land were quantified using spatial pattern metrics over time (1988 to 2002) and space using urban to rural transects.These metrics indicated increasing dispersion of the city in northern and western corridors, while patterns of densification were perceptible in the southwest.A CKNOWLEDGMENTSThis work was funded by NASA grants NGT5-30401abd cgfeFigure 1: Seven corridors analyzed for Chengdu, China, using a moving window of spatial pattern metrics. Stable urban areas are shown in purple, while urban change 1978 to 2002 is shown in red.and NAG5-10534, with additional support from the WorldBank. The authors would like to thank Douglas Webster, CaiJianming and Binyi Luo for their assistance.R EFERENCES[1] Sellers, P.J., et al. “Modeling the exchanges of energy, water andcarbon between continents and the atmosphere, Science, 275, 502-509, 1997.[2] Xia, L., “Measurement of rapid agricultural land loss in the Pearl RiverDelta with integration of remote sensing and GIS,” Environment and Planning B, 25, 447-461, 1998[3] National Bureau of Statistics, China, 2000.[4] Guldin, G.E., Farewell to Peasant China , M E Sharpe, Armonk, NewYork, 1997.[5] Longley, P.A. and V. Mesev, On the measurement and generalization ofurban form, Environment and Planning A, 32, 473-488, 2000.[6] Jensen, J.R. and D.C. Cowen, “Remote sensing of urban/suburbaninfrastructure and socioeconomic attributes,” Photogrammetric Engineering and Remote Sensing, 65, 611-622, 1999.[7] Howarth, P.J. and E. Boasson, Landsat digital enhancements for changedetection in urban environments,” Remote Sensing of Environment, 13, 149-160, 1983.[8] Ehlers, M., Jadkowski, M.A., Howard, R.R. and D.E. Brostuen,“Application of SPOT data for regional growth analysis and local planning,” Photogrammetric Engineering and Remote Sensing, 56, 175-180, 1990. [9] Chan, J.C.W., Chan, K.P. and A.G.O. Yeh, “Detecting the nature ofchange in an urban environment,” Photogrammetric Engineering andRemote Sensing, 67, 213-225, 2001.[10] Herold, M., Scepan, J. and K. Clarke, “The use of remote sensing andlandscape metrics to describe structures and changes in urban land uses,”Environment and Planning A, 34, 1443-1458, 2002.[11] McGarigal, K. and B.J. Marks, “Fragstats: Spatial Pattern AnalysisProgram for Quantifying Landscape Structure,” General Technical Report PNW-GTR-351, Pacific Northwest Research Station, USDA-Forest Service, Portland, Oregon, 1995.[12] McDonnell, M.J., and S.T.A. Pickett, “Ecosystem structure andfunction along urban-rural gradients: an unexploited opportunity,” Ecology, 71, 1232-1237, 1990.[13] Pax-Lenney, M., Woodcock, C.E., Collins, J.B. and H. Hamdi, “Thestatus of agricultural lands in Egypt: the use of multitemporal NDVI features derived from Landsat TM,” Remote Sensing of Environment,56, 8-20, 1996.[14] Quinlan, J. R., C4.5: Programs for machine learnin,. MorganKauffman, San Mateo, California, 302 p., 1993.[15] McIver, D. K. and M. A. Friedl, “Estimating pixel-scale land coverclassification confidence using non-parametric machine learning methods,” IEEE Transactions on Geoscience and Remote Sensing ,” 39, 1959-1968, 2001.[16] Quinlan, J. R., “Bagging, boosting, and C4.5,” Proceedings of the 13thNational Conference on Artificial Intelligence (AAAI-96), 4-8 August, Portland, Oregon, AAAI Press, 725-730, 1996.Figure 3: Results of the moving window analysis for mean patch size (log scale, y axis) along seven corridors in greater Chengdu, starting with the northeast corridor (a), and continuing clockwise (b) – (g). Distance from the city is indicated along the x axis. Corridors (b), (c) and (d) were constrained by mountainseast of the city.Figure 2: Maps of land cover change in Chengdu for (a) 1978 to 1988, (b) 1988 to 1995, and (c) 1995 to 2002. Urban areas are shown in purple, urbanexpansion is shown in red, agriculture is white, and natural vegetation is green.。