HR: 0800h
AN: B11A-0128    [Abstracts]
TI: Mapping Land Cover Changes With Landsat Imagery and Spatio-Temporal Geostatistics; An application to the Pearl River Delta, China.
AU: * Boucher, A
EM: aboucher@pangea.stanford.edu
AF: Stanford University, Dpt of Geological and Environmental Sciences,, Stanford, CA 94305-2115
AU: Seto, K C
EM: kseto@stanford.edu
AF: Stanford University, Dpt of Geological and Environmental Sciences,, Stanford, CA 94305-2115
AU: Journel, A G
EM: journel@pangea.stanford.edu
AF: Stanford University, Dpt of Geological and Environmental Sciences,, Stanford, CA 94305-2115
AB: Accurate quantification of anthropogenic changes are a primordial input to many studies that relate processes (e.g. climate change) to land covers. Satellite images are the principal medium to detect and map changes in the landscape, both in space and time. However, the current image processing techniques do not fully exploit the data in that they do not take simultaneously into account the spatial and the temporal relations between the various land cover types. The method proposed here aims to accomplish that. At each pixel of the landscape, the time series of land cover type is modeled as a Markov Chain. That time series at any specific location is estimated jointly from the local satellite information, any neighboring ground truth land cover data, and any neighboring previously estimated time series deemed well-informed by the satellite measurements. The spatial component of the land cover types is integrated with variograms and indicator kriging. Our proposal has the advantage to account simultaneously for the spatial and temporal patterns of land cover types. Information and knowledge related to those patterns are coded such that expert opinions can easily be integrated in the process. The method is not limited by the time series length or by the number of land cover classes. Mathematically simple, it can be combined with any type of classifier that outputs a probability for a pixel to belong to a specific class, e.g., a probabilistic neural networks, or a maximum likelihood algorithm. The method is applied on six Landsat images spanning nine years to detect anthropogenic changes in the Pearl River Delta, China. This region is under going tremendous growth in population profoundly altering the landscape. The prediction accuracy of the time series improves significantly, the accuracy almost doubles, when both spatial and temporal information are considered jointly in the estimation process. The introduction of spatial continuity through indicator kriging reduces unwanted noise in the classified images, removing the need for post-processing. This improved integration of space and time produces a more accurate change detection map that defines more accurately when and where the landscape has changed.
DE: 1694 Instruments and techniques
DE: 1640 Remote sensing
SC: Biogeosciences [B]
MN: 2004 AGU Fall Meeting