HR: 0800h
AN: B41A-0170    [Abstracts]
TI: Detecting Deforestation In Paraguay From Multi-temporal Landsat Imagery Using A Spatio-temporally Explicit Algorithm
AU: * Liu, D
EM: dsliu@nature.berkeley.edu
AF: University of California at Berkeley, Department of Environmental Science, Policy and Management , 137 Mulford Hall #3114, Berkley, CA 94720-3114 United States
AU: Kelly, M
EM: mkelly@nature.berkley.edu
AF: University of California at Berkeley, Department of Environmental Science, Policy and Management , 137 Mulford Hall #3114, Berkley, CA 94720-3114 United States
AU: Gong, P
EM: gong@nature.berkeley.edu
AF: University of California at Berkeley, Department of Environmental Science, Policy and Management , 137 Mulford Hall #3114, Berkley, CA 94720-3114 United States
AU: Townshend, J R
EM: jtownshe@glue.umd.edu
AF: University of Maryland, Department of Geography, 2181 LeFrak Hall, College Park, MD 20742 United States
AB: Forests in Paraguay have undergone extensive loss in the last decades. Detecting deforestation in this area with the use of satellite remote sensing data has particular scientific interests in a broad range of research fields. Conventional methods addressing this issue in terms of change analysis of difference image or post-classification comparison are incapable of modeling both spatial and temporal contextual information. In this paper, we propose a spatio-temporally explicit algorithm using multi-temporal Landsat imagery to detect the deforestation in Paraguay during the period between 1990 and 2000. In this algorithm, change analysis of difference image and classification of multi-temporal images are combined in a spatio-temporal model. Specifically, this algorithm includes the following three steps. First, a machine learning algorithm, Support Vector Machines (SVM), is trained with spectral observations to initialize the classification and to estimate pixel-wise class conditional probabilities for each individual image. Second, a modified Markov Random Fields (MRF) model accounting for pixel-wise transition probability is used to model the spatio-temporal contextual prior probabilities of images. Finally, an iterative algorithm, Iterative Conditional Mode (ICM), is used to update the classification based on the combination of spectral class conditional probability and spatio-temporal contextual prior probability. The results showed that the proposed algorithm achieved significant improvements over traditional pixel-based single-date approaches. The improvement from the contributions of spatio-temporal contextual evidence indicated the importance of spatio-temporal modeling in multi-temporal remote sensing in general and deforestation in particular.
DE: 0315 Biosphere/atmosphere interactions (0426, 1610)
DE: 0400 BIOGEOSCIENCES
DE: 1615 Biogeochemical cycles, processes, and modeling (0412, 0414, 0793, 4805, 4912)
DE: 1694 Instruments and techniques
DE: 1833 Hydroclimatology
SC: Biogeosciences [B]
MN: Fall Meeting 2005