Classifying multi-temporal TM imagery using Markov random fields and support vector machines

In this paper, we propose a spatial-temporally explicit algorithm to simultaneously classify multi-temporal images for land cover information. This algorithm has three steps: first, a machine learning algorithm Support Vector Machines (SVM), is trained with spectral observations to initialize the classification and estimate pixel-by-pixel class conditional probabilities for each individual image; second, Markov Random Fields (MRF) are used to model the spatial-temporal contextual prior probabilities of images; and finally, an iterative algorithm is used to update the classification based on the combination of the spectral class conditional probability and the spatial-temporal contextual prior probability. Increased accuracies from the contributions of spatial-temporal contextual evidence confirmed the importance of spatial-temporal modeling in multi-temporal remote sensing. In this paper, we propose a spatial-temporally explicit algorithm based on Markov Random Fields (MRF) and Support Vector Machines (SVM) to simultaneously classify multi-temporal images for land cover information. We first review SVM and MRF and present our proposed algorithm based on both of them. We then evaluate the algorithm using a real data set and compare the result with conventional non- contextual and partial-contextual (spatial only and temporal only) approaches.