Unsupervised change detection based on improved Markov random field technique using multichannel synthetic aperture radar images

Abstract

Change detection represents an important remote sensing tool in environmental monitoring and disaster management. In this respect, multichannel synthetic aperture radar (SAR) data offer great potential because of their insensitivity to atmospheric and sun-illumination conditions (over optical multispectral data) and the improved discrimination capability they may provide compared to single-channel SAR. The problem of detecting the changes caused by flooding is addressed by a contextual unsupervised technique based on a Markovian data fusion approach. However, the isotropic formulation of Markov random field (MRF) models causes oversmoothing of spatial boundaries in the final change maps. In order to reduce this drawback, an edge-preserving MRF model is proposed and formulated by using energy functions that combine the edge information extracted from the produced edge maps using competitive fuzzy rules and Canny technique, the information conveyed by each SAR channel, and the spatial contextual information. The proposed technique is experimentally validated with semisimulated data and real ASAR-ENVISAT images. Change detection results obtained by the improved MRF model exhibited a higher accuracy than its predecessors for both semisimulated (average 12%) and real (average 6%) data.