Unsupervised flood detection in X-band SAR data using multi-contextual Markov image modeling on irregular graphs

Abstract

The worldwide increasing occurrence of flooding and the short-time monitoring capability of the new generation of high resolution synthetic aperture radar (SAR) sensors require accurate and automatic methods for the detection of inundations. This is especially important for operational rapid mapping purposes where the fast provision of precise information about the extent of a disaster and its spatio-temporal evolution is of key importance for decision makers and humanitarian relief organizations. In this contribution, a hybrid multi-contextual Markov model for unsupervised near real-time flood detection in multi-temporal X-band synthetic aperture radar (SAR) data is presented. It incorporates scale-dependent as well as spatio-temporal contextual information into the classification process by combining causal with noncausal Markov image modeling related to hierarchical directed and planar un-directed graphs, respectively. Hierarchical Markov modeling is accomplished by hierarchical marginal posterior mode (HMPM) estimation using Markov Chains in scale. This model is initialized by an automatic tile-based thresholding algorithm, to differentiate between open water, flooded vegetation and dry land areas in a completely unsupervised manner. In order to increase computational performance, marginal posterior-based entropies are used for restricting the iterative bi-directional exchange of spatio-temporal information between consecutive images of a time sequence to image elements, exhibiting a low probability to be classified correctly according to the HMPM estimation. The Markov image models, originally developed for inference on regular graph structures of quadtrees and planar lattices, are adapted to the variable nature of irregular graphs, which are related to information driven image segmentation. With respect to accuracy and computational effort, experiments performed on a bi-temporal TerraSAR-X ScanSAR data-set from Lake Liambezi in Namibia during flooding in 2009 and 2010 confirm the effectiveness of integrating hierarchical as well as spatio-temporal context into the labeling process, and of adapting the models to irregular graph structures.