Abstract
Flood extent mapping plays a crucial role in disaster management and nationalwater forecasting. Unfortunately, traditional classification methods are oftenhampered by the existence of noise, obstacles and heterogeneity in spectralfeatures as well as implicit anisotropic spatial dependency across classlabels. In this paper, we propose geographical hidden Markov tree, aprobabilistic graphical model that generalizes the common hidden Markov modelfrom a one dimensional sequence to a two dimensional map. Anisotropic spatialdependency is incorporated in the hidden class layer with a reverse treestructure. We also investigate computational algorithms for reverse treeconstruction, model parameter learning and class inference. Extensiveevaluations on both synthetic and real world datasets show that proposed modeloutperforms multiple baselines in flood mapping, and our algorithms arescalable on large data sizes.