A new approach to segmentation of remote sensing images with Hidden Markov Models

In this work, we present a new segmentation algorithm for remote sensing images based on two-dimensional Hidden Markov Models (2D-HMM). In contrast to most 2D-HMM approaches, we do not use Viterbi Training, instead we propose to propagate the state probabilities through the image. Therefore, we denote our algorithm Complete Enumeration Propagation (CEP). To evaluate the performance of CEP, we compare it to the standard 2D-HMM approach called Path Constrained Viterbi Training (PCVT). As both algorithms are not restricted to a certain emission probability, we evaluate the performance of seven probability functions, namely Gamma, Generalized Extreme Value, inverse Gaussian, Logistic, Nakagami, Normal and Weibull. The experimental results show that our approach outperforms PCVT and other benchmark algorithms. Furthermore, we show that the choice of the probability distribution is crucial for many segmentation tasks.