Hidden Gauss Markov model for multiscale remotely sensed image segmentation

Hidden Markov models are useful tool for tackling numerous problems especially in statistical signal and image processing. This paper presents a wavelet domain approach to remotely sensed image segmentation based on Hidden Markov Tree (HMT) models. The essence of this work is based on capturing the statistical properties of the wavelet coefficients by a tree-structured model. One important drawback to the HMT model is the need for iterative training of the HMT model parameters for a given data set. Following the fast training we perform likelihood computation algorithm for texture classification at different scales and directly segment wavelet-compressed images. We demonstrate the performance of the algorithm with SPOT and RADARSAT images. The findings are found to be encouraging.