Modeling of inhomogeneous Markov random fields with applications to cloud screening
Cloud screening is the process of classifying pixels in satellite images which contain clouds and is an important step in processing remotely-sensed images. This paper applies inhomogeneous statistical spatial models in the form of Markov random field models (MRF) to this problem and develops an efficient algorithm for the estimation of model parameters. The algorithm has a natural parallel decomposition. The model is tested on synthesized images for which ground truth is known and applied to segmentation of clouds in several Advanced Very High-Resolution Radiometer (AVHRR) images. This paper concentrates on the abstract spatial aspects of the models rather than the details of specific remote-sensing mechanisms. The main results are (1) the formulation (in terms of inference and estimation) of the inhomogeneous MRF model, (2) the exact solution of the "pseudo-likelihood" equations used for parameter estimation in this model, and (3) experimental results which indicate that (a) inhomogeneous models perform better than homogeneous models and (b) that spatial models perform better than non-spatial models for cloud-screening problems.