Bayesian cloud classification of multi-source satellite imagery with spatio-temporal dependence

Abstract

This thesis presents a Bayesian classification model for cloud detection in regular half hour time intervals of a 24 hour day. It integrates remotely sensed data from multiple satellites, spatio-temporal pixel dependence information, and a ground cover map from which to make inferences on the classified scene. A Markov random field is utilised to model spatial dependence between neighbouring pixels. Temporal dependence is introduced by the assumption of a first order Markovian process for a succession of images of the scene. Approximate maximum a posteriori estimates of the classified scene are obtained utilising the method of iterated conditional modes. Experimentation demonstrates the strengths and weaknesses of visible, near infrared (channel 3), and infrared imagery in accomplishing the cloud classification task. The model is also shown to be adept at estimation in situations of limited satellite data.