A Class of Cloud Detection Algorithms Based on a MAP-MRF Approach in Space and Time

A recurrent concern in cloud detection approaches is the high misclassification rate for pixels close to cloud edges. We tackle this problem by introducing a novel penalty term within the classical maximum a posteriori probability-Markov random field (MAP-MRF) approach. To improve the classification rate, such term, for which we suggest two different functional forms, accounts for the predictable motion of cloud volumes across images. Two mass tracking techniques are proposed. The first one is an effective and efficient implementation of the probability hypothesis density (PHD) filter, which is based on Gaussian mixtures (GMs) and relies on finite set statistics (FISST). The second one is a region matching procedure based on a maximum cross-correlation (MCC) that is characterized by low computational load. Through extensive tests on simulated images and real data, acquired by the SEVIRI sensor, both methods show a clear performance gain in comparison with classical spatial MRF-based algorithms.