Bayesian cloud detection with pre-clustering of satellite imagery

Cloud detection in satellite data, particularly at night when only thermal imagery is available, continues to present problems to many data applications. The discrimination of clouds within imagery is important to most interpretations of the scene, as pixels can hold information about the cloud, or about the underlying surface, or sometimes a mixture if the two, and the physical meaning of the data cannot be known unless it is known which of these informations is held. Traditional threshold-testing approaches have been criticised for spatial and temporal biases and for lacking a sound physical basis. A new, physically-based probabilistic technique, which uses Bayes Theorem to combine apriori information from Numerical Weather Prediction (NWP) fields with the satellite imagery to calculate a probability of clear (PClr) for each pixel (‘not clear’ is considered analogous to cloud-contaminated) has been formulated at the University of Edinburgh. Where this method has been tested, the results are a significant improvement on the results of more traditional methods, however some further improvements could be made. Highly textural pixels, such as those covering an area of high thermal gradient, for example ocean fronts, are sometimes classified ambiguously, that is, the calculated PClr does not lie at either end of the PClr range (0-1). It is anticipated that spectral clustering of the data prior to applying the detection algorithm may result in more polarised PClr values being calculated. A clustering method which avoids any significant loss of pixel-scale spectral information, and its effects on the calculated PClr values is presented.