SVM-based cloud detection aided by contextual information

Support Vector Machines (SVM) emerge among the classification methods as a very effective tool for separating the illuminated scene in different classes by utilizing multiple features. Typically, a pixelwise classification is performed by employing, as features, the radiances at different bands. This neglects the possibility of accounting for the spatial, and possibly the temporal, correlation within and among the images. Existing methods include the latter through segmentation methods whose output is fused with the SVM classification. We propose here to adjoin the contextual information as a further feature by constructing a penalty map accounting for the correlation among pixels. To illustrate and evaluate the method we present an application of cloud masking on MultiSpectral Images acquired by the SEVIRI sensor.