Geometric and scattering features for ship classification from Sentinel 1 SAR images

Following the evaluation of some ship classification strategies based on geometrical features, this report accounts for the use of scattering measurements in SAR images as additional features, in the hope of improving the classification performance. A set of eight scattering features has been selected and added to the already tested set of eight naive geometric features to explore the discriminating power of the whole feature set or any subset thereof. The algorithm chosen for this investigation is Random Forest, as implemented in the R package randomForest. The basic finding has been that, as opposed to some claims in the literature, the use of scattering features improves the classification performance even from images characterized by a moderate resolution, such as the ones provided by ESA's Sentinel 1 satellite-borne SAR.