Using random forests to classify vessels from naive geometrical features

This report is concerned with the application of Random Forest classification methods to the identification of ship types in moderate-resolution SAR images. After a brief presentation of the theory and and the features of this class of methods, we select an R package useful to train, test and execute the classifier. Some experiments are then reported using naive geometrical features extracted from a few thousands of targets in the OpenSARShip data set. All the ship chips extracted are derived from IW GRD Sentinel 1 C-band SAR images, accompanied by AIS and MarineTraffic ground-truth data. The ideal performance of this classifier is evaluated through the standard classification indices, with respect to the ship types that are sufficiently represented in the subsets considered.