Deep learning for SAR ship classification: focus on unbalanced datasets and inter-dataset generalization

Detection and recognition of vessel targets at sea are tasks of paramount relevance for maritime monitoring purposes. A possible approach to pursue these objectives consists in acquiring and processing Synthetic Aperture Radar (SAR) data related to a given area of interest. Classically, the detection part can be implemented by exploiting statistical properties of the signal to decide whether an image area belongs to background clutter or to a ship (e.g. Constant False Alarm Rate based algorithms). Successively, discriminant features referring to the detected object can be extracted and later fed to a classifier to decide the membership category of the considered target. Recently, thanks to the development of algorithms based on deep neural network architectures, object detection and recognition experienced an unprecedented boost in the observed performances. This work, mainly motivated by the exploration of these novel approaches to the identification of vessel targets, focuses on the analysis of five different deep learning architectures (CNN, pre-trained and non-pretrained versions of ResNet50 and VGG16) trained on two public SAR vessel datasets (OpenSARShip and Fusar). To address the data quantity limitation, a third dataset was created by merging both datasets.