A Ship Detection Model for SAR Data based on YOLOv4: Application to Images from SAOCOM and Sentinel

Abstract

Synthetic Aperture Radar satellites are becoming increasingly important in the field of Earth observation and maritime surveillance. Given the large amount of data generated by satellite platforms, the use of advanced techniques is required to extract useful information from them. Currently, deep learning techniques applied to object detection obtain a high performance, in particular with the use of convolutional neural networks. This work proposes a model with YOLOv4 architecture trained with the HRSID dataset (with offshore and inshore images) using Transfer Learning, which obtains a performance that improves results present in the literature. A suitable set of hyperparameter values is sought and the modification of the architecture is explored in relation to the size of the input image and the structure of the SPP spatial pyramidal pooling layer. Finally, the model is tested against scenes captured with Sentinel 1 and SAOCOM 1A satellites that were not present in the training.