Oil Rig Recognition Using Convolutional Neural Network on Sentinel-1 SAR Images

Recent advances in deep learning for automatic target recognition have inspired studies for maritime surveillance with synthetic aperture radar (SAR). In this context, oil rigs monitoring is of particular interest. There are offshore platforms of different types and sizes which poses challenges for classification methods. In this letter, we compare the recognition of oil rigs on Sentinel-1 SAR images, discriminating them from false alarms, through a convolutional neural network with a different number of layers named models VGG-16 and VGG-19. We used the Sentinel-1 GRD product, which has a resolution of about 20 m, with VV and VH channel image data sets separated to observe polarization influences on target recognition. The experiments show that the VH polarization image embedding contributes with the best classification values, with the mean global accuracy of fifty random sampling tests equal to 86.4% and 84.1% using the VGG-16 and VGG-19 architectures, respectively, to extract features and feed a logistic regression classifier method. Results suggest that these approaches are a useful alternative for maritime surveillance employing SAR images of medium resolution and large swaths. Furthermore, the database balancing for wind conditions did not provide significant contributions to the accuracy of classification algorithms.