Target classification in oceanographic SAR images with deep neural networks: Architecture and initial results

Abstract

Synthetic Aperture Radar (SAR) provides detailed information of Ocean's surface and man-made floating structures. Advances in SAR technology and deployment of new SAR satellites have contributed to an increasing number of remote sensing data available. Handle this large amount of data with human operators is infeasible. Therefore, the use of automated tools to process remote sensing images, identify regions of interest, and select relevant information are needed. The use of neural networks to solve SAR image classification problems is well known. A typical architecture consists of a shallow feed-forward neural network with an input layer, a hidden layer, and an output layer. This type of neural network, combined with back-propagation and a gradient-based training algorithm, is able to solve complex problems in SAR image analysis. However, this architecture is unable to take advantage of unlabeled data during its training process, and in many cases the input features need to be carefully tuned in order to reduce the overall network complexity. This paper proposes the application of Deep Neural Networks (DNN) to perform oceanographic-object classification.