Leveraging Seabed-Context Information for Improved Underwater Target Classification Using Synthetic Aperture Sonar

Recent advances in deep learning have enabled accurate and efficient classification of underwater targets captured in synthetic aperture sonar. The success of these models relies on the availability of large datasets with highly textured and rich images from a variety of environments. However, performance decreases in complex seabeds, where targets are more difficult to detect and false alarms are more common - for example, seagrass drastically increases the difficulty of finding a target, and sand ripples increase difficulty when insonified from an angle orthogonal to the ripples. Moreover, the amount of available training data is often limited in these challenging environments compared to smooth seafloors. High-resolution images enable relatively straightforward visual recognition of seafloor type: smooth, sand ripples, vegetation, clutter, etc.In previous work, we developed an automatic seabed characterization algorithm based on Gaussian processes. In this work, we aim to leverage this knowledge in order to improve target classification performance. To this end, we introduce a new context-dependent classification algorithm to address and exploit the environment. Based on our previous ATR framework, we implemented and trained convolutional neural networks (CNNs) employing two strategies: by injecting the seabed information directly into the decision-maker (i.e. the CNN) and by specializing the CNNs (i.e. fine-tuning) for each class of seabed, with the decision being made using the outputs of the CNNs and the seabed prediction. The effectiveness of our approach will be demonstrated using real data collected during at-sea experiments.