Multimodal Detection of Underwater Gas Seeps Using U-Nets Trained on Simulated Acoustic and Visual Data

Autonomous detection, quantification, and sampling of natural or anthropogenic gas seeps from the seafloor are crucial for sustainable and efficient long-term monitoring of bio-geo-chemical processes and environmental changes. The unstructured and dynamic underwater environment requires robust multimodal approaches for object detection, but the lack of large datasets complicates algorithm development and training. This paper addresses these challenges by developing two U-net models for acoustic and vision-based seep detection, which can be integrated into a multimodal detection framework. The Unreal Engine is used for realistic high-resolution imaging, and a first-principle physics-based sonar simulation is integrated for efficient and accurate training of the acoustic U-net. Preliminary results demonstrate the efficacy of the trained detection models and show that data augmentation based on sonar simulation improves detection performance on real-world sonar datasets.