Detecting Underwater Discrete Scatterers in Echograms with Deep Learning-Based Semantic Segmentation

This paper reports on an exploratory study of the automatic detection of discrete scatterers in the water column from underwater acoustic data with deep learning (DL) networks. Underwater acoustic surveys using moored singlebeam multi-frequency echosounders make environmental monitoring tasks possible in a non-invasive manner. Discrete scatterers, i.e., individual marine organisms, are particularly challenging to detect automatically due to their small size, sometimes overlapping tracks, and similarity with various types of noise. As our interest lies in identifying the presence and general location of discrete scatterers, we propose the use of a semantic segmentation paradigm over object detection or instance segmentation, and compare several state-of-the-art DL networks. We also study the effects of early and late fusion strategies to aggregate information contained in the multi-frequency data. Experiments on the Okisollo Channel Underwater Discrete Scatterers dataset, which also include schools of herring and juvenile salmon, air bubbles from wave and fish school activity, and significant noise bands, show that late fusion yields higher metrics, with DeepLabV3+ outperforming other networks in terms of precision and intersection over union (IoU) and Attention U-Net offering higher recall. The detection of discrete scatterers is a good example of a problem for which exact annotations cannot be reached due to various reasons; in several cases, network outputs seem visually more adequate than the annotations (which contain inherent noise). This opens up the way for utilizing actual detection results to improve the annotations iteratively.