Channel-Extension Augmentation and Attention U-Net for Enhanced Semantic Segmentation of Marine Debris in Forward-Looking Sonar

Marine debris poses an escalating threat to coastal ecosystems and human health, yet accurate detection in turbid underwater environments remains challenging due to speckle noise and low contrast in sonar imagery. In this work, we adopt the open-access Water Tank Marine Debris with ARIS Explorer 3000 dataset, comprising 1,125 training, 126 evaluation, and 617 independent test images, and introduce a robust semantic segmentation pipeline for forward-looking sonar data. We proposed a channel-extension augmentation strategy that integrates three complementary modalities: raw intensity, Laplacian-of-Gaussian edge maps, and local standard-deviation texture channels. These enriched inputs are processed by an Attention U-Net architecture augmented with Squeeze-and-Excitation blocks, enabling dynamic channel recalibration and more discriminative feature learning. Extensive experiments against FCN, UPerNet, and vanilla U-Net baselines show superior performance, with our model achieving $86.2 \%$ mean Intersection-over-Union and $93.5 \%$ mean F-score on the augmented test set, improvements of up to $2.2 \%$ over strong baselines. Qualitative assessments demonstrate markedly sharper boundary delineation and enhanced robustness to speckle artifacts and background clutter.