Benchmarking AI-driven acoustic monitoring for floating marine debris: Challenges in deep learning-based debris extraction

Existing studies have identified a substantial amount of invisible floating debris in low-visibility marine environments, in addition to debris on the surface and seabed. These suspended pollutants represent a persistent and dynamic threat to marine ecosystems and maritime safety. Although sonar technology facilitates debris monitoring in low-visibility waters, the automatic extraction of small and weakly contrasted debris targets remains a critical challenge. Compared with traditional manual statistics, AI-driven debris quantification holds greater promise for large-scale monitoring. This article evaluates the performance of dominant deep learning models for the automatic extraction of floating debris in sonar images. The findings reveal that most segmentation models struggle when applied to floating debris, achieving an average IoU of no more than 50 % due to their low acoustic contrast with the surrounding water and the inherent ambiguity between targets. This study attributes these difficulties not only to data scarcity but also to the inherent physics of acoustic imaging, including range-dependent resolution and viewpoint sensitivity, which further complicate learning-based target extraction. The "Harsh Floating Debris" dataset is introduced and comprises 1000 acoustic images across four categories: plastic bags, bottles, metal cans, and mixed debris. Each image contains two to four manually annotated instances, providing a benchmark for evaluating segmentation and classification models under challenging underwater conditions. This work lays the foundation for developing more robust AI-driven systems for autonomous monitoring of floating marine debris.