Automated monitoring of coastal plastic pollution is critical for environmental management, but is severely hindered by the complexity of shoreline environments. Generic computer vision models fail due to extreme scale variation (from large nets to small fragments), complex background clutter (e.g., seaweed, rocks), and object deformation. We introduce the Shoreline-Adaptive Plastic Segmentation Network (SAPS-Net), a novel dual-cascade framework that addresses these “in-the-wild” challenges. SAPS-Net first classifies the shoreline’s environmental context (e.g., sandy, rocky) to dynamically load a specialized “expert” segmentation model. The model integrates bespoke modules for scale-adaptive attention and noise-gated dynamic convolution to precisely segment both large debris and small fragments. Trained on our new, large-scale Coastal Plastic Litter Data set (CPLD), SAPS-Net achieves a state-of-the-art 70.1% Average Precision (AP), significantly outperforming baselines (e.g., +17.6 AP vs Mask R-CNN). Crucially, it demonstrates a 10.0-point improvement in small-object AP (APs), addressing the most prevalent form of pollution. Empirical validation in a real-world field study (Hebei Province) confirmed the model’s robustness, achieving 0.90 F1-score and 0.95 precision. SAPS-Net provides a validated, scalable, and efficient tool for high-resolution litter mapping, optimizing cleanup logistics, and informing environmental policy.
Chen et al. (Tue,) studied this question.