Interpretable artificial neural network reveals region-specific controls of chlorophyll-a in the Yellow River Estuary and adjacent sea.

This study developed an explainable AI-based framework integrating data augmentation, artificial neural network (ANN) modeling, and SHAP interpretation to investigate chlorophyll-a (Chl-a) variability in the Yellow River estuarine-coastal ecosystem from 2017 to 2024. Random perturbation-based augmentation alleviated data sparsity and improved model robustness. Distinct regional contrasts in environmental factors and Chl-a associations were identified across three dynamically delineated subregions. In the northern part of the Yellow River Estuary (NYRE), dissolved inorganic nitrogen (DIN), dissolved inorganic phosphorus (DIP), and chemical oxygen demand (COD) jointly accounted for 44.9% of modeled Chl-a variability, with response transition thresholds of 0.43, 0.004, and 1.13 mg L-1, respectively. In Laizhou Bay (LB), DIN thresholds were lower (0.32 mg L-1), and the joint above-threshold area (462.7 km2) was substantially larger than in the NYRE (29.5 km2), indicating broader eutrophication potential. In contrast, nutrient-related contributions in the Yellow River minimal-impact Central Bohai Sea (MinCBS) were limited (∼8%), and Chl-a variability was more closely associated with hydrographic and light factors. The framework captures nonlinear and interaction-dependent environmental responses and provides a transferable approach for adaptive nutrient management in complex coastal systems.