Deciphering fish density thresholds under the Yangtze fishing ban: An interpretable hydroacoustic-machine learning framework for adaptive reservoir management.

Effective, non-invasive fish stock monitoring is critical for assessing the ecological outcomes of large-scale management actions, such as China's decade-long Yangtze River fishing moratorium. To address this, we developed a four-year (2021-2024) integrated hydroacoustic and interpretable machine learning (Hydro-ML) framework to quantify fish density dynamics in Zhelin Reservoir, a vital regulated ecosystem within the Yangtze Basin. Among six evaluated ML models, XGBoost showed competitive predictive performance under the random split (NSE = 0.63; RMSE = 0.70; RRMSE = 40.05%). Interpretable SHapley Additive exPlanations (SHAP) analysis revealed water depth (WD), water temperature (WT), transparency (Tran), pH, comprehensive nutrient index (NI), and Chlorophyll-a (Chla) as primary environmental predictors of fish distribution, highlighting non-linear ecological thresholds within the fitted SHAP relationships-most notably habitat-depth (22.96 m; 95% CI: 22.08-23.84 m; CV: 1.1%) and transparency (2.31 m; 95% CI: 2.24-2.38 m; CV: 1.6%) breakpoints. Spatially, high-density fish zones consistently persisted in Xiuhe River tributaries and upstream bays, sharply contrasting with low-density zones in the more downstream reaches of the reservoir impoundment, which are most heavily influenced by dam operations. Our findings indicate that low-density patterns persisted in the downstream sections despite the fishing ban. This framework provides quantitative ecological reference points for adaptive management in Zhelin Reservoir and a reproducible approach for evaluating non-linear fish-environment relationships in other regulated freshwater systems.