An interpretable ensemble-learning framework for total nitrogen retrieval in complex inland waters: application to the Xinfengjiang Reservoir

Monitoring total nitrogen (TN) in drinking water reservoirs is vital for preventing eutrophication, yet remote sensing retrievals often struggle with limited accuracy and opaque decision processes in complex inland waters. This study presents an interpretable framework coupling innovative feature engineering, ensemble learning (XGBoost, CatBoost, LightGBM), the Distance between Indices of Simulation and Observation (DISO) metric, and SHapley Additive exPlanations (SHAP). From 4270 Landsat 8–9 band combinations, feature optimization achieved >99% redundancy reduction. Validated with in-situ data (2020–2024) from the Xinfengjiang Reservoir, LightGBM achieved superior performance (R² = 0.72, RMSE = 0.25 mg/L, BIAS = −0.02 and DISO = 0.35). SHAP analysis revealed (B4 − B3)/(B3 − B6) and (B2 + B1)/(B1 + B4) as the most influential features (SHAP value > 0.1), demonstrating biophysical plausibility linked to algal biomass and water optical properties. Spatiotemporal mapping showed predominantly low TN concentrations (>90% Class II–III surface water standard) alongside persistent upstream hotspots requiring targeted mitigation. By quantifying feature contributions, this interpretable framework enhances retrieval accuracy and methodological transparency, offering actionable tools for reservoir water quality management and sustainable ecological governance.