Inversion of Non-Optically Active Water Quality Parameters in Karst Mountain Reservoirs Based on Multi-Source Data: A Case Study of Pingzhai Reservoir

Accurate monitoring of water quality in karst mountainous reservoirs, where the ecosystems are fragile, is central to sustainable water resource management and risk mitigation. Existing remote sensing inversion studies, however, have mostly relied on spectral reflectance alone, overlooking the synergistic effects of meteorological variables, while machine learning models remain too opaque for meaningful interpretation. Taking Pingzhai Reservoir as a case study, this study develops two machine learning inversion frameworks. The first is driven solely by Sentinel-2 multispectral imagery and compares Random Forest, LightGBM, XGBoost, and CatBoost. The second couples multispectral data with meteorological variables and, based on the optimal model, predicts dissolved oxygen (DO), ammonia nitrogen (NH3-N), total nitrogen (TN), and pH. SHAP is subsequently used to quantify the contribution of each input and to analyze how different features influence model predictions. Multisource data fusion is found to improve inversion accuracy, and Random Forest yields the best performance among the four models, with R2 values of 0.88, 0.68, 0.62, and 0.89 for DO, NH3-N, TN, and pH, respectively. SHAP analysis revealed distinct dominant drivers for different water quality parameters. The framework produced good inversion results at Pingzhai Reservoir, offering an interpretable method for water quality monitoring in karst mountain reservoirs.