Integrating random forest and isotopic tracers to optimize PMF-based source apportionment of watershed pollution

Based on historical data (2020-2024) and intensified sampling in 2024 from the Qujiang River Basin, this study systematically analyzed the variations in water quality parameters across different hydrological periods. By integrating the Random Forest (RF) model, Positive Matrix Factorization (PMF), stable isotope techniques, and phosphorus speciation analysis, an optimized water quality assessment framework was constructed to accurately identify pollution sources. The results indicate that water temperature (T), pH, and permanganate index (COD_Mn) were significantly higher during the wet season, whereas dissolved oxygen (DO), ammonia nitrogen (NH₄⁺-N), total nitrogen (TN), fluoride (F^-), and organic carbon exhibited higher concentrations during the dry season. The RF model successfully reduced the number of key parameters required for Water Quality Index (WQI) evaluation from 10 to 5 (TP, TN, COD, DO, and BOD₅), maintaining high predictive accuracy (R² = 0.9245) while significantly lowering monitoring costs. Stable isotope tracing provided critical validation for the PMF model in identifying pollution sources and accurately constraining their contribution ratios. The results showed that TN primarily originated from sewage (70.7% in the wet season and 40.0% in the dry season) and soil/fertilizer sources. The PMF model identified four major pollution sources: industrial wastewater, agricultural fertilizers, domestic sewage, and seasonal climatic factors. Innovatively, the Random Forest algorithm was applied to weight and optimize the PMF outcomes. To bridge the gap between mathematical solutions and practical management, this study introduces a Random Forest-based weighting calibration for PMF-derived source apportionment, shifting the focus from numerical optimum to environmental accountability. After correction, industrial wastewater was identified as the dominant contributor (39.74% in the wet season and 36.68% in the dry season), a source that had been underestimated in conventional PMF results. Phosphorus speciation analysis further confirmed the influence of land use on pollutant composition. Dissolved organic phosphorus (DOP) dominated in urban areas, dissolved inorganic phosphorus (DIP) was prevalent in agricultural regions, and particulate organic phosphorus (POP) constituted the highest proportion in forested areas. This study reveals the dominant role of anthropogenic drivers in shaping water quality dynamics in rapidly urbanizing river basins and provides a scientific basis for targeted water pollution control strategies.