An integrated SWAT-XGBoost-SHAP framework identifies key drivers of critical source areas during critical periods in...

Non-point source (NPS) pollution has emerged as a critical environmental issue, significantly impacting water quality and ecosystem health at the watershed scale. The identification of critical periods (CPs) and critical source areas (CSAs) is fundamental for formulating effective watershed management strategies. However, the identification of effective management measures remains challenging, primarily due to the complex interplay between diverse pollution sources and dynamic environmental factors. To address this challenge, this study proposes an integrated framework that synergistically combines the Soil and Water Assessment Tool (SWAT) model, the eXtreme Gradient Boosting (XGBoost) machine learning algorithm, and the SHapley Additive exPlanations (SHAP) approach. The framework aims to quantitatively analyze the driving factors responsible for the formation of CSAs of NPS pollution in small watersheds during CPs. SWAT simulated nutrient loads, identifying CPs and CSAs via load-time and load-area curves. XGBoost modeled factor relationships, and SHAP quantified each driver's contribution. Applied to a small watershed, results showed CPs (months 2, 6, 7) contributed 59% of total nitrogen (TN) and 65% of total phosphorus (TP) loads. Within CSAs, 56.2-66.2% of TN/TP loads originated from just 35.2-36.8% of the area. Fertilizer application amount (mean |SHAP|= 1.91) and the proportion of cultivated land (mean |SHAP|= 0.52) were identified as the predominant drivers governing the formation of CSAs. Operationally, the identified thresholds (e.g., runoff 30 mm, fertilizer 100 kg/ha) serve as objective tipping points that trigger the implementation of targeted best management practices (BMPs). Despite limitations in the temporal scope of monitoring data and potential model parameter uncertainties, this framework provides a robust scientific basis and a novel methodology for precision watershed management.