Diagnosing SWAT-Represented Drivers of Low Flow in the Musimcheon Stream Using Explainable Artificial Intelligence (XAI)

Groundwater storage sustains low flows and obscures their controls. We used explainable artificial intelligence (XAI) to diagnose the drivers of low flow as represented by a calibrated Soil and Water Assessment Tool (SWAT) model of the Musimcheon watershed (180.5 km2, Republic of Korea). An XGBoost surrogate was trained on daily SWAT outputs to reproduce the 7-day trailing mean discharge (Q7) from 15 antecedent predictors and evaluated by nested leave-one-year-out cross-validation (NSE of log Q7 = 0.93). SHapley Additive exPlanations (SHAP) showed that short-term (≤30 d) precipitation dominated overall flow variability, whereas on low-flow days the share of long-term (≥90 d) precipitation and deep percolation rose from 40.4% to 51.4% (38.6% → 45.5% out of sample). This shift persisted under alternative low-flow thresholds, explanation methods and predictor sets, emulated groundwater delays of 60–300 days and an extended 2002–2022 record, although its magnitude depended on the delay parameter. The lowest simulated Q7 (June 2017) followed a 180-day percolation of 19% of the seasonal mean. Percolation accumulated over 180 days showed threshold-like behavior (uncertain transition range ≈ 70–83 mm). These results are diagnostic attributions of model behavior based on a short record, not observed causal relationships, and require observational verification before use in early warning.