Interpretable machine learning for river salinity dynamics in arid basins

Abstract Managing salinity in arid rivers is impeded by sparse monitoring, relying on low-frequency grab samples that miss hydrological event dynamics. Here, interpretable machine learning is applied to a 50-year monthly archive (1968–2018; Discharge, major ions, pH) from three stations on Iran’s Karkheh River. Gradient Boosting Regression achieves high predictive skill for Total Dissolved Solids (TDS)/Electrical Conductivity (EC) (test-set R 2 = 0.94/0.97; RMSE = 55 mg L −1 /56 µS cm −1 ), validated via time-aware cross-validation. SHAP-based feature attribution reveals that Na + and SO 4 –2 are the strongest contributors to TDS, while Na + and Cl − dominate EC, consistent with conservative salinity sources under baseflow conditions. A reduced-input decision tree (four predictors) retains R 2 = 0.81–0.87, enabling minimal-sensor monitoring. Flow-regime partitioning and STL (Seasonal-Trend decomposition using Locally estimated scatterplot smoothing)-detrended event composites reveal low-flow salinization and ion-specific post-flood recovery (Cl − : 1–2 months; Na + : 2–3 months), guiding targeted sampling. This framework extracts predictive power, process associations, and operational guidance from legacy grab-sample archives, scalable to data-limited basins worldwide.