Integrating Sentinel-1 Radar Soil Moisture and Explainable Machine Learning to Quantify Irrigation-Induced Groundwater Fluxes in the Hajdúság Region, Eastern Hungary

Abstract. Quantifying irrigation-induced groundwater recharge in water-stressed agricultural systems requires separating satellite-derived surface signals from background hydrological variability. The Hajdúság region (eastern Hungary) — a maize-dominated lowland with shallow sandy aquifers, intensive irrigation, and a dense monitoring network (502 wells) — provides an ideal test case. A statistically significant network-mean groundwater decline of −0.62 cm yr−1 (Sen's slope, 2020–2025, p < 0.01) underscores the need for improved recharge estimation tools. Sentinel-1 C-band relative Surface Soil Moisture (rSSM, 0–100 %) was derived from VV backscatter using the change-detection normalisation of Wagner et al. (1999), operationalised for Sentinel-1 GRD by Bauer-Marschallinger et al. (2019) (Equation 6), and integrated as a SAR-derived auxiliary predictor for maize seasons (April–October, 2020–2025). An XGBoost model with 30 predictors — encompassing autoregressive groundwater-state features, SAR backscatter and rSSM lags, terrain properties, temporal encodings, and auxiliary climate inputs (CHIRPS precipitation; MOD16 evapotranspiration gap-filled via IDW and Hargreaves–Samani) — was trained on 2020–2023 data using a 5-fold Group Cross-Validation scheme grouped by well station to prevent spatial data leakage, and evaluated on withheld data (2024–2025). The target variable was a water-table fluctuation recharge proxy (Sₑ × dH/dt, Sₑ = 0.05). Tree-SHAP decomposition quantified the independent contribution of each predictor group, including the conditional role of rSSM and SAR backscatter variables relative to dominant groundwater-state signals. The model achieved RMSE = 0.27 mm day−1, R2 = 0.788, NSE = 0.788, and KGE = 0.835 on the hold-out test set. Groundwater-state autoregression dominated model skill: dH_roll7 (42.2 %), dH_lag1 (21.3 %), and dH_lag3 (8.9 %) jointly accounted for 72.4 % of total mean absolute SHAP. Seasonal timing encodings provided additional predictive structure, while Sentinel-1-derived rSSM and SAR backscatter variables each contributed a complementary but secondary signal (≤0.9 % mean |SHAP| individually), improving recharge prediction beyond groundwater observations alone. Strongest SAR–recharge coupling occurs in central areas with shallow water tables and intensive irrigation, where SAR-derived features account for an estimated 15–25 % of the model-attributed vertical flux signal during peak growth stages (SHAP-derived model estimate; not a volumetric field measurement). Results are robust to ±50 % variation in Sₑ. Sentinel-1-derived rSSM and SAR backscatter variables serve as physically interpretable auxiliary predictors of shallow groundwater recharge in irrigated sandy aquifers, providing a complementary signal that is most informative during and immediately after irrigation events. Groundwater-state autoregression remains the dominant source of predictive skill. The integrated approach is reproducible, scalable to networks with similar hydrogeological characteristics, and provides a transferable framework for satellite-augmented recharge monitoring in data-rich agricultural regions.