Dynamic critical groundwater depth as a predictor of irrigation-intensified salinization in lowland Hungary

Shallow groundwater in continental lowland environments sustains upward capillary fluxes that transport dissolved salts to the land surface. However, the depth below which this capillary-driven contribution becomes negligible, often parameterized as the extinction depth in groundwater model ET packages, has been treated as a static, soil-dependent parameter. We argue that salinization risk is governed by the dynamic overlap between the seasonal cycles of groundwater level (GWL) and evapotranspiration (ET), and develop a time-varying risk index based on the proximity of the water table to a dynamic, GWL-dependent critical depth, and capillary fringe. The framework is applied to a 4,041 km² sub-region of the TIKEVIR system in southeastern Hungary using 74 years (1951–2025) of gridded GWL observations and FORESEE-HUN climate forcing. The domain-averaged seasonal cycle (shallowest in April at 2.50 m, deepest in October at 3.04 m, amplitude 0.54 m) defines a critical salinization window in March–June when the water table resides within the critical zone (zGWL < dcrit) under rising evaporative demand. A ConvLSTM network with eight input channels, trained on 1971–2019 data and evaluated on 2020–2024, achieves NSE = 0.839 and RMSE = 0.416 m for 1-month-ahead GWL forecasting at 1 km resolution, with 27–40 % RMSE reduction over persistence at 1–12-month lead times. Monte Carlo dropout analysis has verified a narrow prediction uncertainty, with a mean risk credible interval width of 0.014 (dimensionless). By integrating the forecast with the risk index, it was determined that 19.2% of the domain, equivalent to 776 km², is persistently at high or critical risk, predominantly located in discharge zones. A notable secular deepening of the water table, measured at 0.051 m per decade (p = 0.032), has slightly diminished the domain-averaged risk; however, the feedback loop between irrigation and salinization may counterbalance this trend. This methodology holds potential for application to similar lowland aquifer systems where the groundwater level (GWL) and evapotranspiration (ET) cycles are phase-offset.