An integrated diagnosis of agricultural groundwater over-abstraction in a Lake Urmia tributary catchment

Groundwater over-abstraction in agricultural catchments is not only a hydrological decline problem; it is also a management problem shaped by crops, wells, incentives and institutional capacity. This study develops an integrated diagnosis for the Siminehrood catchment, a major agricultural tributary of Lake Urmia, to identify where historical groundwater-decline potential is highest, which farm-level conditions are associated with farmer-reported over-abstraction, and which management instruments are most highly prioritised for restoration-oriented agricultural water management. Groundwater and well records from 117 legally registered wells (1380–1400 Solar Hijri calendar years; approximately 2001–2021; 105 complete cases in the final support vector machine modelling dataset) were combined with Landsat-derived land use, precipitation, elevation, terrain, soil-particle, vegetation, curve-number and distance-to-river indicators, a farmer survey (n = 373), and a two-round Delphi assessment (31 experts in round one; 31 in round two). Random Forest benchmarking showed strong in-sample fit but weak cross-validated performance, whereas the precipitation-included radial-basis-function support vector machine provided modestly better cross-validated performance (coefficient of determination = 0.213; root mean square error = 0.271; mean absolute error = 0.198) and was retained as the final diagnostic Groundwater Decline Potential Index model. Model explanation was implemented using SHAP with the DALEX package. The Heckman two-step model used only questionnaire-based behavioural over-abstraction variables: a binary farmer-reported over-abstraction indicator and a conditional over-abstraction intensity percentage. The Heckman dependent variable does not include the hydrological exposure criterion based on location in a groundwater-level-decline area, making the farmer-behaviour component independent of the groundwater-level records used for spatial modelling. The final framework should therefore be read as a historical diagnostic screening and governance-targeting approach, not as a high-confidence forecast of future groundwater decline or a quantified lake-level restoration model.