Comparative machine learning and deep learning approaches for agricultural drought monitoring

Abstract

Study regionThis study considers Iran, encompassing hyper-arid to humid hydroclimates and major agricultural plains. Using 70 synoptic stations (2001–2022), we collocated station observations with satellite/reanalysis predictors from the Global Precipitation Measurement (GPM) mission, the Moderate Resolution Imaging Spectroradiometer (MODIS), the Famine Early Warning Systems Network Land Data Assimilation System (FLDAS), and the Copernicus Climate Change Service (C3S).Study focusAgricultural drought monitoring benefits from combining indicators of meteorological forcing and land-surface response, yet many studies rely on a single index or combine indices without an operational integration logic. We propose a dual-index framework for Iran integrating the Soil Moisture Deficit Index (SMDI) and the 3-month Standardized Precipitation–Evapotranspiration Index (SPEI-3).New hydrological insights for the regionWe combine stability selection with leakage-safe forward expanding cross-validation and a held-out most-recent test window to compare Light Gradient Boosting Machine (LightGBM), Random Forest, Elastic Net, and a feature-tokenizer Transformer. SMDI is estimated more reliably (best RMSE = 0.80, R² = 0.82) than SPEI-3 (best RMSE = 0.96, R² = 0.55). Uncertainty is quantified from held-out test absolute errors via empirical quantiles (50% and 90%); for SMDI, ∼50% of predictions fall within ∼0.5 index units and ∼90% within ∼1–1.5 units. These quantile error bands are attached as confidence qualifiers to the monthly drought classes in the monitoring framework, where SMDI anchors severity and SPEI-3 supports early-warning escalation.