Azizi et al. (2026) Comparative machine learning and deep learning approaches for agricultural drought monitoring: Dual-index modeling in Iran

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-03-25
• Authors: Masoud Azizi, Ali Abbasi, Mohammad Reza Asli Charandabi
• DOI: 10.1016/j.ejrh.2026.103376

Research Groups

• Department of AI and Land Use Change, Faculty VI, Technische Universität Berlin, Berlin, Germany
• Department of Civil Engineering, Faculty of Engineering, Ferdowsi University of Mashhad, Mashhad, Iran
• Department of Water Management, Faculty of Civil Engineering and Geosciences, Delft University of Technology, Delft, The Netherlands
• Faculty of Civil Engineering, Shahrood University of Technology, Shahrood, Iran

Short Summary

This study develops a dual-index machine learning framework for agricultural drought monitoring in Iran, integrating the Soil Moisture Deficit Index (SMDI) and the 3-month Standardized Precipitation–Evapotranspiration Index (SPEI-3) using multi-source predictors. It demonstrates that SMDI is estimated more reliably (best RMSE = 0.80, R² = 0.82) than SPEI-3 (best RMSE = 0.96, R² = 0.55) and proposes an operational classification system with uncertainty quantification.

Objective

• To develop a nationwide, time-consistent framework for agricultural drought monitoring across Iran by jointly modeling SPEI-3 and SMDI from a harmonized multi-source predictor set.
• To assess the accuracy of SPEI-3 and SMDI estimation from multi-source predictors at monthly scale across Iran's diverse climate regimes.
• To compare the generalization reliability of LightGBM, Random Forest, Elastic Net, and FT-Transformer under time-aware validation.
• To identify the most consistent predictors and lagged features controlling model skill and interpretability for each drought target.
• To develop an operational drought classification linking severity (SMDI) with early-warning signals (SPEI-3).

Study Configuration

• Spatial Scale: Iran, covering approximately 1.6 million square kilometers, using data from 70 synoptic stations distributed across 31 provinces.
• Temporal Scale: Monthly data from January 2001 to December 2022 (22 years).

Methodology and Data

• Models used: Light Gradient Boosting Machine (LightGBM), Random Forest, Elastic Net, Feature-Tokenizer Transformer (FT-Transformer), K-means clustering (for hydroclimatic regimes).
• Data sources:
  • Satellite/Reanalysis:
    • Global Precipitation Measurement (GPM) IMERG (precipitation, 0.1° spatial resolution).
    • Moderate Resolution Imaging Spectroradiometer (MODIS) (Land Surface Temperature (LST), Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Normalized Multi-band Drought Index (NMDI), Normalized Difference Water Index (NDWI5, NDWI6, NDWI7), Normalized Difference Drought Index (NDDI5, NDDI6, NDDI7); 1 km and 500 m spatial resolution).
    • Famine Early Warning Systems Network Land Data Assimilation System (FLDAS) Noah v001 (surface soil moisture (0–10 cm), 0.1° spatial resolution).
    • Copernicus Climate Change Service (C3S) Level-3 Satellite Soil Moisture COMBINED product (root-zone soil moisture, 0.25° spatial resolution).
  • Observation:
    • 70 synoptic stations from Iran Meteorological Organization (monthly mean air temperature, monthly accumulated precipitation).

Main Results

• Predictive skill is consistently higher for SMDI than for SPEI-3 across all models.
• LightGBM achieved the best overall accuracy for SMDI (RMSE = 0.800, R² = 0.821).
• For SPEI-3, LightGBM and Random Forest performed similarly (LightGBM: RMSE = 0.962, R² = 0.547; Random Forest: RMSE = 0.963, R² = 0.545).
• SMDI predictions show dense alignment around the 1:1 line with dispersion widening in tails, while SPEI-3 shows broader spread and understatement of extreme negative values ("regression toward the mean").
• Error behavior is temporally stable, with no sustained drift in residuals over the held-out period.
• Dominant drivers for both targets are short-term persistence (lagged target values), followed by precipitation and land-surface temperature anomalies, then soil-moisture anomalies and vegetation/water indices.
• A three-cluster K-means solution effectively captured hydroclimatic heterogeneity among stations, and cluster indicators contributed modestly but consistently to model predictions.
• An operational dual-index framework is proposed, using SMDI to anchor severity and SPEI-3 for early-warning escalation, with uncertainty communicated via empirical error quantiles.

Contributions

• Provided a national-scale, dual-target modeling system that aligns meteorological and soil-moisture drought on the same temporal scale.
• Introduced a leakage-resistant, forward-chaining validation design for drought machine learning in Iran to ensure realistic prospective performance.
• Leveraged complementary soil-moisture evidence streams (reanalysis-based FLDAS and observation-based C3S) within the same framework to reduce single-product dependence.
• Coupled stability-oriented feature selection (Elastic Net) with interpretable importance profiling and an operational, severity-to-warning classification workflow.

Funding

Funding information for this research is not explicitly provided in the paper.

Citation

@article{Azizi2026Comparative,
  author = {Azizi, Masoud and Abbasi, Ali and Charandabi, Mohammad Reza Asli},
  title = {Comparative machine learning and deep learning approaches for agricultural drought monitoring: Dual-index modeling in Iran},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103376},
  url = {https://doi.org/10.1016/j.ejrh.2026.103376}
}