Verma et al. (2026) An explainable deep learning framework for sub-seasonal to seasonal agricultural drought prediction over India

Identification

• Journal: Agricultural Water Management
• Year: 2026
• Date: 2026-07-07
• Authors: Saurabh Verma, L. Karthikeyan, Subimal Ghosh
• DOI: 10.1016/j.agwat.2026.110597

Research Groups

• Centre of Studies in Resources Engineering, Indian Institute of Technology Bombay, Mumbai, India
• Centre for Climate Studies, Indian Institute of Technology Bombay, Mumbai, India
• Department of Civil Engineering, Indian Institute of Technology Bombay, Mumbai, India

Short Summary

This study presents an explainable deep learning framework to predict agricultural drought (AGD) at sub-seasonal-to-seasonal scales over India using a Stacked Long Short-Term Memory (SLSTM) model. The framework integrates local antecedent land-atmosphere variables and non-local climate drivers, represented by Sea Surface Temperature (SST) anomalies, with Soil Moisture Percentile (SMP) as the drought indicator.

Objective

• Develop a data-driven framework to predict AGD at 1–3-month lead times over India.
• Investigate the impact of local and non-local features on AGD development and predictability across India at various seasons and lead times.

Study Configuration

• Spatial Scale: Grid-wise, with a spatial resolution of 0.25°, covering the Indian landmass from 65°E to 100°E in longitude and 5°N to 40°N in latitude.
• Temporal Scale: Monthly, spanning from June 1981 to May 2022.

Methodology and Data

• Models used: Stacked Long Short-Term Memory (SLSTM) model.
• Data sources:
  • Soil moisture data from the Global Land Evaporation Amsterdam Model (GLEAM).
  • Sea Surface Temperature (SST) anomalies from NOAA-PSL.
  • Precipitation, temperature, and Potential Evapotranspiration (PET) from IMD.
  • Vapour Pressure Deficit (VPD) from Terra-climate.
  • Normalized Difference Vegetation Index (NDVI) from NOAA-CDR.

Main Results

• The SLSTM model outperformed benchmark deep learning models (XGBoost and LSTM) and the SEAS5 dynamical forecast system in predicting major AGD events, including the 2015–16 event.
• At a 1-month lead time, the SLSTM model demonstrated strong performance, with percentage correct and probability of detection between 0.6–1 and false alarm rate between 0–0.3, particularly in Northern, Southern, and Eastern India.

Contributions

• This study presents the first purely grid-wise data-driven stacked-LSTM (SLSTM) framework for predicting soil moisture percentile (SMP)-based AGD across India at 1–3-month lead times.
• The extreme-aware custom loss function introduced in this study improves the prediction of severe drought events and provides a physically interpretable and operationally relevant framework for AGD forecasting over India.

Funding

• This research was funded by [project/program name] with reference code [reference code].

Citation

@article{Verma2026explainable,
  author = {Verma, Saurabh and Karthikeyan, L. and Ghosh, Subimal},
  title = {An explainable deep learning framework for sub-seasonal to seasonal agricultural drought prediction over India},
  journal = {Agricultural Water Management},
  year = {2026},
  doi = {10.1016/j.agwat.2026.110597},
  url = {https://doi.org/10.1016/j.agwat.2026.110597}
}