Spatio-Temporal Deep Learning Models for Forecasting Agricultural Drought in Rain-Fed Regions

Abstract

Droughts pose a severe threat to India’s predominantly rain-fed agriculture, affecting food security and livelihoods. Accurate spatio-temporal drought forecasting is critical for proactive management. This study reviews and advances deep learning (DL) approaches for agricultural drought prediction in India’s rain-fed regions. We leverage high-resolution climate data (e.g., IMD’s gridded rainfall, NASA satellites, soil moisture datasets) and drought indices (SPI, SPEI, PDSI, NDVI/VCI) as inputs. We implement and compare several DL architectures: recurrent neural networks (LSTM, Bi-LSTM), convolutional recurrent models (CNN-LSTM, ConvLSTM), transformer-based models (FourCastNet, EarthFormer), and graph neural networks (GNN-LSTM with attention). Experimental setup includes data preprocessing (e.g. bias correction), training on historical drought indices, and evaluation with metrics (RMSE, MAE, R², accuracy). Our results show that spatio-temporal models (especially transformer and graph-based architectures) outperform simpler models in multi-month forecasts. For example, a GNN-LSTM model yields RMSE≈0.033 on Jaisalmer drought data, significantly lower than CNN-LSTM or ANN baselines. Visualizations (maps, graphs) illustrate model predictions across Indian regions. We discuss model strengths and limitations, highlight challenges in data-scarce areas, and outline future work (e.g. transfer learning, hybrid physical-data approaches). This study underscores the promise of DL for operational drought early-warning in India’s vulnerable rain-fed zones.