Spatio-Temporal Deep Learning for Early Drought Prediction: A ConvLSTM–GNN Framework Using Real-Time Earth Observation Data

Our earlier work introduced an IoT–AI framework for climate-resilient smart farming that used vegetation indices and deep learning to detect crop stress for a farmer-facing mobile and voice Decision Support System (DSS). That framework treated stress detection as a per-location, per-timestep snapshot, and did not model how drought builds up over months or spreads across neighbouring regions. This paper extends it with a hybrid ConvLSTM–Graph Neural Network (ConvLSTM-GNN) framework that fuses real, open-access datasets — MODIS, Sentinel-2, CHIRPS, ERA5-Land, and India Meteorological Department (IMD) gridded products — to forecast district-level drought one to three months ahead. ConvLSTM layers learn the temporal evolution of gridded rainfall, soil-moisture, and vegetation data, while a district-adjacency graph neural network captures how drought propagates spatially between neighbouring districts. Piloted over drought-prone Telangana districts and validated against SPI/VHI-derived drought categories, the framework targets higher accuracy and longer lead time than threshold-based and single-location baselines, feeding its alerts directly into the DSS proposed earlier — moving it from reactive detection toward proactive, spatially aware early warning, in support of SDG 2 and SDG 13.