Integrating Vegetation Health Indices and Machine Learning for Early Prediction of Agricultural Drought

Agricultural drought threatens food security across Nigeria’s Sahel, yet spatially explicit machine-learning-based prediction remains limited in the region. This study developed a spatiotemporal drought monitoring and prediction framework for Yobe State (2000–2025) integrating seven remote-sensing indicators; NDVI, EVI, VCI, TCI, SMI, LST, and rainfall within Google Earth Engine at 3 km resolution. A persistent north–south drought gradient emerged across all indicators, with northern local government areas exhibiting chronic vegetation stress, soil moisture deficit, and elevated surface temperatures throughout the study period. Three machine learning models; LSTM, Random Forest, and XGBoost were trained on Vegetation Health Index-derived drought severity classes. XGBoost achieved the highest predictive performance (accuracy = 86.9%; Kappa = 0.829; RMSE = 2.26), with rainfall and vegetation condition as dominant predictors. All three models converge on the northern LGAs as the highest-risk zone by 2030, supporting targeted early warning deployment and climate-smart agricultural investment across the Sahel.