Machine Learning Based Drought Prediction Using the Standardized Precipitation Evapotranspiration Index (SPEI) in Kebbi State, Nigeria

Drought represents a major threat to livelihoods and economic stability in regions prone to its occurrence. This paper aims to address the gap in applying machine learning techniques for enhanced meteorological drought prediction to support resilience and preparedness. The study focuses on Kebbi State, located in northwest Nigeria, which experiences droughts with devastating agricultural, ecological and humanitarian impacts. The Standardized Precipitation Evapotranspiration Index (SPEI) was used to calculate different drought severity based on rainfall deficit, over varying accumulation periods (3-month, 6-month) over four decades (1980–2022). Different time series meteorological parameters such as mean temperature, maximum temperature, minimum temperature, radiation, wind speed, precipitation was used in training machine learning models to predict and forecast future drought risk across Kebbi’s regions. Four candidate models were evaluated Random Forest (RF), Extreme Gradient Boosting (XGB), 1D Convolutional Neural Networks (CNN), and Long Short-Term Memory Networks (LSTM). Results indicate RF models consistently achieved highest prediction accuracy (R2: 47–67%) for both short and long-term SPEI forecasts across different regions over the other models, while LSTM was not able to make good prediction for drought in Kebbi state. Optimized XGB models also performed reasonably well for specific locations. One-year lead SPEI projections exhibit XGB potential for advancing early warning given forecast reliabilities. This pioneering study provides robust evidence for integrating machine learning for drought prediction in Kebbi state, Nigeria which is located in the sub-Sahara region.