Integrating Meteorological Data for Drought Prediction: A Hybrid Ensemble Machine Learning Approach in Northwestern Bangladesh

ABSTRACT Background Drought is a slow‐onset natural disaster with severe impacts on water resources, agriculture, and public health, particularly in climate‐vulnerable regions like northwestern Bangladesh. Accurate drought forecasting is essential for early warning and proactive mitigation. Methods This study develops a hybrid ensemble machine learning framework for 7‐day‐ahead forecasting of the Standardized Precipitation Evapotranspiration Index (SPEI) in Dinajpur district, Bangladesh, using meteorological data from 2023 to 2025. Seven individual models—random forest (RF), Extreme Gradient Boosting (XGBoost), gradient boosting machine (GBM), support vector regression (SVR), long short‐term memory (LSTM), gated recurrent unit (GRU), and convolutional neural network (CNN)—were developed. Four hybrid configurations were constructed: simple average, weighted average, meta‐RF, and meta‐LR. An ablation study was conducted to evaluate the PM2.5 contribution. Results Among individual models, GBM performed best (root mean square error [RMSE] = 0.0059, R 2 = 0.9999, Nash–Sutcliffe efficiency [NSE] = 0.9999), followed by RF (RMSE = 0.0079, R 2 = 0.9999, NSE = 0.9999) and XGBoost (RMSE = 0.0141, R 2 = 0.9997, NSE = 0.9996). These top three models (GBM, RF, and XGBoost) were selected as base models for hybrid ensembles. Among hybrid configurations, the Meta‐LR ensemble achieved the lowest RMSE (0.0050), whereas the weighted average ensemble (RMSE = 0.0052, R 2 = 0.9999, NSE = 1.0000) was selected as the final model due to its simplicity and statistically comparable performance. The ablation study revealed that PM2.5 did not consistently improve performance; therefore, it was excluded from the final model. The 2027 forecast predicts 75 drought days (20.5% of the year), with April identified as the most critical month (SPEI = −0.976). Conclusions This study demonstrates that hybrid ensemble approaches provide accurate drought forecasts, with the weighted average model offering an optimal balance of performance and interpretability. The framework supports proactive public health interventions, water resource management, and climate‐resilient planning in drought‐prone regions.