An Ensemble machine learning model for drought prediction in Kilifi County, Kenya

Abstract

Drought prediction is critical for mitigating its adverse effects on agriculture, water resources, and livelihoods. This study developed a stacking ensemble learning model for multi-class drought prediction using machine learning techniques, including Random Forest, Support Vector Machine (SVM), XGBoost, and LightGBM. The CRISP-DM methodology guided the research, ensuring a systematic approach to data collection, pre-processing, model training, evaluation, and interpretation. Each model was assessed based on accuracy, recall, precision, and F1-score to determine its effectiveness in classifying drought severity levels. Among the base models, Random Forest achieved the highest accuracy (62.79%) and recall (62.79%), demonstrating its ability to capture complex patterns in the data while maintaining a reasonable balance between precision and recall. In contrast, SVM performed the weakest, with an accuracy of 25.58% and an F1-score of 29.91%, indicating its limited effectiveness in handling the dataset's complexity and class imbalance. The XGBoost and LightGBM models showed moderate performance, with both achieving an accuracy of 55.81% and recall of 55.81%, but their high precision (90.93% for XGBoost and 83.97% for LightGBM) suggests they are effective in minimizing false positives, though they struggle to balance precision and recall effectively. The stacking ensemble model, which integrated these classifiers with Logistic Regression as the meta-model, yielded an accuracy of 65.12% and an F1-score of 62.68%, outperforming all individual base models. To further enhance model performance, the study recommends incorporating additional environmental variables, such as land surface temperature (LST), Standardized Precipitation Evapotranspiration Index (SPEI) and the Palmer Drought Severity Index (PDSI) to significantly improve drought prediction accuracy. Future research should explore deep learning approaches, hybrid models combining physical hydrological and machine learning methods, and explainability techniques like SHAP values or LIME to strengthen trust in AI-driven drought forecasting systems.

Keywords: Drought Severity, Drought Prediction, Stacking, Ensemble Learning, Model Evaluation