Percentage distribution of PDSI classes.

Accurate drought prediction is essential for proactive water management and agricultural planning, especially in regions like Ethiopia that are highly susceptible to climate variability. This study investigates the classification of the Palmer Drought Severity Index (PDSI) using machine learning models trained on TerraClimate data, incorporating variables such as precipitation, temperature, soil moisture, and vapor pressure deficit. We employed several classifiers Logistic Regression, Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Decision Tree, Random Forest, Gradient Boosting, Naive Bayes, AdaBoost, and XGBoost with Logistic Regression serving as a baseline statistical approach for comparison. To address data imbalance across drought classes, we applied a hybrid resampling method combining manual upsampling and SMOTE. Hyperparameter tuning was conducted using grid search and cross-validation. Random Forest outperformed all models, achieving an accuracy of 71.18%, F1-score of 0.71, and ROC AUC of 0.9000. Gradient Boosting and SVM also performed well with ROC AUC values of 0.8982 and 0.8681, respectively. SHAP analysis revealed that soil moisture, precipitation, and vapor pressure deficit were the most influential features in predicting drought severity. For benchmarking, an ARIMA (3,1,2) time-series model was applied but yielded poor performance (RMSE = 1.789, R² = −0.077), confirming the advantages of non-linear machine learning techniques for complex climate data. The results highlight the utility of ensemble learning in environmental modelling, offering valuable insights for drought early warning systems and climate resilience planning in Ethiopia. Future work should explore integrating localized predictors and real-time data to enhance prediction robustness.