Machine learning-based drought prediction using Palmer Drought Severity Index and TerraClimate data in Ethiopia.

Area of Science:

• Environmental Science
• Climate Science
• Data Science

Background:

• Ethiopia faces significant climate variability and drought susceptibility.
• Accurate drought prediction is vital for water management and agricultural planning.

Purpose of the Study:

• To classify the Palmer Drought Severity Index (PDSI) using machine learning models.
• To evaluate various classifiers on TerraClimate data for drought prediction in Ethiopia.

Main Methods:

• Employed Logistic Regression, SVM, KNN, Decision Tree, Random Forest, Gradient Boosting, Naive Bayes, AdaBoost, and XGBoost.
• Utilized a hybrid resampling method (upsampling and SMOTE) for data imbalance.
• Conducted hyperparameter tuning via grid search and cross-validation.

Main Results:

• Random Forest achieved the highest accuracy (71.18%), F1-score (0.71), and ROC AUC (0.9000).
• Gradient Boosting and SVM also demonstrated strong performance (ROC AUC 0.8982 and 0.8681).
• SHAP analysis identified soil moisture, precipitation, and vapor pressure deficit as key predictors.

Conclusions:

• Ensemble learning models, especially Random Forest, are effective for drought severity classification.
• Machine learning offers advantages over traditional time-series models (ARIMA) for complex climate data.
• Findings support drought early warning systems and climate resilience in Ethiopia.