<![CDATA[The growing incidence of drought attributable to climate change poses serious challenges to food systems and agricultural livelihoods in Majene Regency, West Sulawesi. To address this, the present study compares the predictive accuracy of Random Forest and XGBoost in classifying drought occurrences from meteorological variables, using 300 monthly observations recorded by BMKG between 2000 and 2024. Exploratory data analysis revealed that 60% of drought events are concentrated in August and the target variable is severely imbalanced (16.3% drought vs. 83.7% non-drought). At the default threshold of 0.50, Random Forest outperforms XGBoost in all metrics. However, because imbalanced data causes both models to underdetect drought at the default threshold (sensitivity = 0.267), threshold optimization was applied by lowering the decision threshold to τ = 0.15. The result of threshold optimalization was significantly improved Random Forest sensitivity from 0.267 to 0.867, detecting 13 out of 15 drought events in the test set while maintaining an AUC-ROC of 0.916. Feature importance analysis consistently identifies X3 as the dominant predictor in both models. Random Forest with threshold τ = 0.15 is recommended as the primary model for drought early warning systems in Majene.]]>
Copyrights © 2026
