Spatial Analysis of Ensemble Learning Models for Agricultural Drought Early Warning

Drought poses a serious threat to rice production and local food security, triggered by climate anomalies such as El Niño. This study aims to evaluate and compare the performance of Ensemble Learning Models in classifying drought levels and analyze its correlation with periods of climate anomalies. This study uses Landsat 9 image data in the simulation period from June 2024 to July 2025, which is processed with HSV-based pan-sharpening and spectral index extraction (NDVI, NDWI, NDDI, EVI, LST). The modeling process applied undersampling to address class imbalance and hyperparameter tuning optimization using Optuna. The models compared included Random Forest, LightGBM, AdaBoost, XGBoost, and Gradient Boosting. The results showed that Gradient Boosting excelled with a train accuracy of 96,85% in original dataset with split dataset 70:30, whereas rise to 98.98% after tuning. Spatial validation was conducted in other rice field plots, however its steadfastly on research area with same treatment. The classification map shows the dominance of the moderate category, which temporally coincides with the period of rainfall decline associated with El Niño, although a direct causal relationship requires further investigation. These findings confirm that remote sensing combined with machine learning is effective for drought monitoring, with the caveat that the application of undersampling and limited spatial validation that is, confined solely to the research area; needs to be considered in the interpretation of results.