Evaluation of performance of drought prediction in Indonesia based on TRMM and MERRA-2 using machine learning methods

East Nusa Tenggara Province is one of the most vulnerable regions in Indonesia to drought. Drought prediction is definitely needed as a mitigation action to minimize the risk of drought. However, a sparse dataset has led to difficulties in accurately predicting future droughts in areas without meteorological stations, and hence a dataset with a finer resolution is required. This research investigates the performance of a 3-month Standardized Precipitation Index (SPI) derived from the Tropical Rainfall Measuring Mission (TRMM) and Modern-Era Retrospective analysis for Research and Applications (MERRA-2) to predict drought. CART and Random Forest are applied as the classification methods. Using several predictors, the analysis finds that CART has lower predictability than Random Forest. The average accuracy of the prediction using Random Forest reaches 100% with an average Area Under Curve (AUC) of about 0.8. The analysis also shows that predictions using the MERRA-2 dataset lead to higher accuracy and AUC than those using the TRMM. Therefore, using the MERRA-2 dataset predicted by Random Forest can be an optimal way to predict drought in East Nusa Tenggara. The methods confirmed that average soil surface temperature (day and night), Multivariate ENSO Index (MEI), Arctic Oscillation Index (AOI) and Normalized Difference Vegetation Index (NDVI) are strong predictors of drought. The performance of CART and Random Forest is improved with the Synthetic Minority Over-Sampling Technique (SMOTE). The techniques described: • translate drought information and predictors of drought into a base classifier that optimizes the AUC; • allow drought to be predicted for many grid points efficiently and with high accuracy; and • are computationally efficient and easy to implement.