Machine Learning Models for Estimating High Rainfall Potential Using OMI Atmospheric Features as an Indirect Flood Predictor in Pidie Jaya

Abstract

Pidie Jaya and the upstream mountainous area of Peuet Sagoe/Ulu Masen, Aceh, are frequently affected by flooding from overflow of the Krueng Meureudu River following intense rainfall upstream. This study uses Cloud Optical Thickness (COT) as a proxy for high rainfall potential, enabling indirect flood prediction. Four machine learning models, Linear Regression (LR), Random Forest Regressor (RFR), XGBoost Regressor, and Support Vector Regression (SVR), were developed using NASA OMI-Aura Level-2 OMUVB observations (2019–2026), including UV index, erythemal dose, irradiance, tropospheric ozone column, surface albedo, and spatiotemporal variables. Two feature scenarios were evaluated: one including Lamberti an Equivalent Reflectivity (LER) and one excluding it. The LER-inclusive scenario achieved high predictive performance (R² up to 0.986) but showed potential data leakage. In contrast, the LER-excluded scenario yielded much lower R² values (0.026–0.169), in dicating that independent OMI atmospheric features provide limited predictive capability. The most influential predictors were OMTO3ColumnAmountO3, CSErythemalDailyDose, CSIrradiance310/380, and Latitude. Analysis of 15 historical flood events revealed tha t most were preceded by high-to-very-high COT percentiles (>P75 –>P90). These results highlight the importance of integrating actual rainfall observations to improve flood early-warning systems in data-scarce regions.