Enhancing Spatial Resolution of GPM Rainfall Data in Upper Cauvery Basin, India: Machine Learning Approach

Spatial downscaling is an effective way to obtain rainfall with sufficient spatial details. The spatial resolution of the Global Precipitation Measurement (GPM) mission (IMERG) satellite rainfall products is $0.1^{\circ} \times 0.1^{\circ}$, which is too coarse for regional-scale analysis. This study employed two Model averaging methods (Random Forest + XGBoost, Random Forest + CatBoost), Ensemble methods (Random Forest, XGBoost, CatBoost) and Stacked Random Forest + XGBoost model for downscaling GPM IMERG monthly rainfall over the Upper Cauvery Basin from 2015 to 2020 from $0.1^{\circ}(\sim 10$ km) to 1 km resolution. Five land surface variables (auxiliary variables), NDVI, elevation, LST, slope, and aspect, were employed for this purpose. The stacked RFR + XGB model outperformed the model averaging techniques, achieving a higher $\mathrm{R}^{2}$ of 0.694 and a lower RMSE/MAE of 44.57/35.23. While the ensemble method yielded promising results, it struggled to predict extreme rainfall values. The downscaled dataset facilitates improved hydrological applications, including water footprint analysis, hydrological monitoring, and disaster warning systems.