Enhancing satellite-based rainfall estimation in arid regions through AI-driven data fusion

Accurate precipitation estimation is crucial for hydrological modeling and water resource management, especially in arid and data-scarce regions like the UAE. This study presents an AI-driven data fusion framework to enhance satellite-based rainfall estimates by integrating three widely used satellite precipitation products, including CMORPH, IMERG, and GSMaP, with ground-based observations. Using rainfall data from 38 gauge stations across the UAE, four machine learning (ML) models, including Neural Network, Single Layer Perceptron, Support Vector Machine, and Logistic Regression, were employed to merge the satellite datasets. A stacked ensemble approach was further implemented to consolidate the strengths of individual models and produce a refined precipitation product. Model evaluation involved statistical measures (e.g., RMSE, MAE, Correlation Coefficient), categorical indices (POD, FAR, CSI), and extreme climate indices (Rx1day and R20mm). The ensemble model demonstrated superior performance, improving CC to 0.89, and reducing RMSE and MAE by 33% and 40%, respectively. The model also captured extreme rainfall events more reliably, with improvements of 25% for Rx1day and 20% for R20mm. Feature importance analysis identified surface shortwave irradiance and minimum temperature as key predictors of rainfall variability. This research highlights the potential of integrating AI with large-scale satellite and climatic datasets to refine precipitation estimates. The proposed approach is scalable and transferable to other arid regions, offering valuable applications in flood forecasting, drought monitoring, and climate resilience planning.