The accurate prediction of extreme local precipitation is critical for effective coastal flood hazard modeling and risk mitigation, especially in topographically complex regions where coarse-scale atmospheric models fail to capture orographic effects. This research introduces a novel application of the Random Forest (RF) Machine Learning algorithm to statistically downscale daily precipitation data for the humid, mountainous coastal region of the Southern Caspian Sea (Gilan and Mazandaran provinces, Northern Iran).

The methodology utilized high-resolution GPM-IMERG satellite data as the ground truth, trained against large-scale dynamic atmospheric predictors derived from ERA5 reanalysis data across multiple pressure levels (e.g., 500 hPa vorticity, MSLP, and high-resolution topography). The RF model was rigorously validated against independent station data and compared against traditional methods (SVM, KNN, GBR). The results demonstrated the superior performance of the RF framework, achieving an R^2 above 0.81 during unseen test periods against GPM data and significantly outperforming coarser ERA5 outputs at validation stations .

The model’s ability to accurately resolve local intensity, particularly by prioritizing variables related to moisture convergence from the sea and orographic forcing, confirms its strength in capturing the physical mechanisms leading to heavy coastal rainfall events. This downscaled, high-resolution precipitation product directly addresses a key uncertainty in Marine Hazard assessments. The developed AI-driven framework provides a robust, supplementary tool for real-time operational forecasting, enabling water resource managers and civil engineers to enhance the accuracy of coastal flood warning systems and improve the resilience of coastal infrastructure.