Accurately resolving the nonlinear interactions between orographic and biophysical surface forcings in regions of complex topography often surpasses the resolving capacity of conventional observations. The southern coast of the Caspian Sea, characterized by pronounced topographic and land-cover heterogeneity, requires high-resolution reconstruction of steep precipitation gradients to resolve nonlinear atmosphere–surface dynamics adequately. In this study, four machine learning (ML) algorithms—Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Decision Tree (DT)—were employed to downscale CHIRPS annual precipitation data to a 1-km horizontal resolution for the 2001–2025 period. We evaluated the relative importance of predictor variables, including elevation, the Normalized Difference Vegetation Index (NDVI), and land use, using SHapley Additive exPlanations (SHAP). The analyses demonstrate a robust hydrological coupling (r = 0.78) between spatial variations in precipitation (203 to 1536 mm) and latent heat flux (LE). The ensemble averaging approach applied within the RF framework achieved the highest predictive accuracy and physical consistency, achieving a coefficient of determination (R2) of 0.77, a mean absolute error (MAE) of 146.02 mm, and a mean bias error (MBE) of −10.39 mm. In contrast, the SVM model exhibited significant over-smoothing, while prominent blocky spatial artifacts hampered the DT algorithm. SHAP analysis identified elevation as the dominant driver of spatial precipitation variability (absolute SHAP values from 229 to 242), followed by vegetation (representing surface biophysics) as the secondary forcing mechanism; conversely, land use had a negligible marginal influence. These findings suggest that the vegetation cover along the southern Caspian coast functions not merely as a static surface feature, but as an active thermodynamic boundary condition that modifies aerodynamic roughness and modulates moisture fluxes into the atmospheric boundary layer (ABL). Finally, the persistent, systematic negative bias across all evaluated ML models highlights the inherent physical and structural limitations of the coarse parent dataset in capturing complex, sub-grid-scale atmospheric processes.