Abstract
Accurate estimation of precipitation from remote sensing remains difficult due to systematic biases in existing products, particularly over complex terrain where accurate monitoring is key to water resource management. We present a comprehensive framework to improve remote sensing precipitation estimation using attention-based deep learning, along with systematic validation in temporal, spatial, and operational dimensions., In this experimental setup, we inter-compared three major precipitation products (CHIRPS, TerraClimate, ERA5-Land) with 62 ground stations across Iran's topographically diverse Golestan Province for 11 years (2014-2024). ERA5-Land performed best at baseline (R²= 0.437) and was selected to be enhanced by machine learning approaches. Our framework underwent systematic stages: traditional machine learning (Random Forest, XGBoost), attention-based deep learning (TabNet), exhaustive temporal verification, and operational raster execution., TabNet attention networks achieved significant gains over ERA5-Land baseline (R²= 0.801 vs 0.437), with superior spatial generalization to classical approaches. Temporal holdout validation on independent 2023-2024 data resulted in diverging model behaviors: XGBoost exhibited temporal performance deterioration in spite of excellent training performance, whereas TabNet preserved more stable generalization powers. Distance-based evaluation on 20 independent test stations suggested that model performance is a function of distance to training locations, with TabNet exhibiting superior transferability at distances greater than 10 km from training sites., The system …