Estimating precipitation using traditional methods is usually associated with considerable errors, particularly during convective and heavy rainfall events. To enhance the accuracy of precipitation estimation, a deep learning framework integrating satellite observations and reanalysis data was employed. In this study, an attention based U shaped model was developed, in which the inputs consisted of multi channel hourly brightness temperatures from second generation Meteosat satellites, and the outputs were fifth generation land surface reanalysis data (ERA5 Land) for the period 2019–2022, with an hourly temporal resolution and a spatial resolution of 0.1°. To improve model performance, an enhanced adaptive weighted version of the Huber loss function was utilized. By combining the advantages of mean squared error and mean absolute error, this loss function provides greater robustness against extreme precipitation values. In addition, an attention gating mechanism was applied to facilitate information transfer between the encoder and decoder pathways and to suppress irrelevant features, acting as an intelligent filter that selectively passes features most relevant to precipitation prediction. After training, the model achieved a mean absolute error of 0.39 mm, a root mean square error of 0.862 mm, and a correlation coefficient of 0.80, demonstrating strong spatial consistency with ERA5 Land data. This study was conducted over the geographical domain bounded by 28.5°–35.5° N latitude and 52°–59° E longitude. Overall, the results indicate that the proposed multi source data integration approach combined with the developed attention based U shaped architecture can serve as an effective tool for precipitation monitoring and estimation, particularly in regions with sparse or insufficient rain gauge networks.