Accurate precipitation estimation is crucial for hydrological and meteorological applications. IMERG-Late, a widely used global satellite-based precipitation product, exhibits biases in China due to complex topography and climate variability. This study proposes a U-Net-based deep learning framework for calibrating IMERG-Late precipitation estimates using CMPA as the ground truth. The model effectively learns spatial and temporal patterns to reduce systematic errors and improve precipitation consistency. Comparative experiments demonstrate that U-Net outperforms traditional interpolation, statistical correction, and other deep learning methods in enhancing precipitation accuracy. While the approach significantly improves satellite-based precipitation estimates, challenges such as computational costs and generalization in extreme weather events remain. Future work will explore hybrid deep learning-physical model approaches and Transformer-based architectures to further enhance precipitation calibration.