PrecipFormer: Efficient Transformer for Precipitation Downscaling

Precipitation downscaling, which enhances the spatial resolution of gridded precipitation data, remains a critical challenge in climate modeling and hydrological applications. While Vision Transformers (ViTs) have demonstrated remarkable success in various computer vision tasks through their ability to capture long-range dependencies, their application to precipitation downscaling remains largely unexplored due to computational constraints and the challenge of effectively modeling both local and global precipitation patterns. This paper introduces PrecipFormer, a computationally efficient transformer architecture specifically designed for precipitation downscaling. Our model builds upon the Low-to-High Multi-Level Vision Transformer (LMLT) mechanism, enabling parallel processing of features at multiple spatial scales while significantly reducing computational overhead. We enhance the architecture with a Convolutional Block Attention Module (CBAM) in the shallow feature extractor to adaptively focus on critical spatial regions. Through extensive experiments, we demonstrate that the proposed PrecipFormer achieves superior performance compared to state-of-the-art baselines.