KM-UNet: A KAN-Mamba Hybrid Network with Direction-Sensitive Attention and Multi-Scale Fusion for Cloud Mask Nowcasting

Cloud mask nowcasting is a critical yet challenging task in meteorological prediction, which is the prerequisite for precipitation forecasting. Although existing methods have considerably improved prediction accuracy for extensive regions of thin clouds, these methods struggle to maintain comparable accuracy for localized, small-scale, optically thick cloud regions. To address this challenge, We propose a deep learning model called KM-UNet based on the KAN-Mamba hybrid network for heavy cloud mask nowcasting. Specifically, we adapt the KAN Convolution module to enhance the local feature extraction capability on local heavy cloud areas, and introduced the three-dimensional Enhanced ViM module for joint modeling of spatial and channel information to capture the global features on large-scale thin cloud areas. Additionally, we propose the Multi-Scale Fusion and Hybrid Loss function to optimize feature learning. Extensive experiments on two real-world meteorological datasets shown that our model outperforms existing models.