Abstract
A lightweight rainfall nowcasting model is required by Sichuan provincial meteorological bureaus. Deep learning methods such as recurrent, convolutional, and Transformer models have been applied to precipitation prediction. However, recurrent models struggle with suboptimal parallel computational efficiency and error accumulation, convolutional models face challenges in capturing long-range dependencies, and Transformer models are limited by their quadratic time complexity. The Spatial-Temporal Vision Mamba (STVMamba) is proposed, a novel spatiotemporal prediction model specifically designed for precipitation nowcasting. STVMamba achieves high parallel computational efficiency, operates with linear time complexity, and excels at modeling long-range dependencies, overcoming the shortcomings of previous deep learning methods. Specifically, the STVMamba utilises a Spatial-Temporal Selective Scan (STSS) module to capture global spatiotemporal relationships, while a Spatial-Temporal Depthwise Separable Convolution (STDSConv) module is used to learn local spatiotemporal relationships. Furthermore, the two-tier architecture of the STVMamba explicitly learns spatiotemporal relationships across both small and large spatial scales within meteorological data. We evaluate STVMamba on three benchmark datasets: the Sichuan radar echo dataset, the HKO-7 radar echo dataset, and the satellite-based IMERG dataset. On the Sichuan dataset, STVMamba achieves state-of-the-art performance across MSE, SSIM, and CSI-10 metrics. On the HKO-7 dataset, it outperforms existing models in terms of MSE, CSI-10, and CSI-20. On the IMERG dataset, it achieves superior results in SSIM and CSI-0.5, demonstrating its robustness and effectiveness across diverse geoclimatic conditions and data modalities. The code is available at https://github.com/CUITMIR/STVMamba .