SWRVM: Sliding Window Recurrent Vision Mamba Model for Long-Term Spatial-Temporal Prediction

Spatial-temporal prediction, to predict future spatial data for a period with past data, is widely used in precipitation prediction, target motion prediction, and traffic flow forecasting. The data for these tasks typically exhibits multi-scale variability which imposes a great deal of difficulty for long-term prediction. A deep learning model, named sliding window recurrent vision mamba (SWRVM), is proposed for exploiting spatial and long-term temporal information accurately and dexterously to perform effective long-term spatial-temporal prediction in this paper. The proposed SWRVM model combines improved embedding module and sliding window recurrent mechanisms into vision mamba, while the improved embedding module is to retain more spatial and temporal feature, the sliding window recurrent mechanism is the key structure for long-term prediction, and the vision mamba model gives effective results by global receptive field and computational efficiency. We perform extensive experiments on four public datasets and one private dataset in three situations of 10 in 10 out, 20 out, and 40 out. The quantitative and qualitative visualization results demonstrate the SWRVM model outperforms the state-of-the-arts (SOTA) models in multi-scale variations long-term spatial-temporal prediction tasks.