M-Net: Multiscale hierarchical fusion with dual natural patch attention for spatial-Temporal time series forecasting

Spatial-temporal time series forecasting holds significant application value in scenarios such as grid-based weather forecasting and video prediction. However, existing methods often suffer from excessive loss of low-level textures and edge details due to deep semantic encoding, leading to inadequate structural representation in high-contrast regions. Moreover, prediction mechanisms based on single-scale feature maps restrict the model's ability to capture and integrate multi-scale spatial patterns. To address these issues, we propose M-Net, a novel spatiotemporal prediction model that integrates multi-scale perception and multi-level prediction fusion. During preprocessing, the Multiscale Receptive Pyramid (MRP) utilizes a dual-path downsampling strategy to provide multi-scale input representations for different network levels. To mitigate semantic mismatches across levels, the Dual Natural Patch Attention (DPA) treats multi-scale feature maps as semantically explicit natural patches, sequentially modeling channel and spatial dependencies to effectively bridge the semantic gap between features at different scales. Furthermore, the Dual-Axis Attention (DAA) suppresses channel redundancy and spatial noise in deeply encoded features, enhancing the expression of critical channels and increasing perceptual sensitivity to key spatial regions. Finally, the Multiscale Hierarchical Fusion (MHF) leverages the advantages of shallow predictions in transient dynamics and detail reconstruction, alongside deep predictions in long-term evolution and background field modeling, achieving multi-level prediction fusion. Experimental results on ten benchmark datasets show that M-Net achieves state-of-the-art performance on 80% of the evaluation metrics. Our code is available at https://github.com/Dr-BinYang/M-Net.