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

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Existing spatiotemporal forecasting methods struggle with detail loss and limited multi-scale pattern integration.
• Deep semantic encoding often sacrifices low-level textures and edge details, crucial for high-contrast regions.
• Single-scale feature maps hinder the capture and integration of diverse spatial patterns.

Purpose of the Study:

• To introduce M-Net, a novel spatiotemporal prediction model.
• To enhance structural representation and capture multi-scale spatial patterns in forecasting.
• To improve prediction accuracy by fusing multi-level information.

Main Methods:

• Multiscale Receptive Pyramid (MRP) for multi-scale input representations.
• Dual Natural Patch Attention (DPA) to bridge semantic gaps between feature scales.
• Dual-Axis Attention (DAA) to refine deep features and enhance perceptual sensitivity.
• Multiscale Hierarchical Fusion (MHF) for combining shallow and deep prediction advantages.

Main Results:

• M-Net achieves state-of-the-art performance on 80% of evaluation metrics across ten benchmark datasets.
• The model effectively preserves low-level textures and edge details.
• Improved capture and integration of multi-scale spatial patterns were observed.

Conclusions:

• M-Net offers a significant advancement in spatiotemporal forecasting.
• The proposed architecture effectively addresses limitations of existing methods.
• The model demonstrates superior performance in detail preservation and pattern integration.