MDA-MIM: a radar echo map prediction model integrating multi-scale feature fusion and dual attention mechanism

Abstract

To obtain high-quality spatiotemporal features from radar echo maps, an improved MIM (memory in memory) model, MDA-MIM (multi-scale feature fusion and dual attention mechanism MIM) was proposed. Multi-scale feature fusion and a dual attention mechanism were incorporated in MDA-MIM. Dilated convolution was used to extract and integrate multi-scale features. To better capture the non-stationary characteristics of radar echo data, a self-attention mechanism was introduced into the non-stationary module of the MIM model, dynamically adjusting the weights of different time steps and spatial positions. Meanwhile, a local attention mechanism was incorporated into the stationary module, enabling the model to focus on feature correlations within local regions and enhance its ability to extract stationary features. Experiments conducted on real-world datasets demonstrate that MDA-MIM achieves state-of-the-art predictive performance, consistently outperforming baseline models across multiple evaluation metrics, including MSE, MAE, SSIM, and PSNR.

Keywords

• radar echo map
• spatial-temporal prediction
• attention mechanism
• multi-scale feature