An Evolution-Unet-ConvNeXt approach based on feature fusion for enhancing the accuracy of short-term precipitation forecasting

Abstract

Accurate short-term convective weather prediction is crucial for mitigating the impact of natural disasters. Although radar echo extrapolation is a commonly employed forecasting method, traditional optical flow-based approaches face computational accuracy challenges when dealing with rapidly changing weather systems. Additionally, some deep learning models experience degradation in prediction accuracy due to blurring effects over extended forecast periods. In this study, we proposed Evolution-Unet-ConvNeXt, a novel deep learning model that effectively addresses these limitations by incorporating feature fusion of latent physical information. The experiments conducted on MeteoNet dataset have demonstrated a significant enhancement in prediction accuracy and reduction of blurring effects when dealing with complex convective phenomena using this model. The Evolution module successfully extracted both motion and intensity fields from radar images, while the recently developed Unet-ConvNeXt network enhanced the efficiency of feature extraction and processing workflow. Quantitative evaluations indicated that our model achieved substantial improvements in meteorological metrics, such as Critical Success Index (CSI) and Probability of Detection (POD), as well as in image clarity and spatial structure metrics like Tenengrad (TEN) and Structure Similarity Index Measure (SSIM), compared to existing baseline model architectures. Furthermore, qualitative analysis of specific rainfall events demonstrated that the robust predictive capability of this model for the intricated nonlinear dynamics of intense echo weather systems.