# AFTA-Net: Axial Fusion and Triaxial Factorised Attention Network for Nowcasting of Severe Convective Weather

## TL;DR

AFTA-Net is a new deep learning model that improves short-term forecasting of severe weather by better capturing motion and intensity changes in radar data.

## Contribution

AFTA-Net introduces a novel encoder-decoder architecture with axial fusion and triaxial attention to better handle non-linear weather patterns.

## Key findings

- AFTA-Net outperforms existing models in forecasting severe convective weather.
- The model achieves a CSI of 0.2506 and HSS of 0.3430 at the 30 dBZ threshold.
- The AFB and TAFA components improve feature decoupling and noise suppression.

## Abstract

Radar echo extrapolation is a core technique for 0–2 h nowcasting, yet existing deep learning models often struggle with non-linear atmospheric motion and intensity attenuation due to insufficient feature decoupling. To address these limitations, this paper proposes AFTA-Net, a novel encoder–decoder architecture. The model introduces an Axial Fusion Block (AFB) that employs a parallel decomposition strategy to explicitly separate temporal evolution from spatial morphology, preserving structural integrity while capturing motion trends. Furthermore, a Tri-Axis Factorized Attention (TAFA) mechanism is designed to sequentially recalibrate feature representations across Time, Channel, and Spatial dimensions, thereby enhancing sensitivity to high-frequency convective signals and suppressing background noise. Extensive experiments on the Jiangsu radar dataset demonstrate that AFTA-Net significantly outperforms representative baselines. Notably, at the critical 30 dBZ threshold for severe weather, the model achieves a CSI of 0.2506 and an HSS of 0.3430.