The meteorological radar reflectivity data, also known as echo, plays a
crucial role in predicting precipitation and enabling accurate and fast
forecasting of short-term heavy rainfall without the need for complex Numerical
Weather Prediction (NWP) model. Compared to conventional model, Deep Learning
(DL)-based radar echo extrapolation algorithms are more effective and
efficient. However, the development of highly reliable and generalized
algorithms is hindered by three main bottlenecks: cumulative error spreading,
imprecise representation of sparse echo distribution, and inaccurate
description of non-stationary motion process. To address these issues, this
paper presents a novel radar echo extrapolation algorithm that utilizes
temporal-spatial correlation features and the Transformer technology. The
algorithm extracts features from multi-frame echo images that accurately
represent non-stationary motion processes for precipitation prediction. The
proposed algorithm uses a novel parallel encoder based on Transformer
technology to effectively and automatically extract echoes' temporal-spatial
features. Furthermore, a Multi-level Temporal-Spatial attention mechanism is
adopted to enhance the ability to perceive global-local information and
highlight the task-related feature regions in a lightweight way. The proposed
method's effectiveness has been valided on the classic radar echo extrapolation
task using the real-world dataset. Numerous experiments have further
demonstrated the effectiveness and necessity of various components of the
proposed method.

本文介绍了一种新的雷达回波外推算法，该算法利用时空相关特征和 Transformer 技术，通过从多帧雷达图像中提取特征来准确表示降水的非定常运动过程，增强对全局 - 局部信息的感知能力并以轻量级方式突出与任务相关的特征区域，并在真实世界数据集上验证了其有效性和必要性