基于频谱分解和注意力机制的雷达回波外推算法

针对雷达回波数据，首次提出使用频谱分解的方式预先提取多尺度回波信息作为深度学习网络的输入数据，以丰富输入数据的特征信息。模型主框架采用UNet神经网络，并且针对模型退化问题，加入了残差连接结构。基于以上设计，提出SpAt-ResUNet预测模型。试验结果表明，相比于传统雷达回波外推算法SPROG与深度学习网络ResUNet模型，该模型对于未来1 h的雷达回波预测长时间外推图像模糊的问题以及强回波留存能力分别得到改善和增强。

Radar Echo Extrapolation Algorithm Based on Spectrum Decomposition and Spatiotemporal Attention Mechanism

For the radar echo data, the method of spectral decomposition is proposed for the first time to extract multi-scale echo information in advance as the input data of the depth learning network to enrich the feature information of the input data. The main frame of the model adopts UNet neural network, and the residual connection structure is added to solve the problem of model degradation. Based on the above design, SpAt ResUNet prediction model is proposed. The test results show that compared with the traditional radar echo extrapolation method SPROG and the depth learning network RESUNet model, the model has improved and enhanced the problem of long-term extrapolation image blurring and strong echo retention ability for the radar echo prediction in the next hour.