Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (22): 297-304.DOI: 10.3778/j.issn.1002-8331.2104-0106

• Engineering and Applications • Previous Articles     Next Articles

Prediction of Strong Cumulus Clouds by Improved U-Net Network and Self-Attention Mechanism

ZHAI Chenfei, DONG Wenhan, ZHANG Xiaomin, LI Dadong, CHEN Xiaojun   

  1. 1.Air Force Engineering University, Xi’an 710038, China
    2.China Flight Test Academy, Xi’an 710089, China
    3.Unit 94816 of the Chinese People’s Liberation Army
    4.Unit 93135 of the Chinese People’s Liberation Army
  • Online:2022-11-15 Published:2022-11-15



  1. 1.空军工程大学,西安 710038
    2.中国飞行试验研究院,西安 710089

Abstract: Aiming at the problem that the traditional method of radar echo extrapolation of spatio-temporal series is easy to lose the information of low level and strong echo region, a new prediction model of GC-ResUNet network is proposed, which is suitable for the input of feature images of any size. The main framework of the model uses U-Net neural network to solve the problem of information loss at the lower level, and introduces GCNet self-attention mechanism to solve the problem of feature loss in the strong echo region. Taking the 2018—2020 coastal radar echo puzzle as the data sample, and the critical success index, detection rate and false alarm rate as the evaluation criteria, the experiment is carried out. The simulation results show that compared with the traditional optical flow method, the prediction success rate of the model for medium and low intensity echoes within 1 h in the future is improved by about 20%, and the prediction success rate for strong echoes is improved by 33%~70%.

Key words: GC-ResUNet, radar echo extrapolation, self-attention mechanism, ice cloud prediction, spatio-temporal series prediction

摘要: 针对传统时空序列的雷达回波外推方法易出现低层信息及强回波区信息丢失的问题,提出了一种适用于任意尺寸特征图输入的新型GC-ResUNet网络预测模型。模型主框架采用U-Net神经网络解决了低层信息丢失的问题,同时引入GCNet自注意力机制解决了强回波区特征丢失的问题。以2018—2020年间沿海雷达回波拼图为数据样本,以临界成功指数、探测率、虚警率为评价标准进行实验。仿真结果表明,该模型对于未来1?h内的中低强度回波的预测成功率相比于传统光流法提升20%左右,对于强回波的预测成功率提升33%~70%。

关键词: GC-ResUNet, 雷达回波外推, 自注意力机制, 积冰云层预测, 时空序列预测