计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (11): 285-293.DOI: 10.3778/j.issn.1002-8331.2202-0255

• 工程与应用 • 上一篇    下一篇

改进U-Net的风云四号卫星降水估计算法研究

黄杰,张永宏,马光义,朱灵龙,田伟   

  1. 1.南京信息工程大学 自动化学院,南京 210044
    2.南京信息工程大学 气象灾害预报预警与评估协同创新中心,南京 210044
    3.南京信息工程大学 电子与信息工程学院,南京 210044
    4.南京信息工程大学 计算机与软件学院,南京 210044
  • 出版日期:2023-06-01 发布日期:2023-06-01

Research on Precipitation Estimation Algorithm from Fengyun-4 Satellite Based on Improved U-Net

HUANG Jie, ZHANG Yonghong, MA Guangyi, ZHU Linglong, TIAN Wei   

  1. 1.School of Automation, Nanjing University of Information Science and Technology, Nanjing 210044, China
    2.Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science and Technology, Nanjing 210044, China
    3.School of Electronics and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China
    4.School of Computer and Software, Nanjing University of Information Science and Technology, Nanjing 210044, China
  • Online:2023-06-01 Published:2023-06-01

摘要: 针对强对流天气条件下利用卫星图像进行降水量估计时精度不高、时空分辨率低的问题,提出了一种改进U-Net的降水量估计算法。将U-Net模型的编码器通过残差模块与解码器相结合,使得模型参数可以共享,避免深层网络模型梯度消失的情况。在此结构基础上引入了空间金字塔模块进行多尺度特征提取,保留更多的图像特征,加强对细小降水云团信息的特征提取能力;引入了注意力机制模块,提取重要降水特征信息。实验结果表明,该算法在命中率、虚警率、相关系数分别为0.84、0.48、0.59;均方根误差、平均绝对误差分别为1.354?mm/h、0.432?mm/h。与PERSIANN-CNN、U-Net算法相比,有效地提升了降水量估计精度。与其他降水产品对比,能更好地识别出降水区。该算法可以实现近实时的降水估计,能够有效地提升降水估计精度,对低时间分辨率的降水估计研究具有一定的价值。

关键词: 降水量估计, 风云四号卫星, 卷积神经网络, 语义分割

Abstract: Aiming at the problems of low accuracy and low spatial-temporal resolution of precipitation estimation using satellite images under the condition of severe convective weather, an improved U-Net precipitation estimation algorithm is proposed. Firstly the encoder of the U-Net model is combined with the decoder through the residual module, so that the model parameters can be shared to avoid the disappearance of deep network model gradients. Based on this structure, the spatial pyramid module is introduced for multi-scale feature extraction to retain more image features and strengthen the feature extraction ability of small precipitation cloud information. The attention mechanism module is added to extract important precipitation feature information. The experimental results show that probability of detection, false alarm ration, critical success index of the proposed algorithm are 0.84, 0.48 and 0.59, respectively. The root mean square error and mean absolute error are 1.354?mm/h and 0.432?mm/h respectively. Compared with PERSIANN-CNN and U-Net, the proposed algorithm effectively improves the accuracy of precipitation estimation. Compared with other precipitation products, it also has certain advantages. Therefore, the algorithm can achieve near-real-time results and effectively improve the accuracy of precipitation estimation, which is valuable for the research of precipitation estimation with low tempotal resolution.

Key words: quantitative precipitation estimation, Fengyun-4 satellite, convolutional neural network, semantic segmentation