基于深度学习的葵花8卫星资料反演降水

Deep-Learning Based Inversion of Precipitation from Himawari-8 Satellite Data

摘要: 本文构建了2016至2019年Himawari-8静止卫星上9个红外通道亮温数据和GPM IMERG半小时平均降水量的机器学习训练集，然后将其划分为训练集(共5352个样本)、验证集(1784)和测试集(1784)。首先，选用目前深度学习技术中使用最为广泛、稳定可靠的U-Net模型，分析了Himawari-8卫星的不同红外通道对降水估计的贡献，比较了9个单通道和多通道的反演，结果表明采用通道13能获得相对更好的反演效果。随后，本文比较了多个单时刻以及多时刻组合的降水反演精度，发现采用多时刻观测输入，并未能进一步改进反演精度。最后，为进一步提升降水反演效果，本文比较了三个模型的反演精度，包括U-Net、pix2pixGAN和ConvMixer等，进行训练和降水反演试验。结果表明，pix2pixGAN反演的降水分布最优，它比其他模型结果更具有鲜明详实的结构。

Abstract: A machine learning training set of nine infrared (IR) channels of Himawari-8 and GPM-IMERG half-hourly mean precipitation is constructed, and it is divided into a training set (5352 samples in total), a validation set (1784), and a test set (1784). The U-Net model, the most widely used and stable deep learning technique, is chosen to analyze the contribution of different IR channels of the Himawari-8 satellite to the precipitation estimation. The retrieval results of single and multiple channels are compared, and it is found that the Channel 13 yields better results. The accuracy of model outputs with the input of single-moment observation and the input of multiple-moment combinations is compared, and it is found that the latter did not improve the accuracy of the inversions. To further improve the precipitation prediction, the accuracy of three different DL models is compared, including U-Net, pix2pixGAN and ConvMixer. The results show that pix2pixGAN is the best model.