关吉平,高延波,黄从雷,李济,张立凤. 2024. 基于Attention-Unet网络的FY-4A卫星降水估计. 气象学报,82(3):398-410. DOI: 10.11676/qxxb2024.20230037
引用本文: 关吉平,高延波,黄从雷,李济,张立凤. 2024. 基于Attention-Unet网络的FY-4A卫星降水估计. 气象学报,82(3):398-410. DOI: 10.11676/qxxb2024.20230037
Guan Jiping, Gao Yanbo, Huang Conglei, Li Ji, Zhang Lifeng. 2024. Attention-Unet based FY-4A satellite precipitation estimation. Acta Meteorologica Sinica, 82(3):398-410. DOI: 10.11676/qxxb2024.20230037
Citation: Guan Jiping, Gao Yanbo, Huang Conglei, Li Ji, Zhang Lifeng. 2024. Attention-Unet based FY-4A satellite precipitation estimation. Acta Meteorologica Sinica, 82(3):398-410. DOI: 10.11676/qxxb2024.20230037

基于Attention-Unet网络的FY-4A卫星降水估计

Attention-Unet based FY-4A satellite precipitation estimation

  • 摘要: 准确的卫星降水估计是开展天气、气候、水文、生态等研究的重要基础,并且可以降低由降水直接导致的洪水等自然灾害造成的损失。目前业务运行的卫星降水产品主要使用物理反演方法,存在反演过程中降水特征信息较为片面等缺点。近年来随着深度学习技术不断发展,其挖掘隐藏特征信息的能力也逐渐被引入到各种非线性过程研究。以Attention-Unet为核心搭建具备处理卫星多通道数据能力的卫星降水估计网络框架,利用风云4A 卫星(FY-4A)多通道扫描辐射计9—14通道数据构建数据集进行降水估计模型训练。为评估该模型的效果,将Attention-Unet模型的降水估计结果与业务运行的卫星降水产品以及其他成熟深度学习网络模型进行对比。结果表明,Attention-Unet模型的降水估计效果优于使用传统物理反演方法的卫星降水产品FY4A-QPE和CMORPH,也优于作为对比的Unet模型和PERSIANN-CNN模型。在FY-4A多通道数据基础上,在模型训练中加入与降水有较大相关的地形数据,结果表明模型在保持降水区域识别能力的基础上降水量估计误差更小。

     

    Abstract: Accurate satellite-based precipitation estimation is important for research on weather, climate, hydrology and ecology, and can reduce the loss caused by natural disasters such as floods that are directly caused by precipitation. At present, operational satellite-based precipitation products are mainly obtained by various physical retrieval algorithms, which may miss certain hidden precipitation-related features. In recent years, with the development of deep learning, its ability to mine hidden features has been gradually introduced into the research of various nonlinear processes. In this paper, we propose a deep learning model named Attention-Unet for satellite-based precipitation estimation. The high spatiotemporal resolution spectral data of the Fengyun-4A satellite (FY-4A) are used to train the model. To evaluate the effectiveness of the proposed model, we compare model outputs with operational satellite-based precipitation products and products of other mature deep-learning models. Statistics and visualizations of the experimental results show that the proposed model performs better than the operational satellite-based precipitation retrieval algorithms and other mature deep learning models in both precipitation identification and precipitation amounts estimation. The experimental results show that the model has the potential to be applied to meteorological work. Furthermore, based on FY-4A multi-channel data, we add topography data to the training datasets. The experimental results show that after adding topography data to training datasets, the model performs better than before in the estimation of precipitation amount, especially in the mountainous region. The above results shows that Attention-Unet has a great application potential and prospect in satellite precipitation estimation.

     

/

返回文章
返回