基于Attention-Unet网络的FY4A卫星降水估计

Attention-Unet-based Fengyun-4A satellite precipitation estimation

中文关键词:  风云4A卫星，Attention Unet，卫星降水估计，地形

英文关键词:Fengyun-4A satellite, Attention Unet, Satellite-based precipitation estimation, Topography

中文摘要:

      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网络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，利用风云四A 卫星多通道扫描辐射计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进行对比。试验结果表明，Attention-Unet模型的降水估计效果要优于使用传统物理反演方法的卫星降水产品FY4A-QPE以及CMORPH，也要优于作为对比的Unet模型和PERSIANN-CNN模型。其次在风云四A多通道数据基础上，在模型训练中加入对降水有较大相关性的地形数据。试验结果表明模型在保持降水区域识别能力的基础上，降水量估计误差更低。

英文摘要:

      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 which are directly caused by precipitation. At present, the operational satellite-based precipitation products mainly originate from the physical retrieval algorithm, which may miss some 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. To train the model we utilize the high temporal, spatial and spectral resolution data of the FY4A satellite. To evaluate the effectiveness of the proposed model, we compare it with operational satellite-based precipitation products and other mature deep-learning models. Statistics and visualizations of the experimental results show proposed model has better performance than the operational satellite-based precipitation 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. Secondly, based on FY4A multi-channel data, we added topographic data into the training datasets. The experimental results show that after adding topographic data into training datasets, the precipitation amounts estimation performed better than before, especially in the mountainous region.