FY-3E卫星全球降水和降雪数据集（2022-2024）

FY-3E卫星上的被动微波辐射计通过接收到的向上辐射来反演降水情况，接收到的信号包含了地表、大气和水凝物的混合信号。液态降水粒子会在较低频率下增加来自地表的向上辐射，而冰颗粒则会在较高频率下减少向上辐射。反演算法的任务是识别降水相态并将降水信号与地表信号分离。这一数据集基于机器学习方法，通过FY-3E微波湿度计探测器和微波温度计探测器的观测结果来反演降雨和降雪。机器学习算法包括两部分，自组织映射（SOM）用于对降水和底层地表类型进行分类，随后使用人工神经网络（ANN）将亮度温度与从SOM得出的聚类中的降水率联系起来。使用全球降水测量任务中的多源降水融合产品（IMERG）训练ANN。为了解决强降水样本不足的问题，利用RTTOV辐射传输模拟作为大雨样本的补充。在降雨和降雪检索方面，本数据集所使用的SOM-ANN算法的表现优于IMERG和戈达德剖面算法（GPROF）的反演产品。与两年期间约4400个地面雨量计的小时观测结果相比，SOM-ANN的均方根误差分别为降雨和降雪率的1.06毫米/小时和0.34毫米/小时，优于IMERG（分别为1.23毫米/小时和0.42毫米/小时）和GPROF（分别为1.22毫米/小时和降雪率为百分之第四十九）。SOM-ANN算法及评估结果发表在了 https://doi.org/10.1029/2024JD040731

The passive microwave radiometer on the FY-3E satellite retrieves precipitation by detecting upwelling radiation. The received signals consist of mixed contributions from the Earth's surface, atmosphere, and hydrometeors. Liquid precipitation particles enhance upwelling radiation from the surface at lower frequencies, while ice particles reduce upwelling radiation at higher frequencies. The task of the retrieval algorithm is to identify precipitation phases and separate precipitation signals from surface signals. This dataset is developed using machine learning methods to retrieve rainfall and snowfall based on observations from the FY-3E microwave humidity sounder and microwave temperature sounder. The machine learning framework comprises two parts: the Self-Organizing Map (SOM) is employed to classify precipitation and underlying surface types, followed by an Artificial Neural Network (ANN) that correlates brightness temperatures with precipitation rates within the clusters derived from the SOM. The ANN is trained using the multi-source merged precipitation product (IMERG) from the Global Precipitation Measurement (GPM) mission. To address the shortage of heavy precipitation samples, RTTOV radiative transfer simulations are utilized as supplementary data for heavy rainfall cases. In terms of rainfall and snowfall retrieval, the SOM-ANN algorithm used in this dataset outperforms the retrieval products of both IMERG and the Goddard Profiling Algorithm (GPROF). Compared with hourly observations from approximately 4400 ground rain gauges over a two-year period, the root mean square errors (RMSE) of SOM-ANN for rainfall rate and snowfall rate are 1.06 mm/h and 0.34 mm/h, respectively. This outperforms IMERG (1.23 mm/h and 0.42 mm/h, respectively) and GPROF (1.22 mm/h and 49% for snowfall rate, respectively). The SOM-ANN algorithm and its evaluation results have been published at https://doi.org/10.1029/2024JD040731

搜集汇总

背景与挑战

背景概述

该数据集是基于FY-3E卫星被动微波观测的全球降雨和降雪数据，覆盖2022年至2024年，采用机器学习方法（SOM-ANN算法）反演降水率，在精度上优于IMERG和GPROF产品。数据具有小时级时间分辨率和10km-100km空间分辨率，以开放获取方式共享，适用于全球降水研究和气象分析。