基于LSTM的地基微波辐射计湿度廓线反演  

Humidity Profile Inversion of Ground Based Microwave Radiometer Based on LSTM

作　　者：周高进 杨智鹏[1,2] 彭静[1] 杨玲[1] 陶法 茆佳佳[3] ZHOU Gaojin;YANG Zhipeng;PENG jing;YANG Ling;TAO Fa;MAO Jiajia(Chengdu University of Information Engineering,College of Electronic Engineering,Chengdu 610000;Key Laboratory for Cloud Physical Environment of China Meteorological Administration,Beijing 100088;Meteorological Observation Center of China Meteorological Administration,Beijing 100081)

机构地区：[1]成都信息工程大学电子工程学院,成都610000 [2]中国气象局云雾物理环境重点开放实验室,北京100088 [3]中国气象局气象探测中心,北京100081

出　　处：《气象科技》2022年第1期21-29,共9页Meteorological Science and Technology

基　　金：中国气象局云雾物理环境重点开放实验室开放课题(2019Z01610);四川省科技计划项目(2019YFG0496,2020YFG0143)资助。

摘　　要：大气湿度廓线对于研究大气系统的复杂性具有十分重要的作用。地基微波辐射计有着连续观测的特性,能够以高时间分辨率反演出高度至10 km的大气湿度廓线,廓线数据对于气象预报和研究气候系统的变化至关重要。为了提高反演大气湿度廓线的精准度,本文使用时间循环神经网络模型,利用微波辐射计连续探测的信号并使用Ka波段毫米波云雷达数据提高有云时的反演精度,采用LSTM(Long Short-Term Memory)神经网络反演计算大气湿度廓线,用探空仪实测相对湿度验证并分析反演效果。该模型反演的湿度廓线与探空廓线的平均绝对误差为9.80%,均方根误差为13.85%,优于经典的BP(Back Propagation)神经网络模型平均绝对误差11.52%,均方根误差15.66%。通过比较,证明了本文的反演模型利用连续观测的亮温数据能够有效地提高反演精度,特别是对于3~7 km范围内大气湿度廓线分布较为复杂的相对湿度的反演。并验证了该模型加入云观测数据提高了有云时的反演精度。The atmospheric humidity profile is a vital factor in studying the complexity of the atmospheric system. The ground-based microwave radiometer(MWR) can continuously observe and retrieve atmospheric humidity profiles up to 10 km with high temporal resolution. These profiles are essential for understanding the changes in the climate system. In order to improve the accuracy of retrieving the atmospheric humidity profile by MWR, this paper uses a time loop neural network model that uses the continuous detected signals of microwave radiometers. Moreover, Ka-band millimetre-wave cloud radar data is employed to improve the inversion accuracy for cloudy data. LSTM neural network is applied as the inversion method to retrieve atmospheric humidity profile, while radiosonde measures relative humidity as truth-value to verify and analyze the inversion effect. This research has also conducted a detailed comparison with classical inversion methods(BP and support vector machine). The average absolute error of the humidity profile and the sounding profile is 9.80%, the root mean square error is 13.85%, and the BP neural network model’s average absolute error is 11.52%. The root mean square error is 15.66%. The comparison proves that the method using temporal information could effectively improve the inversion accuracy, especially for the inversion of relative humidity in the range of 3 to 7 km, where the atmospheric humidity profile distribution is more complicated.

关 键 词：微波辐射计 循环神经网络 大气湿度廓线 云雷达 

分 类 号：P412.23[天文地球—大气科学及气象学]