Research on neural network training retrieval based on microwave radiometer observed brightness temperature data set

Abstract

In order to improve the accuracy of the ground-based microwave radiometer retrievals of the atmospheric temperature and humidity profile, and enhance the observation performance of locally deployed devices, this study implements two kinds of neural network methods for ground-based radiometer. One is called neural network direct sample retrieval algorithm, and the other is the neural network indirect sample retrieval algorithm based on the preprocessing of the observed brightness temperature. In this paper, these two retrieval algorithms are applied to the HRA002 domestic ground-based microwave radiometer developed by Wuhan Huameng Technology Co., Ltd., and the comparative observation experiments with soundings and three US MP-3000A microwave radiometers are carried out at the Wuhan National Basic Weather Station. The experiments results show that the mean square error of water vapor density and relative humidity retrieved by the improved network in direct sample retrieval is reduced by 0.94 g·m-3 and 5% respectively. The correlation between the observed brightness temperature and the simulated brightness temperature is improved significantly after preprocessing, the mean square error of the retrieval of lower-level temperature, water vapor density and relative humidity compared with sounding are improved from 2.4 K, 3.26 g·m-3 and 18.79% to 1.58 K, 2.18 g·m-3 and 14.55%. The accuracy of the HRA002 indirect sample retrieval method is slightly lower than that of the direct sample retrieval method, but the accuracy of the two methods is close. Compared with the retrieval results of 3 MP-3000A, the temperature profile of HRA002 using the direct sample retrieval method is generally better than MP-3000A, while the average deviations of water vapor density and relative humidity using indirect sample retrieval method of HRA002 is better than others, but the mean square errors are slightly lower. The research results in this paper show that the improved direct sample inversion method is more suitable for radiometer hardware performance and has higher retrieval accuracy, and the brightness temperature preprocessing of HRA002 can effectively improve the accuracy of indirect sample retrieval, making up for the defect that the direct sample retrieval method needs long-term observation data. The results also show that the use of direct sample and indirect sample retrieval algorithms can improve the localized and individualized observation performance of domestic microwave radiometers, and is useful in retrieving atmospheric parameter profiles.

Keywords

• domestic ground-based microwave radiometer
• retrieval algorithm
• brightness temperature preprocess
• co-site observation