An improved retrieval method of atmospheric parameter profiles based on the BP neural network

Surface-based microwave radiometer is used to measure the tropospheric parameter profiles continuously for 24 h. The measurement technology and retrieval methods are described clearly in this study. This paper focuses on the BP network and elaborates on it from a new perspective based on the Jacobian matrices between layers. Gradient descent is achieved by Jacobian matrices to train the network. A layered method is proposed to improve the efficiency and accuracy in training networks to obtain tropospheric water vapor and temperature profiles. Differently from the traditional method, the layered method divides the troposphere of 0-10 km into three layers based on the physical principles of cloud generation. Three networks, named as the bottom, the middle, and the upper network, are developed for the three layers. Therefore, three networks can be trained at the same time,using the same input and different output samples. According to the theories and the radiosonde data of 2012-2015 of Harbin China (45.46°N 126.40°E), a numerical experiment is designed to examine the layered method. The downwelling monochromatic radiative transfer model (MonoRTM) is used to calculate the atmospheric radiation brightness temperatures (BTs) with the radiosonde data. The experimental results show that the RMSEs of temperature and water vapor profiles of the layered method are reduced by 25.6% and 26.2%, respectively, at the altitude above 6 km, respectively, and the efficiency is improved by 20 times compared with the traditional method.