Research On Retrieval Algorithm Of Satellite Microwave Remote Sensing Atmospheric Parameters Using Convolutional Neural Network

Atmospheric temperature and humidity profile are important parameters used in weather forecast and climate research.The spaceborne microwave radiometer can observe the earth all-time and all-weather,and therefore it has become one of the most important profiling methods in the meteorological field.At present,there are many polar orbiting meteorological satellites in the world.China's Feng Yun-3 series satellites are also equipped with microwave payload for detecting temperature and humidity profiles,which is consistent with the US ATMS observation brightness temperature data.However,there is no corresponding retrieval product at present,It is very important to study the retrieval algorithm of microwave remote sensing data.At present,the retrieval algorithms can be divided into statistical regression and physical retrieval.Statistical algorithms mainly include linear regression and neural network algorithms.Neural network algorithms can be used to solve nonlinear problems by adding an activation layer,but its generalization ability to weather and data is weak.The physical meaning of one-dimensional variational algorithm is clear,but its retrieval accuracy is easily affected by the calculation accuracy of radiative transfer model and error covariance matrix,and the retrieval process is computationally heavy and time-consuming.The above two algorithms only use the data of one brightness temperature point and ignore the spatial correlation of brightness temperature data.Based on the above research background,a convolution neural network retrieval algorithm is proposed in this paper.The retrieval data set is constructed by using the temperature and humidity profiles obtained from the observed brightness temperature and WRF short-term forecast of ATMS payload.The preprocessing method of the data set is introduced,and the network structure suitable for microwave remote sensing data retrieval is designed.The loss function is also defined to increase the ability of the network to distinguish weather types.In order to explore the performance of the retrieval algorithm,this paper studies the retrieval of temperature and humidity for ocean and land surface types respectively.At the same time,the influence of weather types on the retrieval results is studied,and the results are compared with the ANN retrieval results and L2 products of ATMS.The test results show that compared with the neural network algorithm,the convolutional neural network retrieval algorithm proposed in this paper has higher retrieval accuracy,stronger data generalization ability,and the overall retrieval accuracy is higher than the L2 product of ATMS.This paper uses the convolutional neural network algorithm to retrieve the profile,where the temperature profile error is within 2K,and the relative humidity error is less than 16%.In terms of the overall retrieval results,the accuracy of convolution neural network algorithm is improved by distinguishing weather types,and the retrieval accuracy of clear sky weather profile is higher than that of cloudy weather,the profile retrieval accuracy of ocean area is higher than that of land area.