The Robustness of an Anti‐Noise BP Neural Network Inversion Algorithm for Ground‐Based Microwave Radiometer

The ground‐based microwave radiometer (MWR) retrieves atmospheric profiles with a high temporal resolution for temperature and relative humidity up to a height of 10 km. These profiles have been widely used in the field of meteorological observation. Due to the inherent fragility of neural networks, one of the important issues in this field is to improve the reliability and stability of MWR profiles based on neural network inversion. We propose a deep learning method that adds noise to the BP neural network inversion (NBPNN) process. Comparison of the radiosonde data and NBPNN results shows that if the error of MWR brightness temperature is in the range of −2–2 K, the root‐mean‐square error (RMSE) of the temperature profile is 2.15 K, and the RMSE of the relative humidity profile is 19.46 % inverted by NBPNN. The results are much less than the errors of the temperature profile and relative humidity profile inverted by the traditional backpropagation neural network inverse method. From the comparison, we demonstrated that NBPNN significantly increases the inversion accuracy and robustness under the condition of errors in brightness temperature, which can reduce requirements for BT accuracy of MWR and achieve MWR long‐term stability. Key Points It's necessary to improve the reliability and stability of inversion for microwave radiometer We propose a deep learning method that adds noise to the BP neural network inversion process The new inversion process significantly increases the inversion accuracy and robustness under the condition of errors in brightness temperature