Effect of Hierarchically Trained BP Neural Networks on The Inversion of Atmospheric Temperature Profiles

The microwave radiometer is a passive remote sensing instrument. The MP3000A microwave radiometer is used in this study. For the inversion problem, this paper uses a neural network algorithm to invert the 0-10 km atmospheric temperature profile. In order to improve the efficiency and accuracy of the neural network inversion of the temperature profile, the 0-10 km is divided into several independent neural networks on the basis of the current common BP (Back Propagation) neural network overall method. The different results of each group were compared and analyzed. The results show that the average absolute error of neural networks trained by dividing them into 0-2km, 2-5km, 5-10 km and 0-3km, 3-7km, 7-10 km, respectively, is reduced by 26.11% and 27.67% compared with the neural networks trained as a whole, and the accuracy is significantly improved. However, the mean absolute error of 0-2.5km, 2.5-5km, 5-7.5km, and 7.5-10 km stratified training was only improved by 17.11% and its mean square error was basically equal to that of the overall method. It shows that the inversion accuracy will be improved under reasonable stratification conditions.