Retrieval of atmospheric temperature profiles by an algorithm based on neural networks: comparison with physical and statistical methods

The retrieval of atmospheric temperature vertical profiles from brightness temperatures measured by a ground-based microwave radiometer in rain absence is performed. Two different neural network (NN) structures are developed: the former, trained with less than 120 historical profiles, gives good results, the latter, called "optimal NN", gives excellent results with respect to radiosonde profiles. In this case, about 250 historical profiles are utilized and an appropriate NN structure is employed. The strength of this method is measured by comparison with the classical inversion methods and the radiosonde profiles. The optimal NN technique exhibits better results than those obtained by classical methods in terms of the error and the elaboration time.