Combining artificial neural networks for parameterization of radiative transfer models

Radiative transfer models are needed for the retrieval of atmospheric parameters like trace gases. The operational use of such of complex and computational intensive models is limited by the computer power available. Even more, it is not possible to use the models in near-real-time applications. This paper investigates the combination of artificial neural networks as a nonlinear regression technique for the parameterization of radiative transfer models. Redundant neural networks are trained using the backpropagation technique and later on they are ensemble in order to improve their overall generalization ability. The ensemble net parameterizes very accurate the radiative transfer model while the processing time and memory requirements are reduced drastically. The resulting multi-net ensemble will be used in an operational near-real-time environment.