Artificial neural networks in environmental sciences. I. NNs in satellite remote sensing and satellite meteorology

Abstract

Two generic satellite remote sensing NN applications are described: NN solutions for forward and inverse (or retrieval) problems in satellite remote sensing. These two solutions correspond to two different approaches in satellite retrievals: variational retrievals (retrievals through the direct assimilation of sensor measurements) and standard retrievals. It is shown that both the forward model and the retrieval problem can be considered as nonlinear continuous mappings. The NN technique is a generic technique to perform continuous mappings. It is compared with regression approaches. Examples of a NN SSM/I forward model and a NN SSIM/I retrieval algorithm are used to illustrate advantages of using neural networks for developing both retrieval algorithms and forward models, and for minimizing the retrieval errors.