Artificial Neural Networks for Spectral Sensitivity Analysis to Optimize Inversion Algorithms for Satellite-Based Earth Observation

Abstract Neural networks (NN) are statistical models that perform nonlinear regression and function approximation, and, as such, can be used as metamodeling tools in sensitivity analysis. Spectral input selection and structure optimization is an important step for the design of optimized retrieval algorithms based on NN for satellite data inversion. This also allows the identification of the most informative input wavelengths for the specific inverse problem under consideration. In the present work, two methods for the input wavenumber selection/reduction and input spectral sensitivity analysis are tested: the extended pruning (EP) and the autoassociative neural network (AANN) techniques. The specific geophysical problem is the retrieval of chemical and microphysical properties (sulfuric acid mixing ratio, number concentration, and mean radius of the particles) of secondary sulfate aerosol layers from high-spectral-resolution thermal infrared satellite observations at the nadir. The Infrared Atmospheric Sounding Interferometer (IASI) instrument is used as a basis for the generation of pseudo-observations. The two methods are found comparable in terms of the performance enhancement of the input-reduced inverse models. The reduced neural schemes have the number of connections reduced by a factor 5 x 103 (EP) and 2 x 104 (AANN) with respect to the maximum dimensionality/fully connected NN (the latter has about 5 x 106 connections, using the maximum spectral resolution of IASI and a fully connected NN model). This has an impact on the training iterations needed for the reduced models to converge, which is reduced of a factor 102 with respect to the maximum dimensionality/fully connected NN (this latter needs about 104 iterations to converge). In addition, the reduced NN are found capable of retrieving aerosol parameters typically with a 1.5–3.0 times better mean square error and smaller biases with respect to the maximum dimensionality NN. The selected reduced spectral information is analyzed and found consistent with the known spectroscopy of the problem. The spectral sensitivity analysis carried out indicates that the most sensitive input wavenumbers for this specific problem are those characterized by the typical absorption band structures for ionic and molecular sulfates present in the sulfate aerosol droplets.