Application of multilayer perceptron to high-resolution infrared measurement retrieval

摘要

Multilayer perceptron (MLP) as universal approximator may be used for fast retrieval of atmospheric parameters such as vertical profiles of temperature, humidity and concentration of absorbing gases from high-resolution infrared spectra measured by satellite sensors. On the one hand, the number of spectral channels even necessary for retrieval of particular atmospheric parameter is very high, so practical use of MLP needs for effective compression of spectral data with tolerable loss of accuracy. On the other hand, algorithm of error back propagation becomes more effective if the input data vector contains uncorrelated values with zero means, their covariance are approximately equal, and information content of training set is maximized. The modified method of principal components (or empirical orthogonal functions expansion) satisfies to all above requirements. The MLP may be constructed using relevant truncated vectors of principal components as input and output data. Such MLP has fewer dimensions (the number of input, output and hidden neurons) and requires less time for training than MLP using the high-resolution spectrum as input vector and set of vertical profiles of atmospheric parameters as output vector.rnThe developed technique was applied to AIRS observations to retrieve temperature, humidity and methane content. The empirical orthogonal functions were obtained as eigenvectors of matrix G = S_e~(-1/2)S_RS_e~(-1/2), where S_R is sample covariance matrix built on real AIRS measurements over given region, and S_e is error covariance matrix characterizing the sensor. The set of measured and model profiles as well as surface temperature and pressure were used for construction of empirical orthogonal functions to represent output data of MLP as truncated expansion. Error profiles and examples of temperature and methane maps are presented.