The principal component analysis and compression of satellite-based hyperspectral IR sounder measurements

The application of the principal component analysis (PCA) techniques is considered for the IR-spectra compression measured by hyperspectral IR sounder IKFS-GS, on-board the future geostationary meteorological satellites of Electro-M series. The volume of IKFS-GS data for one hour of measurements over selected area will be approximately 14,4 GB, accounting for the following design characteristics of the equipment: spectra are recorded in the two spectral bands (680-1210 cm^-1 and 1600-2250 cm^-1), spectral resolution is 0,625 cm^-1, the size of the selected area is 640x640 km², spatial resolution is 4 km at the nadir, about 6,5 million spectra (with 1888 channels) are registered, 10 bits is used per each channel signal recording. Because of such large volume, satellite information compression is necessary. In addition, according to the recommendations of the WMO, measurements of the IR sounders from geostationary meteorological satellites are treated as key in the WMO Integrated Global Observing System and should be distributed to users in terms of the principal component scores (PCs).

Due to the absence of actual satellite-based IKFS-GS measurements and the partial similarity of the IKFS-GS and IKFS-2 sounder characteristics, the compression algorithm has been tested on proxy data – IR spectra measured by the IR sounders IKFS-2 (spectral range 660-2000 cm^-1, 2700 channels) installed on the polar orbiting meteorological satellites Meteor-M No. 2 and No. 2-4. The proposed algorithm performs the decomposition of each IR spectrum using the generalized empirical orthogonal functions (EOF). The set of EOFs is formed from the «senior» eigenvectors of the sample covariance matrix constructed from the measurement data and normalized by the noise covariance matrix [1, 2].

Various EOF ensembles have been derived for numerical experiments. The training set used for the generation of the eigenvectors consists of semi-daily IKFS-2 data files for some dates in 2024 (about 10⁴ spectra) or of «annual» sample including IKFS-2 measured spectra for the 25th day of each month in 2022 (about 10⁶ spectra). The compression technique was applied independently to spectral data in three bands (660-765, 765-1400, and 1400-2000 cm^-1) and to the entire spectrum. The EOF decomposition, is shown, provides reconstruction of the spectra with an error at the instrumental noise level while using 20-30 (first band), 30-40 (second and third bands) and 30-50 (entire spectrum) EOFs, depending on the formed orthogonal basis. Thus, the effectiveness of the proposed compression algorithm has been demonstrated – to reconstruct (with little loss of information) and distribute the satellite-measured IR spectra to the users, it is sufficient to determine no more than 50 PCs.

1. A. B. Uspensky, S. V. Romanov. Use of principal components technique for the analysis of advanced prospective IR-sounders data - Issledovanie Zemli iz Kosmosa, 2003, N 3, pp. 26-33.

2. D. A. Kozlov, I. A. Kozlov, A. B. Uspensky et al. Characterization of the Noise Covariance Matrix of the IKFS-2 Infrared Fourier Transform Spectrometer - Izvestiya, Atmospheric and Oceanic Physics, 2022, Vol. 58, No. 9, pp. 1160–1172.