Robust compression of multispectral remote sensing data

Abstract

This thesis develops efficient and robust non-reversible coding algorithms for multispectral remote sensing data. Although many efficient non-reversible coding algorithms have been proposed for such data, their application is often limited due to the risk of excessively degrading the data if, for example, changes in sensor characteristics and atmospheric/surface statistics occur. On the other hand, reversible coding algorithms are inherently robust to variable conditions but they provide only limited compression when applied to data from most modern remote sensors. The algorithms developed in this work achieve high data compression by preserving only data variations containing information about the ideal, noiseless spectrum, and by exploiting inter-channel correlations in the data. The algorithms operate on calibrated data modeled as the sum of the ideal spectrum, and an independent noise component due to sensor noise, calibration error, and, possibly, impulsive noise. Coding algorithms are developed for data with and without impulsive noise. In both cases an estimate of the ideal spectrum is computed first, and then that estimate is coded efficiently. This estimator coder structure is implemented mainly using data-dependent matrix operators and scalar quantization. Both coding algorithms are robust to slow instrument drift, addressed by appropriate calibration, and outlier channels. The outliers are preserved by separately coding the noise estimates in addition to the signal estimates so that they may be reconstructed at the original resolution. In addition, for data free of impulsive noise the coding algorithm adapts to changes in the second-order statistics of the data by estimating those statistics from each block of data to be coded. The coding algorithms were tested on data simulated for the NASA 2378-channel Atmospheric Infrared Sounder (AIRS). Near-lossless compression ratios of up to 32:1 (0.4 bits/pixel/channel) were obtained in the absence of impulsive noise, without preserving outliers, and assuming the nominal noise covariance. An average noise variance reduction of 12-14 dB was obtained simultaneously for data blocks of 2400-7200 spectra. Preserving outlier channels for which the noise estimates exceed three times the estimated noise rms value would require no more than 0.08 bits/pixel/channel provided the outliers arise from the assumed noise distribution. If contaminant outliers occurred, higher bit rates would be required. Similar performance was obtained for spectra corrupted by few impulses.