Remote sensing hyperspectral data have hundreds or thousands of spectral components from very similar wavelengths. To store and transmit it entails excessive demands on bandwidth and on on-board memory resources, which are already strongly restricted. This leads to stop capturing data or to discard some of the already recorded information without further processing. To alleviate these limitations, data compression techniques are applied. Besides, sensors' technology is continuously evolving, acquiring higher dimensional data. Consequently, in order to not jeopardize future space mission's performance, more competitive compression methods are required. Regression Wavelet Analysis (RWA) is the state-of-the-art lossless compression method regarding the trade-off between computational complexity and coding performance. RWA is introduced as a lossless spectral transform followed by JPEG 2000. It applies a Haar Discrete Wavelet Transform (DWT) decomposition and sequentially a regression operation. Several regression models (Maximum, Restricted and Parsimonious) and variants (only for the Maximum model) have been proposed. With the motivation of outperforming the latest compression techniques for remote sensing data, we began focusing on improving the coding performance and/or the computational complexity of RWA. First, we conducted an exhaustive research of the influence of replacing the underlying wavelet filter of RWA by more competitive Integer Wavelet Transforms (in terms of energy compaction). To this end, we reformulated the Restricted model, reducing the execution time, increasing the compression ratio, and preserving some degree of component-scalability. Besides, we showed that the regression variants are also feasible to apply to other models, decreasing their computational complexity while scarcely penalizing the coding performance. As compared to other lowest- and highest-complex techniques, our new configurations provide, respectively, better or similar compression ratios. After gaining a comprehensive understanding of the behavior of each operation block, we described the impact of applying a Predictive Weighting Scheme (PWS) in the Progressive Lossy-to-Lossless (PLL) compression performance. PLL decoding is possible thanks to the use of the rate control system of JPEG 2000. Applying this PWS to all the regression models and variants of RWA coupled by JPEG 2000 (PWS-RWA + JPEG 2000) produces superior outcomes, even for multi-class digital classification. From experimentation, we concluded that improved coding performance does not necessarily entail better classification outcomes. Indeed, in comparison with other widespread techniques that obtain better rate-distortion results, PWS-RWA + JPEG 2000 yields better classification outcomes when the distortion in the recovered scene is high. Moreover, the weighted framework presents far more stable classification versus bitrate trade-off. JPEG 2000 may be too computationally expensive for on-board computation. In order to obtain a cheaper implementation, we render results for RWA followed by another coder amenable for on-board operation. This framework includes the operation of a smart and simple criterion aiming at the lowest bitrates. This final pipeline outperforms the original RWA + JPEG 2000 and other state-of-the-art lossless techniques by obtaining average coding gains between 0. 10 to 1. 35 bits-per-pixel-per-component. Finally, we present the first lossless/near-lossless compression technique based on regression in a pyramidal multiresolution scheme. It expands RWA by introducing quantization and a feedback loop to control independently the quantization error in each decomposition level, while preserving the computational complexity. To this end, we provide a mathematical formulation that limits the maximum admissible absolute error in reconstruction. Moreover, we tackle the inconvenience of proving the huge number of possible quantization steps combinations by establishing a quantization steps-allocation definition. Our approach, named NLRWA, attains competitive coding performance and superior scene's quality retrieval. In addition, when coupled with a bitplane entropy encoder, NLRWA supports progressive lossy-to-lossless/near-lossless compression and some degree of embeddedness.
- Dades espectrals; Datos espectrales; Spectral data; Compressió de dades; Compresión de datos; Data compression; Codificació piramidal; Codificación piramidal; Pyramidal coding