Low Complexity Prediction Model for Coding Remote-Sensing Data with Regression Wavelet Analysis

Abstract

Fast and efficient coding techniques are being increasingly required to meet the complexity restrictions of on-board satellite compression. The recently proposed Regression Wavelet Analysis (RWA) has proven to be highly effective as a spectral transform for coding remote sensing images. The algorithm is based on a pyramidal prediction, using multiple regression analysis, to tackle residual data dependencies in the wavelet domain. RWA combines low complexity and reversibility and has demonstrated competitive performance for lossless and progressive lossy-to-lossless compression superior to the state-of-the-art predictive-based CCSDS-123.0 and the widely used transform-based principal component analysis (PCA). In this paper we introduce a very low-complexity RWA approach, where prediction is based on only a few components, while the performance is maintained. When RWA computational complexity is taken to an extremely low level, careful model selection is necessary. Contrary to expensive selection procedures, we propose a simple and efficient strategy called neighbor selection for using small regression models. On a set of well-known and representative hyperspectral images, these small models maintain the excellent coding performance of RWA, while reducing the computational cost by about 90%.