Data Compression for Remotely Sensed Hyperspectral Imagery


Due to the high spectral resolution, hyperspectral imagery has more powerful diagnostic capability in material detection, classification, and quantification than the traditional multispectral imagery. However, the vast volume of the resulting three-dimensional (3D) data cube brings about difficulties in data transmission, storage, and analysis. Hyperspectral images typically possess a high degree of spectral as well as spatial correlation. As a consequence, data compression can significantly reduce hyperspectral data volumes to more manageable size. Wavelet-based lossy compression techniques are of particular interest due to their long history of providing excellent rate distortion performance for traditional two-dimensional (2D) imagery. Consequently, a number of prominent 2D compression algorithms have been extended to the 3D hyperspectral imagery. However, it has been argued that such a direct extension from 2D to 3D without the consideration of special characteristics of hyperspectral imagery may be problematic. In this seminar, a more effective compression scheme will be presented, which employs principal component analysis (PCA) for spectral decorrelation followed by wavelet-based JPEG2000 for spatial coding. An approach using only a small set of principal components for compression will be discussed to improve the rate distortion performance. Since the remote sensing society cares about the following data analysis performance, we will also investigate the performance of reconstructed data in detection and classification, and propose an efficient approach in preserving anomalous but important pixels. In addition, strategies for low-complexity PCA will be introduced to fit the requirement of on-board compression.