Feature extraction from hyperspectral images compressed using theJPEG-2000 standard

摘要

We present results quantifying the exploitability of compressedremote sensing imagery. The performance of various feature extractionand classification tasks is measured on hyperspectral images coded usingthe JPEG-2000 Standard. Spectral decorrelation is performed using theKarhunen-Loeve transform and the 9-7 wavelet transform as part of theJPEG-2000 process. The quantitative performance of supervised,unsupervised, and hybrid classification tasks is reported as a functionof the compressed bit rate for each spectral decorrelation scheme. Thetasks examined are shown to perform with 99% accuracy at rates as low as0.125 bits/pixel/band. This suggests that one need not limit remotesensing systems to lossless compression only, since many commonclassification tools perform reliably on images compressed to very lowbit rates