Effect of lossy vector quantization hyperspectral data compression on retrieval of red-edge indices

This paper evaluates lossy vector quantization-based hyperspectral data compression algorithms, using red-edge indices as end-products. Three compact airborne spectrographic imager (CASI) data sets and one airborne visible/infrared imaging spectrometer (AVIRIS) data set from vegetated areas were tested. A basic compression system for hyperspectral data called the "reference" system, and three-speed improved compression systems called systems 1, 2, and 3, respectively, were examined. Five red-edge products representing the near infrared (NIR) reflectance shoulder (Vog 1), the NIR reflectance maximum (Red rs), the difference between the reflectance maximum and the minimum (Red rd), the wavelength of the reflectance maximum (Red lo), and the wavelength of the point of inflection of the NIR vegetation reflectance curve (Red lp) were retrieved from each original data set and from their decompressed data sets. The experiments show that the reference system induces the smallest product errors of the four compression systems. System 1 and 2 perform fairly closely to the reference system. They are the recommended compression systems since they compress a data set hundreds of times faster than the reference system. System 3 performs similarly to the reference system at high compression ratios. Product errors increase with the increase of compression ratio. The overall product errors are dominated by Vog 1, Red rs, and Red rd, since the amplitude of product error for these products is over one order of magnitude greater than those for the Red-lo and Red lp products. The difference between the overall error from the reference and that from system 1 or 2 is below 0.5% at all compression ratios. The overall product error induced by system 1 or 2 is below 3.0% and 2.0% for CASI and AVIRIS data sets, respectively, when the compression ratio is 100 and below. Spatial patterns of the product errors were examined for tha AVRIS data set. For all products, the errors are uniformly distributed in vegetated areas. Errors are relatively high in nonvegeted and mixed-pixel areas.