Hyperspectral Data Processing: Algorithm Design and Analysis

21.7 Experiments for PBDP

As a parallel section to Section 20.6 the same experiments conducted for PSDP were also performed for PBDP in this section for comparison where three applications, endmember extraction, land cover/use classification, and spectral unmixing using three different types of hyperspectral image data sets are considered.

21.7.1 Endmember Extraction

The Cuprite data in Figure 1.12(a) were used for endmember extraction where IN-FINDR was selected to extract the five mineral signatures, A, B, C, K, and M of major interest in the scene. Figure 21.8(a)–(g) plots extracted endmembers by IN-FINDR where seven BP criteria, variance, SNR, skewness, kurtosis, entropy, ID, and neg-entropy were used to prioritize bands to be used by PBDP. VD-estimated value for this scene was n_VD = 22 that was the same used for experiments in Chapter 20. Its twice value 2n_VD = 44 provided an upper bound used by PBDP. Therefore, the x-axis represents the number of prioritized bands used by PBDP starting from 0 to 50, and the y-axis is the number of extracted endmembers. When the number of extracted endmembers is less than 5, the extracted endmembers are specified by particular mineral signatures. As shown in Figure 21.8 the high-order statistics-based BP criteria generally performed better than second-order statistics-based BP criteria such as variance and SNR in the sense that fewer band numbers were required to extract five endmembers. The smallest and largest numbers of bands required to ...