Hyperspectral Data Processing: Algorithm Design and Analysis

10.8 Real Image Experiments

The synthetic image experiments conducted in Section 10.6 demonstrate the effectiveness of REEAs in extracting endmembers. This section further provides evidence of the ability of REEAs in endmember extraction. Due to the availability of the ground truth only the HYDICE 15-panel scene in Figure 1.15(a) and cuprite data in Figure 1.12 were used for experiments.

10.8.1 HYDICE Image Experiments

With the ground truth given in Figure 1.15(b) there are 19 panel pixels marked by red that are assumed to be pure pixels specified by five panel signatures in Figure 1.16. Using this information, a quantitative study of extracted endmembers by various algorithms was performed for analysis. As usual, VD used for the HYDICE data was 9. So, the data dimensionality reduction performed for PPI was reduced to 9, and the number of endmembers required for RPPI and random versions of N-FINDR was set to twice VD, 2n_VD = 18.

10.8.1.1 RPPI

Figure 10.29(a) and (b) shows the results of operating PPI and RPPI using 2000 skewers on the original data and reduced nine-dimensional data sets where data samples marked by circles are those with their PPI counts greater than 0 in Figure 10.30(a) with the total number tallied in the parentheses and the final intersection produced by the RPPI in Figure 10.30(b) with the total number tallied in the parentheses. As noted, the total number of K skewers, 2000, was selected empirically. It is interesting to note that compared to PPI, which could ...