Hyperspectral Data Processing: Algorithm Design and Analysis

15.5 HYDICE Data Experiments

In this section, we used another real hyperspectral data set, HYDICE scene in Figure 1.15 for experiments that will show completely opposite results from those obtained from Purdue Indiana Indian Pine data set. First of all, VD estimated for this scene was used as p = 9 with the false alarm probability . In this case, nine signatures were used for classification. These include five panel signatures in Figure 1.15(b) and other four undesired signatures, referred to as grass, road, tree, and interferer as identified in Figure 1.17. The LSMA performance was evaluated by detection of the 19 R panel pixels shown in 1.15(b) based on their abundance fractions unmixed by LSMA. The same 3D ROC analysis was also used for performance evaluation. Since the experiments on mixed pixel classification conducted for the HYDICE scene are previously reported in Chang (2003a), their results are not included here. Furthermore, the HYDICE experiments were performed in an exactly same manner that was conducted for Purdue data in Section 15.3. In this case, only those results similar to Figures 15.16–15.18 and Tables 15.13–15.15 are shown in Figures 15.19–15.21 and Tables 15.16–15.18 to avoid unnecessary redundancy where the unmixed results of KLSOSP, KNCLS, and KFCLS were obtained from the best empirical selection of their corresponding parameters as they were done for Purdue's ...