Hyperspectral Data Processing: Algorithm Design and Analysis

15.4 AVIRIS Data Experiments

To demonstrate how much benefit can be gained from using kernel-based LSMA in hyperspectral classification, two real hyperspectral data sets were described in Chapter 1, Purdue Indiana Indian Pine test site in Figure 1.13 and HYDICE data scene in Figure 1.15 will be used for experiments for this purpose. In this section, we first study the AVIRIS data of Purdue Indiana Indian Pine test site in Figure 1.15. According to the provided ground truth there are 17 classes in this image scene shown in Figure 1.13(c) including the background labeled by class 17 that includes a wide variety of targets such as highways, railroad, houses/buildings, and vegetation that may not be of interest in agriculture applications. The spatial locations of all the 17 classes are shown in Figure 1.13(d) where the number in a parenthesis after a class label in Figure 1.13(d) is the total number of data samples in that particular class. The total number of data samples in the scene is . Two sets of experiments were conducted based on this scene to evaluate the KLSMA performance. One was mixed pixel classification to show the superior performance of KLSMA to that of LSMA without using kernels where data sample vectors are heavily mixed. The other was to use a three-dimensional receiver operating characteristics (3D ROC) analysis developed in Chapter 3 to conduct a quantitative analysis ...