Hyperspectral Data Processing: Algorithm Design and Analysis

30.3 Experiments

In this section, synthetic and real hyperspectral image experiments were conducted to substantiate and validate the utility of ATGP-FCLS in size estimation of subpixel targets. The data to be used for experiments was the HYDICE image shown in Figure 1.15(a) with ground truth map of Figure 1.15(b).

30.3.1 Synthetic Image Experiments

In order to evaluate our approach, we simulated a synthetic image scene based on Figure 1.15(a) and (b) and the five panel signatures in Figure 1.16. First of all, 400 grass samples extracted from Figure 1.15(a) were used to simulate the image background. Two panel signatures, p₁ and p₂ were used to simulate two sets of target panels, p₁₁, p₁₂, p₁₃ and p₂₁p₂₂, p₂₃, respectively. The three panels in each set had size ranging from 100%, 50%, and 25% of pixel size (i.e., GSD = 1.5 m), namely, 2.25 m², 1.125 m², and 0.5625 m², respectively. Figure 30.1(a) and (b) shows how the target panels p₁₂ and p₁₃ with size being 50% and 25% of pixel size were simulated. For example, in order to simulate the panel p₁₂, we first simulated 2-pixel vectors specified by the grass signature and 2-pixel vectors specified by the panel signature p₁ to form a 4-pixel square panel where each of 4-pixel vectors in the panel had size of 2.25 m² and the resulting 4-pixel square panel had size of 9 m². This 4-pixel square panel was then shrunk to its ¼ size by averaging all 4-pixel vectors to reduce the 4-pixel square panel with size of 9 m² to a 4-pixel square panel ...