Hyperspectral Data Processing: Algorithm Design and Analysis

23.4 Experiments for Progressive Band Selection

When BS is implemented with no prior knowledge, the uniform BS is generally preferred because it attempts to select bands with interband correlation as least as possible by spreading bands as widest as possible in terms of wavelengths. The same reason also applies to PBDP, which prevents it from being used for BS due to the fact that bands highly prioritized by PBDP may also be highly correlated. Consequently, if a spectral band with a high priority score is selected, its neighboring bands may also have similar high priority scores and, thus, the chance for these bands to be selected is also very high. So, in order for PBDP to be implemented as PBS, we need to take care of this issue by including a preprocessing of BD removing highly correlated bands in conjunction with DDA determining the number of bands to adapt to different applications. In this section, the same three applications conducted for experiments in Chapters 20, 21, 22 are also performed for PBS for a comparative study and analysis. Since Gram-Schmidt orthogonalzation is a second-order statistics-based criterion that is not as effective as a high-order statistics-based ID as shown in Liu (2011), only ID is used for BD, referred to as ID-BD, for PBS. Furthermore, according to our extensive experiments, PBS with BD/BP generally does not perform as effectively as PBS with BP/BD. So, experiments were performed only for PBS with BP/ID-BD.