23.6 Land Cover/Use Classification

The Purdue Indiana Indian Pine test site in Figure 1.13 was used in experiments to demonstrate the maximum likelihood classification (MLC) performance resulting from PBS where once again seven BP criteria: variance, SNR, skewness, kurtosis, entropy, ID, and negentropy, were used for BP and SID was used as a BD criterion with ε set to 0.1. Table 23.2 tabulates the first 60 highest prioritized BD de-correlated bands and corresponding removed bands after applying ID-based BD.

Table 23.2 First 60 highest prioritized BP/BD bands and corresponding removed bands after applying ID-based BD on seven BP results of Purdue data with ε = 0.1.

Figure 23.2 shows MLC rates of 16 classes produced by operating PBS on seven different BP/BD prioritized bands where MLC is calculated based on (20.17) and the figures on the right and the left were obtained by PBS with/without using ID-BD, respectively. Also included in the last figure in Figure 23.2 is the overall averaged MLC rate among 16 classes. In order to find DDA, the means of 16 classes are calculated to serve as 16 signatures required by DDA. A vertical line in each figure in Figure 23.2 shows a cutoff threshold for determined and set by DDA = 36 that was obtained from the Huffman coding using 16 class signatures.

Comparing Figure 23.2 with Figure 23.1, MLC rates of PBS using two second-order BP criteria, variance and SNR, benefited significantly from BD as the rates were increased in most of classes. It implies that second-order BP criteria were more appropriate than HOS BP criteria except the entropy criterion. Since MLC classification performance analysis is similar to that in Section 21.7.2, its detailed discussions will not be provided here.

The values of DDA given in Table 22.3 were further used in conjunction with PBS that used prioritized bands with/without ID-BD to perform MLC. Table 23.3 tabulates MLC rates of 16 classes resulting from PBS using various values of determined by VD, DDA, and optimal MLC performance where SID was used to measure spectral similarity and the numbers in parentheses in the last two columns are the true total numbers used for MLC. It should be noted that the column “total” in Table 23.3 is not 202 because some bands are already removed by BD. In this case, the numbers in parentheses were actual total numbers after BD. For example, for variance, skewness, and entropy, these numbers are 84, 86, and 85, respectively.

Table 23.3 MLC rates of 16 classes in the Purdue data using various values of determined by VD, DDA, and optimal FCLS performance.

Comparing Table 23.1 with Table 23.2, it is seen that BP/BD improved MLC rates. Another observation is that variance and entropy criteria provided lower optimal band numbers, while skewness still required more bands in most classes.