17

Unsupervised Linear Hyperspectral Mixture Analysis

Chapter 16 provides the simplest unsupervised means of using hyperspectral measures to analyze data sample vectors for signature discrimination, classification, and identification without appealing for any algorithm. Consequently, its applications are rather limited. Specifically, when it comes to unsupervised linear spectral mixture analysis (LSMA), hyperspectral measures alone cannot do the same tricks as done in Chapter 16. Unsupervised linear spectral mixture analysis (ULSMA) is highly desirable in real-world applications due to the fact that prior knowledge is generally not available. Two of most challenging issues in ULSMA are (1) determining the number of signatures present in the data and (2) finding the signatures needed to perform spectral unmixing, both of which do not occur in supervised LSMA (SLSMA) since the latter generally assumes the target signatures to be known a priori or provided by prior knowledge. With recent advances in hyperspectral sensor technology many unknown and subtle signal sources that cannot be identified by prior knowledge or visual inspection can now be uncovered and revealed. In this case, using preassumed knowledge may not be reliable, accurate, or complete and, thus, the resulting unmixed results may be misleading. This chapter addresses these issues by introducing a new concept of sample spectral statistics generated by interband spectral information (IBSI) among a set of data sample vectors, S, denoted by IBSI(S), which can be used to categorize signatures into background (BKG) and target classes in terms of their sample spectral statistics. In order to extract these two types of signatures, two approaches, referred to as least squares (LS)-based ULSMA and component analysis-based (CA)-ULSMA, are developed to perform LSMA. It turns out that such unsupervised versions of LSMA can perform better than SLSMA in real-image experiments when obtaining true knowledge becomes difficult and incomplete.