One objective of hyperspectral data processing is to classify collected imagery into distinct material constituents relevant to particular applications, and produce classification maps that indicate where the constituents are present. Such information products can include land-cover maps for environmental remote sensing, surface mineral maps for geological applications and precious mineral exploration, vegetation species for agricultural or other earth science studies, or manmade materials for urban mapping. The underlying models and standard numerical methods to perform such image classifications have already been described to some extent in Chapter 12. This chapter formalizes the general classification strategy and underlying mathematical theory, provides further detail on the implementation of specific classification algorithms common to the field, and shows examples of the results of some of the methods. Hyperspectral image classification research is an intense field of study, and a wide variety of new approaches have been developed to optimize performance for specific applications that exploit both spatial and spectral image content. This chapter focuses on the most widely used and reported spectral classification approaches and includes a very limited discussion of the incorporation of spatial features in a classification algorithm.

13.1 Classification Theory

A general strategy for spectral image classification is illustrated in Fig. 13.1. Input data are assumed to be uncalibrated sensor output, calibrated pupil-plane spectral radiance, or estimated reflectance or emissivity derived by an atmospheric compensation algorithm. For some classification algorithms, the level of calibration or atmospheric compensation can make a difference, especially when the approach involves a human operator identifying constituents in the scene to extract training data or label classes. For many algorithms, however, the level of calibration or atmospheric compensation is not as important because the methods are generally invariant to these simple linear or affine transformations of the data.