The main objective of hyperspectral imaging remote sensing is the identification of materials or phenomena from their reflectance or emissivity spectra to serve the needs of different applications. In this chapter, building on the understanding of the phenomenology of spectral remote sensing and the introduced signal processing methods, we develop algorithms for some unique hyperspectral imaging applications: detection of hard targets, gas detection, change detection, and image classification. The emphasis is on algorithms developed based on phenomenologically sound signal models, realistic application-driven requirements, and rigorous signal processing procedures, rather than ad hoc algorithms or trendy theoretical algorithms based on unrealistic assumptions.

Target Detection in the Reflective Infrared

The objective of hyperspectral target detection is to find objects of interest (called “hard targets” or simply “targets”) within a hyperspectral image and to discriminate between various target types on the basis of their spectral characteristics. The advantages are automated signal processing and lower spatial resolution requirements for the sensor. In this section we discuss the defining features of the target detection problem, we explain how to choose target detection algorithms, we investigate the consequences of practical limitations, and we evaluate performance using field data. Predicting detection performance using theoretical models is discussed in Section 11.2.

Definition of the Target Detection Problem

Hyperspectral target detection algorithms search for targets by exploiting the spectral characteristics of the target's surface material by looking at the spectrum of each pixel. Depending on the spatial resolution of the sensor, targets of interest may not be clearly resolved, and hence may appear in only a few pixels or even as part of a single pixel (subpixel target). Thus, the first key attribute of the hyperspectral target detection problem is that a “target present” versus “target absent” decision must be made individually for every pixel of a hyperspectral image. In most applications each target is characterized by its spectral signature and detection algorithms make decisions using the target signature and the data cube of the imaged scene.

Typical “search-and-detection” applications include the detection of man-made materials in natural backgrounds for the purpose of search and rescue, and the detection of military vehicles for purposes of defense and intelligence.