Abstract
Hyperspectral imaging (HSI) has been demonstrated to be useful for estimating spatial distributions of different chemical constituents on surfaces and in imaged scenes, as well as for detecting anomalies and known targets that may be present. In target detection, the objective is to find pixels that contain significant signal attributable to a known “target” spectrum. Unfortunately, measured images are often dominated by “non-target” interference signals with unknown spectra that vary image-to-image. To be effective, target detection algorithms must account for the non-target signal, commonly referred to as “clutter.” The utility of HSI for target detection in spatially complex scenarios can be attributed to two aspects associated with the sensing system: 1) The target can dominate the measured signal in a few individual pixels, although it may have only a minor contribution to the overall signal, and 2) HSI measures many pixels relevant for characterizing clutter that can be used in clutter suppression algorithms to account for interferences.
Clutter suppression enables the detection of minor signal sources in HSI, resulting in a sensing system uniquely suited to highly variable environmental scenarios. Popular clutter suppression methods such as extended least squares (ELS) and generalized least squares (GLS) can be described as extensions of the classical least squares (CLS) model. The main advantage of these methods is that they are fast and easy to train on an image-to-image basis providing highly sensitive, image-specific target detection systems. However, the optimization of such models affects the sensitivity and selectivity of the detection. In an effort to assist practitioners interested in clutter suppression, this paper presents a detailed description of the theoretical aspects of ELS and GLS methods used for target detection on hyperspectral images. The concepts are demonstrated with practical examples of the detection of an acrylonitrile butadiene styrene (ABS) particle in sand based on a single near infrared hyperspectral image.
Clutter suppression enables the detection of minor signal sources in HSI, resulting in a sensing system uniquely suited to highly variable environmental scenarios. Popular clutter suppression methods such as extended least squares (ELS) and generalized least squares (GLS) can be described as extensions of the classical least squares (CLS) model. The main advantage of these methods is that they are fast and easy to train on an image-to-image basis providing highly sensitive, image-specific target detection systems. However, the optimization of such models affects the sensitivity and selectivity of the detection. In an effort to assist practitioners interested in clutter suppression, this paper presents a detailed description of the theoretical aspects of ELS and GLS methods used for target detection on hyperspectral images. The concepts are demonstrated with practical examples of the detection of an acrylonitrile butadiene styrene (ABS) particle in sand based on a single near infrared hyperspectral image.
Original language | English |
---|---|
Article number | 105032 |
Journal | Chemometrics and Intelligent Laboratory Systems |
Volume | 244 |
Number of pages | 12 |
ISSN | 0169-7439 |
DOIs | |
Publication status | Published - 15 Jan 2024 |
Keywords
- Hyperspectral image analysis
- chemometrics
- generalised least squares
- spectroscopy
- Extended least squares
- Generalized least squares
- External parameter orthogonalization
- Classical least squares
- Weighted least squares
- Target detection
- Hyperspectral imaging