A Brief Overview of
Reflectance Spectroscopy and Hyperspectral Imaging
David Leverington
Overview Different materials tend to reflect and emit electromagnetic radiation in different ways. This offers a basis for the detection and mapping of materials over regions of interest. Both calibrated and uncalibrated reflectance data are routinely used in the field of remote sensing to map the extent and abundance of materials exposed at the surface of the Earth and at the surfaces of other solar system bodies. Reflectance Characteristics of Earth Materials For the Earth environment, daytime radiation measured in the visible and near-infrared ranges of the electromagnetic spectrum predominantly consists of reflected solar radiation, whereas radiation measured in the thermal-infrared range is largely related to thermal emission by surface materials. The reflectance, transmittance, and emittance properties of materials can be measured and described using spectral response curves, which depict proportions of reflected, transmitted, or emitted electromagnetic radiation as a function of wavelength. A particular type of material will commonly be characterized by "spectral features" (i.e., troughs and peaks in spectral response curves) that together comprise the "spectral signature" of the material. For example, substantial absorption of incoming radiation in the red portion of the spectrum, and high reflectance in the near-infrared, are both characteristic spectral features of healthy green vegetation (see Figure 3, below). On this basis, the presence and abundance (% surface cover) of healthy green vegetation can typically be inferred from remote sensing imagery through examination of the relative reflectance measurements made in the red and near-infrared portions of the electromagnetic spectrum. The reflectance properties of rock-forming minerals are a function of chemical composition and crystal structure such that many minerals have diagnostic spectral absorption properties (e.g., Hunt, 1980; Goetz and Rowan, 1981; Goetz et al., 1985; Mustard and Sunshine, 1999). Absorption features in the reflectance spectra of minerals in the visible and near-infrared (~400 to 1000 nm) are generally associated with electronic transitions caused by the presence of transition metals such as iron and chromium, whereas absorption features within the near-and shortwave infrared (~1000 to 3000 nm) are typically associated with combinations and overtones of carbonate, hydroxyl, and phosphate fundamentals (e.g., Kendall, 1966; Adams, 1974; Hunt, 1980; Mathews, 1986; Gupta, 1991; Burns, 1993; Gaffey et al., 1993). The products of chemical weathering can strongly influence the spectral characteristics of rocks and soils, even when the depth of alteration is less than several hundred micrometers. This imparts additional absorption features to reflectance spectra that may be otherwise inconsistent with the underlying pristine mineralogy (e.g., Cloutis, 1992; Ferrari et al., 1996). Similarities in the spectral properties of different rock or soil classes, and spatial variability within individual classes, can make the successful discrimination between surface classes especially challenging, particularly if the spectral and spatial resolutions of the input images are limited (e.g., Ager and Milton, 1987; Birnie et al., 1989; Arvidson et al., 1994; Macias, 1995; Chabrillat et al., 2000). In general, hyperspectral datasets offer greater potential for the identification and separation of surface classes than multispectral datasets.
Figure 1: The electromagnetic spectrum (after Leverington, 2001). On the Earth, reflected solar radiation dominates at visible and near-infrared wavelengths, and emitted ground radiation is prominent at thermal wavelengths.
Figure 2: Equipment used to characterize the reflectance properties of field materials. At left is a highly reflective lambertian surface used for system calibration. The spectrometer (an ASD FieldSpec3) sits in a backpack for easy transportation. Measurements are made using a small handheld scope attached to the spectrometer by fibre optic cable, with the spectrometer itself controlled by a small laptop computer.
Figure 3: Reflectance spectra for selected surface classes (vegetation, rock, soil, and water) (Leverington, in prep.). Spectra are given for the range 350-2500 nm (0.35-2.5 µm).
Figure 4: Reflectance characteristics of representative whole-rock samples of lithological units on eastern Melville Island, Nunavut, Canada (Leverington, 2009, in press). Individual spectra are labelled at right: 1 – Degerböls Fm limestone; 2 – Great Bear Cape Fm limestone; 3 – Tingmisut Inlier dolostone; 4 – Canyon Fiord Fm sandstone; 5 – Sabine Bay Fm sandstone; 6 – Bjorn Fm sandstone; 7 – Trold Fiord Fm sandstone; 8 – Assistance Fm mudstone; 9 – gabbro intrusion. Spectra produced using an ASD Fieldspec3 system employing a contact probe. Among the most distinct absorption features associated with materials in the Melville Island study area are those at ~1000 nm and less than 600 nm, related to the presence of ferric iron. Water-related absorption features are located at 1400 and 1900 nm. Carbonate absorption features at ~2300 nm characterize the limestone and dolostone classes, as well as clastic rocks with carbonate cements. Absorption features at ~2200 nm, related to the presence of clay minerals such as kaolinite or smectite, are also present in the spectra of several rock classes (Leverington, in press).
Figure 5: Reflectance spectra of selected geological classes present in Big Bend National Park, Texas: 1 – salt crust; 2 – Javelina Fm tan-weathering sandstone; 3 – Pen Fm tan-weathering shale; 4 – Aguja Fm tan-weathering sandstone; 5 – Chisos Fm ash; 6 – Chisos Fm dark mafic unit; 7 – Chisos Fm tan-weathering basalt; 8 – Tertiary syenodiorite (Leverington, 2008, 2009). Spectra produced using an ASD Fieldspec3 system employing a contact probe. As with the surface materials of eastern Melville Island, the materials of the Big Bend National Park study area are characterized by spectral troughs indicative of the presence of ferric iron, carbonate minerals, and clay minerals. The spectral properties of several sedimentary end members of interest in the study area (Pen Formation shale, Aguja Formation sandstone, Javelina Formation sandstone) are variable across the park and substantially overlap within the study transect. Some end members, including evaporite deposits and mafic igneous units, instead have relatively distinct spectral properties (Leverington, 2009).
Hyperspectral datasets, generated in dozens to hundreds of individual wavelength ranges, potentially allow for improved discrimination between rock and soil classes relative to multispectral datasets (e.g., Goetz et al., 1985). For example, the Hyperion sensor (Pearlman et al., 2003) is one of several instruments mounted on the EO-1 spacecraft, which was launched in 2000 and is now operating on an extended mission in partnership between the United States Geological Survey and the National Aeronautics and Space Administration. Hyperion is an along-track sensor that collects optical data in 242 spectral bands in the visible, near-infrared, and shortwave infrared. Continuous spectral coverage is available throughout the range of ~400 to 2400 nm, with spectral resolutions of individual bands of ~10 nm. The Hyperion sensor operates with a swath width of 7.6 km and a pixel size of 30 x 30 m.
Hyperspectral imagery can be classified or otherwise transformed into measures of class presence or abundance through a wide range of possible techniques. Though standard per-pixel classification schemes can generate useful information (e.g., through maximum likelihood classification, or the use of feedforward backpropagation neural networks), they typically require pre-processing of hyperspectral datasets since classification results degrade substantially when large numbers of input bands are utilized (this is known as the “curse of dimensionality”, or the “Hughes phenomenon”; e.g., Hughes, 1968). Instead, spectral unmixing methods (e.g., Adams and Smith, 1986; Boardman, 1989) are often used with hyperspectral datasets. These methods are better suited for use with hyperspectral data, and offer the additional potential capacity for mapping end members at sub-pixel scales through the generation of abundance maps. Spectral unmixing methods are based on the recognition that the magnitudes of individual pixel values are a function of the combined spectral contributions of one or more end-member classes present at sub-pixel scales. On the basis of image values and the spectral characteristics of pre-defined end-members, spectral unmixing provides the means for determining the fractional abundances of individual end-members based on the composite spectra measured at every pixel location. This enables the generation of a separate fractional abundance image for each end member (e.g., Adams and Smith, 1986; Mustard and Pieters, 1987; Boardman, 1989; Gillespie et al., 1990; Shimabukuru and Smith, 1991; Settle and Drake, 1993; Mustard, 1994; van der Meer, 1996; Maselli, 1998; Huete et al., 2003). Hyperspectral data are well-suited for unmixing, since the larger number of separate bands expands the basis upon which end members can be identified and distinguished, and the larger number of bands mathematically permits the definition of a greater number of end members than datasets of low spectral resolution (it is typical for unmixing algorithms to limit the number of end members to the number of independent input channels minus one). Examples of endmember data and corresponding unmixing results are given below.
Figure 6: Comparison between reflectance measurements derived from six Melville Island (Canada) end members (left), selected Hyperion pixels (center), and selected TM pixels (right) (Leverington, in press). Reflectance values were determined from original Hyperion and TM datasets through the application of standard atmospheric correction procedures. The three sets of reflectance measurements were independently derived, except for the snow end member (left), which is a smoothed version of Hyperion-derived data. Reflectance data are offset for clarity, and y-axes are marked at increments of 10%. The characteristics of image-derived spectra (center and right) are consistent with end-member spectra (left), with differences attributable to the general absence of 100% end-member cover at pixel scales (30 x 30 m) for geological and vegetation classes. The noise that characterizes the Hyperion spectra is present in the original image, and is typical of data generated by this sensor (e.g., Crowley et al., 2003).
Figure 7: Spectral unmixing results for a study area on Melville Island, Nunavut, Canada (Leverington, in press). Left: bedrock geology and Hyperion color-infrared image. Right: An example of unmixing results derived from Hyperion-derived reflectance data for six end members (see figure above for end-member spectra). The values of fractional-component images range from 0.0 (black) to 1.0 (white). Unmixing results in the immediate vicinity of Tingmisut Lake (upper-right corner) were rendered invalid by clouds present at the time of imaging. Apart from low-level overestimations of surface cover, unmixing results are generally valid for the snow, green vegetation, red sandstone, and other sandstone classes, but the presence and exposure of the limestone class is greatly overestimated in the southern half of the study area.
References Cited (c) David Leverington, 2009
|