Spotlight On Hyperspectral. 201 - Shippert P - PDF

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Spotlight on

Hyperspectral

Courtesy of Research Systems, Inc.


Peg Shippert

From analyzing
eelgrass beds
A mong the most significant recent breakthroughs in
remote sensing has been the development of hyper-
spectral sensors and the software to analyze the resulting image
Hyperspectral images are sometimes referred
to as “image cubes” because of the large
number of measured wavelengths. The face
of the cube in this example is an image of an
data. A short time ago, only spectral remote sensing experts
in the Pacific had access to hyperspectral images and the software tools nec-
agricultural region in Australia, which was
collected by the Hyperion sensor. The top and
Northwest to essary to take advantage of them. During the past decade, right side of the cube show hundreds of color-
though, hyperspectral image analysis has matured into one of coded pixel values measured for each pixel
identifying the most powerful and fastest-growing technologies in the field along the top and right edge of the image.
pathfinder of remote sensing.
A hyperspectral image is one in which the reflectance from
minerals for each pixel is measured at many narrow, contiguous wavelength mined by its energy level. Light and
intervals. Such an image provides detailed spectral signatures other forms of electromagnetic radia-
geological for every pixel. These signatures often provide enough infor- tion are commonly described in terms
exploration, mation to identify and quantify the material(s) existing within of their wavelengths. For example, vis-
the pixels. A user could, for instance, employ a hyperspectral ible light has wavelengths between 0.4
hyperspectral image to locate and quantify different types of building mate- and 0.7 microns, whereas radio waves
imagery and rials or minerals that might be present within an area of inter-
est or even within a single pixel.
have wavelengths greater than about 1
millimeter (see Figure 1).
analysis is This article will discuss the difference between hyperspec- Reflectance is the percentage of light
tral and multispectral images, introduce relevant spectral con- hitting a material that is then reflected
proving its cepts, review some recent applications of hyperspectral image by that material (as opposed to being
worth for analysis, and summarize image-processing techniques com- absorbed or transmitted). A reflectance
monly applied to hyperspectral imagery. spectrum shows the reflectance of a
diverse remote material measured across a range of
sensing Spectral image basics wavelengths. Some materials will reflect
To understand the advantages of hyperspectral imagery, it helps certain wavelengths of light, while other
tasks. to first review some basic spectral remote sensing concepts. You materials will absorb the same wave-
may recall that each photon of light has a wavelength deter- lengths. Many materials have unique

Peg Shippert is the earth science


applications specialist for Research
Systems, Inc. (www.researchsystems.
com), a wholly owned subsidiary of
Eastman Kodak (www.kodak.com).
She has a Ph.D. in physical geography
and more than 13 years of experience
analyzing multispectral and
hyperspectral imagery for diverse
applications.
FIGURE 1 The electromagnetic spectrum with the region of visible light expanded.
Hyperspectral 101

FIGURE 2 Here we see the reflectance spectra measured by FIGURE 4 In this figure, we see the reflectance spectra of the same
laboratory spectrometers for three materials — a green bay laurel, three materials seen in Figure 2 as they would appear to the
mineral talc, and silty loam soil. multispectral Landsat 7 sensor.

FIGURE 3 In hyperspectral imagery,


measurements are made at many narrow,
contiguous wavelength bands, resulting in a
complete spectrum for each pixel.

patterns of reflectance and absorption across


wavelengths. The reflectance spectra in
Figure 2, for instance, reveal three differ-
ent materials. Many materials can be iden-
tified by their spectra.
Field and laboratory spectrometers usu-
ally measure reflectance at many narrow,
closely spaced wavelength bands, so that FIGURE 5 This example reveals the reflectance spectra of the same three materials shown
the resulting spectra appear to be continu- in Figures 2 and 4 as they would appear to the hyperspectral Airborne Visible Infrared
ous curves as seen in Figure 2. When a spec- Imaging Spectrometer, or AVIRIS. The gaps in the spectrum are wavelength ranges at which
trometer is used in an imaging sensor, the the atmosphere absorbs so much light that no reliable signal is received from the surface.
resulting images record a reflectance
spectrum for each pixel in the image (see of narrow, contiguous wavelength bands. spectral remote sensing technology.
Figure 3). In fact, the hyper in hyperspectral means Although most hyperspectral sensors
“over,” or “too many.” So you might say measure hundreds of wavelengths, it is not
Hyperspectral sensors that hyperspectral images measure too many the number of measured wavelengths that
Remote sensing has included multispectral wavelengths. When we look at a spectrum defines a sensor as hyperspectral. Rather it
sensors — sensors which collect images for for one pixel in a hyperspectral image it is the narrowness and contiguous nature of
a small number of broad wavelength bands looks very much like a detailed spectrum the measurements. For example, a sensor
— since Landsat 1 was launched in 1972. that would be measured in a spectroscopy that measures only 20 bands would be con-
Although multispectral sensors were revo- laboratory (see Figure 5). This type of sidered hyperspectral if those bands were
lutionary when first introduced, they detailed pixel spectrum provides much more contiguous and, say, 10 nanometers wide.
lack sufficient spectral resolution for pre- information about the surface than a mul- If the sensor measures 20 wavelength bands
cise surface studies because of their tispectral pixel spectrum. Thus, hyper- that are 100 nanometers wide, or that are
low number of spectral bands(see Figure spectral imagery provides the potential for separated by nonmeasured wavelength
4). In contrast to multispectral instruments, more accurate and detailed extraction of ranges, the sensor would not be considered
hyperspectral sensors measure hundreds information than is possible with multi- hyperspectral.
Hyperspectral 101
A passive system. Most hyperspectral the EO-1 satellite and the U.S. Air Force Image analysis broadbrush
imaging sensors, like most multispectral Research Lab’s FTHSI sensor on the Hyperspectral imagery has been used to
imaging sensors, are passive optical sen- MightySat II satellite. Several new space- detect and map a wide variety of materials
sors. This means that they measure the based hyperspectral sensors, though, have having characteristic reflectance spectra.
amount of visible and infrared radiance recently been proposed (see Table 2 on page For example, hyperspectral images have
passively reflected or emitted from the sur- 44). Unlike airborne sensors, space-based been used by geologists for mineral map-
face. Passive optical sensors operating in sensors can provide near global coverage ping (Clark et al., 1992; 1995) and to detect
visible and near- infrared wavelengths repeated at regular intervals. Therefore, the soil properties including moisture, organic
detect sunlight energy reflected by surfaces amount of hyperspectral imagery available content, and salinity (Ben-Dor, 2000).
while the sun is illuminating the Earth. should increase significantly in the near Vegetation scientists have successfully used
There is no reflected energy available from future as new satellite-based sensors are suc- hyperspectral imagery to identify vegeta-
the sun in these wavelength regions at night, cessfully launched. tion species (Clark et al., 1995), study plant
so images cannot be collected at
night. Energy that is naturally emit-
ted (such as thermal infrared), how- TABLE 1 Current hyperspectral sensors
ever, can be detected day or night. Number Spectral range
Satellite sensors Manufacturer of bands (in microns)
All hyperspectral sensors are spec-
trometers, which means that they FTHSI on MightySat II Air Force Research Lab 256 0.35–1.05
include a prism or diffraction grat- www.vs.afrl.af.mil/
TechProgs/MightySatII
ing that can break the incoming radi-
ation into discrete wavelengths, Hyperion on EO-1 NASA Goddard 220 0.4–2.5
which are then dispersed separately Space Flight Center
eo1.gsfc.nasa.gov
to detectors. Thus, hyperspectral sen-
sors are often referred to as imaging Number Spectral range
Airborne sensors Manufacturer of bands (in microns)
spectrometers.
The signal measured by an air- AVIRIS NASA Jet Propulsion Lab 224 0.4–2.5
(Airborne Visible Infrared makalu.jpl.nasa.gov/
borne hyperspectral sensor is gener- Imaging Spectrometer)
ally digitized and then recorded
onboard the aircraft, and retrieved HYDICE Naval Research Lab 210 0.4–2.5
(Hyperspectral Digital Imagery
once the aircraft lands. The signal Collection Experiment)
measured by a satellite hyperspectral
sensor is digitized and then elec- PROBE-1 Earth Search Sciences Inc. 128 0.4–2.5
www.earthsearch.com
tronically transmitted to a ground
receiving station. Most hyperspec- casi ITRES Research Limited up to 228 0.4–1.0
(Compact Airborne www.itres.com
tral data providers will convert hyper- Spectrographic Imager)
spectral image data to radiance units
before distributing the data. Although HyMap Integrated Spectronics 100 to 200 Visible to thermal infrared
www.intspec.com
hyperspectral images are usually not
georeferenced, some data providers EPS-H GER Corporation VIS/NIR (76), VIS/NIR (.43–1.05),
will georeference hyperspectral (Environmental www.ger.com SWIR1 (32), SWIR1 (1.5–1.8),
Protection System)
images to a map projection or lati- SWIR2 (32), SWIR2 (2.0–2.5),
tude/longitude grid on request. Some TIR (12) and TIR (8–12.5)
data providers offer location infor-
DAIS 7915 GER Corporation VIS/NIR (32), VIS/NIR (0.43–1.05),
mation that the end user can employ (Digital Airborne SWIR1 (8), SWIR1 (1.5–1.8),
to georeference the data. Radiometric Imaging Spectrometer)
SWIR2 (32), SWIR2 (2.0–2.5),
and geometric accuracy of the data, MIR (1), MIR (3.0–5.0),
as well as pixel size and image swath TIR (6) and TIR (8.7–12.3)
width, vary significantly among
hyperspectral sensors. It is recom- DAIS 21115 GER Corporation VIS/NIR (76), VIS/NIR (0.40–1.0),
(Digital Airborne
mended that users of hyperspectral SWIR1 (64), SWIR1 (1.0–1.8),
Imaging Spectrometer)
SWIR2 (64), SWIR2 (2.0–2.5),
data explore these issues with their
MIR (1), MIR (3.0–5.0),
data provider prior to purchasing
TIR (6) and TIR (8.0–12.0)
data.
Platforms. For the most part hyper- AISA Spectral Imaging up to 288 0.43–1.0
(Airborne Imaging www.specim.fi
spectral sensors have been airborne Spectrometer)
(see Table 1) with two recent excep-
Footnotes: VIS = visible NIR = near infrared SWIR = shortwave infrared TIR = thermal infrared MIR = mid infrared
tions: NASA’s Hyperion sensor on
Hyperspectral 101
canopy chemistry (Aber and Martin, TABLE 2 Proposed space-based hyperspectral sensors within a spectral image, or individ-
1995), and detect vegetation stress ual pixels. Now let’s discuss some rel-
(Merton, 1999). Military personnel Satellite Sensor Agencies atively common hyperspectral/mul-
have used hyperspectral imagery to ARIES-I ARIES-I Auspace Ltd tispectral image analysis methods.
detect military vehicles under partial (Australian ACRES Whole-pixel. Whole-pixel analysis
vegetation canopy and to achieve Resource Geoimage Pty. Ltd. methods attempt to determine
Information and CSIRO
many other military target detection Environment whether one or more target materi-
Earth Resource Mapping
objectives. The type of quantitative Satellite) als are abundant within each pixel in
Pty. Ltd.
information extraction in the stud- a multispectral or hyperspectral image
ies mentioned above usually requires on the basis of the spectral similarity
accurate preprocessing of the hyper- PROBA CHRIS European Space Agency between the pixel and target spectra.
(Project for
spectral imagery and collection of On Board Whole-pixel scale tools include stan-
accurate auxiliary data. Among the Autonomy) dard supervised classifiers such as
first challenges faced when per- NEMO COIS Space Technology minimum distance or maximum like-
forming quantitative analysis of (Naval Development Corp. lihood (Richards and Jia, 1999), as
hyperspectral data, for instance, are EarthMap Naval Research Laboratory well as tools developed specifically
Observer)
those encountered due to the atmos- for hyperspectral imagery such as
phere. PRISM PRISM European Space Agency spectral angle mapper (SAM) and
(Process
Atmospheric correction. When sun- Research by spectral feature fitting.
light travels from the Sun to the an Imaging SAM. Consider a scatter plot of
Earth’s surface and then is reflected Space Mission) pixel values from two bands of a
to an air- or spaceborne sensor, the spectral image. In such a plot, pixel
intervening atmosphere often scatters some spectra and target spectra will plot as points
light. Therefore, the light received at the (see Figure 6). If a vector is drawn from the
sensor may be more or less than that caused origin through each point, the angle
by reflectance from the surface alone. between any two vectors constitutes the
Atmospheric correction attempts to mini- spectral angle between those two points.
mize these effects on image spectra and is The SAM (Yuhas et al., 1992) computes a
an indispensable step before conducting spectral angle between each pixel spectrum
quantitative image analysis or change detec- and each target spectrum. The smaller the
tion using multispectral or hyperspectral spectral angle, the more similar the pixel
data. Sophisticated atmospheric correction and target spectra. This spectral angle
algorithms are available to calculate con- will be relatively insensitive to changes in
centrations of atmospheric gases directly pixel illumination because increasing or
from the detailed spectral information decreasing illumination doesn’t change the
contained in hyperspectral imagery, with- FIGURE 6 Pixel and target spectra plot as direction of the vector, only its magnitude
out additional data about atmospheric points in this scatter plot of pixel values. (a darker pixel, for instance, will plot
conditions. If a vector is drawn from the origin through along the same vector, but closer to the ori-
Spectral libraries. Another often critical each point, the angle between any two gin). Note that although this discussion
element for spectral analysis is a good spec- vectors constitutes the spectral angle describes the calculated spectral angle using
tral library. Spectral libraries are collections between those two points. a two-dimensional scatter plot, the actual
of reflectance spectra measured from mate- spectral angle calculation is based on all of
rials of known composition, usually in the et al., 1991a; Salisbury et al., 1991b; the bands in the image. In the case of a
field or laboratory. Comparing hyperspec- Salisbury et al., 1994). hyperspectral image, a spectral “hyper-
tral data with data in a spectral library can angle” is calculated between each pixel and
help a user to quickly identify the material Analysis methodology each target.
in question. Spectra from libraries can also Many image analysis algorithms have been Spectral feature fitting. Another approach
guide spectral image classifications or define developed specifically to exploit the exten- to matching target and pixel spectra is to
targets to use in spectral image analysis. sive information contained in hyperspec- examine specific absorption features in the
Many investigators collect spectral libraries tral imagery. Most of these algorithms also spectra (Clark et al., 1991). An advanced
for materials in their field sites as part of provide accurate, although more limited, example of this method, called Tetracorder,
every project to facilitate analysis of multi- analyses of multispectral data. Spectral has been developed by the U.S. Geological
spectral or hyperspectral imagery from those analysis methods usually compare pixel Survey (Clark et al., 2000). A relatively sim-
sites. Several high-quality spectral libraries spectra with a reference spectrum (often ple form of this method, called spectral fea-
are also publicly available (for instance, called a target). Target spectra can be ture fitting, is also available in a commer-
Clark et al., 1993; Grove et al., 1992; derived from a variety of sources, includ- cial image processing software product. In
Elvidge, 1990; Korb et al., 1996; Salisbury ing spectral libraries, regions of interest spectral feature fitting, the user specifies a
Hyperspectral 101
range of wavelengths within not rare, special care must be
which a unique absorption fea- taken when applying and inter-
ture exists for the chosen target. preting matched filter results.
The pixel spectra are then com- This technique filters the
pared with the target spectrum input image for good matches
using two measurements. First, to the chosen target spectrum
the depth of the feature in the by maximizing the response of
pixel is compared with the depth the target spectrum within the
of the feature in the target. Next, data and suppressing the
the shape of the feature in the response of everything else
pixel is compared with the shape (which is treated as a compos-
of the feature in the target (often ite unknown background to the
using a least-squares technique). target). Like complete unmix-
Subpixel methods. Subpixel ing, a pixel value in the output
analysis methods can be used to image is proportional to the
Courtesy of Research Systems, Inc.

calculate the quantity of target fraction of the pixel that con-


materials in each pixel of an tains the target material. Any
image. The technique can detect pixel with a value of zero or less
quantities of a target that are would be interpreted as back-
much smaller than the pixel size ground (none of the target is
itself. In cases of good spectral present).
contrast between a target and its One potential problem with
background, subpixel analysis has matched filtering is that it is
detected targets covering as little The face of this image cube shows the city of Mobile, Alabama with data possible to end up with false
as 1–3 percent of the pixel. collected by the Hymap sensor. Again, the top and right side show color-coded positive results. A solution is to
Subpixel analysis methods include pixel values measured for each pixel along the top and right edge. calculate an additional meas-
complete linear spectral unmix- ure called infeasibility. This
ing and matched filtering. measure is based on both noise
Unmixing. The set of spectrally unique sur- this is almost always true. and image statistics and indicates the degree
face materials existing within a scene are The results of linear spectral unmixing to which the matched filter result is a fea-
often referred to as the spectral endmembers include one abundance image for each end- sible mixture of the target and the back-
for that scene. Linear spectral unmixing member. The pixel values in these images ground. Pixels with high infeasibilities are
(Adams et al., 1986; Boardman, 1989) indicate the percentage of the pixel made likely to be false positives regardless of their
exploits the theory that the reflectance spec- up of that endmember. For example, if a matched filter value.
trum of any pixel is the result of linear com- pixel in an abundance image for the end-
binations of the spectra of all endmembers member quartz has a value of 0.90, then 90 Final frame
inside that pixel. A linear combination in this percent of the area of the pixel contains Hyperspectral sensors and analyses have
context can be thought of as a weighted aver- quartz. An error image is also usually cal- provided more information from remotely
age, where each endmember’s weight is culated to help evaluate the success of the sensed imagery than ever possible before.
directly proportional to the area of the pixel unmixing analysis. As new sensors provide more hyperspectral
containing that endmember. If the spectra of Matched filtering. Matched filtering imagery and new image processing algo-
all endmembers in the scene are known, then (Boardman et al., 1995) is a type of unmix- rithms continue to be developed, hyper-
their abundances within each pixel can be ing in which only user-chosen targets are spectral imagery is positioned to become
calculated from each pixel’s spectrum. mapped. Unlike complete unmixing, one one of the most common research, explo-
Unmixing simply solves a set of n linear doesn’t need to find the spectra of all end- ration, and monitoring technologies used
equations for each pixel, where n is the num- members in the scene to get an accurate in a wide variety of fields.
ber of bands in the image. The unknown analysis (hence, this type of analysis is often
variables in these equations are the frac- called partial unmixing because the unmix- References
tions of each endmember in the pixel. To ing equations are only partially solved). To view the extensive references associated
have more equations than unknowns, it is Matched filtering was originally developed with this article, please see the online
necessary to have more bands than end- to compute abundances of targets that are version at www.geospatial_online.com/
member materials. With hyperspectral data, relatively rare in the scene. If the target is shippert. 

Research Systems, Inc. –– A Kodak Company, is a technical software company that develops software applications for use in remote sensing and GIS, medical imaging,
data visualization, analysis and rapid application development. Among RSI's software products is ENVI (Environment for Visualizing Images), the easiest to use and
most advanced remote sensing software solution available. ENVI includes everything you need to display and analyze virtually any type of high spatial resolution or
high spectral resolution image data, as well as data from any single band and SAR radar sources. Contact Research Systems, Inc. at 303.786.9900 or visit us on the
Web at www.ResearchSystems.com to learn more.

You might also like