REMOTE SENSING
Real-world, Hyperspectral Operations
Drinking from the imaging spectroscopy firehose
By James Sokolowski

There is an increasing interest in hyperspectral remote sensing and a small but growing number of public and private groups are pioneering the implementation of this technology as a fundamental part of their business process. Employing a new technology is never without difficulty or risk, particularly for those at the leading edge of innovation, where there is no body of accumulated experience to draw upon.
    Hyperspectral imaging is to some degree a natural extension of the more familiar panchromatic and multispectral remote sensing, although, it must be cautioned, this interpretation is somewhat of an oversimplification. In dissecting the chain of events comprising any complete hyperspectral application, we use a 4-component model. First is data acquisition, the collection of raw data using airborne or spaceborne sensors; 2nd is data processing, which strips out undesirable artifacts and information from the raw data; 3rd is data analysis, the transformation of processed data into information, knowledge and understanding; the 4th component is the overall business process itself, which establishes the framework within which applications exist, and, ultimately, creates the need for them. Recent advances in the field - predominantly unique to hyperspectral remote sensing - give overwhelming significance to the large-scale processing of hyperspectral data in support of real-world applications.

What is Hyperspectral Remote Sensing?
The underlying power of the spectral perspective has developed since Isaac Newton first studied rainbows and used crude glass prisms to form man-made, rainbow-like spectra with the sun's light. The basic techniques involved in spreading light out into its constituent colors and studying its detailed characteristics and distribution over wavelength are collectively called spectroscopy, and hyperspectral imaging is also called 'imaging spectroscopy,' a moniker still contained in the AVIRIS (Airborne Visible to Infra-Red Imaging Spectrometer) name.
    At its core, spectroscopy depends on the obvious pretext that 'different materials are different because their constituents and structure differ and so they interact differently with light and thus, appear different. Table salt for example, looks nothing like the sodium or chlorine atoms that make up its sodium chloride crystals. Even more dramatically, carbon when laid out in sheets of interlaced hexagons forms graphite (common pencil lead), but arranged in three dimensional structures having 'face-centered, cubic geometry' creates diamonds! Imaging spectroscopy aims to understand the Earth's surface through the detailed analysis of its reflected light, exploiting subtle variations in surface composition and structure in support of real-world requirements. To allow spectroscopic study, a hyperspectral data set must have sufficient spectral detail to discern the subtle differences in color distributions from terrestrial materials. Because molecules of the solids and liquids that populate the Earth's surface have characteristic spectral features generally wider than some tens of nanometers, this naturally establishes a practical definition for the maximum spectral band size for a hyperspectral data set.
    These concepts solidify when typical reflectance spectra from two vegetation species (fir and juniper) along with two minerals (kaolinite and alunite) are plotted. The spectral differences displayed allow easy separation of the vegetation species from the minerals and multispectral applications have successfully exploited this ability. However, spectrally discerning fir from juniper, or kaolinite from alunite is obviously a more delicate process, one which could certainly benefit from hyperspectral's ability to analyze the subtle spectral variations that distinguish these very similar spectra. This is a crude example of the immediate power of hyperspectral technology in the remote sensing context, but beyond this one might well imagine the ability to extract information regarding the health and well-being of a single juniper bush based on the even more subtle spectral variations caused by leaf turgidity or environmentally induced stresses.

The Formation of Hyperspectral Signals
The perspective of a hyperspectral cube as a large stack of images is a natural extension of the smaller number of images within a multispectral set. However, this perspective also begs the obvious and unsettling question, 'how many images does it take to make a hyperspectral cube anyway?' Therefore, a fundamentally different perspective is needed; one, in which a single pixel's 'spectrum' is created by extracting the digital number from each image at that pixel's location and plotting them sequentially. This spectral perspective sees a hyperspectral cube as a collection of individual spectra spanning the area of the imagery. Then, each spectrum within the cube contains the complete color response from its particular pixel.
    In order to understand and appreciate the importance of hyperspectral data processing, it is important to examine the process by which hyperspectral signals are generated and the raw data are acquired. These basics underscore the mixed nature of raw hyperspectral data and the fundamental importance of sifting out the extraneous information and leaving only the pure Earth surface signal that is performed by hyperspectral data processing.
    As with all passive, visible to mid-infrared remote sensing, the process starts when solar illumination strikes the top of the Earth's atmosphere. The incident solar spectrum then propagates down through the atmosphere interacting with atmospheric molecules like oxygen, carbon dioxide, and water, as well as larger particles called aerosols that are suspended by the combination of winds and vertical temperature gradients. The atmospheric gasses absorb radiation in patterns characteristic of their individual spectral fingerprints, while also bouncing a fraction of the radiation around via a process called Rayleigh scattering. Aerosol particles, on the other hand, interact with radiation principally by scattering, but because of their greater sizes, more complex makeup, and distributions, their scattering has a different character than does molecular Rayleigh scattering.
    Interestingly, it turns out that both types of scattering are strongest for the shorter wavelengths (blues as opposed to reds and infrareds), which is why the daytime sky is blue, while at sunset the sky near the horizon takes on a variety of reddish tones. Eventually radiation transmitted straight to the surface (direct illumination) plus that scattered one or more times by molecules and/or aerosols (diffuse illumination) reaches the Earth's surface, where it undergoes both reflection and absorption. The absorbed portion is lost to us, while the reflected part, shaped by the characteristic reflectance spectrum of the particular surface material, travels back up through the atmosphere undergoing yet another round of absorptions and scatterings. Eventually, after traveling this tortuous path, the radiation enters the sensor (hyperspectral or otherwise), where each pixel's continuum of colors is spread apart, converted into an electronic signal, then digitized and stored by a computer.
    The result of the chain of processes described above is that the raw data (digital numbers or DNs) produced by a hyperspectral sensor contain an intricate blending of information about the Sun, which is rich in spectral detail, the Earth's atmosphere, imprinted during both downward and upward passes, the Earth's surface as imaged by the reflection process, and the specifics of the sensor hardware design and operation. From the remote sensing perspective, most of this is superfluous information or background contamination, while only the signal produced during the surface reflection has any value, economic or otherwise. Only by applying hyperspectral data processing can one remove the various background signals from this jumble and ultimately return a data product that allows the hyperspectral analyst to focus attention on the Earth's surface.
    The generic model presented above forms an excellent framework for understanding hyperspectral signal generation, as well as basic hyperspectral data processing and analysis. However, remote sensing experience has shown that one never truly finds any single pixel covered by one pure material, which has always clouded the meaning of remotely sensed classifications that generally label a pixel unambiguously as this, that or the other cover type. When one or the other completely covers a pixel, the pure reflectance signature of that material is discerned by the sensor. Then for varying mixtures of one with the other, a composite signature is generated by that pixel, one having characteristics of both materials.
    The power of hyperspectral remote sensing has been demonstrated by the detection and mapping of the concentrations of trace materials within single spectra, where spectral components have been seen at concentrations below the ten percent level. In fact, hyperspectral's ability to detect and analyze component signatures at low concentrations has led some in the field to use the term 'remote measurement' to describe these techniques.

An Operational Case Study: Noranda Mining & Exploration Inc.
In an ongoing effort to improve productivity and streamline procedures, Noranda Mining and Exploration, Limited in partnership with Falconbridge, Limited has identified hyperspectral imaging as providing as providing a significant beneficial impact for two of its basic business processes. First, is in support of Noranda's mineral exploration activities, where hyperspectral's ability to detect and map concentrations of key mineralogy, even at trace levels, will be used to survey large regions to assist in the prioritization of new mine exploration areas. Second, with increasing world-wide need to ensure the responsible development of the Earth's resources, Noranda's utilization of hyperspectral imaging will assist in the establishment of environmental baseline analyses and periodic monitoring to promote and document its environmental stewardship.
    Noranda has entered into an exclusive contract with Earth Search Sciences, Incorporated (ESSI) to acquire hyperspectral data using their Probe family of 128-band hyperspectral sensors over large-scale survey areas during campaigns lasting typically from two to four weeks each. However, raw instrument data cubes are inappropriate for the direct generation of the information products Noranda requires to support their business objectives. To augment their hyperspectral initiative, Noranda has begun working with Remote Measurement Services (RMS) to first verify the operational characteristics and variability of the Probe-acquired data sets, then to design and implement data processing and analyses strategies and systems in support of this work. To date, Noranda has acquired test data sets over Cuperite, Nevada; Goldfields, Arizona; and a number of sites in Australia.
    With increasing experience and familiarity with these Probe hyperspectral data, at the time of this writing, Noranda has begun implementation of large-scale mapping operations using ESSI's Probe-1 sensor. It is anticipated that this 4-week campaign over three large areas of Mexico, could generate nearly a Terrabyte of raw data encompassing up to 40,000 square kilometers.
    Turning such a flood of raw data into cost-effective information products poses a formidable operational challenge. To overcome these obstacles, RMS has adapted data processing technologies prototyped during our work on NASA's Lewis Hyperspectral spacecraft program, to design the HyDaPS (Hyperspectral Data Processing System). This high-volume data processing environment allows an operator to remove the effects of the sensor's conversion of light into digital numbers by examining and accommodating all variations in the sensor's response and sensitivity. Then after performing various statistical checks on the entire hyperspectral data set to ensure the fidelity and spectral purity of these data, the operator is then able to interactively compensate for the absorption and scattering effects of the atmosphere, as well as the sun's spectrum. This results in what is called a reflectance cube, where only the spectral information regarding the Earth's surface remains. At all steps along the processing pipeline, the operator is able to perform complete quality assurance of the results by examining appropriate images, spectra and various statistical indicators. Depending on the quality of the data acquired, automated processes are also possible which will provide the rapid conversion of data to information relevant to the client. These hyperspectral data processing avenues ultimately ensure that one does not drown while drinking from the imaging spectroscopy firehose.

Summary
The state of hyperspectral remote sensing technologies is advancing rapidly. We have looked at the basics of hyperspectral imaging, what it is and physically how hyperspectral signals are generated. In addition, we have discussed the importance of removing the solar, instrumental, and atmospheric background contamination from raw hyperspectral data so that the analyst can focus on the pure Earth signal. It appears that all components required to support operational hyperspectral remote sensing are coming together and we look forward to the growth of mature hyperspectral imaging applications in support of mining, forestry, agriculture, as well as general environmental analyses and monitoring over the coming years.

About the Author:
James Sokolowski is the founder and president of Remote Measurement Services, LLC. Dr. Sokolowski has a B.Sc. in experimental physics from the University of Michigan and a Ph.D. in astronomy and astrophysics from Rice University. Dr. Sokolowski has over fifteen years of advanced imaging and high-volume data processing experience in the medical, astrophysical, and Earth sciences.

Back