compensation of hyperspectral data for atmospheric effects

26
VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 29 Compensation of Hyperspectral Data for Atmospheric Effects Michael K. Griffin and Hsiao-hua K. Burke Hyperspectral imaging sensors are used to detect and identify diverse surface materials, topographical features, and geological features. Because the intervening atmosphere poses an obstacle to the retrieval of surface reflectance data, algorithms exist to compensate the measured signal for the effects of the atmosphere. This article provides an overview and an evaluation of available atmospheric compensation algorithms for the visible-through-shortwave infrared spectral region, including comparison of operational characteristics, input requirements, algorithm limitations, and computational requirements. Statistical models based on empirical in-scene data are contrasted with physics- based radiative transfer algorithms. The statistical models rely on a priori scene information that is coupled with the sensor spectral observations in a regression algorithm. The physics-based models utilize physical characteristics of the atmosphere to derive water vapor, aerosol, and mixed gas contributions to the atmospheric signal. Treatment of aerosols in atmospheric compensation models varies considerably and is discussed in some detail. A three-band ratio approach is generally used for the retrieval of atmospheric water vapor. For the surfaces tested in this study, the retrieved surface reflectances from the two physics-based algorithms are similar under dry, clear conditions but differ under moist, hazy conditions. Sensitivity of surface-reflectance retrievals to variations in scene characteristics such as the solar zenith angle, atmospheric visibility, aerosol type, and the atmospheric temperature profile is presented in an effort to quantify the limitations of the models. H have been used for more than a decade to help detect and identify diverse surface materials, topo- graphical features, and geological features. Hyper- spectral data are not immune, however, to the effects of the intervening atmosphere. Atmospheric compen- sation refers to the removal of unwanted atmospheric components of the measured radiance. For hyper- spectral data analysis, the general objective of atmo- spheric compensation algorithms is to remove solar illumination and atmospheric effects (predominantly aerosol scattering and water vapor absorption) from the measured spectral data so that an accurate esti- mate of the surface reflectance can be obtained. The retrieved surface-reflectance spectra can then be com- pared with a library collection of spectra representing various materials. Hyperspectral reflectance measurements of the so- lar and near-infrared reflectance spectrum satisfy a host of scientific research applications, including esti- mating atmospheric water vapor, cloud properties and aerosols, agriculture and forest properties, miner- alogy, soil type, snow and ice hydrology, biomass burning, environmental hazards, calibration of air-

Upload: solmazbabakan

Post on 08-Nov-2015

8 views

Category:

Documents


2 download

DESCRIPTION

Compensation of Hyperspectral Data for Atmospheric Effects

TRANSCRIPT

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 29

    Compensation of HyperspectralData for Atmospheric EffectsMichael K. Griffin and Hsiao-hua K. Burke

    Hyperspectral imaging sensors are used to detect and identify diverse surfacematerials, topographical features, and geological features. Because theintervening atmosphere poses an obstacle to the retrieval of surface reflectancedata, algorithms exist to compensate the measured signal for the effects of theatmosphere. This article provides an overview and an evaluation of availableatmospheric compensation algorithms for the visible-through-shortwaveinfrared spectral region, including comparison of operational characteristics,input requirements, algorithm limitations, and computational requirements.Statistical models based on empirical in-scene data are contrasted with physics-based radiative transfer algorithms. The statistical models rely on a priori sceneinformation that is coupled with the sensor spectral observations in a regressionalgorithm. The physics-based models utilize physical characteristics of theatmosphere to derive water vapor, aerosol, and mixed gas contributions to theatmospheric signal. Treatment of aerosols in atmospheric compensation modelsvaries considerably and is discussed in some detail. A three-band ratio approachis generally used for the retrieval of atmospheric water vapor. For the surfacestested in this study, the retrieved surface reflectances from the two physics-basedalgorithms are similar under dry, clear conditions but differ under moist, hazyconditions. Sensitivity of surface-reflectance retrievals to variations in scenecharacteristics such as the solar zenith angle, atmospheric visibility, aerosol type,and the atmospheric temperature profile is presented in an effort to quantify thelimitations of the models.

    H have beenused for more than a decade to help detectand identify diverse surface materials, topo-graphical features, and geological features. Hyper-spectral data are not immune, however, to the effectsof the intervening atmosphere. Atmospheric compen-sation refers to the removal of unwanted atmosphericcomponents of the measured radiance. For hyper-spectral data analysis, the general objective of atmo-spheric compensation algorithms is to remove solarillumination and atmospheric effects (predominantlyaerosol scattering and water vapor absorption) from

    the measured spectral data so that an accurate esti-mate of the surface reflectance can be obtained. Theretrieved surface-reflectance spectra can then be com-pared with a library collection of spectra representingvarious materials.

    Hyperspectral reflectance measurements of the so-lar and near-infrared reflectance spectrum satisfy ahost of scientific research applications, including esti-mating atmospheric water vapor, cloud propertiesand aerosols, agriculture and forest properties, miner-alogy, soil type, snow and ice hydrology, biomassburning, environmental hazards, calibration of air-

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    30 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    craft and satellite sensors, sensor simulation and vali-dation, radiative transfer modeling, and atmosphericcompensation. Figure 1 depicts the relationship be-tween spectral measurement and various features tobe retrieved. As shown in the figure, the contiguousnarrowband character of hyperspectral measurementsprovides a unique source of information for a widevariety of research areas, such as geological, ecologi-cal, and oceanographic endeavors.

    Although numerous airborne hyperspectral plat-forms exist, two platforms have been used extensivelyin this study and are described in the section that fol-lows. Using the data from these platforms, we cantake several different approaches to obtain an accurateestimation of the surface reflectance; a descriptionand comparison of the more popular techniques foratmospheric compensation is presented. Require-ments for atmospheric compensation of hyperspectraldata include the ability to account for stressing atmo-spheric conditions (e.g., high moisture, heavy aero-sol/particulate loading, partial cloud cover, low sunangle) as well as clear sky. Changing the inputs to the

    compensation model can significantly affect the re-trieved surface reflectance. The sensitivity of the re-trieved reflectance to controlled variations in themodel input parameters is analyzed and discussed.Lastly, the computational burden, an important con-sideration for physics-based models, is estimated fortwo of the models.

    Hyperspectral Data Sources

    To meet the requirements for a number of scientificapplications, airborne imaging spectrometers weredesigned to achieve substantial improvements overexisting multispectral imaging systems in the areas ofspatial resolution, sensitivity, and accuracy of absolutecalibration. The Airborne Visible Infrared ImagingSpectrometer (AVIRIS) sensor has been used exten-sively to provide hyperspectral measurements in thevisible, near-infrared, and shortwave infrared (SWIR)spectra [1]. AVIRIS contains 224 different detectors,each with a spectral bandwidth of approximately0.010 m, allowing it to cover contiguously the entirespectral range from 0.38 to 2.5 m. AVIRIS uses a

    FIGURE 1. The locations of spectral regions in the visible, near infrared, and shortwave infrared, which are used to retrieve thefeatures listed.

    Surface

    Atmosphere

    Oceanography

    Ecological

    Vegetation

    Geological

    Surface elevation

    Daytime cirrus detection

    Moisture content

    O3

    CO2

    Other gases

    Aerosols, smoke, fires

    0.7 1.0 1.3 1.6 1.9 2.2 2.50.4 m

    Features

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 31

    scanning mirror to sweep back and forth in a whisk-broom fashion, resulting in 614 pixels for each scanline. Each pixel covers a 20-m 20-m-square area onthe ground (with some overlap between pixels), yield-ing a ground swath width of approximately 11 km foran ER-2 aircraft flight altitude of 20 km. More re-cently, AVIRIS has been mounted on a low-flying (~3to 5 km) aircraft that provides higher spatial-resolu-tion imagery and smaller swaths [2].

    The Hyperspectral Digital Imagery Collection Ex-periment (HYDICE) sensor is a push-broom imagingspectrometer that uses a bi-prism dispersing elementand a two-dimensional focal-plane array (FPA) to en-able a single optical path design [3, 4]. The FPA is a320-element by 210-detector indium antimonide(InSb) array that supports operations over the full0.4-to-2.5-m spectral range. Mounted on a CV 580aircraft, HYDICE has an operating altitude range of1.5 to 8 km and spatial resolutions of 0.8 to 4 m. Atthe lowest flight altitude, the HYDICE swath widthis on the order of 270 m. Calibration is done in-flightthrough the full optical system, and is referenced to aNational Institute of Standards and Technology(NIST) ground standard.

    Atmospheric Properties

    The effects of the intervening atmosphere must beconsidered to estimate the underlying surface reflec-tance from a remote airborne platform. Typically, atwavelengths below 2.5 m, the incident solar flux isimpacted by (1) the absorption by well-mixed gasessuch as ozone (O3), oxygen (O2), methane (CH4),and carbon dioxide (CO2 ); (2) absorption by watervapor; (3) scattering by molecules; and (4) scatteringand absorption by aerosols and hydrometeors. Gas-absorption effects vary strongly with wavelength andgas properties, and can have a significant impact onthe received solar flux. Aerosol absorption, on theother hand, is relatively minor compared to molecu-lar absorption, and is generally viewed as a smoothlyvarying continuous function. Typical absorption lossdue to aerosols is limited to a few percent, with mini-mum absorption occurring for maritime aerosols andmaximum absorption found in urban aerosols thattend to contain significant amounts of carbon. Scat-tering from both molecules in the visible/near infra-

    red (VNIR) and aerosols in the VNIR/SWIR is alsofound to vary slowly and continuously with wave-length. The largest scattering effects are observed inthe visible with decreased effects at longer wave-lengths. Molecular scattering effects are significantout to 0.75 m, while aerosol scattering continues tohave an impact at SWIR wavelengths (1.3 m), pre-dominantly due to the larger aerosol particle size,which can reach 1 m. Under cloud-free conditions,which are usually desired for experimental hyperspec-tral data collections, aerosols provide the bulk of thesolar scattering, while molecular (Rayleigh) scatteringcan be important at visible wavelengths. Models havebeen developed on the basis of the size and near-spherical shape of aerosols to estimate the effects ofscattering on the incident solar flux [5]. Scattering bymolecules can be defined by the well-known Rayleighscattering theorem [6].

    Mixed gases, while making a smaller and discon-tinuous contribution to the transmittance spectrum,can be modeled accurately and play an important rolein the retrieval of specific properties of the earth-at-mosphere system [7]. Figure 2 shows the atmospherictransmittance for a number of important gaseous ab-sorption features observed in the spectral region from0.4 to 2.5 m for two atmospheric conditions: dryand clear, or humid and hazy. Weak ozone absorptionis present from 0.5 to 0.7 m. At 0.76 m, a strong,narrow oxygen absorption line is present; a weakeroxygen line is located at 1.27 m. CO2 absorbsstrongly from 1.9 to 2.1 m. The effect of the CO2absorption is partially negated by a strong overlap-ping water vapor absorption band. CO2 also exhibitsa weak absorption line at 1.43 m. Weak absorptionby methane (CH4 )is present from 2.2 to 2.5 m.

    Water vapor absorption in the VNIR spectrum ischaracterized by a number of bands of variousstrengths and spectral widths. Two very weak absorp-tion bands are located at 0.6 and 0.66 m. Slightlystronger and more significant bands are located at0.73, 0.82, and 0.91 m. At 0.94 and 1.14 m, watervapor absorption is strong enough that measurementsat the band centers and off-center locations can beused to derive the total column water vapor [810].Water vapor absorption near 1.375, 1.9 and 2.5 m isstrong enough to make retrieval of the surface reflec-

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    32 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    tance difficult or impossible. Therefore, these bandsare used to obtain information about high-altitudemoisture and cloud effects [11], since much of thesignal at these wavelengths comes from the mid-to-upper troposphere. The large number of strongly ab-sorbing water vapor bands makes the adjustment forwater vapor a significant focus of atmospheric com-pensation routines.

    While water vapor clearly has the largest effect onthe transmittance, aerosols play an important roleand must be handled properly by compensation rou-tines if an accurate estimate of the surface reflection isto be obtained. Aerosol effects can vary widely withatmospheric moisture, aerosol type, or makeup andoptical depth or visibility [5].

    The application of atmospheric compensation toretrieve surface reflectance can also lead to bettercharacterization of the atmosphere. Secondary prod-ucts of the physics-based atmospheric compensationapproaches can include a map of the integrated col-umn water vapor and scene aerosol type and visibility.The spectral information itself can be used to provide

    quantitative information about atmospheric featuressuch as clouds, dust, or smoke [12]. For an exampleof how spectral information is used to characterize ahyperspectral image, see the sidebar entitled Charac-terization of a Hyperspectral Image.

    Atmospheric Compensation Models

    Imaging spectrometers acquire images in an array ofcontiguous spectral bands such that a complete set ofreflectance or emittance spectra is measured for eachpixel in the image. The two spatial image dimensions(elements and lines) are combined with the spectraldimension to produce the hyperspectral image cube.Typical reflectance measurements from hyperspectralsensors contain information not only about the spec-tral characteristics of the surface but of the interven-ing atmosphere as well. Because these sensors are pri-marily used as a tool to derive spectral-reflectanceinformation for the earths surface, it is advantageousto be able to remove, or compensate for, the effects ofthe intervening atmosphere. Figure 3 shows an ex-ample of the uncompensated, or at-sensor, radiance

    FIGURE 2. Atmospheric transmission plotted for two conditions (dry-clear in green and humid-hazyin red). The contributions to the overall transmission by mixed gases, aerosols, and water vapor areshown in separate plots for the same spectral range.

    Total

    0.5 1.0 1.5 2.0 2.5

    0.8

    0.6

    0.4

    0.2

    0

    1.0

    Gases

    CO2

    CO2O2

    O2

    O3CH4

    0.5 1.0 1.5 2.0 2.5

    0.8

    0.6

    0.4

    0.2

    0

    1.0Aerosols

    0.5 1.0 1.5 2.0 2.5

    0.8

    0.6

    0.4

    0.2

    0

    1.0H2O

    0.5 1.0 1.5 2.0 2.5

    0.8

    0.6

    0.4

    0.2

    0

    1.0

    Wavelength ( m) Wavelength ( m) Wavelength ( m)

    Wavelength ( m)T

    rans

    mis

    sion

    Tra

    nsm

    issi

    on

    MolecularScattering

    Dry-clear

    Humid- hazy

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 33

    and reflectance spectra (radiance normalized by theincident solar flux) for simulated hyperspectral mea-surements of scenes with a spectrally uniform surfacereflectance. The radiance (Figure 3[a]) was computedfor two different atmospheric compositions: humidand hazy, and dry and clear, for two uniform surface-reflectance values (0.2 and 0.4). Differences in thetwo cases are most noticeable in the absorption bandswhere the measured radiance is strongly affected bythe moisture content of the atmosphere. The reflec-tance spectra (Figure 3[b]) display similar featuresthroughout the VNIR region. It is clear that the ef-fects of specific atmospheric constituents, such as wa-ter vapor and aerosol particles, can mask the true na-ture of a remotely sensed surface.

    Various types of atmospheric compensation mod-els exist for application to hyperspectral data. Mostfall into two basic categories: statistical or empirical,and physics based. The former use a priori knowledgeof the reflectance characteristics of specific referenceobjects (such as calibration panels) in a scene to de-velop statistical relationships between the at-sensorobservations and the surface reflectance. The empiri-cal line method (ELM) is a commonly used statistics-based atmospheric compensation model [13, 14].The ELM creates a linear-regression equation for

    each spectral band that provides a relationship be-tween the raw radiance and the surface reflectance.This process is equivalent to removing the solar radi-ance and the atmospheric path radiance from themeasured signal. Gain and offset factors for eachspectral band derived from this relationship are ap-plied to all other pixels in the scene to remove the at-mospheric component from the measurements. TheELM provides good estimates of the surface reflec-tance, assuming that the reference-object spectralreflectances are accurately known. However, informa-tion pertaining to the intervening atmosphere is notderived with this type of approach, and ELM doesnot properly account for topographic variations inthe atmospheric path, which can result in elevation-dependent residual atmospheric absorptions.

    Other statistical techniques are used to adjust ornormalize the raw radiance data. These techniquestend to be more of a calibration tool than a compen-sation method. Two such techniques are the internalaverage relative reflectance (IARR) [15] and the flatfield correction (FFC) [13]. The IARR acts to nor-malize images to a scene-average spectrum, which isparticularly effective for reducing hyperspectral datato relative reflectance in an area where no groundmeasurements exist and little is known about the

    0.5 1.0 1.5 2.0 2.50

    0.1

    0.2

    0.3

    0.4

    0.40.2

    At-

    sens

    or re

    flect

    ance

    Rad

    ianc

    e (W

    /cm

    2 -s

    r-

    m)

    Wavelength ( m) Wavelength ( m)

    0

    0.005

    0.010

    0.015

    0.020

    0.025

    0.5 1.0 2.01.5 2.5

    Surface reflectances

    (a) (b)

    FIGURE 3. Plots of (a) simulated radiance as measured by a hyperspectral sensor and (b) at-sen-sor reflectance (solid lines) for two values of a spectrally uniform surface reflectance of 0.2 and 0.4.Simulated radiances are shown for a solar zenith angle of 30 and for two atmospheric composi-tions: hazy and humid (red lines) and clear and dry (green lines). At-sensor reflectances areshown for the hazy and humid case only, for which dotted lines represent the surface reflectanceused in the simulations.

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    34 THE LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    fromhyperspectral sensors contain in-formation not only about thespectral characteristics of the sur-face but of the intervening atmo-sphere as well. The imageryshown in Figure A is from theAVIRIS sensor, where the mea-sured radiance has been con-verted to at-sensor reflectance bydividing by the incident solar ir-radiance corrected for sun angleand earth-sun distance. Thescene was collected in the foot-hills east of Linden, California,on 20 August 1992 and consistsof a forest fire emitting a thickplume of smoke toward the east(north is toward the top of theimage). A cloud produced by thestrong thermal properties of the

    fire overlies the smoke plume.Toward the northwest, two smol-dering fires emit a thin veil ofsmoke that covers much of theupper half of the scene. Thesouthwest portion of the scene isfree of clouds and smoke. Shad-ows are also observed just to thenorth of the cloud. The left im-age was generated via a naturalcolor rendition by using threebands in the visible portion ofthe spectrum (red: 0.65 m,green: 0.55 m, blue: 0.45 m).The right image is formed by us-ing visible and shortwave-infra-red bands (red: 2.14 m, green:1.62 m, blue: 0.55 m). Whilethe left image displays the visiblecharacteristics of the scene, theright image emphasizes features

    C H A R A C T E R I Z A T I O N O F AH Y P E R S P E C T R A L I M A G E

    present in longer wavelengths,where the atmosphere is moretransparent to smoke. The darkbrown patch of burnt vegetationin the region of the smoke fromthe smoldering fires and thebright red and yellow colors de-picting the location of the firehot spots are visible. The radi-ance measured near the fires isdue to a combination of reflectedsolar and emitted thermal radi-ance (from the intense heat of thefire) that is most pronounced inthe shortwave-infrared bands.

    Figure B shows a number ofspectral features that are promi-nent in these images. Regions ofmoderate to strong water vaporabsorption are found where theobserved reflectance is low. High

    FIGURE A. AVIRIS sensor red-green-blue (RGB) imagery of a scene collected in the foothills near Linden, California.The left image is a pseudo-natural color depiction. The right image is a false-color display emphasizing details visible atshortwave-infrared band signatures.

    Smokesmall particles

    Cloud

    Shadow

    Soil

    GrassLake

    Hot area

    Smokelarge particles

    Forest

    Burnt vegetation

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 THE LINCOLN LABORATORY JOURNAL 35

    reflectance features such asclouds, the thick smoke plume,and highly reflective vegetationin the near infrared have promi-nent spectral signatures. Notethat while clouds exhibit highspectral reflectance across allwavelengths, the reflectance ofthe thick smoke plume decreasesdramatically beyond the firststrong water vapor absorptionband (1.37 m). Also note the re-flectance difference between thesmoke (large particles) that over-lays burnt vegetation and the un-disturbed forest vegetation in thenear infrared, and to a greater ex-tent in the shortwave infrared.

    FIGURE B. Spectral curves for nine features identified in the AVIRIS sensorimagery of Figure A.

    scene. It works best for arid areas with no vegetation.An average spectrum calculated from the entire sceneis used as the reference spectrum, which is dividedinto the spectrum at each pixel of the image. TheFFC works similarly, normalizing the hyperspectraldata to an area of known flat reflectance, and again re-ducing the hyperspectral data to relative reflectance.

    When true atmospheric compensation is desiredand direct information about a scene is unknown, sta-tistical algorithms may not be applicable. For thesecases, a more robust algorithm is needed to retrievethe surface characteristics. Typically, physics-basedmodels are chosen for this task. Two of the more com-monly used models are Atmospheric Removal(ATREM) [16] and Fast Line-of-Sight AtmosphericAnalysis of Spectral Hypercubes (FLAASH) [17].While ATREM and FLAASH apply different meth-ods of atmospheric compensation, both use variationsof the three-band ratio techniques [8, 9] to accountfor the effects of water vapor on the hyperspectralmeasurements. The methodology of the three-bandratio technique and the model implementation dif-ferences are described in more detail in a later section.

    Atmospheric Compensation Theory

    For physics-based models, the relationship betweenthe surface reflectance s and the at-sensor reflectance* can be represented by the following equation:

    * ,= +

    +

    ( )

    a gs g

    adj

    adj s g

    adjT

    T T T T T T

    S Sr

    1 1 (1)

    where Tg is the gaseous transmittance, a is the atmo-spheric reflectance, T T and are the upward anddownward scattering transmittances, r is the diffuse-to-total transmittance ratio for the ground-sensorpath, and S is the spherical albedo of the atmosphere.The parameter adj represents the surface reflectanceaveraged over a region around the pixel to account forscattering into the pixel-to-sensor path, known as theadjacency effect. While FLAASH has the option ofperforming adjacency calculations, ATREM does notincorporate scattering from adjacent pixels into itscomputations. In ATREM, a simplification to Equa-tion 1 is made of the form adj = s . Using this ap-

    0.1

    1.0

    10

    At-

    sens

    or re

    flect

    ance

    0.4 0.7 1.0 1.3 1.6 1.9 2.2 2.5

    Wavelength ( m)

    GrassForestLakeSoil

    CloudHot areaShadowSmoke (small particles)Smoke (large particles)

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    36 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    proximation, and given a measured radiance spec-trum, we can derive the at-sensor reflectance from thesensor radiance and then compute the surface reflec-tance from the following equation:

    *

    *

    *.=

    +

    T

    T T ST

    ga

    ga

    (2)

    Figure 4 shows the schematic flow of the ATREMprogram. ATREM uses a narrowband spectral model[7] to derive the gaseous transmittance and the Sec-ond Simulation of the Satellite Signal in the SolarSpectrum (6S) code [18] to compute the necessaryscattering terms by using a user-selected aerosolmodel. The individual gas concentrations are as-sumed to be uniform across the scene except for watervapor, which the algorithm treats separately. In thismanner, only a single scene transmittance spectrum iscalculated for each uniform gas. Water vapor trans-mittances, however, are calculated on a pixel-by-pixelbasis. ATREM provides a simple input parameter in-terface and relatively fast execution. Input parametersare limited to basic atmospheric parameters (aerosolvisibility and model type, atmospheric model, andozone concentration) and solar and viewing geom-etry. The model processes both AVIRIS andHYDICE spectral hypercubes and has features for

    processing hyperspectral data cubes from other sen-sors such as Hyperion.

    FLAASH [17] utilizes the full Moderate Resolu-tion Transmittance (MODTRAN)* radiative-transfercode [19] functionality, which includes a completepackage of transmittance and scattering methods, in-cluding a full accounting for adjacency effects (thescattering from adjacent pixels into the current pixel-sensor line of sight) associated with atmospheric scat-tering. Figure 5 shows the schematic flow of theFLAASH code. FLAASH has multiple spectral reso-lution options to increase processing speed. In addi-tion to retrieving the surface reflectance and the col-umn water vapor amount (for each pixel in a scene), italso offers the options of retrieving the aerosol opticaldepth (using a dark-pixel approach) and the terrainelevation (using the 0.76-m O2 line), and of mask-ing out clouds. The technique used by FLAASH toderive the surface spectral reflectance, which includesthe adjacency contribution, employs Equation 1.

    Model Comparisons

    For all atmospheric compensation models, the pri-mary output product is the surface reflectance. Forphysics-based models to derive an accurate estimateof the surface reflectance, the effects of aerosols, watervapor, and other mixed gases must be taken into ac-

    * Throughout this article, all mentions of MODTRAN refer to ver-sion 4 of the software.

    FIGURE 4. Schematic flow of the physics-based Atmospheric Removal (ATREM) program. The diagram shows thevarious steps used to convert the radiance data to surface-reflectance data.

    Surface-reflectance

    cube

    Input: sensor information, solar

    zenith angle, aerosol

    Water vaporlook-up table

    Radiative transfer code with standardatmosphericconstituents

    Top-of-atmosphere

    solar irradiance

    Imageradiance

    cube

    At-sensor reflectance

    Averageintegrated

    water vapor

    Pixel-by-pixel total

    transmittance

    Three-channel water vapor

    ratio @ 0.94 m and 1.14 m

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 37

    count. Because atmospheric concentrations of mixedgases such as O2, O3, and CO2 are generally wellknown, their absorption properties can be computedquite accurately. The models must treat other gases,for which the concentration is unknown and/orhighly variable, such as water vapor, separately. Thefollowing section details the approach that bothATREM and FLAASH take to estimate the atmo-spheric water content of a scene. It was found that thecurrent technique used by ATREM to estimate col-umn water vapor could be improved with the inclu-sion of a weaker absorption band into the computa-tion. This enhancement to ATREM is discussedbelow. A comparison of the retrieved surface reflec-tance from both models for three different scenes fol-lows the water vapor discussion.

    Water Vapor Retrieval Methodology

    Both ATREM and FLAASH utilize a three-band ra-tio method to estimate the water vapor transmittanceof the intervening atmosphere. The three-band ratiotechnique selects channels in the 0.94-m and1.14-m water vapor absorption bands [10, 20]. Asshown in Figure 6, the water vapor bands are wellcharacterized by the narrowband hyperspectral chan-nels (shown as red horizontal bars in the image). Typi-cally, three to five hyperspectral channels are chosen

    to represent the strongest portion of the water vaporband. Two more sets of channels located in the off-band window (high transmission) regions, where ab-sorption due to water vapor is small are selected (oneset on each side of the water vapor absorption band).

    A mean radiance is computed for each of the threechannel sets. The two window radiances are com-bined by using a spectrally weighted average, and a ra-tio of the water vapor band radiance to the overallwindow radiance is computed, producing a pseudo-transmittance value. This value is compared with aprecomputed table of three-band ratios for the sameband specifications over a range of column water va-por amounts, as shown in Figure 7. Ratios are shownfor three water vapor bands: the default 0.94-m and1.14-m bands used in both ATREM and FLAASHand the weaker 0.91-m band. Each band has a dis-tinct array of ratio values, with the 0.91-m bandhaving the smallest range of ratios for the given spanof column water vapor values. Figure 7 shows that the1.14-m band provides the greatest sensitivity to wa-ter vapor variations for amounts less than 3 cm. Forcolumn water vapor amounts greater than about 4 or5 cm, the ratio values change very little. Thus smallvariations in the ratio values can lead to large changesin the retrieved column water vapor. Significant er-rors are possible when dealing with very moist scenes.

    FIGURE 5. Schematic flow of the Fast Line-of-Sight Atmospheric Analysis of Spectral Hypercubes (FLAASH)code. The diagram defines the basic steps used to convert the sensor measured radiance to surface reflectance.Secondary products such as the column water vapor and the aerosol optical depth can also be obtained. The dot-ted connection lines indicate an optional selection in FLAASH.

    Water vapor absorption band (e.g., 0.94 m and/or 1.14 m)

    Input:solar sensor angle,sensor information,

    and other a priorimeasurements

    Radiance values

    Imageradiance

    cube Spatialsmoothing

    foradjacencycorrection

    Column water vapor

    Aerosoloptical depthTerrain

    height

    Look-uptable

    Look-uptable

    Look-uptable

    Look-uptable

    Absorption band of well-mixed gas(e.g., O2 @ 0.76 m)

    Dark or known pixel (e.g., 0.45 m or 0.65 m)

    Surface-reflectance

    cube

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    38 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    FIGURE 7. Three-band ratios for three water vapor bands inthe near-infrared region as a function of the total water va-por content. The ratio value is computed as the water vaporband radiance relative to the adjacent window radiance. Ver-tical dashed lines represent the column water vapor for threestandard atmospheric models.

    Water vapor transmittance and column water va-por amounts for ATREM are determined from theclosest match to the measured three-band ratio valuesand are used to define the water vapor characteristics

    of the pixel. This procedure is repeated for every pixelin the image. In ATREM, the operation tends to befast, since it involves a simple ratio calculation and alook-up table (LUT) search for each pixel.

    FLAASH uses a slightly different approach to ob-tain the column water vapor. FLAASH computationsdepend on separate within-band and window radi-ances and not the ratio of the two. This process re-quires a onetime calculation of a two-dimensionalLUT and an interpolation step for each pixel. TheLUT is generated from MODTRAN calculations ofthe full gaseous absorption characteristics and aero-sol/molecular scattering for each spectral band; there-fore, taking the ratio of the two values is not required.

    Test Scenes for Water Vapor Retrievals

    Hyperspectral data collected from three different lo-cations, summarized in Table 1, were selected forevaluating ATREM and FLAASH retrievals of watervapor. Figure 8 depicts the scenes as false-color imagesobtained by displaying three channels, one in each ofthe primary colors (red, green and blue). The chan-nel-color combinations for these figures were chosen

    FIGURE 6. Atmospheric transmittance in the region of three water vapor absorption features(lines and continuum) in the near infrared. Nominal Airborne Visible Infrared Imaging Spec-trometer (AVIRIS) channels are represented by their 10-nm bandwidths (red horizontal barspositioned near the representative mean channel transmittance).

    Water vapor continuum

    1.14

    0.91 0.94

    0.8 0.9 1.0 1.1 1.2 1.3

    0.4

    0.2

    0

    0.6

    0.8

    1.0

    Tra

    nsm

    ittan

    ce

    Wavelength ( m)

    Total water vapor

    AVIRIS channels

    Tropical

    Ratio 0.91 m

    Ratio 0.94 m

    Ratio 1.14 m

    Midlatitudesummer

    Subarcticwinter

    Column water vapor (cm)

    0.4

    0.5

    0.6

    0.7

    0.8

    0.9

    1.0

    0.2

    0.3

    0.1

    00 1 2 3 4 5 6 7 8 9 10

    Rat

    io

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 39

    to enhance surface feature identification in the dis-played images. Different channel combinations canbe used to highlight other features such as haze/aero-sol clutter and atmospheric moisture content.

    Figure 9 displays the derived water vapor images,based on the FLAASH and ATREM retrievals, of col-umn water vapor for the Jasper Ridge case. As shown,the atmosphere was relatively free of water vapor, andresults from both algorithms were similar, agreeingwell with values derived from nearby National Oce-anic and Atmospheric Administration (NOAA)Na-tional Weather Service (NWS) radiosonde observa-tions (0.67 cm of water vapor at 12 UTC, 3 April1997 and 1.13 cm of water vapor at 00 UTC, 4 April1997). Similar spatial variability is found in both de-rived images; the standard deviation is less than 10%of the mean value, which is indicative of a spatiallyhomogeneous water vapor field. Surface features suchas roads, a lake, and urban structures are apparent inthe ATREM water vapor image and to a lesser extentin the FLAASH image. The appearance of these fea-tures is an indication of the failure of the three-bandretrieval approach to eliminate the surface-reflectanceeffect. Retrieval of the column water vapor over watersurfaces is a challenging task for atmospheric com-pensation models because of the extremely low sur-face radiance signal. Water surfaces are often excludedin the processing for this reason. A key assumption ofthe retrieval approach is that the surface reflectancevaries linearly across the water vapor band such thatan average of adjacent window reflectances simulatesthe value for the absorption band. For some surfacefeatures, such as urban areas, this assumption may notbe valid. Also, terrain-elevation variations can induce

    water vapor values to follow the terrain contours.Both ATREM and FLAASH assume a constant eleva-tion across the image; large variations in terrain eleva-tion can result in misleading water vapor features, ascan be observed in the left side of the column watervapor images in Figure 9.

    In contrast to the Jasper Ridge scene, the Keystonescene represents a moderately moist atmosphere, asshown in Figure 10. The derived water vapor imagesfrom ATREM and FLAASH were quite different forthis scene. ATREM retrieved unreasonably high wa-ter vapor content, while the FLAASH retrieval is closeto radiosonde observations and is also climatologi-cally more realistic. One reason for this overestima-tion of the water vapor may be due to the exclusion ofthe effects of the water vapor continuum in theATREM calculations. This possibility is exploredmore in the next section. Again, for both codes, lessthan 10% spatial standard deviation was observedacross the scene.

    The Moffett Field scene, shown in Figure 8, is lo-cated approximately 20 km from the Jasper Ridgescene. Both scenes consist primarily of a hilly ridgeand an urban area. However, in the FLAASH-derivedwater vapor image for the Moffett Field scene shownin Figure 11, a moisture gradient can be observedwith largest values in the lower left corner, decreasingtoward the upper right. The southern portion of theSan Francisco Bay is located just to the left (west) ofthe image. The foothills provide a boundary for theflow of moist air from the bay. This geography is evi-dent both in the image and from the histogram of thewater vapor, which clearly shows a bimodal distribu-tion corresponding to the low-lying urban area

    Table 1. Test Scenes for Water Vapor Retrievals from Hyperspectral Data

    Hyperspectral Image Location Scene Dimensions Date/Time Pixel ResolutionSensor (km2) UTC (m)

    AVIRIS Jasper Ridge, California 11 10 3 April 1997/2010 20

    HYDICE Keystone (Able), Pennsylvania 0.25 0.25 10 April 1997/1730 ~2

    AVIRIS Moffett Field, California 11 10 20 June 1997/1850 20

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    40 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    (moist) and the higher-altitude foothills region (dry).Figure 12 displays the retrieved column water vaporamounts for three cross sections in the image. The leftside of the abscissa corresponds to the top of the im-age. The curves in Figure 12 are consistent with thelocal topography and perceived moisture flow fromSan Francisco Bay. While the trend from top to bot-tom of the image along these cross sections is towardincreasing water vapor amounts, the fine scale struc-ture displays many features associated with both ter-rain changes and local water vapor variations.

    Enhancements to Water Vapor Retrievals

    The apparent overestimation of water vapor byATREM prompted further analysis. We found thatby using channels in the ratio computation offsetfrom the 0.94-m water vapor band, specifically inthe weaker 0.91-m band, the ATREM ratio calcula-tion produced more reasonable values for moistscenes. ATREM does not include a contribution fromthe water vapor continuum in its computation of thewater vapor effects. From Figure 6, it is clear the con-tinuum provides a significant contribution to the wa-ter vapor absorption, which may in part be respon-sible for the high retrieved column water vaporvalues. The continuum contribution is negligible at0.91 m, so any error due to its omission is mini-mized. In addition, since the absorption is weaker at0.91 m, this band is less sensitive to higher amountsof moisture. When we used the 0.91-m band ratio,the mean water vapor amounts that ATREM com-puted for the Keystone case reduced from 6.14 to5.49 cm, a modest reduction in the right direction,but still well above the observed value.

    To investigate further the use of the 0.91-m bandratio, we generated simulated AVIRIS data fromMODTRAN calculations for clear-sky, uniform sur-face-reflectance scenes under various moisture andrural aerosol conditions. As shown in Table 2, retriev-als of the column water vapor amounts from ATREMfor the simulated data with the 0.91-m band wereclose ( < 0.5 cm) to the input model values for thesesimulated test cases and to those obtained byFLAASH. The near-perfect retrieval of water vaporby FLAASH is somewhat misleading in that both thesimulated AVIRIS data and the FLAASH model are

    FIGURE 8. AVIRIS and Hyperspectral Digital Imagery Col-lection Experiment (HYDICE) test images. The top andmiddle AVIRIS images were collected at Jasper Ridge andnear Moffett Field, California, respectively. The bottomHYDICE scene was collected at Keystone, Pennsylvania.

    11 km

    ~250 m

    11 km

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 41

    dependent on MODTRAN computations. It istherefore not unexpected that the FLAASH resultswere accurate under these circumstances.

    Surface-Reflectance Comparison

    In this section, we compare the model retrieved sur-face reflectance for the three test scenes listed in Table1. For these three scenes, depicted in Figure 8, sets of

    three to four pixels were chosen whose locationmatched specific surface features. For the JasperRidge AVIRIS image, pixels representing the radiancefrom a paved road, an urban area, ridgeline vegeta-tion, and a lake were examined and the spectral sur-face reflectances were retrieved by using bothFLAASH and ATREM atmospheric compensationmodels. Identical input parameter specifications de-

    FIGURE 9. FLAASH-computed and ATREM-computed column water vapor images for the AVIRISJasper Ridge Test Scene (2010 UTC on 3 April 1997). Mean and standard-deviation column water vaporvalues are shown for both models. Surface features such as roads, a lake, and urban structures are ap-parent in the ATREM water vapor image and to a lesser extent in the FLAASH image.

    FLAASH water vapor ATREM water vapor

    Mean = 0.72 cm, = 0.07 cm Mean = 0.75 cm, = 0.05 cm

    Table 2. Retrieved Water Vapor for Simulated AVIRIS Test Cases*

    Model Surface Visual Water Derived Column Water Vapor (cm)Atmosphere Albedo Range (km) Column (cm) ATREM FLAASH

    Wavelength** 0.91 m 0.94 m 1.14 m 0.94 m 1.14 m

    Subarctic winter 0.2 23 0.42 0.42 0.47 0.39 0.42 0.41

    Subarctic winter 0.4 23 0.42 0.43 0.48 0.40 0.42 0.42

    Midlatitude summer 0.2 10 2.93 2.98 3.79 4.00 2.92 2.94

    Midlatitude summer 0.4 10 2.93 3.12 4.01 4.21 2.92 2.92

    Tropical 0.2 5 4.11 4.40 5.60 6.10 4.11 4.19

    Tropical 0.4 5 4.11 4.66 6.04 6.52 4.11 4.16

    * All cases were generated by using MODTRAN with a rural aerosol model and a 30 solar zenith angle.** These wavelengths correspond to the water vapor band-ratio curves shown in Figure 7.

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    42 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    scribing the ambient scene characteristics were pro-vided to both models. The Jasper Ridge image wascollected on a clear, dry day at 2010 UTC (1210PST). Figure 13 shows the resulting retrieved surface-reflectance comparison. The models compare well, es-pecially for the water pixel and the vegetation pixel

    (outside the near-infrared water vapor absorption re-gion). Both the road and urban comparisons showslightly higher retrieved reflectance differences (stillless than 0.02), most noticeably near the weak watervapor absorption bands and the 1.6-m atmosphericwindow. In Figure 13 and subsequent figures, the re-

    Column water vapor (cm)

    800

    400

    200

    0

    600

    1.51.00.5

    Num

    ber o

    f pi

    xels

    FIGURE 11. Retrieved integrated water vapor AVIRIS image from the Moffett Field test case(1850 UTC on 20 June 1997) and the associated histogram showing a bimodal distribution ofwater vapor in the image. Vertical lines correspond to column water vapor cross sectionsshown in Figure 12.

    FLAASH water vapor ATREM water vapor

    Mean = 3.40 cm, = 0.25 cm Mean = 6.14 cm, = 0.43 cm

    FIGURE 10. FLAASH-computed and ATREM-computed water vapor images for the HYDICE KeystoneTest Scene (1730 UTC on 10 April 1997). Mean and standard-deviation column water vapor values areshown for both models; nearby radiosonde values of 2.97 and 3.25 cm were observed.

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 43

    trieved surface reflectance is zeroed out near the 1.38-and 1.88-m water vapor absorption bands, since thecomponent of the measured signal emanating fromthe surface is almost completely masked by the strongwater vapor absorption.

    In the Keystone HYDICE image, the four pixelschosen represented the radiance from a bare soil sur-

    face, an unpaved road, and two different types oftrees. The Keystone data were collected in Augustduring a typical warm, hazy day. Identical input pa-rameters were again provided to the models. Figure14 shows the spectral surface reflectances retrieved byusing ATREM and FLAASH. The comparison isgood, especially above 1.5 m. The largest differencesoccur near the water vapor bands where exaggerationsof reflectance values, or spiking, occurs. Spiking maybe due to inaccurate band wavelength assignment.The spiking artifact is largest for ATREM retrievals,presumably because of its difficulty in processingscenes where moist atmospheric conditions prevail.Inaccurate spectral wavelength values tend to enhancethe moisture retrieval errors. FLAASH also displayssome reflectance spiking, but to a lesser degree thanATREM.

    Data for the third test scene were collected nearMoffett Field, California. Ambient conditions ob-served at the time of the collection were somewhatwarmer and more humid than the Jasper Ridge scene(located approximately 20 km to the northwest), butstill much dryer than the HYDICE scene discussedpreviously. Figure 15 shows the retrieved surfacereflectances modeled by ATREM and FLAASH forthree surface features (paved road, green ridge, and

    FIGURE 12. FLAASH-computed column water vaporamount (cm), shown as a function of pixel location alongcross sections from top to bottom of the AVIRIS image inFigure 11. The three curves correspond by color to thecross-section lines in that image.

    FIGURE 13. Comparison of retrieved surface reflectances from ATREM andFLAASH for the Jasper Ridge AVIRIS image. Four surface typesroad, ur-ban, ridge, and waterwere located in the image and their spectral surfacereflectances retrieved.

    Pixel location

    1.10

    1.00

    0.90

    0.80

    0.70

    0.60

    100 200 300 400 500 600

    Col

    umn

    wat

    er v

    apor

    (cm

    )

    FLAASHATREM

    Wavelength ( m)

    0.4

    0.2

    0.3

    0.1

    0

    0.5 1.0 1.5 2.0

    Ref

    lect

    ance

    RoadUrbanRidgeWater

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    44 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    brown grass). The ATREM and FLAASH resultscompare well, with absolute model differences in re-trieved reflectance of less than 0.01 across the spec-trum for all three surface features. Interestingly, thereflection spikes seen in the other two comparisonsare less prominent here; the reason is not apparent.

    For all three cases, the comparison betweenATREM and FLAASH was quite good. The largestdifferences were observed in the 0.6-to-1.1-m spec-tral region, where numerous water vapor absorptionbands can cause the retrieved surface reflectance toexhibit sharp spikes with differences greater than 0.05

    FIGURE 14. Comparison of retrieved surface reflectances from ATREM andFLAASH for the Keystone HYDICE image. Four surface typessoil, unpavedroad, and two types of treeswere located in the image and their spectralsurface reflectances retrieved.

    FIGURE 15. Comparison of retrieved surface reflectances from ATREM andFLAASH for the Moffett Field AVIRIS image. Three surface typesgreenridge, brown grass, and paved roadwere located in the image and theirspectral surface reflectances retrieved.

    SoilUnpaved roadTrees 1Trees 2

    FLAASHATREM

    Wavelength ( m)

    0.5

    0.2

    0.3

    0.4

    0.1

    0

    0.5 1.0 1.5 2.0

    Ref

    lect

    ance

    0.4

    0.3

    0.2

    0.1

    0

    0.5 1.0 1.5 2.0

    Ref

    lect

    ance

    FLAASHATREM

    Wavelength ( m)

    Green ridgeBrown grass Road

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 45

    from the accepted truth value. There was also a smallincrease in the model differences for the moist, hu-mid HYDICE case over the drier AVIRIS cases. Thisdifference was not believed to be an instrument arti-fact, but rather a known difficulty with ATREM indetermining the column water vapor for vapor-ladenscenes and the subsequent effects on retrievalcomputations.

    Sensitivity to Input Parameter Specification

    As discussed previously, atmospheric compensationmodels have been developed to isolate the surface re-flectance signal and remove unwanted atmosphericand illumination affects. These models all requiresome a priori information about the atmospheric andsurface characteristics, which is used in the compen-sation process. Incomplete knowledge or inaccurateestimation of certain input parameters adds a degreeof error to the retrieval of the surface reflectance. Toillustrate the sensitivity of the retrieved surface reflec-tance s to mis-specification of FLAASH input pa-rameters, we performed a study by using estimatedinput parameters to depict the atmospheric state forthe Moffett Field AVIRIS image shown in Figure 8.The study compared retrieved surface reflectance atthree specific locations in the image obtained fromFLAASH runs by using two sets of physically reason-able input specifications. These results, along with theinput specifications for both runs, are given in Figure16. Under these circumstances, differences in the re-trieved s values can range up to 0.11 (absolute differ-ence), depending on the surface feature and the wave-length. For most remote sensing applications, theseerrors are unacceptable. Accurate specification of theinput parameters is clearly an important part of thehyperspectral image analysis process.

    Further sensitivity analysis utilized simulatedAVIRIS data generated by using MODTRAN calcu-lations for clear-sky, uniform surface reflectancescenes under three different moisture and visibilityconditions. The conditions ranged from a relativelyclear and dry atmosphere to one with hazy, wet condi-tions, as listed earlier in Table 2. Using these test-datacubes, we can vary specific input parameters to the at-mospheric compensation models and determine theeffect they have on the retrieved surface reflectance.

    Comparison to the known s provides an absolutemeasure of the sensitivity to the varied input param-eter. The following subsections attempt to isolate thedegree to which errors in each of the following atmo-spheric compensation model-input parameters canaffect the final retrieved s: visibility, aerosol model,atmospheric model, and solar zenith angle.

    Visibility

    Atmospheric visibility is inversely proportional to theoptical depth. The optical depth varies in part as afunction of the aerosol and moisture content of thelower 2 to 3 km of the atmosphere. Water vapor ab-sorbs and aerosols scatter the radiance in proportionto their concentration. In general, the higher theaerosol concentration (or optical depth), the lowerthe visibility. Hence visibility is an indicator of theamount of attenuation in the lower atmosphere andtherefore a useful characterization for the atmo-spheric compensation process.

    To examine the sensitivity of the models (FLAASHand ATREM) to uncertainty in visibility, we used twomodel surface-reflectance values0.2 and 0.4both

    FIGURE 16. Sensitivity of FLAASH to atmospheric charac-terization for surface features. The spectral surfacereflectances are retrieved for three surface types in theAVIRIS Moffett Field scene by using two different sets ofmodel-input conditions. The input specifications for the tworuns are (1) midlatitude summer atmosphere, urban aerosolmodel, and 16-km visibility, and (2) U.S. Standard atmo-sphere, rural aerosol model, 60-km visibility, and time offseterror of 20 minutes (equivalent to a 2 error in the solar ze-nith angle).

    0.5 1.0 1.5 2.0 2.5

    Wavelength ( m)

    0

    0.10

    0.20

    0.30

    0.40

    0.50

    Ret

    riev

    ed s

    urf

    ace

    refl

    ecta

    nce Green ridge

    Brown grassRoad

    MidlatitudeU.S. Standard atmosphere

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    46 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    spectrally invariant. Figure 17 displays the absolutedifference in the surface reflectance (shown as the re-trieved reflectivity error) for FLAASH runs usingthree assumed input visibilities (5, 10, and 23 km) atthe two model s values. The error is computed bysubtracting the model s values (0.2 and 0.4) from theretrieved s values, for three FLAASH runs with thethree different visibility values. All runs utilized amidlatitude summer atmosphere with a rural aerosolmodel. The truth visibility value is 10 km. BecauseMODTRAN and FLAASH share the same radiative-transfer code, we expect FLAASH will accurately re-trieve the test cube reflectances, given the correct in-put parameters. Such is the case, as denoted by thegreen line in Figure 17.

    The reflectivity errors for both the 5-km and 23-km visibility runs are within + 0.03 of the true valuesexcept near the 0.94-m and 1.14-m water vaporabsorption bands. The sloping error lines in theshortwave region are the result of the increased effectof aerosol and Rayleigh scattering. At these wave-lengths, the reflectance due to aerosol and Rayleighscattering dominates the measured reflectance signal(for lower s values, 0.2). When the visibility of ascene is underspecified (hazier), the resulting increasein perceived atmospheric reflectance is compensatedfor by an underestimation of s. In the near infrared,where atmospheric scattering is not as dominant, theperceived reduction in the atmospheric transmissionis compensated for by an overestimation of the sur-face reflectance. An analogous situation results for anoverspecification of the visibility (clearer), althoughthe error values are not quite as large.

    Figure 18 shows the retrieved reflectance errors forthe ATREM runs. ATREM does not reproduce thesurface-reflectance values as accurately as FLAASHfor a number of reasons. ATREM model parameteri-zations are based on less accurate molecular bandmodels and different aerosol models than those usedin MODTRAN. Also, ATREM decouples the ab-sorption from the scattering processes in the radia-tive-transfer model calculations. This effect can beseen in Figure 18(a), in which the model truth s val-ues are, as before, subtracted from the retrieved s val-ues for the three optimal visibility values (5, 10, and23 km). Here the water vapor absorption line struc-

    ture is still quite apparent in the spectral reflectancecurves. To remove the absorption and other model ef-fects from the reflectance, we subtract the ATREM-retrieved s values for the 10-km visibility run (thetruth visibility) from the retrieved s values for theother visibility runs, with results shown in Figure18(b). These error curves, which are in effect com-pensated for model-induced differences, are muchmore similar to what FLAASH produces.

    For surfaces with moderate or high reflectances, ifthe value assumed for the visibility, as input to themodel, is greater (clearer) than the actual truth vis-ibility of the scene, the retrieved surface reflectance isunderestimated. The converse is also true: an inputvisibility smaller (more opaque) than the truth vis-ibility results in an overestimate of the surface-reflec-tance values.

    Aerosol-Model Type

    There are a variety of predefined aerosol models thatare available for use in ATREM and FLAASH. Themodels typically represent the characteristics of aero-sols found in the lowest 2 km (within the atmosphericboundary layer) for a set of basic topographic types:desert, maritime, rural, and urban. Each model con-sists of a weighted mixture of four basic kinds of aero-sol particles: dust, oceanic, water-soluble, and soot.Urban, for example, has more soot and water-soluble

    FIGURE 17. FLAASH-retrieved surface reflectance minusmodel-input surface reflectance for three input visibilityconditions. The model-input surface-reflectance values areeither 0.2 (dotted curves) or 0.4 (solid curves) and are spec-trally invariant. The true scene visibility is 10 km.

    0.03

    0.02

    0.01

    0

    0.01

    0.02

    0.03

    0.04

    Ret

    riev

    ed re

    flect

    ance

    err

    or

    5-km visibility10-km visibility23-km visibility

    0.5 1.0 1.5 2.0 2.5Wavelength ( m)

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 47

    particles than maritime, which consists of predomi-nantly oceanic particles such as salt and sea-foamevaporates. The rural (or continental) aerosol modelconsists of a large percentage of dust particles. Withsuch a diversity of scatterers, the choice of an aerosolmodel for a particular scene can have a significant ef-fect on the radiative transfer in the lower atmosphereand on the retrieved surface reflectance.

    Figure 19 displays the sensitivity of the FLAASH-retrieved reflectance to various aerosol models for twohighly different atmospheres. For the dry and rela-tively clear rural aerosol case in Figure 19(a), the ef-fect of varying the aerosol model produces errors inretrieved reflectance of less than + 0.02, except at vis-ible wavelengths, where the effect gradually increaseswith decreasing wavelength for the urban model only.The results obtained with the urban aerosol modelseem to be most sensitive to the atmospheric compo-sition, as can be seen for the tropical, hazy test case inFigure 19(b). While the other model errors in re-trieved reflectance are generally less than 0.04, the er-rors for the urban model results are within 0.10 atwavelengths greater than 1 m but increase rapidly atshorter wavelengths. This behavior seems to indicate

    a strong sensitivity of the urban model results tomoisture and visibility. The significant degree of sootin the model can have a strong absorbing effect underlow visibility (high aerosol density) conditions mask-ing the signal from the surface and producing errone-ous results from the model.

    Atmospheric Model

    In MODTRAN, there are six choices for model at-mospheres. Each model atmosphere includes a set ofprofiles defining the pressure, temperature, air den-sity, water vapor, and ozone characteristics representa-tive of the seasonal conditions for a geographic re-gion. Surface temperatures for these atmospheres varyfrom 257 K for a subarctic wintertime atmosphere to300 K for a tropical atmosphere. Table 2 shows thatthe amount of moisture an atmosphere can supportvaries greatly over the range of standard atmospheresused in MODTRAN. This variation is a function ofthe temperature profile: i.e., colder temperaturessaturate at lower levels of moisture than higher tem-peratures, which limits the choices of model atmo-spheres for a specific scene. Errors result if a modelatmosphere incapable of supporting the moisture

    FIGURE 18. ATREM-retrieved reflectances with (a) model-input surface reflectance subtracted and (b) the ATREM retrieved re-flectance using the true visibility as input subtracted, for three input visibilities. Truth scene values are midlatitude summer at-mosphere with a rural aerosol model and 10-km visibility. The model-input surface-reflectance values are either 0.2 (dottedcurves) or 0.4 (solid curves) and are spectrally invariant.

    5-km visibility10-km visibility23-km visibility

    5-km visibility10-km visibility23-km visibility

    0.5 1.0 1.5 2.0 2.5

    Wavelength ( m)0.5 1.0 1.5 2.0 2.5

    Wavelength ( m)

    0.09

    0.12

    0.06

    0.03

    0

    0.03

    0.06

    0.09

    0.12

    0.09

    0.12

    0.06

    0.03

    0

    0.03

    0.06

    0.09

    0.12

    Ret

    riev

    ed re

    flect

    ance

    err

    or

    (b)(a)

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    48 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    FIGURE 19. FLAASH-retrieved reflectance errors for four aerosol model types. Values are shownfor two atmospheric scenes, (a) dry and relatively clear (subarctic winter atmosphere, rural aerosolmodel, 23-km visibility) and (b) humid and hazy (tropical atmosphere, rural aerosol model, and 5-kmvisibility). The model-input surface-reflectance values are either 0.2 (dotted curves) or 0.4 (solidcurves) and are spectrally invariant.

    content of the scene is used. A set of model atmo-spheres was used with this limitation in mind as inputto FLAASH to study the sensitivity of the model re-trievals to changes in the atmospheric profiles.

    Results of varying the atmospheric model type(and associated moisture and temperature profile)seem to have little effect on the retrieved reflectancevalues as long as the input model temperature profilecan support the total column water vapor of thescene. In Figure 20, retrieved reflectance errors aregenerally less than + 0.01, except near the water vaporabsorption bands. FLAASH performs well in estimat-ing the actual water vapor content when using anymoderately moist atmospheric profile. Similarly, inrelatively dry scenes, FLAASH internally adjusts themoisture content of the model to provide the closestapproximation to the scene moisture amount. How-ever, the converse does not always apply. Using a cold,dry atmospheric model as input when processing amoist scene can lead to erroneous results.

    Solar Zenith Angle

    For a typical AVIRIS scene, the date, time, and loca-tion are known accurately; therefore, the solar zenithangle can be computed with high precision. The plotsin Figure 21 show the errors incurred by using a con-stant solar zenith angle (SZA) assumption for all

    scenes taken at different times during a flight. For aforty-minute flight centered in time at solar noon, theSZA varies by approximately 2 (from a mid-flightvalue of 30) and the retrieved reflectance errors arewithin + 0.04. As would be expected, these errors in-crease for a two-hour flight exceeding 0.10 at visiblewavelengths. Specifying input of a larger SZA thanthe true value results in an overestimate of the re-trieved reflectance value; the converse is also true. It is

    FIGURE 20. FLAASH-retrieved reflectance errors for runsmade with three atmospheric models (truth atmosphere ismidlatitude summer). The model-input surface-reflectancevalues are either 0.2 (dotted curves) or 0.4 (solid curves) andare spectrally invariant.

    0.5 1.0 1.5 2.0 2.5

    Wavelength ( m)0.5 1.0 1.5 2.0 2.5

    Wavelength ( m)

    0.10

    0.10

    0.20

    0.30

    00

    0.10

    0.10

    0.20

    0.30

    Ret

    riev

    ed re

    flect

    ance

    err

    orUrbanRuralMaritimeDesert

    UrbanRuralMaritimeDesert

    (a) (b)

    TropicalMidlatitude summerSubarctic summer

    0.02

    0.01

    0

    0.01

    0.02

    Ret

    riev

    ed re

    flect

    ance

    err

    or

    0.5 1.0 1.5 2.0 2.5Wavelength ( m)

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 49

    also useful to note that errors tend to be smaller at lowSZAs (high sun elevation) than for high SZAs (lowsun elevation) because of the differing rates of changein the SZA at those times.

    Computational Complexity

    Until now, this discussion on the merits of the atmo-spheric compensation models has not considered theeffect these operations have on the overall computa-tional budget for processing hyperspectral data cubes.The decision on how best to use these models requiresan inquiry into the benefits versus computationalcosts. The primary metric for computation time/complexity is the number of floating-point opera-tions (FLOPs) per data processing task. The FLOPsmetric has the advantage of being independent of theprocessor speed. On the other hand, it does not takeinto account the cost of I/O (file reading and writing)time, which can be a significant portion of the atmo-spheric compensation processing time.

    Basic Atmospheric Compensation Computational Costs

    ATREM calculates the reflectance for each pixel ana-lytically from Equation 2. The solution to this equa-

    tion requires 6 FLOPs and 3 one-dimensional LUTs(each comprising 3 FLOPs), for a total expectednumber of FLOPs per pixel per channel of 15. For theFLAASH model, the computational complexity is di-rectly related to the selected complexity of the modelanalysis. With the adjacency option turned off,FLAASHs calculation is mathematically equivalentto ATREMs, so it is possible to use equivalent codeand achieve the same number of FLOPs (15 per pixelper channel). With the adjacency option activated,the number of FLOPs per pixel per channel is ex-pected to increase by a factor of two or three. [21]. Analternative method that involves tabulating param-eters from Equation 1 in multidimensional LUTs canresult in a computational savings of one-third, yield-ing a value similar to that for the ATREM model.

    Computational Costs of EnhancedAtmospheric Compensation

    While ATREM is limited to the basic atmosphericcompensation process described above, a number ofenhancements to the basic FLAASH processing canhave a considerable impact on the net computationalburden. These enhancements include compensation

    FIGURE 21. FLAASH-retrieved reflectance errors for a constant solar zenith angle (SZA) in the model calculations for twoflight lengths (40 minutes and ~2 hours). Values are shown for two atmospheric scenes, (a) dry and relatively clear (subarcticwinter atmosphere, rural aerosol model, 23-km visibility) and (b) humid and hazy (tropical atmosphere, rural aerosol model and5-km visibility), both with a true SZA value of 30. The SZA for each of the offset times is shown in the graphs.

    t = 1.2 hrs t = 20 min

    t = 0 t = +20 min

    t = +1.1 hrs

    t = 1.2 hrs t = 20 min

    t = 0 t = +20 min

    t = +1.1 hrs

    43

    32

    30

    28

    20

    43

    32

    30

    28

    20

    0.5 1.0 1.5 2.0 2.5

    Wavelength ( m)0.5 1.0 1.5 2.0 2.5

    Wavelength ( m)

    0.10

    0

    0.10

    0

    0.10

    0.20

    0.10

    0.20

    Ret

    riev

    ed re

    flect

    ance

    err

    or

    (a) (b)

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    50 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    Table 3. Increase in Computational BurdenAssociated with FLAASH Code Enhancements

    FLAASH FLOPs per pixel Enhancement per channel

    Aerosol compensation Double

    Sensor angular Doublefield-of-view correction

    Thin cloud (cirrus) Add 510compensation

    Spectral polishing Add 625

    Smile correction Add 032

    for nonuniform aerosol content, sensor angular field-of-view variations, cirrus clouds, spectral polishing,and channel wavelength variations. While each en-hancement can provide improved surface-reflectanceretrievals under certain scenarios, they all require fur-ther processing of the data.

    After clouds and water vapor, aerosol amount (asindicated by the visibility parameter) is the next mostvariable constituent in the atmosphere. FLAASH uti-lizes an approach that finds dark (low reflectance) pix-els in an image and performs an extended radiativetransfer calculation to obtain a number of visibilityvalues for the scene [22]. While the computational re-quirements increase in proportion to the number ofvisibility values determined, a doubling of the num-ber of FLOPs is not an unrealistic estimate.

    Across a scene, the line of sight (LOS) to the sensorvaries. For near-nadir views, most atmospheric com-pensation codes assume a single LOS associated withthe center of a scene. However, for off-nadir measure-ments, it may be necessary to perform the radiative-transfer calculations at multiple viewing angles to ac-count for the variations in absorber column densities.The increased FLOPs required to account for thisadded computational burden are similar to the aero-sol content computation.

    The ability to identify thin cirrus clouds has beendemonstrated by using spectral channels located inwater vapor bands, and the potential exists for com-pensating the surface reflectance to account for thepresence of clouds. One method involves the subtrac-tion of a cloud radiance or reflectance signal from theobserved radiance [11]. An approximate reflectanceor radiance post-processing method requires only afew additional FLOPs per pixel channel. A similarmethod is being implemented in FLAASH [23].

    Spectral polishing is a mathematical method for re-moving artifacts from reflectance spectra. Whenproperly implemented, it dramatically reduces spuri-ous, systematic spectral structure due to wavelengthregistration errors and molecular absorption residualswhile leaving true spectral features intact. A numberof processing steps are involved, including spectralsmoothing, reference-pixel selection, scaling param-eter determination, and application of the lineartransformation to the data cube. Since polishing is a

    post-processing step, the computation time simplyadds to that required to generate the original reflec-tance spectra, an increase of 6 to 12 FLOPs per pixelper channel.

    Spectral smile refers to a wavelength calibrationproblem such as that experienced with the HYDICEspectrograph sensor. The problem arises from the ten-dency of spectrographs to have a slight variation indispersion along the dimension of the entrance slit.This means that each row of the array has a slightlydifferent wavelength calibration, which translatesinto small, known spectral shifts in the data that de-pend on the pixel location or sample number alongthe cross-track direction of the data cube. Onemethod of compensating for smile is through spectralpolishing. Another method is to compensate explic-itly for the wavelength shifts in the atmospheric com-pensation algorithm. The latter can result in an in-crease in FLOPs analogous to the aerosol visibilitydetermination.

    Table 3 shows a summary of the increase in com-putational burden due to the enhancements to theFLAASH code. We see that correcting for aerosol orLOS angle variations can significantly increase thecomputational burden. Other enhancements such asspectral polishing and smile compensation are rela-tively efficient and provide little additionalcomputation.

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 51

    FLOPs Measurements

    The computationally significant components for at-mospheric compensation include the radiative-trans-fer calculations to set up the basic reflectance re-trieval, perform the basic retrieval, and make anyadditional computations needed to enhance the basicretrieval. Below we break down each of these compo-nents into computational tasks to determine the totalcomputational burden.

    Radiative-transfer calculations for ATREM runswere obtained by computing the number of FLOPsfor scenes with various pixel dimensions and extrapo-lating the curve (assuming a linear change in FLOPswith number of pixels) to the limit as the number ofpixels approaches zero. This value depicts the compu-tational burden of the radiative-transfer calculationsand was found to be 8 108. While FLAASH has anumber of variations regarding how the radiative-transfer calculations are obtained, for a typical modelrun, the computed FLOPs count was 7 1010, ap-proximately two orders of magnitude greater than theATREM run.

    The total FLOPs may be approximated as the sumof the radiative-transfer FLOPs and the reflectance-retrieval FLOPs, of which the latter scales with thenumber of pixels and channels. Estimated results for a200-channel sensor are shown in Figure 22 as a func-tion of the number of pixels for the current FLAASHand ATREM codes, for an optimized FLAASH, andfor an optimized version of FLAASHs adjacency-cor-rected method. Figure 22 shows that the performanceof optimized versions of FLAASH is similar in com-putational burden to ATREM.

    Implications and Alternatives

    Compensating for the atmospheric effects for a fullframe of hyperspectral data can mean processing mil-lions of pixels of spectral information, a time-con-suming task. Various methods have been explored toreduce the amount of compensation needed to ana-lyze a scene. Anomaly identification is a method thatinvolves identifying small numbers of pixels that arespectrally unique. These unique groups of pixels takeup only a very small portion of the image. The tech-nique can be applied to data prior to correcting for at-

    mospheric effects. Typically, anomaly identification isused as an initial filtering step to select a small subsetof potentially significant pixels from the scene. Theatmospheric compensation can be applied to thatsubset of pixels rather than to the entire data cube,saving computation time.

    With FLAASH, there are several possibilities forspeeding up the calculations. Using a coarser bandmodel (15 cm1 versus 1 cm1) can reduce the FLOPsfor a typical run by an order of magnitude. Coarseratmospheric layering can save an additional factor oftwo. Another approach is to eliminate the radiative-transfer calculation step completely by interpolatingfrom a comprehensive precalculated database thatcovers all anticipated geometries and atmosphericconditions.

    Water vapor computations are normally per-formed on a pixel-by-pixel basis. Some computationtime can be saved if this operation is performed on aregional scale. For example, a clustering type of algo-rithm could be applied to the scene radiance (or onthe measured three-band radiance ratios) to grouppixels with similar water vapor characteristics and toreduce the amount of processing time required.

    Currently, physics-based and pixel-level atmo-spheric compensation algorithms are computation-

    FIGURE 22. Estimated floating-point operations (FLOPs)versus data-cube size for selected atmospheric compensa-tion methods.

    109

    1010

    1011

    1012

    FLO

    Ps

    105 106 107

    Pixels

    1998 FLAASH betaOptimized FLAASH, adjacency

    ATREM 3.0Precalculated radiative transfer,

    no adjacency

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    52 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    ally burdensome. Achieving efficient atmosphericcompensation may require the use of a multiproces-sor computing system. Many aspects of the atmo-spheric compensation process can be applied simulta-neously to individual pixels by dividing the workloadamong the processors, which could speed up the ex-ecution time by an order of magnitude or more.

    Summary

    In this article, hyperspectral measurements in the vis-ible, near-infrared, and shortwave-infrared reflectedspectrum collected by the AVIRIS and HYDICE air-borne sensors have been used to compare retrievedvalues of column water vapor and surface reflectanceobtained from the ATREM and FLAASH atmo-spheric compensation models. ATREM andFLAASH, two commonly used physics-based atmo-spheric compensation models, were also evaluated forsensitivity to the specification of scene characteriza-tion input parameters and respective computationalcomplexity.

    Both models retrieved similar column water vaporfields for two relatively dry scenes while a notable dif-ference was observed for a scene collected during awarm and humid day. The mean ATREM-retrievedcolumn water vapor for this warm humid scene wassignificantly larger than those observed by a nearbyradiosonde. The exclusion of the effects of the watervapor continuum in the ATREM model calculationsis believed to contribute in part to this error, whichwas confirmed by the improved results from using theweaker 0.91-m water vapor band in the water vaporcalculations. However, the approach used by ATREMfor column water vapor estimation displays limitedsensitivity for column water vapor values over 3 cm(small variations in the computed band ratio result inlarge changes in the estimated column water vapor),which would also contribute to the overestimation.The effect of the inaccurate water vapor estimationwas most notable in the retrieved surface reflectancefor the near infrared region, where spectral variationsin water vapor absorption are common.

    To retrieve the surface reflectance, both ATREMand FLAASH require some knowledge of the scenecharacteristics, which in general are not preciselyknown and must be estimated from available scene

    information. We assessed the sensitivity to incom-plete knowledge or inaccurate estimation of specificinput parameters in terms of the degree of error to theretrieval of the surface reflectance. Results of ourstudy show that certain variations in the inputs to theatmospheric compensation models produce retrievedreflectance differences near 0.1; these errors can becumulative as we showed in one example. Collectionof ambient atmospheric and surface information dur-ing hyperspectral data acquisition is the best strategyfor avoiding this type of processing error.

    An examination of the computational budget forprocessing hyperspectral data cubes showed that opti-mized versions of the more complex FLAASH modelare similar in computational costs to those ofATREM model. However, enhancements to theFLAASH processing, including computations withnonuniform aerosol content, sensor angular LOSvariations, spectral polishing, and channel wave-length corrections can add significantly to the overallprocessing burden. Several possibilities for speedingup the calculations were highlighted in the text. Theoptimization of the atmospheric compensation pro-cess to take advantage of multiprocessor computersprovides the best possibility of reaching the goal of ef-ficient processing of hyperspectral data cubes.

    Acknowledgments

    The authors wish to thank Gail Anderson at the AirForce Research Laboratory for making available ver-sions of MODTRAN and FLAASH for our studiesprior to their official release and the reviewers fortheir insightful comments. We also want to thankCarolyn Upham and Kris Farrar for their help withthe data processing.

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    VOLUME 14, NUMBER 1, 2003 LINCOLN LABORATORY JOURNAL 53

    R E F E R E N C E S1. G. Vane, R.O. Green, T.G. Chrien, H.T. Enmark, E.G.

    Hansen, and W.M. Porter, The Airborne Visible/Infrared Im-aging Spectrometer (AVIRIS), Rem. Sens. Environ. 44 (2/3),1993, pp. 127143.

    2. J.W. Boardman, Precision Geocoding of Low-AltitudeAVIRIS Data: Lessons Learned in 1998, Proc. 7th Ann. JPLAirborne Earth Science Workshop, JPL Publication 99-17, Pasa-dena, Calif.,1999, pp. 6368.

    3. L.J. Rickard, R. Basedow, E. Zalewski, P. Silvergate, and M.Landers, HYDICE: An Airborne System for HyperspectralImaging, SPIE 1937, 1993, pp. 173179.

    4. R.W. Basedow, D.C. Carmer, and M.E. Ander-son, HYDICE System, Implementation and Performance,SPIE 2480, 1995, pp. 258267.

    5. E.P. Shettle and R.W. Fenn, Models for the Aerosols of theLower Atmosphere and the Effects of Humidity Variations onTheir Optical Properties, AFGL TR 79-0214, Air Force Geo-physics Lab., Hanscom AFB, Mass., 20 Sept. 1979.

    6. E.J. McCartney, Optics of the Atmosphere: Scattering by Mol-ecules and Particles (Wiley, New York, 1976).

    7. W. Malkmus, Random Lorentz Band Model with Exponen-tial-Tailed S1 Line-Intensity Distribution Function, J. Opt.Soc. Am. 57 (3), 1967, pp. 323329.

    8. B.-C. Gao, and A.F.H. Goetz, Column Atmospheric WaterVapor and Vegetation Liquid Water Retrievals from AirborneImaging Spectrometer Data, J. Geophys. Res. 95 (D4), 1990,pp. 35493564.

    9. B.C. Gao, and Y.J. Kaufman, The MODIS Near-IR WaterVapor Algorithm, Algorithm Technical Background Doc. ID:MOD05: Total Precipitable Water, NASA Goddard SpaceFlight Center, Greenbelt, Md., 1997.

    10. Y.J. Kaufman and B.-C. Gao, Remote Sensing of Water Va-por in the Near IR from EOS/MODIS, IEEE Trans. Geosci.Rem. Sens. 30 (5), 1992, pp. 871884.

    11. B.-C. Gao, Y.J. Kaufman, W. Han, and W.J. Wiscombe, Re-moval of Thin Cirrus Path Radiances in the 0.41.0 mSpectral Region Using the 1.375-m Strong Water Vapor Ab-sorption Channel, Proc. 7th Ann. JPL Airborne Earth ScienceWorkshop, JPL Publication 97-21, Pasadena, Calif., 1998, pp.121130.

    12. M.K. Griffin, S.M. Hsu, K. Burke, and J.W. Snow, Charac-terization and Delineation of Plumes, Clouds and Fires in Hy-perspectral Images, SPIE 4049, 2000, pp. 274283.

    13. A.D. Roberts, Y. Yamaguchi, and R.P. Lyon, Calibration ofAirborne Imaging Spectrometer Data to Percent ReflectanceUsing Field Spectral Measurements, Proc. Nineteenth Int.Symp. on Remote Sensing of Environment 2, Ann Arbor, Mich.,2125 Oct. 1985, pp. 67988.

    14. J.E. Conel, R.O. Green, G. Vane, C.J. Bruegge, R.E. Alley,and B.J. Curtiss, Airborne Imaging Spectrometer-2: Radio-metric Spectral Characteristics and Comparison of Ways toCompensate for the Atmosphere, SPIE 834, 1987, pp. 140157.

    15. F.A. Kruse, Use of Airborne Imaging Spectrometer Data toMap Minerals Associated with Hydrothermally Altered Rocksin the Northern Grapevine Mountains, Nevada, and Califor-nia, Rem. Sens. Env. 24 (1), 1988, pp. 3151.

    16. B.-C. Gao, K.B. Heidebrecht, and A.F.H. Goetz, Derivationof Scaled Surface Reflectances from AVIRIS Data, Rem. Sens.Environ. 44 (2/3), 1993, pp. 165178.

    17. S.M. Adler-Golden, A. Berk, L.S. Bernstein, S. Richtsmeier,P.K. Acharya, M.W. Matthew, G.P. Anderson, C. Allred, L.Jeong, and J. Chetwynd, FLAASH, A MODTRAN4 Atmo-spheric Correction Package for Hyperspectral Data Retrievalsand Simulations, Proc. 7th Ann. JPL Airborne Earth ScienceWorkshop, JPL Publication 97-21, Pasadena, Calif., 1998, pp.914.

    18. E. Vermote, D. Tanre, J.L. Deuze, M. Herman, and J.J.Morcrette, Second Simulation of the Satellite Signal in theSolar Spectrum (6S): 6S Users Guide Version 2, NASAGoddard Space Flight Center, Greenbelt, MD, 1997.

    19. A. Berk, L.S. Bernstein, G.P. Anderson, P.K. Acharya,D.C. Robertson, J.H. Chetwynd, and S.M. Adler-Golden,MODTRAN Cloud and Multiple Scattering Upgrades withApplication to AVIRIS, Remote Sens. Environ. 65 (3), 1998,pp. 367375.

    20. Y.J. Kaufman and C. Sendra, Algorithm for Automatic At-mospheric Corrections to Visible and Near-IR Satellite Imag-ing, Int. J. Rem. Sens 9 (8), 1988, pp. 13571381.

    21. M.W. Matthew, S.M. Adler-Golden, A. Berk, S.C.Richtsmeier, R.Y. Levine, L.S. Bernstein, P.K. Acharya, G.P.Anderson, G.W. Felde, M.P. Hoke, A. Ratkowski, H. Burke,R.D. Kaiser, and D.P. Miller, Status of Atmospheric Correc-tion Using a MODTRAN4-Based Algorithm, SPIE 4049,2000, pp. 199207.

    22. S.M. Adler-Golden, M.W. Matthew, L.S. Bernstein, R.Y.Levine, A. Berk, S.C. Richtsmeier, P.K. Acharya, G.P. Ander-son, G. Felde, J. Gardner, M. Hoke, L.S. Jeong, B. Pukall, J.Mello, A. Ratkowski, and H.-H. Burke, Atmospheric Correc-tion for Short-Wave Spectral Imagery Based on Modtran4,Proc. 7th Ann. JPL Airborne Earth Science Workshop, Pasadena,Calif., JPL Publication 97-21, pp. 914.

    23. S.M. Adler-Golden, R. Y. Levine, A. Berk, L.S. Bernstein, G.P.Anderson, and B. Pukall, Detection of Cirrus Clouds at 1.13m in AVIRIS Scenes over Land, SPIE 3756, 1999, pp. 368373.

  • GRIFFIN AND BURKECompensation of Hyperspectral Data for Atmospheric Effects

    54 LINCOLN LABORATORY JOURNAL VOLUME 14, NUMBER 1, 2003

    - . leads the Sensor Technologyand System Applicationsgroup at Lincoln Laboratory.In the area of remote sensing,she specializes in applicationsof optical sensors, particularlydata analysis and algorithmdevelopment. A recent area ofinterest is the application ofhyperspectral technology toDepartment of Defense needs,including fusion of hyperspec-tral data information withother sensors. She received herPh.D. degree in atmosphericphysics from Rice University,and joined Lincoln Laboratoryas a staff member in 1981.

    . is a staff member in the SensorTechnology and System Appli-cations group. His currentresearch focuses on applica-tions of hyperspectral technol-ogy to meteorological andenvironmental applications.A recent effort involved thedesign and employment of acloud cover detection algo-rithm for the Hyperion hyper-spectral sensor to run on-board the Earth Observer(EO)-1 space-based platform.He came to Lincoln Labora-tory in 1997 to work on dataanalysis, algorithm develop-ment, and assessment ofhyperspectral sensors. Prior tojoining Lincoln Laboratory hewas an atmospheric physicistat the Air Force ResearchLaboratory where he wasinvolved in the developmentof cloud algorithms frommultispectral imagery andworked on the retrieval ofhumidity, precipitation, andsurface properties from micro-wave sensor measurements. Hereceived a B.S. degree inphysics and mathematics fromCalifornia State University,Chico, and M.S. and Ph.D.degrees in meteorology fromthe University of Utah.