analyzing hyperspectral images · images are covered on pages 24-29. pages 30-36 introduce methods...

Analyzing Hyperspectral Images

page 1

Tutorial

AnalyzingHyperspectral

Images

with

TNTmips®

The hyperspectral analysis techniques described in thistutorial can be applied FREE by anyone with AVIRIS orother hyperspectral imagery using TNTmips FREE

HYPERSPECTRAL


page 2

Before Getting Started

You can print or read this booklet in color from MicroImages’ Web site. TheWeb site is also your source for the newest tutorial booklets on other topics.You can download an installation guide, sample data, and the latest versionof TNTmips.

http://www.microimages.com

This booklet introduces the Hyperspectral Analysis process in TNTmips®, whichprovides specialized methods for visualizing, processing, and analyzing imagesproduced by the new generation of hyperspectral remote sensors. These exer-cises will familiarize you with basic procedures in Hyperspectral Analysis, includingacquiring and saving image spectra, working with spectral libraries, calibratingimages to reflectance, and spectral matching and classification.

Prerequisite Skills This booklet assumes that you have completed the exercisesin the tutorial booklets entitled Displaying Geospatial Data and TNT ProductConcepts. Those exercises introduce essential skills and basic techniques thatare not covered again here. Please consult those booklets for any review youneed. Additional background on hyperspectral analysis can be found in the bookletIntroduction to Hyperspectral Imaging.

Sample Data The exercises presented in this booklet use files in the CUP97 datacollection. Because of the large size of this dataset, it is not distributed with theTNT products. You can download the CUP97 dataset from microimages.com orcontact MicroImages to obtain a free copy of these Project Files on CD.

More Documentation This booklet is intended only as an introduction to theHyperspectral Analysis process. Background information on hyperspectral im-ages and their analysis can be found in the booklet Introduction to HyperspectralImaging. Further information can be found in a variety of tutorial booklets, Tech-nical Guides, and Quick Guides that are available from microimages.com.

TNTmips® Pro and TNTmips Free TNTmips (the Map and Image ProcessingSystem) comes in three versions: the professional version of TNTmips (TNTmipsPro), the low-cost TNTmips Basic version, and the TNTmips Free version. Allversions run exactly the same code from the TNT products DVD and have nearlythe same features. If you did not purchase the professional version (which re-quires a software license key) or TNTmips Basic, then TNTmips operates inTNTmips Free mode.

All the exercises can be completed in TNTmips Free using the sample geodataavailable from MicroImages.

Randall B. Smith, Ph.D., 23 August 2013©MicroImages, Inc., 2000-2013


page 3

STEPSlaunch TNTmipschoose Image /Hyperspectral from theTNTmips menu

Welcome to Hyperspectral Analysis

The exercises on pages 4-10 show you how to opena hyperspectral image,examine image bands andband information, and usethe Hyperspectral Explorerto find a useful RGBcombination of bands forvisual reference duringanalysis. On pages 11-14you learn how to acquire,view, and save imagespectra. Classification ofthe hyperspectral imageusing image spectra isintroduced on pages 15-21,and pages 22-23 introducedimensional reductionmethods. Procedures forfinding spectral end-members for analyzingimages are covered onpages 24-29. Pages 30-36introduce methods ofcorrecting the raw imagevalues to reflectance anddiscuss the use of labora-tory reflectance spectra andspectral matching proce-dures. Pages 37-39discuss continuum-removaland some additional imageprocessing options.

Hyperspectral remote sensors collect image data inmany narrow, contiguous spectral bands. The re-sulting datasets contain numerous image bands, eachdepicting the scene as viewed within a narrow wave-length range. Although this multiband conceptual-ization is a natural outgrowth of our experience withmultispectral images, an alternative approach is moreuseful for analyzing and interpreting hyperspectraldata. The hyperspectral scene can be thought of in-stead as a single image with a spectrum of bright-ness values stored for each image cell. These imagespectra can be compared with field or laboratoryspectra in order to recognize and map the spectralsignatures of ground materials such as green veg-etation or particular minerals.

The prototype Hyperspectral Analysis process inTNTmips provides specialized interactive analysistools that allow you to fully exploit the spectral rangeand spectral resolution provided by hyperspectraldatasets. The process allows you to view and saveimage spectra and to compare image spectra withlaboratory spectra stored in spectral libraries. Sev-eral methods of spectral classification are alsoprovided.

To illustrate the use of the Hyperspectral Analysisprocess, we will work with a sample scene fromNASA’s Airborne Visible/Infrared Imaging Spec-trometer (AVIRIS). This sensor collects data in 224contiguous spectral bands with a bandwidth of 0.10µm. Each 20-m square cell in the scene has a con-tinuous spectrum over the range from 0.4 to 2.5 µm.The scene covers the Cuprite mining district in west-ern Nevada, USA. Rocks in several areas of thescene have been altered and mineralized by hot wa-ter solutions (hydrothermal alteration), creatingconcentrations of alteration minerals that providegood targets for spectral analysis.


page 4

STEPS click the Open

icon button in theHyperspectral Analysiswindow andselectHyperCubeObject from thedropdownmenu

use thestandard SelectObject dialog tochoose theCUPRITE97 objectin the CUP97CUB

Project File

Open a Hyperspectral ImageA hyperspectral image can be stored in a TNTmipsProject File in one of two ways: as a set of separateraster objects, with one object for each wavelengthband, or as a single hyperspectral object. A hyper-spectral (or hypercube) object includes raster valueand wavelength information for all of the bands in ahyperspectral image. The hyperspectral object in-corporates lossless image compression, and so offersa significant reduction in file size in comparison tothe multiraster format. Image display in the Hyper-spectral Analysis process is also much faster whenusing a hypercube object.

When you import ahyperspectral imageyou can choose eitherthe multiraster orhypercube format. Youcan also convert animage from one formatto another within theHyperspectral Analysisprocess using optionson the Image menu.

NOTE: The sample filesused in these exercises arenot distributed on theTNTmips CD. You candownload these files fromthe MicroImages web site(www.microimages.com) orcontact MicroImages toobtain a free copy of thefiles on CD-R.


page 5

3D Hypercube DisplaySTEPS

choose 3D Hyper CubeDisplay from the Imagemenu in the Hyperspec-tral Analysis windowmove the Band Numberslider at the bottom ofthe 3D Hyper CubeDisplay window to viewdifferent wavelengthbands on the front faceof the cubechoose Close from theFile menu in the 3DHyper Cube Displaywindow

The 3D Hypercube Display emphasizes the highspectral content of the hyperspectral image while al-lowing you to quickly examine any of the individualwavelength bands. The 3D Hypercube Display win-dow portrays the hyperspectral image as athree-dimensional “image cube”. The top and rightpanels of the cube show the corresponding edge cellsof each wavelength band, with wavelength increas-ing toward the back of the cube. The shortestwavelength band is displayed by default on the frontof the cube. You can select any of the wavelengthbands for display on the front of the cube by usingthe Band Number slider or the arrow icon buttons.

You can also apply a color map tothe image cube. Select Edit ColorMap from the File menu to open astandard Color Palette Editorwindow. You can use the Palettemenu to choose one of the manypredefined color palettes, or designyour own color palette. Theillustration to the right shows theCuprite image cube with thepredefined Beach color palette.

The “fuzzy” zones on the edgesof the hypercube indicate twogroups of very noisy imagebands. Each of the two groupsspans a wavelength region inwhich atmospheric water vaporabsorbs almost all incoming andreflected solar radiation. As aresult the bands in these wave-length regions (near 1.4 and 1.9micrometers) have little imagecontent and are dominated bysensor noise.


page 6

Select Bands for RGB DisplaySTEPS

in the toolbar in theHyperspectral Analysiswindow, click onthe Select Bandsicon buttonin the Select Bandswindow, use the Red,Green, and Blue menusto change the bandsselected for display andpress [OK]

The Hyperspectral Analysis process automaticallyselects three bands for RGB display in theHyperspectral Image window. The bands are ini-tially chosen to span most of the spectral range ofthe data while avoiding bands with excessive noiseor no data. The selection of bands for display hasno impact on analysis processes, so you can usethe Select Band window to manually change the com-bination of bands you want to use as a visualreference. The Raster Layer Controls window showsthe band numbers that have been selected for Red,Green and Blue color components along with theirassigned wavelengths. You can also use this win-dow to select any of the standard automaticcontrast-enhancement methods for each color chan-nel. Your band selections and contrast settings areautomatically saved as part of the display param-eters for the hypercube object.

The Hyperspectral Explorer procedure (described onthe following pages) provides a rapid means forevaluating many RGB band combinations.

Cuprite image with bands selected toapproximate natural color view. TheCuprite district is a desert area withvery little vegetation. It includesseveral hills surrounded by alluvialfans, and is bordered on the east by adry lake bed (Stonewall Playa).Hydrothermally altered rocks areexposed in the hills on either side ofthe highway.

Areas of hydrothermal alteration

Stonewall Playa(salt flat)

Highway

in the Layer Managerwindow, click onthe icon for theCuprite hypercubelayernote the Band numberand wavelengthreported for Red, Green,and Blue, then click [OK]


page 7

STEPSpress the Hyper-spectral Explorericon button on theHyperspectral Imagewindowhold the left mousebutton as you drag thecursor over the image todraw a rectangularinspection area; releasethe button to ceasedrawingresize or move therectangle to match theillustrationclick the right mousebutton to accept theinspection area and loadthe image bands (thismay take severalminutes)

Start the Hyperspectral ExplorerThe Hyperspectral Explorer is a unique, automatedtool for visualizing the hyperspectral image. It al-lows you to automatically create a large number ofpotential RGB band combinations for a sample area.You can then preview the combinations as an ani-mated sequence and select one to use as a referenceimage during further analysis.

After you select the Explorer tool, use the mouse todrag open a rectangular inspection box in the Hy-perspectral Image window. When you right-click toaccept the box, the portion of each band includedwithin it is loaded in memory so that different RGBcombinations can be displayed rapidly. (The size ofthe inspection box cannot exceed the product of 256by 256 raster cells.) Other controls for the proce-dure are accessed from the Explorer window, whichis explained on the following pages.

This exercise continueson the next page.

Hyperspectral Image window with completed Explorerinspection box. (If the box reaches the Explorer size limitas you are drawing, the initial box disappears, and a newbox begins to draw at the current cursor location.)

Explorer window aftersample images are loaded.


page 8

STEPSpress theParameters buttonon the Explorer windowchange the InterbandInterval value in theParameters window to15, and press [Close]

Explorer Window Display

You can also drag thecolored sliders for the Red,Green and Blue compo-nents to set an initial RGBcombination. This setsboth the wavelength orderand interband intervals forall RGB combinations.

The Hyperspectral Explorer creates a suite of RGBband combinations using constant Red-Green andGreen-Blue intervals (in number of bands). You canset a single value for both intervals using the Inter-band Interval field in the Parameters window. Whenyou use this option, the shortest-wavelength bandin each combination is assigned to the Red displaychannel and the longest-wavelength band to theBlue channel. For example, an Interband Interval of15 creates RGB combinations using bands 1, 16, and31; 2, 17, and 32; 3, 18, and 33; and so on until eachband has been used in at least one RGB image.

To understand the lower part of the Explorer win-dow, imagine that all of the prepared RGB images ofthe inspection box are stacked on top of each other,with the lower wavelength combinations at the top.Then make a vertical slice down through the stackalong a line through the middle of the inspectionbox (shown as the dashed cross section line in theview window). The Explorer window displays thisvertical slice. Each horizontal line of pixels in the

window pane shows the colors produced byone of the RGB combinations along the crosssection line. The display gives an overview ofthe range of colors produced by each of theband combinations, allowing you to focus onuseful ranges of combinations.

These fields show the wavelength(in micrometers) of each compo-nent in the current RGB set.


Because AVIRIS bands are narrow and have nowavelength gaps between them, adjacentbands appear very similar to each other. Whena small value (such as 3, the initial default) isused for the Interband Interval, each RGB setconsists of three very similar bands. Conse-quently, each RGB image cell has nearly equalcontributions of Red, Green, and Blue. Theresulting color images (and the Explorer windowview) are dominated by gray tones (see theillustration on the previous page). Increasing theInterband Interval reduces the similarity betweencomponent rasters, producing more colorfulRGB combinations.


page 9

When you press the Play button on the Explorerwindow, the process cycles through the prepared setof RGB combinations in increasing wavelength or-der and displays them at one second intervals in theinspection box. You can use the other icon buttonsto pause, reverse, or restart the sequence, and thevertical slider to change position in the sequence.When you find a combination that highlights fea-tures of interest, press the Full View icon button toapply this combination to the full image.

STEPSclick the Play iconbutton on the Explorerwindowpress the Pauseicon button whena color combinationsimilar to the one in theillustration is displayedpress the FullView icon button

Run the Hyperspectral Explorer

Bands 172 (1.989 µm), 187 (2.139 µm), and 202(2.288 µm) displayed respectively as R, G, and B.This band combination highlights the hydrothermal-ly altered rocks in the Cuprite scene and distin-guishes different types of alteration.

Portion of the original RGB imagewith a new RGB combinationdisplayed in the inspection box.

When you play the sequence of Explorerimages, the colored sliders move to indicatethe band wavelengths used in the currentcombination, and the vertical slider shows thecurrent position in the stack of combinationsdepicted in the lower part of the window.

press the Zoom Boxicon button in theHyperspectral Image windowto hide the Explorer window

Restart

Reverse

The “fuzzy” horizontal stripes inthe Explorer window representgroups of RGB sets that includeone or more bands withsignificant image noise. Youshould avoid including very noisybands in analysis operations,since they may not provide usefulinformation.


page 10

The Band Number slider on the Image panel allowsyou to quickly select any one of the componentbands and review its spectral characteristics. A his-togram of the current band is shown at the bottomof the panel. The Wavelength field shows the wave-length in micrometers of the band center, while theBandwidth field shows the width of the spectralband. For AVIRIS images, the import process auto-matically incorporates this information with thespectral bands. If you import hyperspectral datafrom another source not yet supported by TNTmips,you will need to enter the spectral information foreach band. If the wavelength value (in micrometers)is included in the object description, you can usethe Auto Define utility to assign the appropriatewavelength to each band; approximate bandwidthsare then calculated when needed during process-ing.

STEPSclick on the Image tabon the HyperspectralAnalysis windowmove the Band Numberslider to examine theWavelength andBandwidth values andthe histogram for variousbands

Examine Band Characteristics

The Display menu provides several modes for viewinghistograms, including outline (shown), bars, and 3D strips.

Use the AutoDefine option onlywhen spectralinformation hasnot yet beenassociated withthe spectralbands.

Use the BandNumber slider orthe arrow iconbuttons to selectthe desiredband.

Hyperspectral analysisprocedures require spectralinformation for each band inthe hyperspectral image.Controls for viewing andediting this information arefound on the Image tabbedpanel in the HyperspectralAnalysis window.


page 11

The Spectral Profile tool allows you to select an im-age cell and display its spectrum in the Spectral Plotwindow. You can also specify a small area and dis-play its average spectrum. A spectral plot is a graphof spectral brightness value (vertical axis) versuswavelength in micrometers (horizontal axis). TheSpectral Plot window can display one or more spec-tra drawn from the hyperspectral image, or spectrastored in a spectral library. The value and wave-length ranges and grid intervals for the plot arecomputed automatically from the selected spectrumor spectra. You can also choose to set the range foreither plot axis manually using the controls at thebottom of the Spectral Plot window.

STEPSclick the Info tab on theHyperspectral Analysiswindowturn on the DisplaySpectral Plot togglebuttonselect the SpectralProfile tool fromthe Hyperspectral Imagewindow toolbarmove the cursor to themiddle of StonewallPlaya (see illustration)and left-click to place theselection cursorright-click to read anddisplay the spectralprofile

View Image Spectra

You can choose the dimen-sions of a square sample areafrom the Sample Windowoption menu. The default valueis 1, which displays thespectrum for a single imagecell. An AVIRIS image cellrarely includes only a singlematerial, so most single-cellspectra are composite innature. In spectral analysis it isoften advantageous to restrictthe sample area to a single cellthat is the “purest” example ofa particular material of interest.

You can left-click in the spectral plot to placecrosshairs which you can move around the plot.The values for the crosshair location are shownin the Wavelength and Reflectance fields belowthe plot.

When you turn on the Multiple Spectra togglebutton, you can show a series of spectra indifferent colors in the Spectral Plot window.


page 12

The characteristic absorption features of many ofthe alteration minerals in the Cuprite scene lie in thewavelength range from 2.0 to 2.46 µm. The Subsetoption on the Use Wavelength Range menu allowsyou to restrict processing to the hyperspectral im-age bands within a specified wavelength range.Once you designate the subset range, image spec-tra and all spectral analysis functions are limited tothat group of image bands. Specifying a wavelengthsubset can speed processing and increase the qual-ity of spectral classification results. You can changethe subset range at any time during an analysis ses-sion.

STEPSclick on the Image tab inthe HyperspectralAnalysis windowselect Subset from theUse Wavelength Rangeoption menuenter 1.989 in the Fromtext field and 2.457 inthe To text fieldmove the cursor into theHyperspectral Imagewindow and click theright mouse button torecompute the currentimage spectrum

Set a Wavelength Range

The Exclude Atmospheric Absorption Bandstoggle excludes from processing two presetwavelength ranges: 1.33 to 1.41 µm, and 1.81to 1.95 µm. Atmospheric water vapor absorbsmost of the energy in these spectral regions,so the included image bands contain muchrandom noise and little information aboutsurface materials. These ranges are shown asgaps in spectral plots, as in the plot to the left(highlighted here by the red bars).

Recomputed image spectrum forthe wavelength range 1.989 to2.457 micrometers (µm). Rangevalues that you enter are auto-matically adjusted to reflect actualimage band wavelengths.

crobbins

Highlight

crobbins

Sticky Note

Use Image tab settings to calibrate and/or limit the input bands for processing. Thus classification, principal component analysis, and other operations use the input specified here.


page 13

In some instances you may want to acquire a num-ber of image spectra for use in analyzing andclassifying the hyperspectral image. To do so, youneed to save the spectra in a spectral library object.Choosing the New Spectral Library option opens atemporary library. After one or more spectra areadded to the library, you can save it as an object in aProject File.

If you save an image spectrum acquired from a sub-set of the image bands, only the portion of thespectrum corresponding to the selected wavelengthrange is saved in the library. Hyperspectral analysisoperations that compare spectra automatically com-pensate for differences in spectral range and spectralresolution. Before comparison functions are applied,overlapping spectra are truncated to their commonspectral range, and spectra with differing bandwidths are resampled to the coarsest input spectralresolution.

When you extract an image spectrum with the Spec-tral Profile tool, a default spectrum name is createdwith the form ImageXX,YY; XX is the raster columnnumber and YY is the raster line number. You canrename the spectrum if you like by editing the Spec-trum Name text field before saving the spectrum inthe library.

Save Image Spectra in a Spectral LibrarySTEPS

click the NewSpectral Library iconbutton on the Hyper-spectral Analysis windowmove the cursor into theHyperspectral Imagewindow and right-click toreacquire the imagespectrumclick the Info tab andnote the default name inthe Spectrum Name fieldclick [Save Spectrum]click the Save Asicon buttonuse the standard SelectObject procedure toname a new Project Fileand library object

The spectral values in animage spectrum areadjusted automaticallybefore plotting or process-ing according to the optionsselected on the Calibrationsubpanel of the Imagetabbed panel. Theseoptions are used to adjustradiance values measuredby a sensor to valuesapproximating the spectralreflectance of the groundmaterials; they are dis-cussed in a later exercise.

The names of spectra in the open library are listed in the paneon the left side of the Hyperspectral Analysis window.

The name of the file containing the currently openspectral library is shown in the text field.

The spectrumcurrentlyhighlighted in thelibrary listing canbe deleted usingthe DeleteSpectrum button.


page 14

Save a Spectrum as TextSTEPS

click the Save as Textbutton on the Info paneluse the standard SelectFile dialog to name anoutput text file for thespectrumopen the OUTPUT .CSV filein a text editor toexamine its formatexit the text editor whenyou are finishedexamining the file

wavelength,reflectance,standard deviation(bandwidth) 1.98867, 0.91535, 0.010990 1.99869, 0.88486, 0.010970 2.00870, 0.85609, 0.010950 2.01872, 0.89134, 0.010940 2.02873, 0.92012, 0.010920 2.03874, 0.93310, 0.010900 2.04875, 0.93234, 0.010880 2.05876, 0.93215, 0.010870 2.06876, 0.94171, 0.010850 2.07877, 0.94504, 0.010830 2.08877, 0.94841, 0.010810 2.09877, 0.95240, 0.010800 2.10876, 0.96267, 0.010780...

You may wish to export spectra for use in other soft-ware for further analysis and plotting. The Save AsText button saves the data for a selected spectrumin an text file with the file extension .csv (comma-separated values). The format of the file is illustratedby the sample below. The file has one line for eachspectral band, with the data lines preceded by aheader line. There are three columns of data that areseparated by commas: wavelength, reflectance (oruncalibrated brightness value), and standard devia-tion (bandwidth). Files in this format can be openeddirectly by most spreadsheet programs. If you wantto export more than one spectrum, each one must besaved separately to its own file.

Use the Save As Text button to save theselected spectrum in a CSV file.


page 15

STEPSclick [Spectral Library...]select object SAMPLIB1from the SAMPLIBS ProjectFileon the Info panel, turn onthe Multiple Spectratoggle button for spectralplots

If you know the locations of key surface materials inportions of the hyperspectral scene, you can use im-age spectra from these locations to identify otheroccurrences in the scene. This procedure is in someways similar to supervised classification, and sev-eral methods are available on the Classificationtabbed panel.

The SAMPLIB1 spectrallibrary contains imagespectra from relativelypure occurrences of threeof the common alterationminerals in the Cupritescene. However, most cells in the image probablyinclude more than one material or mineral, each ofwhich has its own characteristic spectrum. The spec-trum of such a cell has a composite shape thatdepends on the spectral shapes and relative abun-dances of the included materials. The hyperspectralclassification methods produce a raster with cell val-ues that measure the similarity between each imagespectrum and the target spectrum. Be cautious wheninterpreting the values in this raster, because eithera high abundance of the target material or the pres-ence of another material with a similar spectrum canproduce equivalent “similarity” values.

Use Image Spectra for Classification

select each of the threespectra listed in thepane on the left side ofthe HyperspectralAnalysis window to addthem to the Spectral Plotwindowclick the Classify tab

This exercise continues onthe next page.

337,336_Alunite

529,267_Chalcedony

550,94_Kaolinite

Locations and spectral plotsof image spectra included inthe SAMPLIB1 spectral library.

Spectral Plot

Val

ue

0.5

0.6

0.7

0.8

0.9

Wavelength2.1 2.2 2.3 2.4


page 16

STEPSselect Matched Filteringfrom the Algorithmoption menuselect spectrum337,336_Alunitepress [Add] to add thisspectrum to the EndMember listrepeat the last two stepsfor the remaining twospectra in the librarypress [Classify]use the standard SelectObjects dialog to nameeach output raster anddirect them to anew Project File(processing maytake severalminutes or more)use the LayerManager for theHyperspectralImage windowto view theoutput rasters

The Matched Filtering algorithm uses a ConstrainedEnergy Minimization technique to assess the spec-tral composition of each image cell. Each imagespectrum is assumed to be a linear mixture of a tar-get spectrum and multiple unknown spectral signa-tures. The process identifies what proportion (if any)of each composite image spectrum could be pro-duced by the target spectrum (end member). A per-fect match yields an output value of 1, while poormatches yield low positive or even negative values.The process creates a separate floating-point outputraster for each target spectrum (end member).

Matched Filtering result for spectrum337,336_Alunite with linear contrastenhancement. Brightness indicatesdegree of match to the target spectrum;only the brightest areas indicate a con-fident identification of the target material.

Matched Filtering

Two variants of the standard matchedfiltering procedure are also available onthe menu: Derivative Matched Filteringand Locally Adaptive Matched Filter-ing. The Derivative method uses thefirst derivative (slope) of the spectralcurves rather than the actual spectral val-ues, providing a similarity check that ismore sensitive to curve shape and lesssensitive to overall brightness differ-ences. The Locally Adaptive methoddetermines a local unknown backgroundsignature for each image cell, and is bestsuited for identifying occurrences ofmaterials that are relatively rare in thescene.

crobbins

Sticky Note

Before selecting Classify, check your input calibration and wavelength limits ... For example, set Image tab, Use Wavelength Range: Subset 1.989 to 2.457 as specified on page 12.

crobbins

Sticky Note

The displayed results shown here don't use the default auto-linear contrast. For similar appearance, limit the contrast range to 0.1 and 99.9.


page 17

Spectral Angle Mapper

Portion of the spectral angle raster for thetarget Alunite spectrum. Low values(darkest tones) indicate areas with the bestspectral match.

Class raster displayed over the SpectralAngle raster, with cells assigned to theend member class shown in red.

The Spectral Angle Mapper treats target spectra andimage spectra as vectors in n-dimensional spectralspace. Each spectrum defines a point in spectralspace, and this point can also be treated as the endof a vector that begins at the origin of the coordinatesystem. The angle between a pair of vectors is ameasure of the similarity of the spectra; smaller spec-tral angles indicate greater similarity. This methodis insensitive to differences in average brightness be-tween spectra that may be due to topographic orsensor gain effects, because these factors change thelength of the spectral vector, but not its orientation.

This process produces two outputrasters: a Spectral Angle raster thatcontains values of the spectral anglefor each image cell, and a Class ras-ter in which cells are assigned toend-member classes based on theangle value that you set for the Threshold Value pa-rameter. If you select multiple end-member spectra,the spectral angle is calculated between an imagespectrum and each end member spectrum, and theminimum value is shown in the Spectral Angle ras-ter. However, this raster is more readily interpretedif only a single end member is used.

STEPSselect the Kaoliniteimage spectrum in theEnd Member list andpress [Remove]repeat for theChalcedony imagespectrumselect Spectral AngleMapper from theAlgorithm option menupress [Classify]name each output rasterand direct them to yourclassification Project File

Cells with a spectral anglevalue less than theThreshold Value areassigned to the endmemberclass in the Class raster.You will need to experimentto find the appropriate valuefor a particular targetspectrum.


page 18

The Cross-Correlation matching method calculatesthe linear correlation coefficient between each im-age spectrum and the target spectrum for differentmatch positions (spectral band shifts). The correla-tion values in the best matching position for eachimage cell are shown in the Correlation raster (val-ues between 0 and 1, with 1 indicating a perfectmatch). Cells in the Class raster are assigned to anend-member class using the correlation value thatyou specify for the Threshold Value parameter. Thismethod is relatively insensitive to differences in av-erage brightness.

Comparing the results of these classification exer-cises, note that the output rasters produced by the

Cross-Correlation and SpectralAngle Mapper methods look like acoherent image, while much of theMatched Filtering (MF) image ap-pears “blurry”. The MF results areactually the best, because most of

the scene has low match values for a given targetspectrum, and there is little spatial correlation be-tween these low values. The MF method does abetter job of recognizing the target spectrum withinvarious composite (mixed material) image spectra.

STEPSremove the Aluniteimage spectrum from theEnd Member listadd the Kaolinite imagespectrum to the EndMember listselect Cross Correlationfrom the Algorithmoption menuenter 0.87 in theThreshold Valueparameter fieldpress [Classify]name each output rasterand direct them to yourclassification Project File

Cross-Correlation

Correlation raster for the target kaoliniteimage spectrum with auto-linear contrastenhancement. High values (bright tones)indicate areas with the best spectral match.

Class raster (threshold value = 0.87)displayed over the Correlation raster, withcells assigned to the end-member classshown in yellow.


page 19

The Linear Unmixing method uses a suite of end-member spectra to assess the spectral compositionof each image cell. Cells are assumed to containvarying proportions of the selected endmember ma-terials, and the resulting image spectra are assumedto be linear combinations of the endmember spec-tra. The process produces a raster for eachendmember that shows the fractional abundance ofthat endmember material for each image cell. Theprocess also creates an Error raster.

Selection of appropriate endmember spectra is im-portant to the success of this process. The selectedendmembers should represent “pure” (single-mate-rial) cells, if they exist in the image. Procedures foridentifying such image cells are outlined on pages21-27. These procedures were used toidentify the spectra included in spectrallibrary SAMPLIB2. On the basis of spectralshape and published information about theCuprite area, four of these spectra can beattributed to single minerals (alunite, cal-cite, kaolinite, and muscovite). The othertwo spectra in the library probably repre-sent spatial mixtures of two or moreminerals that do not occur in pure formthroughout any single image cell in thescene.

Linear UnmixingSTEPS

remove the Kaoliniteimage spectrum from theEnd Member listselect Linear Unmixingfrom the Algorithmoption menuclick [Spectral Library...]select object SAMPLIB2from the SAMPLIBS ProjectFileadd all six spectra to theEnd Member listpress [Classify]name each output rasterand direct them to yourclassification Project File

Fraction image forthe IEM-Aluniteendmemberdisplayed withautonormalizedcontrastenhancement.


page 20

Assessing Unmixing Results

Fraction images for theIEM_Kaolinite (top),IEM_Calcite (middle), andIEM_Muscovite (bottom),endmembers, all displayedwith autonormalizedcontrast enhancement.

In the ideal linear unmixing case, fraction values forindividual endmembers should range from 0 to 1,and the sum of the fractional abundances for eachimage cell should equal 1. But if you supply too fewendmembers, or inappropriate endmembers, the frac-tional abundances for each cell may not sum to thedesired value of 1. If a selected endmember is amixed cell, and “pure” cells representing its compo-nents are present in the image, the pure cells will be

assigned negative abundance values or val-ues greater than one for some end members.If there are a large number of such “out ofrange” values in the fraction images, youmay need to add or replace endmember spec-tra.

The linear unmixing equations include anerror term that is minimized to identify thebest-fit set of endmember abundance val-ues. The computed error value for each imagecell is incorporated in the Error raster. Brightcells in the Error raster (high error values)indicate areas that were not adequately mod-eled by the selected set of endmembers, soyou may need to concentrate your searchfor additional endmembers in those areas.

Error raster for the Cuprite linear unmixing runusing the endmember spectra in SAMPLIB2(displayed with autoexponential contrastenhancement). Bright areas probably containmaterials not included in the endmember set.


page 21

Self-Organizing Map ClassifierSTEPS

select Self OrganizingMap Classifier from theImage menuuse the standard SelectObjects dialog to nameeach output raster anddirect them to yourclassification Project File

The Self-Organizing Map Classifier performs anempirical (unsupervised) classification of the hyper-spectral image, so you don’t have to specify targetspectra. Like the Spectral Angle Mapper, imagespectra are treated as points in n-dimensional space.The process uses neural network techniques to finda best-fit set of 512 class-centers, then assigns allimage spectra to one of these classes using spectralangle as the measure of similarity.

The neural network used to determine the class-center vectors is a square 16 x 32 array of nodes,each representing one class. Initially the spec-tral values for each node are assigned randomly(producing randomly-oriented class vectors). Alarge sample of image spectra are compared one byone to the full set of nodes. For each test spectrum,the node with the most similar spectrum has its spec-tral values adjusted to improve the match. Inaddition, nearby nodes in the array are also adjustedto a lesser extent. After many sample spectra havebeen processed, the node values converge to a set ofclass-center vectors that approximate the distribu-tion of all image spectra in N-dimensional space.Similar class vectors also lie close together in thenode array, and the greater the area that similar ma-terials cover in the hyperspectral image, the moreclasses are used to representthem. This property of the clas-sifier ensures adequate spectraldiscrimination of different va-rieties of common, widespreadmaterials.

Class raster with 512 classescreated by the Self-OrganizingMap classifier, shown with arandom color map.

The Distance raster is agraphical representation ofthe final state of the Self-Organizing Map, the 16 x 32array of nodes used todetermine class centers.Each cell represents oneclass, and the cell value isthe average spectraldistance to neighboringclasses. Dark areasrepresent tightly groupedclass centers.


page 22

Principal Components TransformSTEPS

click the PCA/MNF tabclick [CalculateEigenvectors] (process-ing will take severalminutes)choose each of the fouroptions on the Displayoption menu and studythe resulting plotspress [ForwardTransform] to create aset of PC componentrastersuse the standard SelectObjects procedure todirect the output rastersto a new Project Fileuse the Layer Managerfor the HyperspectralImage window to viewthe PCA rasters

Adjacent hyperspectral image bands are visually andnumerically similar, and therefore contain much re-dundant information. The Principal Componentstransform is a standard method for deriving a newset of images with reduced spectral redundancy. Theprocess is a linear transformation that projects eachimage cell’s spectrum to a new set of orthogonalcoordinate axes. These axes are chosen so that theoutput images are uncorrelated and ordered by de-creasing variance, with the first principal componentaxis corresponding to the direction of maximum vari-ance in spectral space. Important image informationis generally concentrated in the low-order compo-nents, while noise increases with increasingcomponent number. Use of low-order PC rasters inplace of original image bands can speed visual analy-sis and classification of the hyperspectral image.

The eigenvalue plot shows the imagevariance for each component axis. Thesum of the eigenvalues equals the sumof the band variances in the originalimage. You can also display plots ofComponent Variance and Band Variance(in percentage of total variance).

Second principal component raster

The eigenvector plot for each principalcomponent shows the set of weightingcoefficients (one for each image band)used in the coordinate transformationfor that component. Use the slider orarrow buttons to choose the desiredcomponent.


page 23

Minimum Noise Fraction TransformSTEPS

click the Minimum NoiseFraction radio button atthe top of the PCA/MNFpanelclick [CalculateEigenvectors]display the various plots

If bands in a hyperspectral image have differingamounts of noise, standard principal componentsderived from them may not show the usual trend ofsteadily increasing noise with increasing componentnumber. The Minimum Noise Fraction transform(MNF) is a modified version of the Principal Com-ponents transform that orders the output componentsby decreasing signal to noise ratio.

The MNF procedure first estimates thenoise in each image band using the spatialvariations in brightness values. It then ap-plies two successive principal componenttransforms. The first uses the noise esti-mates to transform the dataset to acoordinate system in which the noise is un-correlated and is equal in each component.Then a standard principal componentstransform is applied to the noise-adjusteddata, with output components ordered by decreas-ing variance. This procedure produces a componentset in which noise levels increase uniformly withincreasing component number. The low-order com-ponents should contain most of the imageinformation and little image noise.

RGBI display of the first 4MNF components (3 = Red,2 = Green, 4 = Blue,1 = Intensity).

Creating MNF component rasters using the Forward Transform option can requireseveral hours or more, depending on the number of input components. You can findthe first 5 MNF component rasters for the Cuprite scene in the MNF_CUP Project File.

You can use the Save and Load buttonsto save and reload the statisticscalculated by the Principal Componentsand MNF transforms. For example, youmight calculate the transform in onesession, and reload the statistics later tocreate component rasters. For the MNFtransform, you can save the imagestatistics and noise statistics using theoptions on the Statistics button.

The plots produced by the MNFtransform show the noise statisticscalculated from the original input bands.


page 24

Pixel Purity IndexSTEPS

click the Openicon button andselect RVC Image Listfrom the dropdown menuselect all raster objectsin the MNF_CUP Project Fileon the Image panel,select None from theoption menus for bothAdditive Offset andAtmospheric Correctioncalibration methodsselect Compute PixelPurity Index from theImage menuin the Pixel Purity Indexwindow, select 200 fromthe Test Directions PerPass option menu, andchange the value in theThreshold Value field to0.10click [Run] and direct theoutput PPI raster to anew Project File

The set of low-order MNF components provides a“distilled” version of the hyperspectral image thatcan be used to rapidly identify relatively “pure” im-age spectra for use as spectral endmembers in theLinear Unmixing operation. The Pixel Purity Index(PPI) operation is the first step in identifying theseendmember spectra.

When image spectra are plotted as points in n-di-mensional space, endmember spectra should lie atcusps along the margins of the data cloud. (This istrue regardless of whether we use the original spec-tral bands or the transformed MNF components).The PPI creates a large number of randomly-orientedtest vectors emanating from the origin of the coor-dinate space. The spectral points are projected ontoeach test vector and spectra within a threshold dis-tance of the minimum and maximum projected valuesare flagged as extreme. As directions are tested, theprocess tallies the number of times an image cell isfound to be extreme. Cells with high values in theresulting PPI raster should correspond primarily tothe locations of “edge” spectra in the image.

A two-dimensional illustration of how the PixelPurity Index identifies potential extreme imagespectra (large red dots).

When the PPI operation is run on raster sets withinteger values, the default threshold value of 1.00corresponds to a distance of 1 digital number. Whenthe PPI is run on floating-point rasters with non-integervalues, like MNF component sets, the default value isinterpreted as 1 Standard Deviation. A threshold closeto 0.10 is more appropriate for MNF component sets.

You can select both the number of passes that thePPI operation makes through the full set of imagespectra and the number of test directionsexamined per pass. You can use the Resume PPItoggle to rerun and add to an existing PPI raster,continuing until the number of new cells identifiedper pass falls to a value close to 0.

Component 1

Co

mp

on

ent

2

Test Vector

Thresholdvalue

Max.

Min.


page 25

Start the n-Dimensional VisualizerTo assess the results of the PPI operation we willuse the n-Dimensional Visualizer tool. This toolallows you to create and view an n-dimensionalscatterplot of spectra for an area in the hyperspec-tral image or MNF component set. You can viewthe scatterplot from different viewpoints and rotatethe plot manually or automatically in real time. Youdesignate the selected area by drawing a polygon inthe Hyperspectral Image window. You can use thistool to investigate the spectral properties of differ-ent materials in the image and to search for imagecells with extreme spectral values that might repre-sent pure endmembers.

When you use the PPI raster as a mask, only cellswith PPI values above the specified threshold value(and within the selection polygon) are displayed inthe scatterplot. The threshold eliminates cells thatwere only flagged as extreme in a few spectral di-rections. When no mask is used, you should use asmaller selection polygon to obtain a manageableset of spectral points.

STEPSin the Image Masksection of the Info panel,click [Raster...]select the PPI raster youmade in the previousexercise, or the oneprovided in the PPI_CUP

Project Filechange the ThresholdValue field for the imagemask to 25.00

click then-DimensionalVisualizer icon button inthe Hyperspectral Imagewindowuse the Line / PolygonEdit Controls to draw apolygon that includesmost of the image, asshown belowclick [Apply] (or pressthe right mouse button)make sure theStyle icon buttonon the Visualizerwindow is toggled outturn on the LabelAxes icon button

This exercise continues on the next page.

The Style icon button toggles the pointdisplay style between two modes: singlepixel (button in) and box (shown here).

The grayscalePPI mask rasteris displayedover the MNFimage with 0-value (null) cellstransparent.


page 26

Rotate the n-Dimensional ScatterplotSTEPS

move the Speed slideron the Visualizer windowto speed 2press [Spin]; the buttonchanges to read Stopafter watching the plotautomatically rotate for awhile, press [Stop]use the direction arrowicon buttons to rotate theplot by incrementspress the ChangeViewpoint iconbutton severaltimesplace the mouse cursorin the view pane of theVisualizer window, holddown the left mousebutton, and drag in thedesired direction torotate the plot manually

You can use the n-Dimensional Visualizer to findand mark the most extreme cells among those flaggedby the PPI operation. A typical n-dimensional scat-terplot shows better separation of point clusterswhen viewed from certain spectral directions thanfrom others. Usually you need to rotate the plot inorder to find these directions and to understand thedistribution of points. The rotation options in then-Dimensional Visualizer provide a great deal of flex-ibility, including manual, incremental, and automaticrotation modes.

The Spin button starts an animated rotation of thescatterplot in successive, randomly-chosen direc-tions. You can use the Speed slider to vary therotation speed. The rotation speed is also depen-dent on the number of spectral points and the numberof dimensions. The animation works best with asmall number of dimensions (less than 15). You canalso rotate the plot by small increments in a specificdirection using the arrow icon buttons. With thewindow mode set to Rotate, you can rotate the plotmanually inany directionby draggingthe cursor inthe Visualizerwindow.

The arrow iconbuttons rotate theplot by a small,constant amountin the specifieddirection.


The Change Viewpoint icon buttonalters the viewing direction by 90degrees. Clicking it several timesallows you to rapidly view the plotfrom a number of different directions.

With Rotate mode turned on you can rotatethe plot in any direction by dragging thecursor in the Visualizer window.


page 27

As you rotate the n-dimensional scatterplot, look forclusters of points that define the edges of the distri-bution in various directions. In some cases theseare isolated small clumps of points, in others the endsof elongated clusters of points that define a mixingline between two endmembers. Stop rotating whenyou can distinguish several edge clusters. The n-Dimensional Visualizer allows you to select eachsuch cluster of spectral points and group them into aclass. The points you assign to a class are shown inthe window in the selected color. You can repeatthe grouping operation for other points to assign themto the same class, or choose a new color to createadditional classes. Class colors are maintained whenyou rotate the plot so that you can better visualizethe distribution of the spectral points. The ShowClasses option enhances visibility of classes whenonly a few spectral points are assigned to each. Ituses solid lines to connect the spectral points in theclass to the class mean.

Mark Endmember ClassesSTEPS

when you haveidentified severaledge clusters, press theClassify icon button tochange the cursor modemake sure thatthe ShowClasses iconbutton is toggled inleft-click in the plot areato place verticesdefining a polygon thatsurrounds a pointcluster (as in illustration1 below)right-click to accept thepolygon (illustration 2)click on the coloredbutton labeled Classand choose anothercolor such as cyandraw a polygon aroundanother cluster of points(2) and right click (3)press [Spin] andobserve how themarked clusters relateto the remaining pointspress [Stop]

Press the Clear Class iconbutton to release all pointsassigned to the currentlyselected class color.

Use the Class coloroption menu to select acolor for marking a class.

In Classify mode you canuse the mouse cursor todraw a polygon to select acluster of points to assign toa class.

1 2 3

Example first class

Polygon forsecond class

Press the Clear All Classesicon button to release allclassified points from theirrespective classes.

This exercise continues on the next page.


page 28

Save Endmember Classes as a MaskSTEPS

continue rotating theplot and grouping pointsuntil you have markedsix or seven classeswhen you are satisfiedwith the set of classes,press the Classesto Source iconbutton on theVisualizer windowclick [Yes] in the firstVerify window thatopens (to create a newmask raster containingthe class cells)click [No] in the secondVerify window that askswhether to save thecurrent mask raster (nochanges have beenmade that requiresaving)

Example endmember classes for the Cuprite MNF component set shown fromdifferent viewpoints in the n-Dimensional Visualizer. Your classes may belarger or smaller or use different colors. Note that a given endmember clusterlies on the edge only when seen from certain viewpoints.

to save the temporaryclass mask, click [SaveAs]save the class mask inthe same Project File asthe PPI raster youcreated earlier

A temporary mask rastershowing the locations ofcells in the marked classeshas now been substitutedfor the PPI raster you havebeen using as a mask.

Continue changing viewpoints and rotating the plotto find additional endmember clusters. As you viewthe plot, resist the natural impulse to visualize it as arotating three-dimensional volume. At any momentthe plot is showing a two-dimensional projection ofan n-dimensional hyperspace (5 dimensions in thisexample). The spatial relationships of point clusterscan only be determined by examining projectionsfrom many viewpoints. For example, a point clusterthat always appears to lie on a straight line betweenthe same two endmember clusters is probably a mix-ture of the two, rather than an additional endmember.

The Classes to Source option allows you transferthe class cell locations to the current mask raster, orto a new mask raster. Each class is assigned a uniqueraster value, and a Colormap subobject containingthe class colors is automatically created. A newmask created by this procedure is a temporary ob-ject that is automatically substituted for the currentmask. To reuse this object you must save it usingthe Save As button on the Image Mask section ofthe Info panel.


page 29

STEPSrepeat the stepson page 4 toreopen the CUPRITE97hypercube objecton the Image panel,select Subset from theUse Wavelength Rangeoption menu and set therange to 1.98 to 2.46In the Layer Managerwindow, click theAdd Raster iconbutton and selectSingle from thedropdown menuselect MNF_1 from theMNF_CUP Project Filein the Visualizer ImageMask section of the Infopanel, click [Raster...]select the Classes maskraster you saved in theprevious exercise, or theone provided in thePPI_CUP Project Fileuse the tools in theHyperspectral Imagewindow to zoom in onthe northeast quadrantof the scene and locateone of the collections ofcolored mask cells thatmark the members of anendmember classclick on theSpectral Profileicon button on theHyperspectral Imagewindowon the Info panel, turn onthe Display Spectral Plottoggle buttonmove the cursor to theindividual class cells andright-click to view theirspectra

In the final step in defining image endmembers, wecan use the endmember class raster as a mask withthe original Cuprite hyperspectral image to selectand save endmember spectra. The colored cells inthe mask identify the locations of candidate imagespectra. Other cells in the mask are transparent, re-vealing the underlying image. To enhance visibilityof the colored class cells, you can add one of thegrayscale MNF component rasters as a referencelayer. You may also want to use the controls in theLayer Controls window to hide temporarily all lay-ers except the mask in order to locate all of the classcells.

For each endmember class, use the Spectral Profiletool to extract and display the image spectra for theindividual class cells. (Use the Multiple Spectra op-tion for the Spectral Plot window.) Choose the mostrepresentative spectrum and save it in a spectral li-brary. Repeat until you have saved a spectrum foreach endmember class.

Select Endmember Spectra

Spectral Plot

0.6

0.7

0.8

0.9

Va

lue

2.1 2.2 2.3 2.4Wavelength

Zoomed viewof part of thenortheastquadrant of theCuprite sceneshowing threespatial clustersof class cellsrepresentingdifferentendmemberclasses.

Plot of several individual spectra for the endmemberclass represented by the color cyan.


page 30

Import a Field or Laboratory SpectrumSTEPS

in the Layer Managerwindow, right-click onthe entry for the MNF

reference layer andselect Remove Layerfrom the dropdownmenuin the Image Masksection of the Info panel,click [Clear...]on the Image panel,select Full from theWavelength Rangeoption menuclick the NewSpectral Libraryicon button

Spectral reflectance is the ratio of radiant energyreflected by a material at a particular wavelength tothe energy incident on the material at that wave-length. Variations in spectral reflectance withwavelength constitute a material’s reflectance spec-trum, and arise from its chemical make-up andphysical structure. Instruments called spectrometersare used to measure reflectance spectra of materialssuch as soil, rock, and vegetation. Reflectance spec-tra can be measured in the field under naturalillumination, or in the laboratory under more con-trolled conditions. Field and laboratory spectraprovide known spectral signatures that can be usedto identify different materials in a hyperspectral im-age.

You can import spectra from externalsources into a TNTmips spectral li-brary by using the Import SpectralCurve procedure in the HyperspectralAnalysis process. You can choosefrom several common spectrometer fileformats, or select the generic ASCIIformat. ASCII files must be in the .csv

format described in the earlier exercise entitled “Savean Image Spectrum as Text”. Each spectrum mustbe in a separate external file.

from the File menu onthe HyperspectralAnalysis window selectImport Spectral Curve,then ASCII Formatuse the standard SelectFiles window tonavigate to and selectfile GRNGRASS.CSV

turn on the DisplaySpectral Plot togglebutton and select thenew library spectrum


page 31

Open the USGS Spectral Libraryclick the OpenStandard SpectralLibrary icon button onthe HyperspectralAnalysis windowopen the Spectral Plotwindow if it is notcurrently open, and turnon the Single Spectrumtogglein the spectrum list,select the next-to-lastAlunite spectrum

Keep the SpectralPlot window openfor the nextexercise.

Plot of the selectedAlunite reflectancespectrum.

The United States Geological Survey’s SpectroscopyLab has measured the spectral reflectance of about500 minerals (basic components of rocks and soil)under standard laboratory conditions. They havecompiled the measurements in the USGS SpectralLibrary, which is freely available to the public.MicroImages has imported the USGS Spectral Li-brary and incorporated it within the HyperspectralAnalysis process as a Standard Spectral Library.Other public spectral libraries may be added to theStandard library list in the future.

The reflectance spectra in the USGS Spectral Librarycover a wavelength range from 0.2 to 3.0 microme-ters, with spectral bandwidth less than 0.01micrometers (comparable to the spectral resolutionof an AVIRIS image spectrum). In addition to min-eral spectra, the USGS spectral library also includesa few spectra acquired from different types of veg-etation foliage. The vegetation spectra are groupedtogether following the alphabetical listing of mineralspectra.

You can find additionalinformation about the USGSImaging Spectroscopyprogram and the USGSSpectral Library on theWorld Wide Web at:http://speclab.cr.usgs.gov


page 32

Correction to ReflectanceSTEPS

on the Image panel,choose None from boththe Additive Offset andthe AtmosphericCorrection option menus

select theSpectral Profiletool from theHyperspectral Imagewindow toolbarobtain an imagespectrum from theStonewall Playa areanear the east edge ofthe image (see page 4)

Raw radiance spectrum forStonewall Playa. Theoverall shape of thespectrum, with maximumbrightness in the visible lightrange, replicates the solarirradiance curve and thedropoff in atmosphericscattering with increasingwavelength.

An airborne or satellite-mounted hyperspectralimager measures spectral radiance: the upwellingradiant energy at the sensor for many narrowwavelength bands. These radiance values are onlypartly determined by the spectral reflectance

properties of surface materials inthe scene. A number of otherfactors contribute to the measuredradiance, with effects that vary withwavelength. For example, theamount of solar energy illuminatingthe surface varies with wavelength

(described as the solar irradiance curve) and withthe sun’s height in the sky when the image wasacquired (solar zenith angle). Other factors includewavelength-selective absorption and scattering byatmospheric gases and particulates, and illuminationdifferences due to topography (direction and angleof slope, and shadowing).

In order to make meaningful comparisons betweenimage spectra and laboratory reflectance spectra,the image radiance values must be corrected (cali-brated) to reflectance by removing these othercontributing effects. The Hyperspectral Analysisprocess in TNTmips offers several commonly-usedempirical correction methods. These methods areapplied to the raw hyperspectral image on-the-fly,so you do not need to recompute and store different

versions of the image. The steps in thisexercise disable these calibrations so thatyou can examine raw radiance spectra.

AB

CD

Several major lows in the radiance spectrum(labeled A through D) result from strongatmospheric absorption at specific wave-lengths. If the atmosphere was uniform overthe scene and elevation differences are notextreme, raw spectra from all image cellsshould include about the same effects fromatmospheric scattering and absorption, aswell as the same solar irradiance spectrum.

crobbins

Highlight


page 33

The Flat Field correction method reduces atmo-spheric and solar irradiance effects using the meanspectrum of an area you designate with the Flat FieldCalibration tool. The area you select should be to-pographically and spectrally flat: it should have auniform spectral reflectance at all wavelengths, with-out significant absorption features. The averagespectrum from such a “flat field” will be an almostpure signature of solar irradiance plus superimposedatmospheric scattering and absorption effects. TheFlat Field method divides each image spectrum bythe flat field spectrum. With an appropriate flat fieldspectrum, this procedure will largely remove atmo-spheric and solar irradiance effects. (The selectedarea should be bright in order to reduce the effectsof random image noise on the correction.) If thescene includes large variations in elevation or theatmosphere was not uniform over the scene, someresidual topographic and atmospheric effects willremain in the calibrated image spectra.

Choosing an appropriate flat field isobviously a critical step in this cor-rection method. No natural materialhas a completely flat reflectancespectrum. Bright man-made materials such as con-crete have a fairly flat spectral response and thuscan provide appropriate targets in hyperspectralscenes of urban areas. In the Cuprite scene, thebright, salt-encrusted surface of Stonewall Playa hasno significant mineral absorption features, so it pro-vides an approximate flat field spectrum.

The Maximum correction method assembles an ap-proximate flat field spectrum automatically using themaximum value in each spectral band. Image spec-tra are divided by the maximum spectrum to producethe corrected image spectra. This method may beappropriate if the scene includes a variety of materi-als of differing overall brightness.

Flat Field CorrectionSTEPS

use the standard Viewwindow tools to zoom inon Stonewall Playaselect the FlatField Calibrationtool from theHyperspectral Imagewindow toolbaruse the standard Line/Polygon Edit Controls todraw a polygon coveringthe brightest part ofStonewall Playa (seeillustration)

choose Flat Field (ByRegion) from theAtmospheric Correctionoption menu

Draw a polygon with theFlat Field Calibration tool toindicate the area to be usedto determine the flat fieldspectrum.

reacquire an imagespectrum from StonewallPlaya

crobbins

Sticky Note

Click on the Spectral Plot icon in view window and right-click. This will re-acquire the spectrum from previously selected point.

crobbins

Highlight

crobbins

Sticky Note

Use mouse pointer to draw and define the area to use for calibration, then righ-click to apply.


page 34

STEPSselect Equal AreaNormalization from theAtmospheric Correctionoption menuselect Minimum from theAdditive OffsetCalibration option menureacquire an imagespectrum from StonewallPlaya

Equal Area NormalizationEqual Area Normalization is the default method ofatmospheric correction in Hyperspectral Analysis;it is enabled automatically when you start the pro-cess. This calibration method does not require anyknowledge of surface materials in the scene. Theradiance values in each image spectrum are firstscaled so that the sum of the values for each imagecell is constant over the entire scene. This proce-dure shifts all image spectra to nearly the same

relative brightness, removing differ-ences in overall brightness betweenmaterials as well as illumination differ-

ences caused by topography. An average spectrumfor the entire scene is then calculated from the nor-malized spectra. For a scene that includes manydifferent land cover types, averaging may largelyremove spectral features attributed to the surfacematerials. The average spectrum should then re-semble the flat field spectrum described in theprevious exercise. Finally, each normalized imagespectrum is divided by the average spectrum. Theresulting spectral values represent reflectance rela-tive to the average spectrum, and in ideal casesshould be comparable to true reflectance spectra.

Large elevation differences or variations in the at-mosphere across the scene may cause atmosphericeffects on the image spectra that are not removed byEqual Area Normalization. In addition, if there are afew widespread materials in the scene, the averagespectrum may include absorption features related tothese surface materials instead of representing onlyatmospheric and solar irradiance effects. In thesecases the calibrated spectra produced by Equal AreaNormalization may differ significantly from true re-flectance spectra.

NOTE: If the hyperspectral image has already been processed to reflectance, you canchoose the Manual options on the two Calibration menus and set numeric values for offsetand scaling. For example, sample AVIRIS scenes calibrated to reflectance by NASA useinteger values from 0 to 10,000 to correspond to the reflectance range 0 to 1.0. In thiscase set the Additive Offset value to 0 and the Atmospheric Correction value to 10,000.

The methods on theAdditive Offset Calibrationoption menu correct foratmospheric backscatterand instrument artifacts thatraise the measured radiancevalues by a constantamount for each cell in animage band. The defaultMinimum method deter-mines the minimum value ineach band and subtracts itfrom all cells in the band.The Dark Field Calibrationsubtracts from each cell theaverage determined from anarea that you outline usingthe Dark Field Calibrationtool on the Hyper-spectral Imagewindow. The areayou select should be dark inall bands; a deep waterbody or heavily shadowedarea would be the besttarget.


page 35

STEPSpress the Fullicon button on theHyperspectral Imagewindowselect Subset from theUse Wavelength Rangeoption menu on theImage panel, and set thewavelength range from1.98 to 2.46in the spectrum list forthe USGS SpectralLibrary, reselect thenext-to-last Alunitespectrumpress [Add] on theClassify tabbed panel toadd the Alunite spectrumto the End Member listuse the default SpectralAngle Mapper algorithm,and set the ThresholdValue to 15press [Classify] anddirect the output rastersto your classificationProject File

Classification with Reflectance SpectraOnce the hyperspectral image has been calibratedto reflectance, you can use reflectance spectra ofthe materials in the USGS Spectral Library (or anyother library of reflectance spectra) to locate andclassify image cells with similar spectra. You use thesame classification methods and follow the same pro-cedures outlined in the previous exercises in thisbooklet on spectral classification, but select libraryreflectance spectra instead of archived image spec-tra as end members. The spectral classifiers allowyou to produce a series of material maps (class ras-ters) showing the distribution of image cells matchingthe target spectrum within the threshold you haveset.

With the empirical calibration methods currentlyimplemented, the Spectral Angle Mapper and CrossCorrelation methods produce the best results usingreflectance spectra. For the Cuprite scene, the de-fault Equal Area Normalization does not appear toachieve a sufficiently accurate reflectance calibra-tion to produce good results with the MatchedFiltering method using reflectance spectra.

Class raster displayedover the SpectralAngle raster, with cellsassigned to theAlunite class shown inred.


page 36

Spectral MatchingThe spectral matching procedure is useful for iden-tifying specific unknown materials in a hyperspectralimage. You can save an image spectrum for the tar-get material into a copy of a reference spectral library,then search the library for the reflectance spectrathat best match the unknown spectrum. (In this ex-ercise you use a library containing the reflectancespectra in the USGS Spectral Library along withthree Cuprite image spectra.) The matching processfinds the twenty spectra that best match the targetspectrum and ranks them by closeness of fit.

Two spectral matching methods are available. TheSpectral Angle Mapper method is identical to theclassification method described previously. TheBand Mapping method compares spectra on the ba-sis of the positions and relative depths of absorptionfeatures (local minima in the spectral curves). Bothmethods are relatively insensitive to differences inaverage brightness (note the large vertical separa-tion between the target spectrum and best matchingspectrum in the spectral plot).

In this example, the top five matchingspectra are alunite laboratory spectra.

337,336_Aluniteimage spectrum.

Best matchingspectrum from theUSGS SpectralLibrary: an alunitereflectancespectrum.

STEPSclick [Spectral Library...]on the HyperspectralAnalysis windowselect object SAMPLIB3from the SAMPLIBS ProjectFilein the Plot controls onthe Info panel, turn onthe Multiple Spectraradio buttonin the Spectral Plotwindow, turn on theupper Manual Rangetoggle button and set theWavelength Range from1.98 to 2.46scroll to the bottom ofthe spectrum list andselect 337,336_Aluniteclick the Math tab, andselect Band Mappingfrom the Algorithmoption menu in theSpectral Matchingcontrol panelpress [Find Match]select the firstspectrum in thelist of matchingspectra


page 37

Remove Continuum from ImageMany minerals can be recognizedby the position and shape of oneor a few narrow absorption featuresin a particular part of the spectrum.These features (commonly called“bands”) appear as valleys in thespectrum. In a composite spectrumfrom a hyperspectral image, theshape of the spectral curve sur-rounding an absorption band mayvary. This broader curve shape,or continuum, is the aggregate re-sponse of all the materials thatcontribute to the spectrum. A sloping continuumcauses an apparent shift in the location of an ab-sorption band’s reflectance minimum. Spectralmatching of individual absorption bands (such as inthe Band Mapping method) is more accurate if thecontinuum is removed from the spectra prior tomatching. This involves dividing the spectral valuein each channel by the continuum value.

You can use the Remove Con-tinuum option on the Image menuto create a continuum-removed ver-sion of the hyperspectral image (forthe current wavelength subset).The operation computes the con-tinuum for each image cell, thencreates and saves the correspond-ing continuum-removed spectrumfor the cell. Spectra you extract fromthis processed image can then bedirectly compared to continuum-re-moved library spectra. The nextexercise discusses how to removethe continuum from library spectra.

Spectral plot showing the kaolinite imagespectrum (bottom) from the SAMPLIB3 spectrallibrary, the continuum computed for it (smoothline), and the continuum-removed kaolinitespectrum (top).

STEPSselect RemoveContinuum from theImage menu on theHyperspectral Analysiswindowuse the Select Fileprocedure to name anew Project File tocontain the continuum-removed image bands

RGB display of continuum-removed bands172, 187, and 202. To examine the imagewithout interrupting the continuity of theseexercises, launch another instance of theHyperspectral Analysis process and openthe continuum-removed image there.

NOTE: the continuum-removed imageis created with a separate rasterobject for each band rather than as asingle hyperspectral object.

crobbins

Sticky Note

As shown here, a continuum line for the selected library spectrum can be added to the Spectral Plot via Edit tab, Operation setting.


page 38

Remove Continuum from Library SpectraYou can remove the continuum from individualspectra in a library by using the controls on the Edittabbed panel. The spectrum you select from thelibrary is shown in tabular form in this panel. SelectContinuum from the Operation menu to merelycompute the continuum for display in a spectral plot,or Remove Continuum to compute a continuum-removed spectrum from the original library spectrum.

When you apply any of the mathematical transfor-mations selected from the Operation menu, thespectrum table on the Edit panel is updated and theresult is stored temporarily in memory. Use the Re-store button to remove the transformation or theSave Spectrum button to save the result as a newspectrum in the library. This spectrum is assignedthe same name as the parent spectrum. You can usethe controls on the Info panel to edit this name andresave the spectrum.

STEPSselect the550,94_Kaolinitespectrum from theSAMPLIB3 spectral libraryin the Plot portion of theInfo panel, turn on theSingle Spectrum radiobutton, then the MultipleSpectra buttonin the Spectral Plotwindow, turn on thelower Manual Rangetoggle button and set theValue Range from 0.5 to1.0click on the Edit tabfrom the Operationmenu, select Continuum,then press [Apply]press [Restore]select RemoveContinuum from theOperation menu, thenpress [Apply]press [Save Spectrum]to save the continuum-removedspectrum in thelibrary

The steps in thisexercise create aspectral plot similarto the one illustratedon the precedingpage.


page 39

Log residuals for bands 172, 187, and 202displayed respectively as R, G, and B. (Theoriginal RGB combination is shown on page 8).

Other Image Processing OptionsThe Image menu on the Hyperspectral Analysis window offers several additionalprocessing options which are summarized briefly below.

Log Residuals is a calibration technique that models the observed radiance valuefor each image cell as the product of the surface reflectance, a topographic factor,and an illumination factor. The topographic factor (slope direction and shadow-ing) is constant for all wavelength bands for any image cell, but varies spatially.The illumination factor (solar irradiance) is constant over the image (assuming acloud-free sky), but is unique for each wavelength. In order to determine reflec-tance from the measured radiance value, we need to remove the effects of thetopographic and illumination factors, neither of which is usually known. TheLog Residuals algorithm mathematically eliminates these factors by performinga series of divisions for each image cell using the measured band values, thegeometric mean over all wavelength bands for each cell, and the spatial geomet-ric mean for each image band. Thegeometric means are computedfrom the arithmetic means of thelogarithms of the measured values.These calculations are too in-volved to be done on-the-fly, so theLog Residuals process produces atransformed set of image bands.The resulting log residual imagespectra may lack some of the origi-nal image information. If anymaterial spectrum is present in allimage spectra, that spectral com-ponent will be removed by theprocess. Log residuals images alsotend to accentuate image noise.

VQ Filtering is an image transformation procedure that is related conceptuallyto the Self-Organizing Map (SOM) classifier discussed earlier in this booklet. Theprocedure applies the SOM classifier to the image to derive 256 spectral classes,and assigns each image cell to the class with the closest mean in spectral space.Then the mean spectrum for the assigned class is subtracted from each observedspectrum. The resulting residual (filtered) image of each input band shows howwell (or how poorly) each cell fits its assigned class at that wavelength. Youmight use the results to pinpoint the locations of unusual (but potentially inter-esting) materials in the scene.


page 40

Advanced Software for Geospatial Analysis

MicroImages, Inc.

HYPERSPECTRAL

MicroImages, Inc. publishes a complete line of professional software for advanced geospatialdata visualization, analysis, and publishing. Contact us or visit our web site for detailed productinformation.

TNTmips Pro TNTmips Pro is a professional system for fully integrated GIS, imageanalysis, CAD, TIN, desktop cartography, and geospatial database management.

TNTmips Basic TNTmips Basic is a low-cost version of TNTmips for small projects.

TNTmips Free TNTmips Free is a free version of TNTmips for students and profession-als with small projects. You can download TNTmips Free from MicroImages’ web site.

TNTedit TNTedit provides interactive tools to create, georeference, and edit vector, image,CAD, TIN, and relational database project materials in a wide variety of formats.

TNTview TNTview has the same powerful display features as TNTmips and is perfect forthose who do not need the technical processing and preparation features of TNTmips.

TNTatlas TNTatlas lets you publish and distribute your spatial project materials on CD orDVD at low cost. TNTatlas CDs/DVDs can be used on any popular computing platform.

IndexAuto Define wavelength.......................10Band mapping (spectral matching)..........36continuum, removal.........................37-38classification

of image......................................15-21in n-Dimensional Visualizer...............27using reflectance spectra...................35

continuum.. . . . . . . . . . . . . . . . . . . . . . . . . . . . .37,38cross-correlation.....................................18equal area normalization..........................34flat field correction to reflectance............33Hyperspectral Explorer.......................7-9hyperspectral (hypercube) object..............4

3D display of.......................................5RGB display of...........................4,6-9

illumination effects...............................32imaging spectrometer.....................4,11,18import, spectrum................................30irradiance, solar..................................32,33linear unmixing..................................19,20log residuals calibration method..............39

matched filtering..........................16,18minimum noise fraction transform.......23n-Dimensional Visualizer.................25-28pixel purity index.................................24principal components transform.........22self-organizing map classifier................21spectral angle mapper.................17,18,36spectral libraries....................13,15,29,31

USGS spectral library......................31spectral matching................................36Spectral Profile tool........................11,29spectral radiance..................................32spectral reflectance........................30,31

correcting radiance to...............32-34spectrum (spectra)

endmember...................19,20,24-29image...............................3,11,13-20reflectance...............................32-36saving in library..............................13saving as text...........................14viewing......................................11

analyzing hyperspectral images · images are covered on pages 24-29. pages 30-36 introduce methods...

Documents