Rochester Institute of Technology
RIT Scholar Works
Theses
Thesis/Dissertation Collections
8-1-1990
Merging panchromatic multispectral images for
enhanced image analysis
Curtis K. Munechika
Follow this and additional works at:
http://scholarworks.rit.edu/theses
This Thesis is brought to you for free and open access by the Thesis/Dissertation Collections at RIT Scholar Works. It has been accepted for inclusion in Theses by an authorized administrator of RIT Scholar Works. For more information, please contactritscholarworks@rit.edu.
Recommended Citation
MERGING PANCHROMATIC AND
MULTISPECTRAL IMAGES
FOR
ENHANCED IMAGE ANALYSIS
by:
CURTIS K. MUNECHIKA, Captain, USAF
A thesis submittedinpartial fulfillment of the requirements for the degree of Master of Science
in the Center for Imaging Scienceinthe College of Graphic Arts and Photography at the
Rochester Institute of Technology
August 1990
Signature of the Author ....lo.(...1JLLrtLLiS:>...JKL>...l.lMUlI..u.IO.JJ:e:.Jo...cL1JbiL.Do.ka..a _
Accepted by _
Center for Imaging Science College of Graphic Arts and Photography
Rochester Institute of Technology Rochester, New York
CERTIFICATE OF APPROVAL
M.S. DEGREE THESIS
The M.S. degree thesis of Curtis Munechika has been examined and approved by the thesis committee as satisfactory for the thesis requirement
for the Master of Science degree
Dr. John R. Schott, Thesis Advisor
Lt. Col. Phil Datema
Mr. Carl Salvaggio
d3
406
/0
Center for Imaging Science
College of Graphic Arts and Photography Rochester Institute of Technology
Rochester, New York
THESIS RELEASE PERMISSION FORM
Title of Thesis: Merging Panchromatic and Multispectral Images for Enhanced Image Analysis
I, Curtis K. Munechika grant pennission to the Wallace Memorial Library of the Rochester Institute of Technology to reproduce this thesis in whole or in part
provided any reproduction will not beof commercial use or for profit.
MERGING PANCHROMATIC AND
MULTISPECTRAL IMAGES FOR
ENHANCED IMAGE ANALYSIS
by
Curtis K. Munechika
Submittedto theCenter forImagingScience in
partialfulfillmentoftherequirementsforthe
MasterofScience degreeatthe
Rochester InstituteofTechnology
ABSTRACT
This study evaluates severalmethods that enhance the spatial resolution of multispectral
images using afinerresolutionpanchromatic image. Theresultanthybrid,highresolution, multispectral data set has increased visible interpretation and improved classification
accuracy, while preserving the radiometry of the original multispectral images. These
ACKNOWLEDGEMENTS
Thiswork could nothave been completed without the supportofmanypeople. I would first like to thank theUnited StatesAir Force forallowingmethe opportunity to study at
theCenterofImaging Science. Theexperiencehas beenexceptionallyrewarding.
Iwouldalsoliketo thankmycommittee membersforalltheirassistance:
toDr. JohnSchott,for providing theoverall guidance,focus, and an occasionalkick in the pants;
to Lt Col Phil Datema, for his time, patience, and support in maintaining a long distance workingrelationship; and
toCarl Salvaggio, for answering allofmynot-so-quick questions.
Specialthanks alsoto: Jim Warnickwhofirst startedthis study andwasmy best sounding board; toDr. Roger Eastonwhosecompanyattheoddesthours wasappreciated; to Steve Schultzwhorescuedmyworkfromaterminallyillcomputer; andtoLauraAbplanalpand
Leslie Gentherwhohelped in puttingthis together.
Lastly I wish to acknowledge my comrades-in-arms fortheir encouragement, assistance,
and abuse: RanjitBhaskar, Betsy Fry,GopalSundaramoorthy, WendyRosenblum, Steve
DEDICATION
Thisworkis dedicated tomy family...
To myparents,forinstillingmy desiretolearn (and fortheirbabysitting);
To mybrother,for his insightandtime (as wellasuseofhis computer);
andespecially
to
Table of Contents
ListofFigures ListofTables
1.0 Introduction 1
1.1 Spatial, Spectral, Radiometric Resolution 2
1.1.1 Resolution and Classification Accuracy 3
1.1.2 Resolution Trade-offs 5
1.2 Overview on Merging Images 7
1.2.1 Merging for Display 7
1.2.2 Merging bySeparate Manipulationof
Spatial Information 9
1.2.3 Merging and Maintaining Radiometry 10
2.0 Merging Methods and Modifications 12
2.1 DIRS Method for Merging 12
2.1.1 Results Using the DIRS Merging Method 13
2.1.2 Concerns on the DIRS Merging Method 13
2.2 General Modifications to the DIRS Method 17
2.2.1 Creating a TM Panchromatic Image 17
2.2.2 Interpolated TM Input Images 24
2.2.3 A Technique for Radiometric Post-Correction 25
2.3 DIRS Enhancement 1 Method for Merging 27
2.4 DIRS Enhancement 2 Method for Merging 29
2.4.1 A Modification to DIRS Enhancement 2 Method 36
2.5 Price's Method for Merging 37
2.5.1 Case 1 - Correlated Input Images 37
2.5.2 Case 2
-Weakly Correlated Bands 40
2.6 Price's Modification ~ A Method
ofHandling
Weakly Correlated Bands 42
3.0 Experimental Approach 46
3.1 Selection and Preparation of Test Imagery 46
3.1.1 Registration of Images 47
3.1.2 Blurred Image Sets 49
3.1.3 Figures of Images 50
3.2 Classification of Test Imagery 54
3.2.1 Training Samples for Classification 54
3.2.2 Measuring Classification Accuracy 55
3.2.3 Testingthe SignificanceofOverall Classification Accuracy 56
3.3 Radiometric Error Analysis 57
3.4 Running the Methods and Modifications 59
3.4.1 Running DIRS Enhancement 2 59
3.5 Experimental Procedure 61
3.5.1 Pan 1
-Merging with Coarser Resolution Data Sets 61 3.5.2 Part 2
-Merging withCoarser Resolution DataSets with a
Lower Registration Error 64
3.5.3 Part 3 Merging with Original Resolution Data Sets 64
4.0 Results 65
4.1 The New Synthetic Panchromatic Image 65
4.2 Computing the Error in Reflectance 70
4.3 Results for Part 1
-Merging with Coarser Resolution Data Sets 75 4.3.1 Observations from Part 1
-MergingwithCoarser
4.4 Results for Part 2 Mergingwith CoarserResolutionDataSets
with a Lower Registration Error 80
4.4.1 Observations from Part 2 Mergingwith Coarser
Resolution Data Sets withaLowerRegistration Error 84
4.5 Results for Part3
-Merging withOriginal Resolution DataSets 85 4.5.1 Observationsfrom Part3
-Merging with
Original Resolution Data Sets 90
4.6 Other Comparisons Between Merging Methods 90
4.6.1 Visual Comparisons 90
4.6.2 Implementation Comparisons 99
5.0 Conclusions 101
6.0 Recommendations 103
6.1 Improvements to the Price Methods 103
6.2 Improvements to the DIRS Methods 104
6.2.1 Recommendations in Processing Correlated Bands 104 6.2.2 Recommendations in Handling Weakly Correlated Bands 106 6.2.3 Improvements to the Interpolated Inputs 107 6.2.4 Improvements to the DIRS Enhancement 2 Method 107
6.3 General Considerations 109
Appendix A Classification Results for Original TM Data
Appendix B Classification Results for Hybrid Data(30m)using
Blurred TMandSPOT Data
AppendixC- Radiometric Results fortheHybrid Data
(30m)using Blurred TMandSPOT Data
AppendixD Classification Results for Hybrid Data(30m)using
BlurredTM andRe-registered SPOT Data
Appendix E~ Radiometric ResultsfortheHybrid Data
(30m)using Blurred TMandRe-registered SPOT Data
Appendix F Classification Results for Hybrid Data(10m)using
TMandRe-registered SPOT Data
AppendixG - Statistical Tests
AppendixH~
Input ParameterstoComputetheWeightingFactors
foraSynthetic Panchromatic Image
Appendix I
List of Figures
Figure 1-1 Pradine'sMergingMethod
Figure2-1 Superpixeland subpixels
Figure2-2 Reflectance SpectraforSelectedUrbanTargets Figure 2-3 Reflectance SpectraforSelectedSoilTargets Figure2-4 ReflectanceSpectra for SelectedWaterTargets
Figure2-5 Reflectance SpectraforSelected TreesTargets
Figure 2-6 Reflectance SpectraforSelected Grass Targets Figure 2-7 Pixel Replication
Figure 2-8 InterpolatedInput
Figure2-9 DIRS NearestNeighbor Correction Routine Figure2-10 DIRS Enhancement1 Correction Routine
Figure2-1 1 Connectionsbetweenthe
center subpixel
Figure2-12 Flow DiagramofDIRS Enhancement2MergingMethod Figure 2-13 Linear RegressionofTM Bands
andthePanchromatic Image Figure 2-14 Price'sMergingMethod- Case 1
Figure2-15 Creationof aLook-upTable (LUT)
Figure 3-1 Original TM (30mGIFOV)Bands 5,3,2 in RGB
Figure3-2 Blurred TM (90mGIFOV)Bands 5,3,2 in RGB Figure 3-3 OriginalSPOT (10mGIFOV)
Figure4-1 PlotofTM 1 DCversusEstimatedReflectance
Figure 4-2 PlotofTM2 DCversusEstimated Reflectance
Figure 4-3 PlotofTM3 DCversusEstimated Reflectance
Figure4-4 PlotofTM4 DCversusEstimated Reflectance
Figure4-5 ComparisonofCoarser Resolution Images
Figure 4-6 ComparisonofOriginal ResolutionImages
Figure 4-7 Replicated TM Bands 5,3,2 (30mGIFOV) (magnification4x)
Figure4-8 Interpolated TM Bands 5,3,2 (30mGIFOV)(magnification4x) Figure 4-9 DIRS MethodwithInterpolated Input (magnification4x)
Figure4-10 DIRSMethodwithInterpolatedInputandPost-fixing(magnification4x) Figure4-11 DIRSEnhancement 1 withPost-fixing (magnification4x)
Figure4-12 DIRSEnhancement2withTM4,TM5modification(magnification 4x)
Figure4-13 Price's Method (LUT)(magnification4x)
List of Tables
Table 1-1
Table 3-1 1
Table3-2 Table3-3
Table 3-4 ]
Table3-5 I Table 4-1
Table 4-2
Table 4-3
Table 4-4 !
Table4-5 I
Table 4-6
Table 4-7a
Table 4-7b I
Table4-7 ! Table 4-8a I
Table4-8b :
Table4-8c
Table 4-8 I
Table 4-9a
Table 4-9b
Table 4-9c !
AComparisonofMultispectral Scanners(MSS) and
Thematic Mapper(TM)
SceneAcquisition andEphemeris Data
Scene Statistics
Transformation Coefficients
New Transformation Coefficients SummaryTableofMergingMethods
SummaryofWeightingFactors usedtoGenerate theTM Panchromatic Image
SummaryoftheHistogram StatisticsforthePanchromatic Images
SummaryoftheLinearCoefficientsusedfor Adjustment
SummaryoftheError Differences Selected Control Points
Averaged DCandEstimated Reflectance
SummaryofClassification Breakdown byPercentage (30mhybrid) SummaryofClassificationAccuracy with
Independent DataSet 1 (30mhybrid)
SummaryTable for 30m HybridResults
SummaryofClassification BreakdownbyPercentage
(re-registered 30mhybrid)
SummaryofClassificationAccuracywith
IndependentData Set 1 (re-registered 30mhybrid)
SummaryofClassificationAccuracywith
aRandomData Set 1 (re-registered 30mhybrid) SummaryTable for Selected 30m Hybrid Sets usinga
Re-registered SPOT imageasInput
SummaryofClassificationBreakdownbyPercentage (10mhybrid) SummaryofClassificationAccuracywith a
Random DataSet 1 (10mhybrid)
SummaryofClassificationAccuracywith
Table 4-9d SummaryofClassificationAccuracywith Independent Data Set 2 (10mhybrid)
1.0 INTRODUCTION
Spatial resolution is an importantparameter in image interpretation. However, to
capture multispectral images or images within a narrower spectral bandpass, spatial
resolution is often diminished. In general, if a sensing system has fine spectral discrimination,thenitis physicallydifficulttoalsohave fine spatial resolution.
The emphasis of this study is to enhance the spatial resolution of multispectral
images using data from another sensor. In particular, this study evaluates several
techniques tomerge the medium resolutionThematic Mapper(TM) multispectral images
with a high resolution panchromatic image captured from the SPOT satellite. The
performance ofeach merging technique is measured by classification accuracy on the hybrid images, as well as on how well the hybrid images maintain the radiometric and
spectralinformationoftheTM data.
The results of this study show that these merging techniques can produce high
resolution multispectral images for enhanced image analysis. Not only is visible
interpretation clearly improved, butclassification accuracy isalsoincreased. In addition,
theresultanthybridimagesadequatelymaintainthe TMradiometry.
Fortheremainderofthisintroductorychapter, section 1.1 discusses some concepts
on resolution in theremote sensing arena. It covers the effects that different types of
1.1 Spatial, Spectral, and Radiometric Resolution
Ingeneralterms,resolution can bethoughtof as theabilityofa systemtodistinguish
fine detail. Most often, we think of resolution in terms of "spatial resolution", orhow
well we can resolve the spatial detail in an image. As such, there are many ways to
measure orquantify the spatial resolution of an imaging system (i.e. by themodulation
transferfunction, ground resolvabledistance,etc).
For simplicity, the ground instantaneous field-of-view (GIFOV) is used in this
study. TheGIFOV is theprojection ofthelimitingdetector aperture ontothe ground. Itis
similar to the ground sample distance and has units of length. Thus the smaller the GIFOV, the smaller the sampling distance on the ground, and the finer the spatial
resolution ofthe system.
However, in the field of remote sensing, and especially with multispectral data,
thereareotherformsofresolutionthatareequally asimportant.
"Radiometric resolution"
is determinedby thenumberof effective grey levels that
are availableto the systemtorepresentthescene brightness. Forexample, an8-bitsystem
would be ableto record an image in 256 levelsofbrightness. Its radiometric resolution
wouldbe finerand more precisethan a6-bitsystem which canonlywork with64 levels.
"Spectral resolution"
can be thought of as the width of the bandpass over which
radiance is measured. The more narrow this wavelength interval, the finer the spectral
resolution. Intuitively, thefineryourspectral resolution,themore spectralbandswe could
more spectralinformation willbeavailable.
1.1.1 Resolution and Classification Accuracy
Theeffect ofthese three types of resolution onclassifyingimageshas been studied,
particularlysince theintroductionoftheThematic Mapper.
The ThematicMapper(TM) was launchedaboardLandsat-4 onJuly 16, 1982. In
comparison with the older Multispectral Scanners (MSS), TM provided finer spatial
resolution, narrower and more optimally placed spectral bands, and finer radiometric
precision. Acomparison chartis shownbelow.
Table 1-1
AComparisonofMultispectral Scanners(MSS)and ThematicMappej(TM)
MSS TM
GIFOV
(m) 80
30 (bands1-5,7)
120 (band6)
Spectral
Bandpass
(micron)
0.5 - 0.6 0.6- 0.7 0.7-0.8 0.8-1.1
0.45- 0.52 0.52- 0.60 0.63- 0.69 0.76- 0.90 1.55
- 1 .75 10.40 - 12.50
2.08- 2.35
Quantization Levels 64 (6Ms) 256 (8bits)
determine their effect on classification accuracy. Even before the TM was launched,
Sadowski et al [77] used simulated data to find that classification accuracy would be
significantlyenhanced with theadditionalTM spectral bandsandtheincreasednumber of quantization levels. However, they alsodiscovered thatclassification accuracy actually
decreased as spatial resolution became finer. These results were corroborated by
Morgenstern et al [77], Latty and Hoffer [81], Markham and Townshend [81], and
Williamset al[84].
The reason behind the apparent lack of effect that finer spatial resolution has on
classification is becauseoftwooffsettingeffects. Improving spatial resolution will sharpen
boundaries and reduce theamountofmixed pixels in an image. This reduction in mixed
pixels contributesto an increase inclassification accuracy. However, improving spatial
resolution also increases the within-class spectral variance causing a decrease in
classification accuracy [Landgrebe77, and Irons et al 85]. Thus depending on the input
image, increasing spatial resolutionmayormaynot aidincomputerclassification - even
though the advantages of increased spatial resolution appearobvious when conducting
manual photointerpretation.
Many of these studies concluded that the commonly-used per-pixel Gaussian
maximumlikelihood(GML)classifierdidnoteffectivelyusetheadditionalinformationthat
increased spatial resolution provides. However, in these cases, only spectral bands were
usedasinputto theGMLclassifier.
To better use this information in high spatial resolution images, current classifiers
have beenfollowingeitheroftwopaths,one usingthetextural features found in theimage
and the other using context to support classification. For two of the more well-known
papers on using textureand context forclassification, referto Haralick [79] and Gurney
Using these types of classifiers which apply spatial information, classification
accuracieshave increasedwith morespatial detail [DiZenzoet al 87,HjortandMohn 84, Warnick et al 89]. Rosenblum [90] found that the optimal set of input features for
classification ofhighresolutionairphotos containbothspectral andtexturalinformation.
Generally, if a classifier can effectively use all spatial, spectral, and radiometric
information, classification accuracies can be improved with finer spatial, spectral, and
radiometric resolution.
1.1.2 Resolution Trade-offs
Unfortunately, asin manyrelationships, therearetrade-offsbetween thedifferenttypes
of resolution. Forexample, ifwe wished tohave a systemwith finespatialresolution, it
would be very difficult for this system to additionally have fine spectral resolution. To improvethespatial resolution of ourscanningsystem, we wouldrequirea smallerGIFOV. A smaller GIFOV means the energy reaching the sensor has originated from a smaller
ground area and ifother parametersremain constant-- this
meansless energy reaching the sensor. Lower levels ofenergy means a lower signal level available to the sensor,
resulting in a lower signal-to-noise ratio (SNR). To produce useful output, the sensor mustbe ator above athresholdSNRtodistinguisha signalfromthenoise. Tocompensate thislowerSNR due toa smallerGIFOVwe couldbroaden thespectral bandpassto allow
moreenergythrough
-i.e., spectral resolution is sacrificedto compensateforthe gainin
spatial resolution.
Conversely, ifwe wished ourdata tobe captured within a narrower spectral band,
again less energywouldbe incidentonthedetector. Increasing theGIFOV(losing spatial
acceptableSNRismaintained.
Another trade-off mentioned by Green [88] is a simple data handling problem.
Supposewehad 8Mbitsofdigitalstorage. Wecould approximatelystore either a:
1000x 1000pixel image,8bits/pixel, 1 spectralband;
or
380x 380pixelimages, 8 bits/pixel,7 spectralbands;
or
1000x 1000pixelimages,4bits/pixel,2 spectralbands.
The first system provides high spatial resolution but limited spectral information. The
second example provides a significant amount of spectral data, but at a low spatial
resolution. The third system provides some spectral resolution athigh spatial resolution, but at much lower radiometric precision than the first two systems. Other similar data
handlingconstraintsthatcoulddrive informationtrade-offsarelimitations in transmission
and acquisitiontime.
All these system trade-offsare dependent on thecurrent state oftechnology. With
timeandtechnologicaladvances,information trade-offswillbecomelesssevere. Butgiven
existing data andexistingremote sensing systems, we can overcomethese current trade
offs bycombining data from two complementary systems. The resultant hybrid data set
would contain the "best"
1.2 Historical Overview on Merging Images
Toimprovetherelatively coarse spatial resolution of multispectralimages,theidea
of merging data from different sensors has been attempted. Early studies reported
successful merging with the LandsatMSS images (80 m GEFOV). Radar images from
airborne systems and the Shuttle Imaging Radar
(SIR-A) have been merged with MSS
images toenhance geologicalinterpretation [Daily 79; Chavez 83]. LauerandTodd [81]
combined MSS and RBV images to obtain a hybrid product, while Schowengerdt [82]
mergedMSSwiththeHeatCapacity MappingMission(HCCM)data.
With theintroduction oftheTMand SPOTsensor systemswhich have finerspatial
and spectral resolutions, further studies on merging werereported. All ofthese merging
techniques, as well as those done with.MSS images, could basically be classifiedinto 3
categories.
1.2.1 Merging for Display
Price [87] coined the first category of merging techniques as the "ad hoc"
approaches. Since the primary concern ofthese methods was to optimizeimage display,
some unusual operations were done on the data. However, theresultanthybrid images
appeared tohave increased spatial resolution. Welshet al [87] summarizedthese ad hoc
methods intotwoequations:
M'i = aj x (m- x p)1/2
+ b; (1-1)
or
where:
M|'
= thedigitalcount
(DC)fora pixelinthei-th bandofthemergedimage;
M-= DC forthe
correspondingpixelinthei-thmultispectralimage;
P = DC for the
correspondingpanchromatic referenceimagepixel;
(>l ,a>2 = weightingfactors;
ai ,
bj
= scaling factorstooptimizethedynamicrange; and= operatorwhich couldbe
addition, subtraction, multiplication, ratio,etc.
Using these methods and simulated SPOT data, Cliche [85] integrated the
panchromatic channel into the multispectral channels to significantly improve visible
interpretation. Chavez [84] added edge enhanced (simulated) SPOT data toTMimages.
Chavez [86] also mergedTMdatawithadigitizedpanchromatic photograph. Hashim [88]
simply
"overlaid"
registeredTMandMSS imagestoobtain ahybridproduct.
Currently, many Geographical Information Systems (GIS) which contain layersof
information intheirdatabase,implementsome sortofdatacombination fordisplay to the
user [Welsh 85; Walsh et al 87]. Since a high priority of a GIS is to optimize image
display, "ad hoc"
approaches to merging are frequently used. One of the more simpler
methodsforanenhancedRGB displayistoplacethehighresolution panchromatic image
intothe green channel, and two lowresolution multispectral bands into thered and blue
channels. Since the green channel contributes the most to the intensity component, the
1.2.2 Merging by separate manipulation of spatial information
Amultispectralimagecanbethoughtof ashavinga spectral component and a spatial
component. Thesecondtypeofmergingalgorithmfirsttries toseparate thesecomponents,
then manipulates the spatial component to obtain spatially enhanced images without
touchingthe spectralinformation.
One way to separate spatial/spectral information is by looking at the spatial
frequenciesoftheimage. Animage,accordingtoSchowengerdt,can alsobe consideredto
bethe sum of alow spatialfrequency component and ahighspatial frequencycomponent
[Schowengerdt 80].
image = lowpass
(image ) + highpass (image ) (1-3)
The primaryassumptionin this techniqueisthat thespectralinformation iscontained
inthelowpasscomponent andthatthe spatialinformation is inthehighpass component. If
theedges,orthehigh spatialfrequenciesoftheimagestobemerged arecorrelated, thenthe
highpass component of the finer spatial resolution image could be substituted for the
highpasscomponent ofthe lowerspatial resolution image. Assumingthat the majorityof
the spectralinformation iscontained in the low frequencycomponent, then theresultant
hybridimagewould maintain itsspectral content whilegaining improvedspatial resolution.
Schowengerdt used this technique to reconstruct compressed MSS images by
extrapolating the edge information from the high resolution bands to the low resolution
(compressed) bands. Tomet al [85] usedavariationofthis technique to sharpenTM band
Another way to separate spatial and spectral information is based on the
intensity-hue-saturation (IHS) color transformation [Hayden et al 82]. The IHS transformation
allows spatialinformation, contained in theintensitycomponent, tobe treated separately
from the spectral information which is embedded in the hue and saturation components.
The user can then manipulate the spatial information via the intensity component while
maintainingoverall colorbalanceoftheoriginal scene.
With the IHS method, threecarefully selected multispectral bands are transformed
into the IHS domain. The digital counts (DC) of the intensity component can then be
modifiedusingthehighresolution(panchromatic)referenceimage DCs. Themodifieddata
isthen transformedbackintothered-green-blue (RGB)colordomain anddisplayed.
UsingIHS methods,Carperet al [87] andWelsh and Ehlers [87] obtainedvisually
superiorresults than thoseobtained using Cliche's (adhoc)methods. Care hasto be taken
though,toensurethat thehighspatialresolution referenceimage ishighly correlatedto the
otherinput images. Having high correlation between two images means that the linear
relationship between theimages is very strong.
1.2.3 Merging and Maintaining Radiometry
All of these methods up to now produced visually enhanced images. Spatial
resolution appearedto takeonthereference (panchromatic)imageresolutionwhileretaining
mostofthe spectralinformation. However,in allcases, thespecific radiometric values of
themultispectralimageswerelost. Thethird typeofmergingalgorithmissimilarto thead
hocapproach,butismorestatistically-basedand attemptstomaintain radiometricintegrity.
One method suggested by Pradines [86] to keep radiometric quality of the
P1 P2
P3 P4
XP1 XP2
XP3 XP4
where XP(J) = X x P(J)
PI +P2+ P3+P4 , J
= 1..4
Figure 1-1. Pradine'sMergingMethod
3y using this method, the aggregate of the four hybrid pixels will return the
radiometry oftheoriginalmultispectralimage. However,thehigh spatial resolution image
(panchromatic channel)mustbe correlatedwith theindividual multispectral images. For
thosebands not correlatedwiththereferenceimage, thismergingalgorithmcannotbeused.
Two other merging techniques which try to maintain the radiometry of the
multispectral images are the Price [87] and the DIRS methods [Warnick 89]. These
2.0 MERGING METHODS AND MODIFICATIONS
Thisstudyevaluates the third typeofmergingtechniques
-those techniqueswhich
attempt topreserve the integrity ofthe multispectral data. The two primary techniques
undertest are the DIRS and the Price methods. In addition, several modifications and
enhancements have been incorporated to address some ofthe known short-comings of
these techniques. Thesemodified versions are alsoincludedforcomparison.
2.1 The DIRS Method
The DigitalImaging andRemoteSensing Laboratory (DIRS)conducted a
proof-of-concept study on merging multi-date-multi-sensor-multi-resolution images forenhanced
image analysis [Warnick, et al, 1989]. Specifically they merged a high resolution
panchromatic image (SPOT-1 panchromaticchannel, 10m GIFOV)with medium resolution
multispectral images(Landsat-5,TM bands 1-5, 7, 30m GIFOV).
The DIRSmethod canbe summarizedinthefollowingsteps:
(1) TheSPOT image is geometricallyregisteredto theTM images.
(2) Amediumresolutionpanchromaticimage is createdfroma weighted
average ofTM bands 1 through 4. This synthetic image approximates the same spectral
characteristics asthehighresolution SPOTpanchromaticchannel.
(3) The histogram of the SPOT panchromatic image is then linearly
adjustedto the histogramofthe syntheticTMpanchromatic image. This transformation
will, to thefirstorder, accountforthedifferingatmospheric and sensor effectsbetweenthe
(4) The images are then merged to create a high resolution, multiband
hybridimage. The mergingalgorithmis:
DCHybridMultibandO) =
DCSPOTPan
L
^C(l)1
(2-1)\UL.SynTMPan/
where:
DCHybrid
Multiband (-) isthedigitalcountofthei-th band inthehybridmultibandimage;
DCSPOT
Pan isthe digitalcountintheadjusted panchromaticSPOTimage;DCjj^O)istheDCinthei-th bandoftheoriginalmultispectralimage; and
---'CsynTMPan -s^ie digitalcountinthesyntheticTMpanchromaticimage.
The DIRS method is applied on apixel-by-pixel basis, and therefore each ofthe
aboveterms alsohas apixel location. These locationsareleftoffforeasierreading.
2.1.1 Results Using the DIRS Merging Method
In addition to the visible improvement of the images, the overall classification
accuracywasapproximately 10percentagepointshigherforthehybrid datasetthan forthe
originalTMimages.
2.1.2 Concerns on the DIRS Merging Method
Because the DIRS method was developed only as a proof-of- concept
study, the
authors identifiedseveralareasfor furtherstudy.
The synthetic, medium resolution panchromatic imagewas produced as a weighted
average ofTM bands 1 through4. The weighting factor foreach band is determined by
findingthecommon area oftheTM spectral response curve andthe SPOTpanchromatic
spectral response curve. Thecommon area under the two curves isthen divided by the
totalareaundertheTMcurve considered. Theresultant quotientistheweighting factor for
that TM band. By using integration to determine areas, the weighting factor can be
representedas:
I
min(TMRSR,i(?i), SPOTRSR(X.))d(X)w;'
=
^ (2-2)
TMRSR,i(X) dX
1
where:
i refersto theTM bandnumber, i= 1 ..4;
w-'
istheweighting factor forthei-thband;
TMRSR
yk) is therelative spectral responsecurveofthei-th bandoftheTMscene; and
SPOTRSR(X) istherelative spectral response curve oftheSPOTpanchromatic
channel.
Once the weighting factors are determined for the 4 TM bands, the factors are
w
W;'
1 "
4
X
Wi'i=l
ThesyntheticTMpanchromaticimagecan nowbecreated as:
(2-3)
TMsynPan =
W; X i (2'4)
i=l
However, itwasfoundthat the syntheticTMpanchromaticimage wasnot sensitive
in the 700 to 750 nm range, while the SPOT panchromatic sensor is still relatively
responsive. Since the reflectivity of vegetation is active in this region, the SPOT
panchromaticimagerespondsto thisvegetation reflectanceinformationwhile thesynthetic
TMpanchromaticdidnot.
Tocorrectthisdiscrepancy, thespectralresponse curvefor TM band4 was
"shifted"
from a bandpass of (760, 900) nm down to (710, 850) nm. This shift increased the
weighting factor for TM band 4, thereby increasing the responsivity in this region and
providinga
"truer"
approximation oftheSPOTpanchromatic image. However,theauthors
suggestedthatanother method ofproducinga syntheticTMpanchromatic imageshould be
explored.
(2) Concernson theweak correlationbetweenthe SPOTpanchromatic channel andTM
bands4, 5, and7:
The mergingalgorithm used in theDIRS method aspreviously shown in equation
DCHybridMultiband(i) = DCSpOT
Pan '
L.^
\U<~SynTMPan
where i represents the TMbands 1 through 5and 7. Theconcern iswhen the images for
TM bands4, 5, and7 are merged with the SPOTpanchromatic image. These bands are
only weakly correlated with the SPOT panchromatic channel and should probably be
mergedinadifferent fashion.
(3) Concernson Radiometry:
The high resolution, multiband hybrid images can be thoughtof as TM multiband
images thathave been spatiallyenhanced. Theenhancementistheresult oftheintegration
oftheSPOTpanchromaticimage,buttheresultantmeanradiances should still bethesame
asthe original TM data. As illustratedbelow, theoriginal TM images haveone pixel to
describea30mby 30marea,whilethehybrid image has9pixels todescribethe same area.
Forconsistent nomenclature, the 30m by 30m area will be referredto as a "superpixel";
and a 10mby 10mpixel as a "subpixel".
30m
TMpixel
"superpixel"
(a)
1 2 3
4 5 6
7 8 9
Nine highresolution
"subpixels"
(b)
Figure 2-1 Superpixeland subpixels
Assuminglinearmodeling, thefollowingcondition mustholdtomaintain radiometric
integrity.
9
DC =
t
E
DCHybrid(j) (2-5)j=i
However,no such check was made or enforcedin theDIRSmethod.
2.2 General Modifications to the DIRS Method
Thissectiondescribes threegeneral modificationsto theDIRS methodwhichaddress
some ofthe concerns raised in the previous section. These modifications are grouped
together because they donot alterthe merging algorithm. Instead, they affect eitherthe
input imagesto themergingalgorithm,orthehybridoutputimages.
2.2.1 Creating a TM panchromatic image
In the DIRS study, the weighting factors for each TM band were obtained by
computing the overlapping areas ofthe SPOTpanchromatic and the TM bands spectral
response curves. However,the syntheticTMpanchromaticimage hadalackofsensitivity
in the 700-750 nm region. Consequently the TM band 4 spectral response curve was
"shifted"
toprovide a strongerweighting factorand moreresponsivity inthatbandpass.
Anothermethod to obtain the weightingfactors has been used by Suits [Suitset al
Suit's approach,newTM weighting factors werecomputedusinga multivariateregression on estimated sensor signals.
EstimatingSensor Signals
Given the reflectance spectrum of a target, some atmospheric parameters, and an atmospheric model (such as LOWTRAN) we can estimate the radiance that reaches the sensor in a bandpass of interest. This radiance parameter is designated L^ , and is a
function of wavelength, X. With
L^
we can cascade the sensor's spectral responsefunction, $(X),to gettheeffective radianceseen bythe sensor. Thiseffectiveradiance, Ls
iscomputed as:
(
Lx p(?0 d^Ls = ^ (2-6)
P(X) dX
Jo
Theoutput signal (inDCs) can nowbeestimatedusingLs andtheknown gain and offset of each sensor.
DCs = Ls-i
-offseti
(2-7)
garni
where: DQ = estimatedDCofTM band i;
L = effective radiance seenbythei-th band; and
S-l
gainj, offset;
These simulatedsignals are thenregressed todeterminethe
weighting factors. This
regression canberepresentedas:
DCspot-1
DCspot-2
DCspoT-3
DCspQT-n _
DCxMl-1 DC-TM2-1 DCTM3-1 DCtM4-1
DCtmI-2 DCtM2-2 DCjM3-2 DCTM4-2
DCtmI-3 DCtM2-3 DCTM3-3 DCTM4-3
_ DCTMl-n DCTM2-n DCTM3-n DCTM4-n
CO! co2 (03 C04. El 2 3 -nJ (2-8)
where: DC istheestimated signal ofthesubscriptedsensor;
com , m = 1..4 is the weightingfactors forTM1, TM2, TM3,andTM4
respectively;
n is thenumberofsamplesin theregression; and
e istheerrorvector whichis minimized when solving forco.
To compute the simulated signals from TM1, TM2, TM3, TM4 and SPOT,
LOWTRAN7 (an atmospheric propagationmodel)was modified. The firstmodification to
LOWTRAN 7 was to allow the program to access a target reflectance spectrum.
LOWTRAN 7 in itsoriginalformusesonlyone reflectancevalue (calledthesurface albedo
or'SALB')forall wavelengths. This limitation impliesthatalltargetreflectance spectra are
uniform and flat. The modified version now allows LOWTRAN 7 to access a file
containingreflectanceinformation. Areflectance valueis then accessed(orcomputed via
interpolation)foreachwavelengthrunbyLOWTRAN.
Thesecond modification was availablefrompreviousworkby CarlSalvaggio. His
modification integrated the sensor's spectral response curves with the radiance values
equation (2-7).
Having modified LOWTRAN 7 to produce simulated sensor DCs, 25 target
reflectance spectra were chosen. These 25 targetswere comprisedof5 majorclasses
-urban, soil, water, trees, and grass with 5 samples each. These spectra are plotted in
figures 2-2through2-6onthefollowingpages.
^ 40
30
-20
-10
-1 1 1 1 1 1 ' r~
0.3 0.5 0.7 0.9 1.1
wavelength (um)
asphaltave concrete ave
slate ave gravel
roofasphalt
ou
-4n-
.-
30-
20-/
VS*-""
*&r<
y'"
10
-0 - '
1 * 4?
clayave
loamdryave
loamwet ave
sand ave
soil ave
0.3 0.5 0.7 0.9 1.1
wavelength (um)
Figure 2-3 Reflectance SpectraofSelected Soil Targets
0.2
waterl
water2
water3
water4
water ave
0.4 0.6
wavelength (um)
0.3 0.9
wavelength (um)
ash
beech
maple
oak
pine
r
1.1
Figure2-5 ReflectanceSpectraofSelected Tree Targets
clover
coarse grass
grassave
orchardgrass
swampgrass
0.7
wavelength
The modified LOWTRAN 7 was then run with these 25 reflectance spectra at 3
differentatmospheres ofvarying haze. These 75 samples werethenregressedvia equation
(2-8) and new weighting factors forTM1, TM2, TM3 and TM4 were obtained. These
results are described in section 4. 1. For further information on the LOWTRAN input
2.2.2 Interpolated Input Images
In the original DIRS method, each TM 30m superpixel was replicated into nine
smaller 10mpixels tomatchtheSPOT'Spixel resolution(see figure 2-7).
512
pixels
512
pixels
->
K
30m pixels
-> 3 9 15 21 1536 pixels Y 1536 pixels h-(a)
10mpixels
(b)
->
3 3 3 9 9
3 3 3 9 9
3 3 3 9 9
15 15 15 21 21
15 15 15 21 21
Figure 2-7 PixelReplication
(a) 512 by512originalTMimagewith 30mpixels.
(b) Afterreplication, theimage isnow 1536by 1536with 10mpixels. Thecovered ground areaisthesame.
ThesereplicatedTMimageswere thenused asinputto themergingalgorithm. In
hopesofreducingthe subsequent
"blocky"
appearancein thehybrid images (caused inpart
input images. This simple filtering technique has been used to reduce the blocky
appearance when enlarging images [Bernstein 79], and when merging multi-resolution
images [Chavez 84, 86]. Thus, given a replicatedimage as in Figure 2-7, the resultant
input image isshownin Figure 2-8 below.
1 1 1
1 1 1
1 1 1
3 3 5 7 9
3 3 5 7 9
7 7 9 11 13
11 11 13 15 17
15 15 17 19 21
(a) 3by3 averaging kernel (b) resultantinterpolated input
Figure 2-8 Interpolated Input
By using these averaged TM inputs, the merging algorithm does not change, althoughthe termDCj^ri) nolonger hasthesame value overtheentiresuperpixel.
Forthisstudy,thesereplicatedimagesthatwere smoothedbytheaveraging filterare referredtoasthe"interpolated" inputimages. This termisusedtodistinguishtheseimages
fromotherimagesthatwillbe"blurred", "smoothed", or "averaged".
2.2.3 A Technique for Radiometric Post-correction
Although the interpolation of the input data reduces the blocky appearance, the
the original data radiometric integrity is lost. In addition, the DIRS method does not
check or enforcewhethertheaverageDCwithintheSPOTpanchromaticsuperpixelequals
thecorrespondingsuperpixelin thesyntheticTMpanchromaticimage. Again, theprecise
radiometric valueislost.
Ratherthanalterthebasic mergingalgorithm oftheDIRS method (or anymethod),
radiometric integrity canbe restored at the superpixel level by multiplying each hybrid
superpixel (a block of 9 10m pixels) by a correction factor. This correction factor is definedas:
fjp _ TMj
. 30m
(2-9)
(i)
Hybrid(j)i-iOm j=iwhere:
TM;.30m
= DC fromone original30m pixelfromTMband i
Hybrid(j)i. 10m
= DC fromthej-th 10m pixel (outof9)fromthehybrid
superpixelthatcorrespondsto
TMi
.Thispost-fix operationcorrectsfor differences intheSPOTpanchromaticand thesynthetic
TMpanchromatic images,as wellas theradiometric changeto theinterpolated TM input
images. Thiscorrection, however, re-introducessomeoftheblockappearances sincethe
2.3 The DIRS Enhancement 1 Merging Method
In a simple efforttoreduce the "blocky" appearance ofthe hybrid data, the DIRS
report suggested implementing a nearest neighbor-type of correction scheme. In this
correctionroutine,each subpixel was comparedtoitsown superpixel'scenter valueandto
its neighbors center value(see Figure 2-9).
::::^
\
\
-y
-O-^:::: >::^
..^7. \J
~
comersubpixel
side subpixel
center subpixel
Figure2-9 DIRS nearest neighbor correction routine.
Comer subpixels compared values with the diagonal neighbor, and side subpixels
compared to the adjacent neighbor superpixel's center value. If the subpixel's DC was
closer toitsown center pixel's value,then themergingalgorithm remained thesame as in
equation(2-1):
DCHybridMultibandO) = DCsPOT Pan
DCtmG)
|
Ifthe subpixel's DCwas closerto theneighbor's DC,than the neighbor's ratiowas
usedinthemerging algorithm:
T\r Ji\ - T\r
I
DCNeighborTM(i)\
^T-HybridMultiband(l) = ^SPOT
Pan
-f-7; s
\UL,NeighborSyn TM PanI (2-10)
The DIRS Enhancement 1 merging methodfollows this same correction scheme,
except that the comer pixels also compare theirvalues against its adjacent superpixel's
center values (notjust the diagonal superpixel). Thecomparison scheme forone corner
subpixel looksasbelow:
::::^ m
\
1/
/
K
::::>^
*>;;;!
Figure2-10 DIRS Enhancement 1 Correction Routine
Thesuperpixel whose center pixelvaluewas closestto thesubpixel value contributed
theratiointhemergingalgorithm.
subpixelsis because thecorrelationbetween adjacent pixelsismuch higher thanbetween
diagonalpixels[Ahem 86]. This bettercorrelation meansthereisan increasedchance that
a comer subpixel willfind abetter,more accurate matchforsubstitutingratios. Correcting
thecomer subpixelsobviouslyreducestheoverallblockappearance.
Despitethesechanges,theDIRS Enhancement 1 method stillhasseverallimitations.
These include:
- The
centersubpixel of a superpixel isnever changed andwill always use itsown
ratio.
- Each
andevery non-center subpixel is checked,regardless if itresides in a
non-varyingsuperpixel, orifthe neighboringsuperpixelshavea highvariance.
- TM bands
4,5, and 7 are still handled as though they are correlated to the SPOT
panchromaticimage.
-Radiometry isstill notpreciselypreserved.
2.4 DIRS Enhancement 2 Merging Method
DIRS Enhancement 2 is an attempttocorrect some ofthedeficiencies in previous
DIRS methods. This method is described below and can be followed in the flow chart
diagramin Figure2-12attheend ofthis section.
The basic premise ofthis method is that the high resolution SPOT panchromatic
imageprovidesthe spatialinformationon whichsegmentationand computationaldecisions
images. The SPOTpanchromatic image is handled in 3 by 3 blockswhich correspond to
thereplicatedTMinput imagesuperpixels.
Thestepsforeach3by3panchromaticblockare outlined as follows:
1. Check to see ifthe SPOT block is "mixed"
or "pure". A block is considered
mixed ifthe variance among its nine pixels is higher than an established threshold. A
"pure"
block wouldbeone with a variancelowerthan this threshold.
a) If the superpixel block is pure, then the DIRS original merging
algorithm (section2.1)with a post-fix correction routine (section2.2.3)isrun tocreatethe
corresponding hybrid block. This merging operation will be referred to as the "pure
merge"
inthisalgorithm.
b) If the block has a high variance, then the algorithm continues
processingthemixedblock.
2. Check ifall the neighboring blocks are also mixed. Ifall are mixed, then the
algorithmdefaults to do the pure merge (the original DIRS method) over the superpixel
block of interest. If some of the neighboring superpixel blocks are "pure", then the
algorithm continues.
3. Connectthepanchromatic superpixel. Eachsubpixel within thepan superpixel is
compared toitsadjacent subpixels. Ifthedifference betweenthe2subpixels is lowerthan
a setthreshold, then they areconsidered"connected". Iftwo subpixelsareconnected,then
each subpixel is allowed to compare to the others adjacent subpixels for further
connections. Thus anetwork ofthesuperpixelcan be constructed. An exploded view of
Figure 2-11. Connectionsbetween thecentersubpixel
4. Findthe subpixels within the mixedSPOT block that can be matched to a pure
neighboring superpixel. A match can occur when the subpixel DC and the neighboring superpixel's mean value are within a set threshold. Rather than limit the subpixels to
compareonlywiththose neighboringsuperpixelsthey cantouch (as in DIRS Enhancement
1), this method also allows the subpixels to compare to any of the eight neighboring
superpixels providedthat: (a) the neighboringsuperpixel ispure; and(b) thesubpixelcan
be "connected" to thesuperpixel. Ifthe subpixelhas morethan one superpixel to whichit
couldbematched, then thesuperpixel whose mean valueistheclosesttothe subpixelvalue ischosen. Thesematched subpixels are thenreferencedas
"pure"
subpixels.
5. Createthe corresponding TM hybridsubpixels bymergingthesepure subpixels
usingthe matching superpixel's ratio andcorrection factor. This merge is similarto the
6. Derive a newTM value forthe remaining (mixed) subpixels
by removing the
hybridvaluescomputed inthestep before. Thiscanbedescribedas:
9-(TM0
-hybridiO)
NewTMi = -J^.
(2.U)
where:
NewTM;
= New DC fortheremaining (mixed)subpixels within
thesuperpixel ofinterestin TM bandi;
TMj
= Original DCofthesuperpixel ofinterestin TM bandi;
n = number of pure
subpixels;
hybrid;(j) = those hybrid
subpixels computed fromthe pure
sub-pixels;
Ifall the subpixels werefound to bepure, (n = 9), then this
step is skipped to step
10.
7. Checkto seeifthenewTMvalues are valid. ThenewTMvalues cannot be less
than 0or greaterthan 255.
a) Ifthevaluesareinvalid,then the leastpurehybrid subpixel isremoved
and reclassified as a mixed subpixel. The leastpurehybridsubpixelis thesubpixel which
hasthelargest differencebetween itselfandits matchingsuperpixel mean. The leastpure
hybridsubpixel will continuetoberemoveduntilthereare no purehybridsubpixelsleft,or
new valid valuesfor TMare computed. Ifthereare no pure hybrid subpixelsleft, then the
algorithmdefaults todoa pure merge onthesuperpixel.
TM DCs. Using the new valid TM DCs and the weighting factors developed in section
2.1.1,thenew syntheticTMpanchromaticcanbedescribedas:
NewTMSynPan =
j
wi x NewTM; (2-12)
i=l
9. The remaining subpixels are now merged using the new TM DC and the new
synthetic
TM^
DC usingthesamebasic DIRS algorithm:DCHybrid(i)
= DCspOTPan-L^^^I
(2-13)
\!JUNewTMSynPan/
10. Radiometricallycorrectthesemixed hybridsubpixels generatedin step 9 using
thepost-fix operation.
11. Check to see if this final hybrid block (pure and mixed subpixels) is
"reasonable". Some checksforbeingreasonable are: (a) ifthehybrid band iscorrelated
withtheSPOTpanchromaticimage,theorder ofthe9 hybridsubpixels should bethesame
as the order of the original SPOT subpixels; (b) the standard deviation among the
subpixels of the hybrid superpixel should be within a threshold factor of the standard
deviationoftheoriginalSPOTsuperpixel.
12. Ifthehybrid superpixelis foundtobe unreasonable, then theleastpure subpixel
is removed and the algorithm returns to step 6. Ifthe hybrid superpixel is found to be
Note thatwith thisnewmethod, thefollowingimprovementstoDIRS Enhancement
1 weremade:
(1) The center subpixel is not constrained to only use its superpixel ratio and
correction factor. If itcan match toaconnecting superpixel, then thecenter subpixel can
usethematching superpixel's parameters. In addition, theother subpixels can also match
withanyoftheeightneighboringsuperpixels providedthata connection exists between the
twoandthe superpixelis pure.
(2) Rather than running comparisons for all subpixels, only those SPOT pan
superpixel blocks thatareidentified as
"mixed"
(havea largevariance) will beprocessed.
In addition,only thoseneighboringsuperpixelsthatare purewillbeable tocontribute their
ratioandcorrection factor. Thesechanges should helpin speed andin avoiding improper
substitutionsforsuperpixelparameters.
(3) Radiometry is preserved. Those subpixels that are identified as 'pure'
(are
matched to a pure neighboring superpixel) produce the pure hybrid subpixels. The
remainingsubpixels are thencomputedtomaintaintheradiometric value oftheoriginalTM
value. Insimplisticterms, itcanbe describedas:
known_answer =
pure_valuessubpix + mixed_valuessubpix
Givenaknown answer(original TMvalue) andthepuresubpixel component(determined),
themixedcomponentissimplywhat'sleft.
However, DIRS Enhancement 2 still does not handle the weakly correlated TM
CD
Get Hi-Res PanchromaticSuperpixelO*- No
JL Connectthesubpixels
Findthepure subpixels
-thosewhichare closetoa
pure neighbor's mean value
>0
Mergepuresubpixelsby
usingthepure neighbor's
ratio and correctionfactor
Removetheleast
purehybridsubpixel
DerivenewTM DCs for
remainingmixed sub-pixels
2.4.1 A Modification to the DIRS Enhancement 2 Method
This small modification is a quick fix to handle the TM bands that are weakly
correlated to thepanchromaticimage. Forthismodification, theonlychange to theDIRS Enhancement 2 method is to the "puremerge"
operation fortheweaklycorrelated bands.
Insteadof puremergingthesebands usingtheDIRSmethod:
DCHybridMultibandO) =
DCsPOTPan '
\DCsynTMPan)
thehybridvalue nowsimplytakes thevalue oftheoriginalTM superpixel:
DGHybrid
UncorrelatedBandW = DCTM(i)Thus if a subpixel is matched to a neighboring pure superpixel, the hybrid subpixel for
theseweaklycorrelatedbandswould contain theTMvalueforthatmatchingsuperpixel.
Allother(correlated)bands aremanipulatedintheidenticalfashion asbefore.
Thesemethodsdescribedsofar insection2arevariationsofthebasic DIRS merging
algorithm. Theremainderofthissectiondescribes Price's mergingmethod and some ofits
2.5 Price's Merging Technique
Price'stechniquefor mergingmultispectral and panchromaticimagesis splitintotwo
cases. The firstcase is formultispectral bands thatare correlatedwith the panchromatic
image. Thesecond caseisfor weaklycorrelatedbands.
2.5.1 Case 1 -- Correlated Input Images
Becausetheimagesarecorrelated, therelationship canbe describedaslinear. On an
individual band basis, the linear coefficients can be determined using a simple linear
regression. This is illustrated in Figure 2-13, where the DCs of an averaged SPOT
panchromaticimageare plotted againsttheDCsofaThematic Mapper(TM) band.
TM TM = ( a x Pan ) + b
i i ave i
Pan ave
Figure2-13. Linear RegressionofTM bandsandthePanchromaticImage
where: TM- = DCofasuperpixelintheTM i-thband;
Panave
= theaverage oftheDCs inthecorresponding superpixeloftheSPOTpanchromaticimage;
a-,
After solving forajandb- we cancompute ahighresolution estimate ofthei-th band
using theoriginalhighresolutionpanchromatic channel:
fMj
=^ Paniom + b; (2-14)
The mergingalgorithmthenusesthehighresolutionestimatetocreatethehybrid image:
Hybrid, - LLi (2-15)
1Mi_ave
where: Hybrid- = DCofthehybrid i-thband;
TMj = DC fromtheoriginal TMi-thband;
TMi = DC fromthehighresolutionestimateofTM i-thband;
TMi.ave = average oftheDCs inthecorrespondingTM;image
correspondingto the
TM^
pixel.Forexample, themergingofTMmultispectral images (30mGIFOV) witha SPOT
T
Original
Panchromatic
30m
1
PI P2 P3
P4 P5 P6
P7 P8 P9
Original TM
Tl T2 T3
T4
*
TJ T8
?
High Resolution EstimateofTM
HI H2 H3
H4 H5 H6
H7 H8 H9
Hybrid.
Figure 2-14 Price'sMergingMethod - Case 1
where:
H(J); = TM, T(J)j
TMj.ave
, J = 1...9
and TMi.ave =
i
T(J)-y j= i
Notethe similarity toPradines' method,except that Price isusing an estimate ofa
multispectral band insteadofthepanchromaticdatadirectlyforthemergingoperation. In
multispectral band, while Pradine has the sum of the hybrid DCs equal the original
multispectralvalue. Finally,Price suggests a methodtohandlemultispectral bandsthatare
weaklycorrelatedwiththepanchromaticchannel.
2.5.2 Case 2
-Weakly Correlated Bands
Becausesomebandsare notlinearlyrelated(ievisible andnearinfrared channels),a
more general relationship was used by Price. After registering the images, Price again
averaged the panchromatic channel to the same spatial resolution as the multispectral
images. Pricethen computedtheexpected (mean)valuefortheweaklycorrelatedband for
eachgivenvalueinthe 30mpanchromaticimage.
Forexample,Price wouldfind allthepixelsin the30mpanchromaticimage thathad
aDCofX (0to 255). He would then find the average oftheDCs ofthe corresponding
pixelsintheweaklycorrelated band. Thusa simplelook-up-table (LUT)can begenerated.
This process is depicted in Figure 2-15 on the following page. Using this LUT, a high
resolution estimate of the multispectral band can be computed from the original high
resolution panchromatic image. This high resolution estimate is now used as in the
correlatedcase,andismergedwiththeoriginal multispectraldata.
2.5.3 Price's Results
Pricetested hisprocedureusing SPOTsimulationdatawitha 10mpanchromatic and
three 20m multispectral channels. He then averaged the channels to obtain 20m
panchromatic and 40m multispectral channels. Applying the above procedures to the
averageddata,Priceproducedhybrid 20mvalues whichwere thencomparedto theoriginal
0 0 0
0 0 2
1 1 2
Average Pan Image (30m)
Average Pan DC
8 8 8
7 9 16
20 24 14
WeaklyCorrelatedTM Band (30m)
CorrespondingDCs from
WeaklyCorrelatedTM Band
0
1 2
8, 8, 8, 7, 9, 20, 24, ...
16, 14, ...
255
Computethemean ofthe
corresponding TM DCs
tocreatetheLUT
Ave Pan DC Mean TM DC
0
1 2
255
8 22
15
Multispectralchannels 1 and2werehighlycorrelatedwith the panchromaticchannel,
and produced residual errors of aroundtwodigitalcounts. This isquite accurate sincethe
standarddeviationoftheoriginaldatawas around20 DCs.
For the thirdchannel which is notcorrelated, the procedure was
only "moderately
successful". Thepredicted values accounted only for about 75% ofthe variance in the
originaldata,as comparedto99%inthecorrelated case.
Price didnottest theeffecthistechniquehadon classification accuracy.
2.6 Price Modification -- A Method of
Handling Weakly Correlated Input
Images
One ofthe biggest concerns in all ofthemerging techniques to date has been the
mergingof multispectral imagesthatareweaklycorrelated withthereference panchromatic
channel. In our case, TM bands 4, 5, and 7 are not strongly correlated with the
panchromatic channel.
This modification to Price'stechnique implementsa method byTom et al [85]. In
hispaper,Tomimprovedthespatial resolutionoftheTM band 6 from 120mto30musing
theotherTMbands. His techniqueis basedon an adaptive multibandleastsquares method
which computes an optimal image estimate. His approach relies onthe assumption that
registeredTMdataare correlated acrossthebands in small local areas. Byusingthelocal
correlationproperty, Tomusedvisible and IRbands topredictthe thermal IR image data
(band 6)in 30m resolution.The thermalimageestimate wasformed byaweightedlinear
image.
By modifying Tom's technique, we could obtain high resolution estimates ofTM
bands 4,5, and 7. To compute the high resolution TM band 7 estimate, the adaptive
multiband enhancement procedure wouldtake thesegeneral steps:
(1) Average filterthepanchromatic channeltoobtain30mresolution.
(2) Using theaveraged 30 m panchromatic image and the originalTM 1, 2, and 3
images asinput,we can implementtheadaptiveleastsquares methodtogeneratethelinear
prediction coefficientsfor band 7 at each pixellocation. Fora givenlocationofthesliding
3by 3window,wecan representthisas:
TM7i TM72 TM73
TM79 _
1 TMh TM2] TM3i
Pan-1 TM12 TM22 TM32 Pan2
1 TM13 TM23 TM33 Pan3
1 TM19 TM2g TM39 Pan9
~b0" "ei"
bi e2
b2 + e_3
b3
-b4_ _e9_
(2-16)
The least square solution is computed by solving fortheset of coefficients (thevector b)
thatminimizes theerror vector(e). Notethat each pixellocation will have its own set of
coefficients. One waytohandle thisdata istohave matchingcoefficient
"images"
foreach
input/referenceimage.
(3) Obtain 10m resolution estimates ofTM bands 1, 2,and3byprevious methods.
(i.e.,PriceorDIRS methods)
(4) GenerateanoptimalestimateofTM band 7 usingthepredictioncoefficients and
the high resolution input images (panchromatic, Hybrid bands 1, 2, 3). For each 10m
TM7 =
(br
Hybi)+(b2- Hyb2)+(b3- Hyb3)+
(b4-Pan,om)
+ b0 (2-17)where
Hybj
,Hyb2
andHyb3
are the 10m hybrid pixels derived from usingPrice's techniqueforcorrelatedbands.
(5) Mergethishighresolution estimate ofTM band 7 withtheoriginal band 7 using
Price'smethod with correlatedimages(section2.5.1)toobtain a 10m
Hyb7
image.After obtaining Hyb7, we can usethis band as another inputreference to compute
Hyb5. As before, avector b is determined, this time with one more dimension for the
additionalinputreferenceimage.
TM5i
TM52 TM53
TM5g
1 TMh TM2i TM3i
1 TM12 TM22 TM32
1 TM13 TM23 TM33
TM7i Pam
TM72 Pan2 TM73 Pan3
1 TM19 TM2g TM3g TM79 Pan9
"bo" -, bi ei e2 b, + e3 b3 b4
Lbs.
.e9_ (2-18)With this vector b, the high resolution estimate for TM5 can be computed similar to
equation(2-17):
TM5 =
(by
Hybi)+ (b2- Hyb2)+(b3- Hyb3)+(b4- Hyb7)+
(b5Paniom)
+ b0(2-19)
The hybrid image for TM5 is then obtainedby merging this highresolution estimate of
TM5with theoriginalTM5 imageusing Price'smethodwith correlatedimages. Oncethe
hybrid for TM5 (Hyb5)is found, this band as well as
Hyb7
are used to compute the highTheorder ofcomputing band 7, 5, and then4ischosen becausethis is theorder of
decreasing band correlation with the panchromatic image. Tom et al [85] additionally
showedthat increasing thenumber ofinputreference bandsdecreased thehighresolution
estimate error. Therefore,band 4 the weakest correlatedband~ is
computedlastto use
the2additional referencebands.
Section 2 described the twoprimarymethods ofmergingthatareinvestigated inthis
study the DIRS and the Price methods. In addition, several modifications and
enhancements have also been introduced. The following section describes how these
3.0 EXPERIMENTAL APPROACH
In comparing these mergingmethods, several variations ofthe input images were
used. Thesevariations aredescribed insection 3.1. Thetwoprimarytestsforevaluation
-classification accuracy and radiometric error
-are described in sections 3.2 and 3.3
respectively. Finally,theprocedureforthis studyispresentedin section3.4.
3.1 Selection and Preparation of Test Imagery
The SPOTandTM imagesselected forthis studyare theidentical imagesusedinthe
DIRS proof-of-concept study [Warnick 89]. The images were selectedfrom a scene of
greaterRochester, NY,