an author's
https://oatao.univ-toulouse.fr/27001
https://doi.org/10.1016/j.sigpro.2020.107905
Vincent, François and Besson, Olivier Robust adaptive target detection in hyperspectral imaging. (2021) Signal
Processing, 181. 107905. ISSN 0165-1684
Robust
adaptive
target
detection
in
hyperspectral
imaging
François Vincent
∗, Olivier Besson
ISAE-SUPAERO, 10 Avenue Edouard Belin, Toulouse 31055, France
Keywords: Hyperspectral imagery Robust Detection GLRT AMF
a
b
s
t
r
a
c
t
Oneofthemainissueindetectingatargetfromanhyperspectralimagereliesonproperlyidentifyingthe background.Manyassumptionsaboutitsdistributioncanbeadvocated,eveniftheGaussianhypothesis prevails.Nevertheless,thehugemajorityoftheresultingdetectionschemesassumethatthebackground distributionremainsthesamewhetherthetargetispresentornot.Inpractice,becauseofthespectral variabilityofthetargetand thenon-linearmixingwiththebackgroundradiance,thishypothesisisnot strictlytrue.Inthispaper,weconsiderthatanunknownbackgroundmismatchexistsbetweenthetwo hypotheses.Undertheassumptionthatthismismatchissmall,wederiveanapproximationofthe Like-lihoodRatiofortheproblemathand.ThisgeneralformulationisthenappliedtothecaseofGaussian distributedbackground,leadingtoarobustAdaptiveMatchedFilter.Thebehaviourofthisnew detec-torisanalysedandcomparedtopopulardetectors.Numericalsimulations,basedonrealdata,showthe possibleimprovementincaseoftargetsignaturemismatch.
1. Introduction
Anhyperspectralimage isatwo-dimensionalmap,whereeach pixeliscomposed ofhundreds ofspectral bands.Thisspectral in-formation allows to characterize the different materials present in the image and hence hyperspectral imaging has encountered a large field of applications, ranging from remote sensing to medicine [1–8].One of theseapplications consistsin detecting a target whose spectral signature could be known fromlaboratory experiments,orunknown.Inthislattercase,onetalksofanomaly detection.Inthispaper,wewillfocusonthefirstcase,namely tar-getdetection.
Thedifficultyofthisproblemliesinthefactthatthesignature ofinterest(SoI) t isburiedinabackground b withunknown statis-tics.Evenifitsdistributionwereknown,theparametersdescribing it (for instance meanand covariancematrix) are not known and mustbeestimatedfromtheavailabledata.Consequently,detection oftheSoIinapixelundertest(PUT) y requiresusingotherpixels (socalledtrainingsamples)tolearnthebackgroundpresentinthe PUT.Ithastobenoticedthatinthehyperspectraldomain,wedeal withrealpositivedataleadingtononzeromeansignals.
Certainlydueto itseasy handlingproperties,theGaussian as-sumptionprevailstomodelthebackground.Inthiscontext,many algorithms have been developed, such as the Adaptive Matched
∗Corresponding author.
E-mail addresses: francois.vincent@isae-supaero.fr , francois.vincent@isae.fr (F. Vincent), olivier.besson@isae-supaero.fr (O. Besson).
Filter (AMF) [9],Kelly’s Generalized Likelihood Ratio Test (GLRT)
[10],theOrthogonal SubspaceProjection(OSP) [11],or the Adap-tiveCoherent/CosineEstimator(ACE)[12],tonamethemost pop-ularones. However, withrealhyperspectral data,Gaussian distri-butionsrarelyoccur [13–16],andmorerepresentativebackground distributions have been considered. One ofthe most common is the EllipticallyContoured (EC)t-distributed modelthat allows to extendtheGaussian distribution toa broaderclass ofprobability densityfunctions(p.d.f.).Differentdetectorshavebeenderived un-dersuchanhypothesis,astheEC-GLRT[17]forinstance.
Nevertheless,althoughconsideringabroadclassofdistribution tomodel thebackground,thehuge majorityofthedecisiontests relyontheassumptionthatthebackgroundremainsthesame un-derthetwohypotheses.
H0:y=b (1)
H1:y=
α
t+bwhere
α
istheso-calledfill-factor.Hence, if considering the SoI and
α
as deterministic, which isa widespread assumption, thismodel amountsto consider the same distributionunder the two hypotheses. The only difference betweenthetwohypothesesisameanshift.Unfortunately,thisstrongassumptionisusuallynotmetinreal hyperspectraldetectionschemes.Indeed,thefirstreasonisrelated tothedataacquisitionitself.Theradianceresultingfromthe mix-ture of a background and a target are driven by non-linear ef-fects [18]. These effects are due to multi-reflections or masking effects. Thereby, the additive model from Eq.(1), is over
fied. A morerepresentative model, namelythe so-called replace-ment model, can be used [1]. This model assumes that the tar-get, ifpresent,replaces a partofthebackground,asit willmask an areaproportionaltoits fill-factor
α
.Somedetectorshavebeen derived undersuch a model,as the FTMF [19]or ACUTE[20]. A moregeneralmodel,relaxingtheunitaryconstraintsonthe ampli-tudes, onlyassumesa linearmixingbetweenthebackgroundand the target [21].Butevenifthesetwo models aremore represen-tative,they donot takeintoaccountthemultiplelight reflections conductingtonon-linearmixing,andconsequentlyamore compli-catedbackgroundbehaviourwhenthetargetexists.The second main reasonconductingto achange inthe distri-bution between the two hypotheses is related to mismatches in the target SoI.Indeed, thisso-calledspectral variability isdueto changesintheconditionsbetweenthelabmeasurementsandthe data acquisition [9,22]. Infact, the image isacquired by a differ-entinstrumentandata largerdistancefromthescene.Then,the rawradiancemeasurements havetobeconvertedintoreflectance data using a so-called radiometric compensation [23]. Different techniques canbe used [24,25],neverthelessall these compensa-tion techniques are based on models inducing possible SoI mis-matches. Then, when assuming the SoIasdeterministic and per-fectly known, thisspectral variability resultsin anoise measure-ment increase. Again,the consequenceis that thePUT exhibitsa differentdistributionbeyondasimplemeanshift.
Thereby, in this paper, we consider a more general detection model that may include a large class of mismatches, including those statedhere-above. In this model,the background distribu-tionmayvarybetweenthetwohypotheses.Morepreciselywe as-sumethatwhenthetargetispresent,thebackgroundundergoesa possibleperturbationwithunknownstatistics:
H0:y=b (2)
H1:y=
α
t+b+b
where
b isthemismatch.
Conventionally, the distribution of b can be estimated from trainingsamples x k.Ontheotherhand,
b hasanunknown distri-bution,possiblylinkedto
α
.Thisunknownperturbationcanmodel a large class of possiblemismatches induced by non-linear mix-turesbetweenthetargetandthebackgroundorfromtarget signa-tureerrors.Underthehypothesis that
b issmallcomparedto b ,we de-rivea robustLikelihood Ratio (LR),andshowthat itcan be writ-ten asthesumofthe standard LRanda correction factor.When applied to the popularcase where b is supposed to be Gaussian distributed, we derivean approximationof theGLRT. Thisrobust detector is shownto be the popular AMF, corrected by a simple additiveterm.
The remaining of this paper is organized asfollows. We first derivetheLikelihoodRatioforthegeneralcaseofanybackground distribution insection2.Then, wefocuson thepopularGaussian caseinSection 3,wherewe propose asimpleadditive correction fortheAMFdetectortoimproverobustness.InSection4,we com-pare the behaviour of the proposed robust AMF to popular de-tectors, thanks to their thresholdlimit plottedin a Matched Fil-ter Residual(MFR)diagram.Tofinish,wecomparethesedetectors using a Monte-Carlo simulation based on realdata, inSection 5. FinallyconcludingremarksendthispaperinSection6.
2. Robust likelihood ratio
As alreadysaid inthe introduction,we assume thatthe back-groundunder H1 writes b 1= b +
b ,where theprobability
den-sity function (pdf) of b , pb
(
.)
is supposed to be known, andits parameterscanbeestimatedfromthetrainingsamples,thatisthestandard hypothesis.Hence, forany
b pdf, p
(
.)
, thepdf of b 1canbewrittenas
pb1
(
u)
=(
pb∗ p)(
u)
=p
(
z)
pb(
u− z)
dz=E[pb(
u− z)
](3)
Now, using the approximation derived in the Appendix, for any given u ,wehave E[pb
(
u− z)
]pb(
u−μ
)
+12Tr∂
2p b(
z)
∂
z∂
zT|
u−μC (4)whereTr
{}
standsforthetraceofthematrixbetweenbraces,μ
and C denoterespectivelythemeanandcovariancematrixofthe unknown background mismatchb , and [∂2pb(z)
∂z∂zT
|
u−μ] denotestheHessian of pb
(
z)
evaluatedat(
u −μ
)
.It should bepointedoutthat thisapproximationdoesnotrequirecompleteknowledge ofthepdfof
b ,butonlyofitsmeanandits covariancematrix, whichisanappealingfeature.
Hence, forany PUTmeasurement y ,the Likelihood Ratio (LR) associatedwiththedetectionproblemfromEq.(2)canbewritten inthefollowingform,assoonasthebackgroundmismatch
b is small: LRr= pb1
(
y−α
t)
pb(
y)
pb(
y−α
t−μ
)
pb(
y)
+ Tr{
∂2pb(z) ∂z∂zT|
y−αt−μC}
2pb(
y)
(5) = pb(
y−α
t−μ
)
pb(
y)
1+1 2 Tr{
∂2pb(z) ∂z∂zT|
y−αt−μC}
pb(
y−α
t−μ
)
(6)Moreover,
b beingsomeperturbationaround b ,weassume that
μ
=0,sothatwesimplyhave LRrLR 1+1 2Tr{
∂2p b(z) ∂z∂zT|
y−αt pb(
y−α
t)
C}
(7) whereLR= pb(y−αt)pb(y) istheLRinthestandardcase,i.e.whenthere
isnobackground mismatch(
b =0). Hence,we can seethatfor any zero-mean mismatch distribution, we can approximate the modifiedLR by the sumof theLR withoutmismatch anda cor-rectiveterm,dependingonthepdfof b andthecovariancematrix of
b .
3. Robust adaptive matched filter
In this section, we now investigatethe casewhere the back-grounddistributionunderH0isGaussian, b ∼ N
(
μ
,C)
.Letusstartwiththenominalcase,torecalltheAMFderivation.
Ifnomismatchispresent(b 1=b ),thestandardLRwrites LRG= pb
(
y−α
t)
pb
(
y)
=e−12(y−αt−μ)TC−1(y−αt−μ)
e−12(y−μ)TC−1(y−μ)
(8)
Assuming that the background parameters are known from the trainingsamples,theonlyunknown parameteris
α
, whose Max-imumLikelihood (ML) writesα
ˆ= tTC−1(y−μ)tTC−1t .Then, the logarithm
oftheGeneralizedLR(GLR)isshowntobe:
logGLRG=log
pb(
y− ˆα
t)
pb(
y)
=1 2|
tTC−1(
y−μ
)
|
2 tTC−1t = 1 2AMF (9)wherewerecognizethepopularAMF= |tTC−1(y−μ)|2
tTC−1t .
Fig. 1. Typical MFR plot.
Now, in the mismatched case (b 1=b ), the Hessian in
Eq.(7)writes
∂
2p b(
z)
∂
z∂
zT|
y−αt=C− 1 2C−12
(
y−α
t−μ
)(
y−α
t−μ
)
TC−12− I C−12 ×pb(
y−α
t)
(10)sothatthemodifiedLRbecomes
LRG r LRG ·
1+12Tr C −12
(
y −α
t −μ
)(
y −α
t −μ
)
TC −12 − IC −12C C −12 (11)
First, we havetoestimate
α
.As thebackground mismatchunderH1 is supposed to be small,we can usethe ML estimation of
α
obtainedhereabove inthenominalcase, asa firstorder approx-imation, namely
α
ˆ=tTC−1(y−μ)tTC−1t . Hence, taking the logarithm and
substituting
α
,theGLRwrites: logGLRG r logGLRG+log ·1+1 2Tr C −12
(
y − ˆα
t −μ
)(
y − ˆα
t −μ
)
TC −12 − IC −12C C −12 (12)
or,thankstoEq.(9) 2logGLRG r AMF+2log
1+1 2Tr
uuT− IC−1 2CC− 1 2 (13) where u =P ⊥t cy c, with t c=C −1 2t and y c=C − 1 2
(
y −μ
)
, the whitenedversionsof t and(
y −μ
)
respectively.Therefore,wecanseethatassumingaGaussianbackground un-derH0,andforsmallmismatchesunderH1,theGLRisthepopular
AMF, plus a correction. Thisadditive term tends to measure the gap betweenthe supposed background covariance matrix C and the actual one underH1.Indeed, u isan estimation ofthe
back-groundunderH1,whitened bythesupposedcovariancematrix C ,
sothat
(
uu T− I)
isakindofmismatchmeasurement.Inthesameway, C −12C C −12 isalsoanormalizedmeasurementofthisgap. Unfortunately, this expression of the correction is difficult to use in practice,except ifone has an idea of C . In orderto ad-dressthisissue,wenowderiveanapproximationofthiscorrection inordertogetamoreconvenientexpression.
The correction depends on T=Tr
{
uu T− IC −12C C −12
}
that canbewrittenasfollowsT=Tr
{
C˜}
uTC˜ u Tr{
C˜}
− 1 (14)Fig. 2. Typical MFR plot.
Fig. 3. Complete RGB view of the Viareggio test scene.
whereC ˜=C −12C C −12 isthecovariancematrixofthebackground mismatch whitened by C . Diagonalising this covariance matrix, ˜
C =
ku ku Tk, the quadratic term here above writes
uTC˜u Tr{C˜} =
(
kk
)(
uT
ku
)
2.Thereby,itcorrespondstoaweightedsumoftheenergyof u .Unfortunately, inmostcases,we donot haveaccess to C andtheweightsare unknown.Tocircumventthisproblem, we simply propose to use a non-weighted sum to estimate this energy,namely uTC˜u
Tr{C˜}
uTu
N .Thisapproximationamountsto
con-siderthatalltheeigenvaluesofC ˜areequal,namelythat C =
C . In other words, there is no approximation when the covariance matrix ofthe mismatch is proportional to the covariance matrix ofthebackground C .Thishypothesissometimesappearsinthe lit-eraturewhenconsideringtargetsignaturemismatches,evenifthe rationalebehinditisnotobvious[26].Nevertheless,this hypothe-sisincludesthepopularcasewherebothbackgroundandthe mis-matcharesupposedtobewhite.
On the other hand, Tr
{
C ˜}
=Tr{
C −21C C −12}
=Tr{
C C −1}
=Tr
{
Eb
b TC −1
}
=Eb TC −1
b representstheexpectationof
theenergyofthemismatch,whitenedby C .Assumingthat
b and b aredecorrelated, we canestimate thisenergyasthe difference betweenthe energyof b 1 andtheenergyof b ,bothwhitened by C .Thefirstonecanbeestimatedby u Tu ,as u isan estimationof
Fig. 4. Receiver Operating Characteristic for V 5 and V 6 targets, for r = 0 (no mismatch).
Fig. 5. Receiver Operating Characteristic for V 5 and V 6 target, for r = 0 . 2 .
thebackgroundunderH1,whitenedby C ,whereasthesecondone
issimplyequaltoN,inaverage. Substituting both C ˜ and uTC˜u
Tr{C˜} by their estimation in Eq.(14)wehave TN
uTu N − 1 2 (15)Then,fromEq.(13),wecandefinearobustAMFas
AMFr=AMF+2log
1+N 2 uTu N − 1 2 (16)At thisstage,we canmake afew commentsontheproposed ro-bust AMF. Whereas the standard AMF only measures the energy ofthedataprojectedonthetargetsubspace,theproposedscheme alsomeasures theremaining energy,projectedonthebackground subspace. It checks ifthisenergy corresponds to the background
energyestimatedfromthetrainingsamples,andmeasuresthegap. Then,itcorrectstheAMFwithrespecttothisestimatedmismatch energy.Otherpopulardetectorsalsousethedataenergyprojected onthebackgroundsubspace toimproveperformance,likeACE or Kelly’s GLRT,as we will see hereafter. In the following part, we compare our robust AMF with thesepopular detectors regarding their behaviour withrespect to thisresidual energy. Tothisend, we plot both the data and the detectors threshold on the so-calledMatched FilterResidual(MFR) diagrams.These informative 2D plots show both the energyprojected on the target subspace (MF)andtheenergyprojectedonthesubspaceorthogonal tothe target(R).
4. Insights
Just like for the AMF, the majority of popular detectors are basedonthewhitenedandcentredversionsofboththedataand 4
Fig. 6. P f a gain for P d = 0 . 5 , for V5 target, with α= 0 . 1 and KN = 9 .
the target signature, namely y c and t c.As noticedpreviously,the
standard AMF simply consistsinprojecting thedata onto the1D
targetsubspaceandinmeasuringtheenergy,asAMF=
P tcy c2= y T
cP tcy c, where P tc= tctTc tT
ctc is the projection onto the target sub-space. Nevertheless,theglobal energy,
y c2,or theresidualen-ergy,namelyR=
P ⊥tcy c 2 =y T cP ⊥tcy c=(
y c 2 − AMF)
,where P ⊥tc= I − Ptc istheprojectionontotheso-callednoisesubspace,arealso informative. Indeed, when thetarget amplitudeis unknown, this last quantityisan estimationoftheLoglikelihoodofthedata,up to a scaling factorand an additive constant. Hence plottingdata ina2Dspace, wherethex-axis istheAMFandthey-axisisR,is convenientandvaluablebothfordataanddetectorsanalysis.SuchrealdataMFRanalysesexistintheliterature[27–29]and show thatthescatter plotsforall thepixelsaredistributedalong linesconvergingtotheoriginoftheplot,eachlinecorresponding to a given material on theground. Fig.1 showsa representative situation with just3 pixels classes(for instance, forest in green, roads in blue,andbuildings inred), considering thebuildings as thetargetfortheAMFaxis.
Observing such a MFR plot, itis clear that the standard AMF couldnotgetgooddetectionperformance,asitonlythresholdsthe x-axis, asshownon Fig.2.Toimprovedetection, oneshould also usethey-axis.Thatiswhatpopulardetectionschemesdo,suchas Kelly’s GLRTorACE. Indeed, thesetwo detectors share thesame formulation,namely
AMF
(
a+1KyTcyc
)
(17)
where a=0 forACE, anda=1 forKelly, K being thenumber of trainingsamples.
In both cases, comparing this quantity to a threshold
γ
amounts to comparing AMF to a limit dependingon the square normofthedata,y c2,sothatthedecisionlimitisdefinedbyAMF =
γ
a+1 Ky T cyc(18)
Now, as
y c2=AMF+R, the thresholds of these two detectorsdrawlinesintheMFRplot,whoseequationare
R=AMF
Kγ
− 1 − Ka (19)As represented on Fig. 2, the threshold line corresponding to ACE goesthroughtheoriginoftheplot,asa=0.Inthiscase,
an-otherintuitive waytoseethatthethresholdregionisaline,isto refertothestandardinterpretationofACEasthesquarecosineof the anglebetween y c and t c.Thresholding ACE amounts to limit
thisangleas shownonFig. 2.In thecaseof Kelly, thethreshold is still a line,but shifted along the AMF axis,depending on the numberofsecondarydataK.
Now,referring to the definitionof u inEq. (13),we haveR=
u 2,so that thethreshold limitof therobust AMF, AMFr intheMFRplotisdefinedas AMF=
γ
− 2log 1+N 2 R N− 12 (20)
Then, this threshold limit is characterized by a second order polynomial equation in the MFR plot, as represented on Fig. 2. The threshold limit corresponding to the larger AMF being ob-tained when R=N. In the case of a white noise
(
C =σ
2I)
, thispoint corresponds to the case where the residual energy equals the noise energy estimated from the training samples, namely
P ⊥t(
y −μ
)
2=σ
2N. Hence, as can be seen on Fig. 2, the AMF rhasapproximatelythe samebehaviour asACE or Kellywhen the residual energy is smaller than that estimatedfrom the training samples.
In the opposite case, the behaviour of AMFr is not the same.
Indeed,whentheresidualenergyislarger thanthetraining sam-plesone,thethresholdreducestoimprovedetectionwithrespect to the datamodel from Eq.(2). Infact, according to thismodel, anynoiseenergyvariationwithrespecttotheestimatedonefrom thetrainingsamplesisacluefortheH1hypothesis.Obviously,this
possibledetectionimprovementcanalsoincrease thedetectionof confuserelements.Indeed,improvingdetectionaccordingtomodel
(2),goesagainsttheso-calledselectivityofthedetector.
Insituationswhere possibleunwantedtargetsexit,one hasto supplementthedetectionschemebyaso-calledFalseAlarm Miti-gation(FAM)step.Thesetwo-stepprocedureshavebeenaddressed inthe hyperspectral literature [27–30], butalsoin the radar one
[31].Themainideabehindthisconceptistoimprovethe selectiv-ityissuesbyaddingaseconddetectionstagebasedontheresidual energy.Indeed,asnoticedpreviously,theresidualenergyisrelated to thelikelihood ofthe data withrespect toa given target type. Then, the possible false alarms existing after the first stage can berejectedusinganappropriatethresholdontheresidualenergy. Thisstep amounts toaddinga second thresholdon the y-axisof theMFRplot.Inthehyperspectralcontext,thisprocedurehasbeen firstdevelopedtotackletheselectivityissuesoftheAMF,withthe MatchedFilterwithFalseAlarmMitigation(MF-FAM)[27]as rep-resented on Fig. 2.This algorithm is very similar to the popular MixtureTuned Matched Filter(MTMF)approach from ENVI®soft-wareenvironment.Inthislastcase,theresidualenergyisreplaced byanequivalentcriterion,calledinfeasibility[30].
5. Performance evaluation
Inordertoassessthevalidityoftheproposed robustdetector, wenowconducta Monte-Carlosimulationbasedonareal exper-iment, namelytheairborneViareggio 2013trial[32].This bench-markinghyperspectraldetectioncampaigntookplaceinViareggio (Italy), in May 2013, with an aircraft flying at 1200 meters. The open data consist in a [450× 375] pixels map composed of 511 samplesintheVisibleNearInfraRed(VINR)band
(
400− 1000nm)
. Thespatialresolutionoftheimageisabout0.6meters.Differentkindsofvehiclesaswellascolouredpanelsservedas knowntargets.Foreach of thesetargets, aspectral signature ob-tainedfromgroundspectroradiometermeasurements isavailable. Moreover,ablackandawhitecover,servingascalibrationtargets, were also deployed. As can be seen onFig. 3,the scene is com-posedofparkinglots,roads,buildings,sportfieldsandpinewoods.
Fig. 7. P f a gain for P d = 0 . 5 , for V5 target, with α= 0 . 1 and KN = 9 .
Asforthemajorityofhyperspectraldetectionschemes,thefirst step ofthe processingaims atconverting theraw measurements into a reflectance map, namely removing all atmospheric effects andnon-uniformsunillumination.Tothisend,we usethe Empir-ical Line Method(ELM) [23,25], considering the black and white calibrationpanels.Thenaspectralbinning[33]isperformedto re-ducethevectorsizedimensiontoN=32.
Based onthis map,we willconduct a Monte-Carlosimulation byinsertingasynthetictarget t ,initiallynotpresentintheimage, intoarandomlychosenpixel y ofthemap(y →
(
α
t +y)
,whereα
is thetarget amplitude).The target signatureissubjectto uncer-taintiesandiscomposedofthesumoftheassumedtarget signa-ture¯tandarandomzero-meannormallydistributedwhitevector t ,whoseenergyisr timestheenergyof¯t:t=¯t+t (21)
where t ∼ N
(
0 ,r¯tTN¯tI)
.Figs. 4 and 5 represent the Receiver Operating Characteristic (ROC) of different detectors, for both the so-called V5 and V6
targets and for r=0 (no mismatch) and r=0.2, respectively. In both cases,weusea 17× 17windowto estimatethe background mean and the sample covariance matrix. The target’s amplitude are
α
=0.1 for V5 andα
=0.2 forV6, as this last one is moredifficulttodetect.
First of all, in the nominal case, where no mismatch exists (r=0), all the detectors exhibit approximately the same perfor-mance, with a slight loss for the proposed AMFr. On the other
hand,whenthereisamismatchonthetarget’ssignature,the ro-bustAMFoutperformstheotherdetectors.Wecannoticethatboth ACEandKellyexhibit alargerlossthanthestandardAMF.Indeed, the existenceofamismatchonthetarget signatureincreasesthe residualpart(R)ofthedata.Then,referringtoFig.2,itisobvious thatbothACEandKellywillbethemoreaffectedinthissituation. Wehighlightherethecompromisebetweenselectivityand robust-ness.ThatiswhytheFAMtechniques,usingfirstarobust scheme toavoidnon-detection,followedbyaselectiveschemetosortthe targetsandreducefalsealarm,areworthyofinterest.
Now,we willsuccessivelyanalyse theinfluenceofthe fill fac-tor
α
andofthelevelofuncertaintiesr.Tothisend, wecompute the gain provided by AMFr compared to AMF interms ofproba-bility offalse alarm (Pfa) for a givenPd, i.e., 10log10 Pf a(AMF) Pf a(AMFr), for
Pd=0.5.ThisPfagain allowstomeasuretheimprovement(if pos-itive) ortheloss(when negative)oftherobustAMF withrespect
tothe standardone. Weconsiderthe so-calledV5 target forboth
theseanalysis.Fig.6representsthisfalsealarmgainwithrespect to
α
.OnthisfigurewekeepthesameconfigurationasinFig.4.We roughlyobservealinearperformanceincreaseasthetarget ampli-tudeincreases.Tofinishwith,wenowmakevarythelevelofuncertaintieson t ,namelyr, inFig.7.As expected, weobservethat the improve-ment proposed by the robust AMF improves asthe level of un-certaintiesincreases,withlessthan1db loss whenthetarget sig-natureisperfectlyknown(r=0). Therefore,assoonasthetarget signatureisnotperfectlyknown,theproposed robustAMFhasto beconsidered.
6. Conclusions
Inthispaper,weconsideredthedetectionofatargetinan hy-perspectralimage when thebackgrounddistribution mayslightly varybetweenthe nullhypothesis andthe alternateone, duee.g., tospectralvariability.UndersuchanassumptionwederivedtheLR foranybackgrounddistribution,andthe GLRTunder thepopular Gaussiancase.Thislastdetectorisshowntobeasimplecorrection oftheAMF.Weprovidedinsightsintothebehaviourofthisrobust schemeandshoweditsgoodperformancethroughnumerical sim-ulations.
Declaration of Competing Interest
Theauthorsdeclarethattheyhavenoknowncompeting finan-cialinterestsorpersonalrelationshipsthatcouldhaveappearedto influencetheworkreportedinthispaper.
Appendix A. Proof of (4)
Inthisappendix,wederivetheprooftoEq.(4).Foranyrandom vector u ,wecan write u =E
{
u}
+u ,where
u isa zeromean vectorwiththesamecovariancematrixas u ,namely C u.Then,we
canwritethefollowingTaylorseriesexpansionaroundE
{
u}
:f
(
u)
= f(
E{
u}
+u
)
f(
E{
u}
)
+∂
f(
u)
∂
uT|
E{u}u +12
uT
∂
2f(
u)
∂
u∂
uT|
E{u}u (A.1)
Takingtheexpectationofthisexpressionleadsto
E
{
f(
u)
}
f(
E{
u}
)
+1 2EuT[
∂
2f(
u)
∂
u∂
uT|
E{u}]u (A.2) =f
(
E{
u}
)
+1 2Tr{
[∂
2f(
u)
∂
u∂
uT|
E{u}]Cu}
(A.3) CRediT authorship contribution statementFrançois Vincent: Conceptualization, Methodology, Validation, Formal analysis, Software, Writing - original draft, Writing - re-view & editing. Olivier Besson: Methodology, Validation, Writing -review&editing.
References
[1] D.G. Manolakis , R.B. Lockwood , T.W. Cooley , Hyperspectral Imaging Remote Sensing, Cambridge University Press, 2016 .
[2] E. Keith , H. Dan , O. William , J. Shridhar , B. Eustace , L. Dereniak , Hyperspectral imaging for astronomy and space surveillance, in: Proc. SPIE, vol. 5159, 2004 .
[3] S. Michel , P. Gamet , M.-J. Lefevre-Fonollosa , HYPXIM A hyperspectral satel- lite defined for science, security and defence users, Processing 3rd Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing (WHISPERS), 2011 .
[4] H. Kwon , N. Nasrabadi , Kernel rx-algorithm: a nonlinear anomaly detector for hyperspectral imagery, IEEE Trans. Geosci. Remote Sens. 43 (2) (2005) 388–397 .
[5] E.M. Winter , M. Miller , C. Simi , A. Hill , T. Williams , D. Hampton , M. Wood , J. Zadnick , M. Sviland , Mine detection experiments using hyperspectral sensors, SPIE Int. Soc. Opt. Eng., Orlando, FL, United States, 2004 .
[6] C.C. Funk , J. Theiler , D.A. Roberts , C.C. Borel , Clustering to improve matched fil- ter detection of weak gas plumes in hyperspectral thermal imagery, IEEE Trans. Geosci. Remote Sens. 39 (7) (2001) 1410–1420 .
[7] D.-W. Sun , Hyperspectral Imaging for Food Quality Analysis and Control, Else- vier, 2010 .
[8] R. Koprowski , Processing of Hyperspectral Medical Images, Applications in Der- matology Using Matlabë, Springer International Publishing, 2017 .
[9] D. Manolakis , G. Shaw , Detection algorithms for hyperspectral imaging appli- cations, IEEE Signal Process. Mag. 19 (1) (2002) 29–43 .
[10] E. Kelly , An adaptive detection algorithm, IEEE Trans. Aerosp. Electron.Syst. 22 (2) (1986) 115–127 .
[11] C. Chang , Orthogonal subspace projection (OSP) revisited: a comprehensive study and analysis, IEEE Trans. Geosci. Remote Sens. 43 (3) (2005) 502–518 .
[12] S. Kraut , L.L. Scharf , L.T. McWhorter , Adaptive subspace detectors, IEEE Trans. Signal Process. 49 (1) (2001) 1–16 .
[13] M.J. Carlotto , A cluster-based approach for detecting man-made objects and changes in imagery, IEEE Trans. Geosci. Remote Sens. 43 (2) (2005) 374387 .
[14] D.B. Marden , D. Manolakis , Using elliptically contoured distributions to model hyperspectral imaging data, in: Proceedings of SPIE - The International Society for Optical Engineering, vol. 5425, 2004, p. 558572 .
[15] D.G. Manolakis , D. Marden , J.P. Kerekes , G.A. Shaw , Statistics of hyperspectral imaging data, in: Proceedings SPIE 4381 Algorithms for Multispectral, Hyper- spectral, and Ultraspectral Imagery VII, vol. 4381, 2001, pp. 308–316 .
[16] S. Matteoli , M. Diani , G. Corsini , A tutorial overview of anomaly detection in hyperspectral images, IEEE Aerosp. Electron. Syst. Mag. 25 (7) (2010) 5–27 .
[17] J. Theiler , B.R. Foy , EC-GLRT: Detecting weak plumes in non-Gaussian hyper- spectral clutter using an elliptically-contoured generalized likelihood ratio test, in: Proceedings IGARSS, vol. 1, 2008, pp. 221–224 . Boston, MA.
[18] N. Dobigeon , J.-Y. Tourneret , C. Richard , J.C.M. Bermudez , S. McLaughlin , A.O. Hero , Nonlinear unmixing of hyperspectral images, IEEE Signal Process. Mag. (2014) 82–94 .
[19] A. Schaum , A. Stocker , Spectrally-selective target detection, in: Proceedings of ISSSR, vol. 12, 1997, pp. 2015–2018 .
[20] F. Vincent , O. Bessson , One-step generalized likelihood ratio test for subpixel target detection in hyperspectral imaging, IEEE Trans. Geosci. Remote Sens. 58 (6) (2020) 4 479–4 489 .
[21] F. Vincent , O. Bessson , Generalized likelihood ratio test for modified replace- ment model in hyperspectral imaging detection, Signal Process. 174 (2020) 107643 .
[22] R. DiPietro , D. Manolakis , R. Lockwood , T. Cooley , J. Jacobson , Performance evaluation of hyperspectral detection algorithms for sub-pixel objects, in: Pro- ceedings of SPIE - The International Society for Optical Engineering, 2010 .
[23] G. Ferrier , Evaluation of apparent surface reflectance estimation methodolo- gies, Int. J. Remote Sens. 16 (1995) 2291–2297 .
[24] T. Popp , Correcting atmospheric masking to retrieve the spectral albedo of land surfaces from satellite measurements, Int. J. Remote Sens. 16 (1995) 34833508 .
[25] G.M. Smith , E.J. Milton , The use of the empirical line method to calibrate re- motely sensed data to reflectance, Int. J. Remote Sens. 20 (1999) 2653–2662 .
[26] D. Manolakis , R. Lockwood , T. Cooley , J. Jacobson , Is there a best hyperspectral detection algorithm? in: Proc. of SPIE, vol. 7334, 2009 .
[27] R.S. DiPietro , D. Manolakis , R.B. Lockwood , T. Cooley , J. Jacobson , Hyperspectral matched filter with false-alarm mitigation, Opt. Eng. 51 (1) (2012) 1–8 .
[28] E. Truslow , Performance evaluation of the adaptive cosine estimator detector for hyperspectral imaging applications, Northeastern University Boston, Mas- sachusetts, 2012 Ph.D. thesis .
[29] R.S. DiPietro , The detection of sub-pixel objects and mitigation of false alarms in hyperspectral imagery, Northeastern University Boston, Massachusetts, 2010 Ph.D. thesis .
[30] J.W. Boardman , F.A. Kruse , Analysis of imaging spectrometer data using n-di- mensional geometry and a mixture-tuned matched filtering approach, IEEE Trans. Geosci. Remote Sens. 49 (11) (2011) 4138–4152 .
[31] N.B. Pulsone , C.M. Rader , Adaptive beamformer orthogonal rejection test, IEEE Trans. Signal Process. 49 (2001) 521–529 .
[32] N. Acito , S. Matteoli , A. Rossi , M. Diani , G. Corsini , Hyperspectral airborne ǣViareggio 2013 trial ǥ data collection for detection algorithm assessment, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 9 (6) (2016) 2356–2376 .
[33] M. Shi , G. Healey , Hyperspectral texture recognition using a multiscale oppo- nent representation, IEEE Trans. Geosci. Remote Sens. 41 (5) (2003) 1090–1095 .