IJJSS2014 :The6thIndonesiaJapan JointScientificSymposium Yogyakarta,28‐30October2014 Theme code: E5 Applicationofhyperspectralcameraforaerosol characterization NaohiroManago1*,HayatoSaito1,YoheiTakara2,MakotoSuzuki3,HiroakiKuze1 1CenterforEnvironmentalRemoteSensing,ChibaUniversity,1‐33Yayoi‐cho,Inage‐ku, Chiba‐shi,Chiba,263‐8522,JAPAN 2EBAJapanCo.,Ltd.,#3012‐17‐12Takanawa,Minato‐ku,Tokyo,108‐0074,JAPAN 3InstituteofSpaceandAstronauticalScience,JapanAerospaceExplorationAgency, 3‐1‐1Yoshinodai,Chuo‐ku,Sagamihara,Kanagawa,252‐5210,JAPAN Abstract Optical observation of aerosol particles in the atmosphere is indispensable for various purposessuchasatmosphericcorrectionofsatellitedata,basisforclimatechangestudies, and air pollution monitoring. The progress in the remote sensing instrumentation has madeitpossibletocarryoutaerosolstudiesonthebasisoftwo‐dimensionalimagery.As comparedwithconventionalmethodologyusingaskyradiometerorasunphotometer,the obvious advantage of the image formation is the elimination of necessity for precise trackingofthesolarposition,inadditiontotheeaseofcloudrecognition.Inthisstudy,we describethecharacterizationofanewlydevelopedhyperspectralcamerathatprovides uswithimageswith752×480pixelsoverthewavelengthrangeof350‐1100nmwitha typicalresolutionof6nm.Thewavelengthandradiancecalibrationandtheuniformity checkoftheinstrumentaredescribedindetail.Wealsoreportonthepreliminaryresults oftheaureoleandskylightanalysisconductedattheCenterforEnvironmentalRemote Sensing, Chiba University. The present results have shown that the use of a compact hyperspectralcameracanbepromisingfordetailedaerosolcharacterizationstudies. Keywords Hyperspectralimaging;Skylight;Troposphericaerosol;PM2.5; 1.Introduction Recently, aerosol has become a popular concern because of the hazardous concentrationof“PM2.5”occasionallyoccurringinlargecitiesworldwide.Sincephysical andchemicalcharacteristicsofaerosolsvaryspatiallyandtemporallyinasignificantway, aerosolcharacterizationisindispensabletounderstandingtheatmosphericenvironment precisely.Skylightintensity(spectralradianceofscatteredsolarradiation)dependson theseparationanglefromthesolarposition,andisstronglymodulatedbyintermediate substancesintheatmosphere,especiallybytroposphericaerosols. *Correspondingauthor.Tel.:+81‐43‐290‐3841;fax:+81‐43‐290‐3857. E‐mailaddress:manago.naohiro@chiba‐u.jp Inthisstudy,physicalandopticalpropertiesoftroposphericaerosols,suchassize distribution and complex refractive index, are retrieved from the analysis of hyperspectral (HS) imagery of skylight in conjunction with the information on the wavelengthdependentaerosolopticaldepth(AOD)derivedfromdirectsolarirradiation. Compared with conventional approaches such as those based on a sunphotometer (Holbenetal.,1998)oraskyradiometer(Nakajimaetal.,1996),thepresentmethodis advantageousbecauseofthefollowingfourreasons;1)usingaHScamera,skylightclose totheSuncanpreciselybemeasuredwithagoodspatialcontext,whichresultsinbetter determinationofsizedistribution;2)cloudypixelscanberejectedaccurately;3)column amountofmoleculessuchaswatervaporcandirectlybemeasuredfromtheirabsorption features;and4)simultaneousmeasurementofskylightintensitycanbemadewithinthe view angel of the camera (ca. 20 deg). We explain the methodology for calibrating instrumentsintermsofparameterssuchaswavelength,radiance,viewingangle,etc.Also, wedescribetheformulationhowtheretrievalofaerosolpropertiesisimplemented,with someexamplesshowingtheusefulnessofthepresentapproach. 2.Methodology 2.1Instrument Inthisresearch,weuseaHScamera,NH‐2,manufacturedbyEBAJapanCo.,Ltd.(Takara et al., 2012). Major specifications of the camera are listed in Table 1. It has a built‐in scanningmechanism,sothatnootherexternaltranslationdeviceisrequiredtotakeHS images,anditiscapableofscanning16º×10ºwideareain4seconds.Theopticsisbased ontransmissiondiffractiongrating,andthereforelesssensitivetothepolarizationstate ofincominglightcomparedwithanacousto‐optictunablefilter(AOTF)oraliquidcrystal tunablefilter(LCTF). Table1.SpecificationofEBAJapanNH‐2 Item Value SensorType CMOS Optics Transmissiondiffractiongrating Max.imagesize 752×480pixel FieldofView 16º×10º Spatialsamplinginterval 0.02º/pixel Spatialresolution 0.07º Max.wavelengthsize 480pixel Wavelengthrange 350∼1100nm Wavelengthsamplinginterval 1.6nm/pixel(minimum) Wavelengthresolution 6nm(typical) Colordepth 10bit(sensor),12bit(camera) Dimension 76×62×193mm Weight 850g 2.2Calibration ThereareNx×Nw=752×480pixelsintheCMOSsensorarrayinsideNH‐2.Incidentlight from an oblong light source is distributed into the x direction on the sensor, while chromaticlightdispersedbythediffractiongratingisdistributedintothewdirection.In addition,thebuilt‐inscannermovesthe“linesensor”verticallyinNy=480stepstoform aHSimagewith2spatialdimensionsplus1wavelengthdimension.Thedataacquisition software enables us to pick out wavelength pixels arbitrarily, and spatial pixels in a 2 rectangularregionofinterest.Themovingdistanceofthescanner’ssteppingmotorcan bechangedbyreconfiguringthesettingssavedintheonboardROM.Thepixelnumbers inxandydirectionsofaHSimage(pxandpy)arerelatedtohorizontalandverticalangle (θx and θy), respectively, while the pixel number in w direction (pw) is related to wavelength(). Similarly, the digital number output from the sensor (z) is related to spectral radiance (L). In order to obtain coefficients to convert these instrumental values to physical values, we have to perform viewing angle calibration, wavelength calibration, andradiancecalibration.Wavelengthcalibrationandradiancecalibrationaredoneusing calibrationcoefficientsderivedatthecentralregionofaHSimagefirst,thenuniformity correction is performed to take the partial difference in an image into account. Also, temperaturecorrectionsareneededtocompensatethetemperaturedependenceofthe conversion coefficients. After proper calibrations and corrections, measured data are comparedwithcalculateddatareproducedbyradiativetransfersimulationtoretrieve aerosolproperties.Sinceinthisprocedure,thespatialresolutionandopticalresolution ofthecalculateddatashouldbematchedwiththatofthemeasureddata,itisnecessary toestimatetheseresolutionsbeforehand.Itisexpectedthatpolarizationdependenceof NH‐2issmall,butitneedstobeconfirmedexperimentally. Viewing angle calibration coefficients (Dx = dθx/dpx|r=0 and Dy = dθy/dpY|r=0) are optimized so that the result of the simulation that modeled the optical system of the camera can reproduce actual HS image. Here, we assume an image distortion model expressedas rd r (1 K1 r 2 K 2 r 4 ) , (1) andoptimizethedistortioncorrectioncoefficients(K1andK2).Theparameters,randrd, are the radial distance from the optical axis to image on the sensor before and after correctingimagedistortion,respectively.Across‐sectionpaperonawallapartfromthe cameralensisagoodtargetoftheHSimagefortheviewinganglecalibration.Oncewe gettherelationshipbetweenpixelnumberandviewingangle,wecanestimatethespatial resolutionoftheHScamerafromthespreadofanimageofapoint‐likelightsourceata fardistance. The relationship between wavelength pixel number (pw) and wavelength (λ) is approximatedwithapolynomialfunctionas c0 c1 p w c 2 p w 2 c3 p w 3 c 4 p w 4 c5 p w 5 . (2) The coefficients (wavelength calibration coefficients) in eq.(2) are optimized using HS imagesoflightsourceshavingnarrowemissionlineswithknownpeakwavelengths,such asdischargelampsorlasers. WhenacquiringHSimages,atransmissivediffuserisplacedinfrontofthelensso thattheincidentlightcanilluminatetheentiresensoroftheHScamera.First,weobtain the spectral intensity (z) around an emission peak, as a function of wavelength pixel number (pw), averaged in the central region of a HS image. Then, we find the peak wavelengthpixelnumberandrelateittotheknownwavelength().Here,byusingspline interpolation,thespectrumisinterpolatedonafinegridtoincreasetheprecision.After pairingasufficientnumberofpwand,weobtainwavelengthcalibrationcoefficients(c0, c1,etc.)byleastsquaresfitting.Next,wefindthepeakwavelengthpixelnumberinthe entireregionoftheHSimageandcalculatethedifferencefromthecentralregionforeach pixel.ThewavelengthresolutionoftheHScameracanbeestimatedfromthefullwidthat halfmaximum(FWHM)ofintrinsicallynarrowemissionlines. Radiance calibration coefficients (R=dz/dL) are obtained from a HS image of a lightsourcewithknownspectralradiance(L)bycomparingsensoroutput(z)withL.The spectral radiance of the light source can be evaluated if an already‐calibrated spectroradiometerisavailable.Inthiscase,itisnecessarytomatchtheFOVbetweenthe spectroradiometerandtheHScamera.Also,weneedtoconsiderthedifferenceofoptical resolution, though it is not necessary to match the optical resolution if we use a light source with smooth wavelength dependence like a halogen lamp. The linearity of the sensoroutputcanbecheckedbycomparingradiancecalibrationcoefficientsobtainedat variousinputlevelschangingeitherexposuretimeoftheHScameraortheintensityof thelightsource.Inordertoobtainuniformitycorrectioncoefficients,alightsourcewith uniformluminancedistribution(orknownluminancedistribution)withintheFOVofthe HScameraisrequired.Integratingspheresandtwilightflat(skylightaroundthezenithin cloudlessearlymorningorevening)areconsideredtobesuchlightsources. TemperaturedependenceoftheHScameracanbeinvestigatedbyputtingitina temperature‐controlled chamber and observing a light source outside the chamber, changingtheinsidetemperaturestepbystep.Thefluctuationofthelightsourceintensity should be monitored and corrected using a spectroradiometer. Apparent wavelengths obtained by eq.(2) will change with temperature, and the difference of apparent wavelengths()correspondingtotemperaturechange(t)canbemodeledas c0 c1 c22 c3 c4 c52 t c6 c7 t 2 t . (3) A light source with some spectral features like a discharge lamp or a xenon lamp is suitableforwavelengthcalibration.Asforsensoroutput(radiancecalibration),thedark signal and dark‐subtracted signal have individual temperature dependence. It can be considered that there is no wavelength dependence of dark signal itself, and it can be modeled using a 5th order polynomial. The temperature dependence of the dark‐ subtractedsignalcanbemodeledusing2‐dimensional5thorderB‐splineinterpolation based on measurements at several temperatures and wavelengths. A light source with smoothwavelengthdependencelikeahalogenlampissuitableforradiancecalibration. 2.3Observation The solar and sky radiation measurements were conducted at the Center for EnvironmentalRemoteSensing(CEReS,35.62ºN,140.10ºE),ChibaUniversityunderclear skyconditions.ThelocationoftheChibacityisabout30kmsouthofthecentralTokyo, along the east coast of the Tokyo Bay. We pointed the HS camera toward the sun and recordedHSimagesoftheaureole.Weputalightshield(steelbar)about2mapartfrom the camera to block the direct sunlight. Almost at the same time, we measured solar irradianceusingacalibratedspectroradiometer(EKO,MS‐720)withabaffletubelimiting theFOVto5degrees(ManagoandKuze,2012). 2.4Analysis Oncetheinstrumentsarecalibrated,wecancomparetheobservedspectrawith simulatedones,andbyupdatingtheinputparametersofthesimulation(MODTRAN),we can obtain optimal set of parameters describing tropospheric aerosols. The aerosol properties that can be optimized by this method are AOD, extinction coefficient, absorption coefficient, phase function, etc. Optimizing these parameters directly will causeanill‐posedproblemsincetherearetoomanyunknownquantities,soweintroduce three component aerosol model (Manago and Kuze, 2012). Especially for the test observation,weusesimplifiedmodelswithoneortwocomponents.Inthesesimplified models, we assume that aerosol shape is spherical, and the size distribution can be describedbythelognormalmodel.Eachcomponenthasacomplexrefractiveindex(real 4 and imaginary part may be wavelength dependent, but must be controlled by two parameters)andsizedistributionparameters(moderadiusandwidthofthedistribution). The total amount of aerosol is controlled by AOD and relative abundance of the two componentsiscontrolledbyamixingparameter.Wemayfixsomeparametersifwedo nothaveenoughinformation. 3.Resultsanddiscussion 3.1Wavelengthcalibration Thewavelengthcalibrationcoefficientsweremeasuredusingdischargelampsof mercury,cadmium,krypton,argon,etc.andaNd:YAGlaser.TheresultisshowninFigure 1.Weused14emissionlinesinthewavelengthrangebetween400and1064nm.The errorofthe5thorderpolynomialapproximationtoconvertwavelengthpixelnumberto wavelengthwas0.29nm(RMS,theerrorofthe1storderwas1.87nm).Weestimatedthe opticalresolutionfromwidthsofseparatedemissionlinestobe5nmatwavelength470 nmand8nmatwavelength1064nm. Figure1.Resultofwavelengthcalibrationusingvariouslamps. 3.2Radiancecalibration The radiance calibration coefficients were measured by the following three methods;i)Weusedhalogenlampsorskylightasalightsource,andcomparedoutput fromtheHScameraandaspectroradiometer(EKO,MS‐720)calibratedbytheLangley method. ii) We used halogen lamps and an integrating sphere as a light source, and comparedoutputfromtheHScameraaswellasanotherHScamera(EBAJapan,NH‐7) calibrated with the JAXA integrating sphere system (Yamamoto et al., 2002). iii) We calibratedtheHScamerawiththeJAXAintegratingspheresystem(withhalogenlamps asthelightsource).Theresultsexhibitgoodagreementsinthewavelengthrangebetween 450 and 700 nm. In the shorter wavelength, however, the results with halogen lamps seemstooverestimatethesensitivityoftheHScamera,whileinthelongerwavelength theskylightresultstendtooverestimatethesensitivity.Itturnsoutthattheinfluenceof stray light increases for wavelengths < 450 nm and >700 nm, and the influence of secondarylightappearsforwavelength>700nm.Sincehalogenlampsandskylightare dimmerintheshorterandlongerwavelengths,respectively,largererrorsareexpected for these wavelength ranges. A composite sensitivity curve made from the results of halogenlampsinthelongerrangeandthatofskylightintheshorterrangeisshownin Figure2.Theinfluenceofsecondarylightwillberemovedbyusingalongpassfilter. Figure2.SensitivityoftheNH‐2hyperspectralcamera. 3.3Uniformitycheck The wavelength calibration and sensitivity analysis were done using 10 × 10 pixelsaroundthecenteroftheFOV.Furthermore,weinvestigatedtheuniformityofthe entire pixels. We measured the spectrum of a fluorescent lamp using NH‐2 with a transmissivediffuserin frontofthelenssothatthewholesurface ofthe detectorwas homogenouslyilluminated.Wesearchedtheapparentpeakwavelengthsaround620nm. For each pixel, the spline interpolation was employed to derive the peak wavelength. Judgingfromahistogramoftheapparentpeakwavelengthsforthewholeimage,theRMS of the distribution is 0.41 nm, which is much smaller than the 1.6 nm width of each wavelengthbin. Next,toinvestigatetheuniformityoftheradiantvaluesmeasuredbyNH‐2,we observedthetwilightflatbeforesunset.Sincewecannotignorethenon‐uniformityofthe twilight flat in the FOV of NH‐2 (16º×10º), we rotated the camera horizontally, and averaged4imagescorrespondingtotherotationanglesof0º,180º,90º,and270º.When takingtheimageattherotationangleof0º,thedirectionoftherotationaxiswaschosen soastomaketheimageasuniformaspossible.Sincethebuilt‐inscannerofthecamera movesintheydirection,verticalstripes(fixedpatternnoise:FPN)canbeseenintheflat imageduetothedifferentgainofindividualelementofthelinesensor.Toremovethe noise,theflatimagewasdividedbytheFPN.Theremainingnon‐uniformityofthisimage is presumably ascribable to the optical system itself and/or the non‐uniformity of the twilightflat.TheRMSoftheradiantvaluedistributionis1.4%,anditis1.2%iftheFPNis removed. 3.4Darknoise Whentheobservationtargetisnotbrightenough,thedarknoiseofthecamera becomesrelativelyimportant.Weobtainedadarkimagebyaveragingimagestakenwith thetypicalsettingsforskylightmeasurements.ThemeanandRMSofthedarkimagewere estimatedtobe70±15count.ThedarkimageiscontaminatedbytheFPNarisingfrom theoffsetvaluesofthesensorelements,andaftersubtractingtheFPN,theresultingvalue was0±5count.Notethatalthoughthecolordepthofthesensoris10bits,themaximum signalofthecamerais4095counts(i.e.theunitcountis4). 3.5Viewingangle TheviewinganglewasestimatedfromaHSimageofa1cm‐griddedcross‐section paperonawalllocated1.2mapartfromthecameralens.Bycomparingtheimagewith optical simulation results, the viewing angles in x and y directions are estimated to be 0.022 deg/pixel and 0.020 deg/pixel, respectively. The viewing angle in x direction is determinedbythepixelsizeoftheCMOSsensor,whiletheviewingangleinydirectionis determinedbythemovingdistanceofthescanner’ssteppingmotor.Thedistortionatthe 6 outersideoftheimagewasnegligiblysmallcomparedwiththemeasurementerror.From a HS image of distant streetlights taken at night‐time, the spatial resolution of the HS camerawasestimatedtobe0.07deg. 3.6Temperaturedependence TemperaturedependenceoftheHScamerawasmeasuredusingatemperature‐ controlledchamber(Espec,MC‐710T).Wechangedthetemperatureinsidethechamber between 0 and 40ºC by 10 K step. We waited at least 150 min after changing the temperaturetoobtainHSimages.ThetemperatureoftheHScamerawasmeasuredbya hygrothermal sensor (Sensirion, SHT21) installed inside the camera, and it was approximately2Khigherthanchambertemperaturewhenitreachedequilibrium.The lightsourcewasplacedoutsidethechamberwherethetemperaturewascontrolledtobe 22±1ºCbyanairconditioner,andthelightenteredinsidethechamberthroughaglass window.Weusedamercurydischargelamptomeasurethetemperaturedependenceof apparentwavelength.Withareferencetemperatureof22ºC,thedifferenceofapparent wavelengths is less than 1 nm below 32ºC. However, the higher the temperature, the larger the difference, which becomes as large as 2 nm at 42ºC. The fitting error of the polynomialmodelis0.1nm(RMS). Formeasuringthetemperaturedependenceofsensoroutput,weusedahalogen lamp. The dark signal increases monotonically as temperature rises. The temperature correctionfactorofdark‐subtractedsignalissmall(lessthan5%)between500and900 nm.Outsidethiswavelengthregion,however,thefactorcanbeaslargeas30%.Thefitting errorofthe2‐dimensionalB‐splinemodelis2%(RMS,0.7%at<40ºC). From measurements with polarized filters, we confirmed that polarization dependenceoftheHScamera(NH‐2)isnegligiblysmall(lessthen±1∼2%). 3.7Aerosolcharacterization We conducted test observations using NH‐2 under clear sky conditions at the CEReSsiteseveraltimessinceApril,2013.Throughtheanalysis,ithasbeenfoundthat theuseofaonecomponentaerosolmodelbasedonasetofcomplexrefractiveindexand sizedistributionparameterswasgenerallyinsufficientforreproducingtheobserveddata (radiancedistribution).Thisisduetothesteepincreaseoftheaureoleintensitynearthe sun:intheregionwithseparationanglessmallerthan3deg.Recently,ithasbeenfound that the separation angle dependence can be reproduced by introducing the two component aerosol model with different particle sizes. An example of observation and analysis results is shown in Fig. 3. In Fig. 3(a), the contour plot depicts the aureole intensityat550nm.Thesunwaslocatedattheupperrightpositionindicatedbytheblack star.Thus,thisimagerepresentsthelowerleftquadrantregionaroundthesun.There wasasteelbaratthesolarpositiontoblockthedirectsolarradiation,butitwaserased byimageprocessingandshadedwithcrosssigns. All pixel values except the shaded region were plotted against the separation angle in Fig. 3(b). It can be seen that pixel values at the same separation angle are distributedinaverynarrowregion.Inthisfigure,thesimulationresultsusingthetwo componentaerosolmodelareshownwithadashedline.Thesimulationresultsreproduce theobserveddataevenatthesmallestseparationangle,assmallas1deg. (a) (b) Figure3.(a)2‐dimensionalaureoleintensitydistributionaroundthesunmeasuredwiththeNH‐2HScamera. (b)Separationangledependenceoftheaureoleintensity. Theparametersofthetwocomponentaerosolmodelandtheiroptimizedvalues arelistedinTable2.Thevaluesinsidebracketswerefixedduringthefittingprocedure. In addition, we optimized AOD and water scale factor (WSF) using the direct sunlight spectrum observed with MS‐720. Then, the imaginary part of the refractive index was optimizedusingtheskylightspectrum.Finally,themoderadiusofthecomponent1and the mixing ratio (weighted by extinction coefficient) were simultaneously optimized usingtheaureoleintensitydistribution.Wefedbacktheresultstore‐optimizethevalues ofAODetc.andthewholeprocedurewasrepeatedseveraltimes.Themoderadiusofthe component2waschosenfromthefollowingsetofvalues:0.01,0.05,0.1,or0.5m:after optimizingotherparameters,thevaluewiththesmallestresidualwaschosen. Table2.Parametersofthe2componentaerosolmodel Parameter Value Aerosolopticaldepth 0.17 Realpartofcomplexrefractiveindex (1.53) Imaginarypartofcomplexrefractive index 8.8×10‐5 Moderadius 2.61mand0.5m Sizedistributionwidth (0.26) Mixingratio* 1:5.32 Waterscalefactor** 0.25 *Mixingratioweightedbyextinctioncoefficientat550nm **Scalefactortothedefaultwatercolumnamount(2.92g/cm2). 4.Conclusion WehavedescribedtheresultsofourrecentmeasurementsconductedwithNH‐2 hyperspectralcameradevelopedforvariousremotesensingpurposes.Sinceonlyaslight changeintheskyspectracanleadtoconsiderabledifferenceintheaerosolpropertyin the atmosphere,careful calibrationisneededbeforeapplyingthecamera totheactual solar radiance measurements. Thus, we have conducted calibration measurements in termsofthewavelengthreadout,radiancestability,uniformity,darknoise,viewingangle, andtemperaturedependenceoftheoutputofNH‐2.Apreliminaryanalysisoftheaureole andskyradiancedatahasindicatedthatbyusingatwo‐componentaerosolmodel,theill‐ posednatureoftheaerosolretrievalcansuccessfullybecircumventedandtheobserved imagery was well fitted to the simulation result based on detailed radiative transfer calculation.Inthefuture,wewillapplythemethodologyforaerosolmeasurementsunder 8 various conditions, with detailed comparison with collocated measurements with a sunphotometerandaskyradiometer. Acknowledgements ThisworkwasfinanciallysupportedbyJSPSKAKENHIGrantNumber13401088. References Holben, B. N., et al. (1998), AERONET – A Federated Instrument Network and Data Archive for AerosolCharacterization,RemoteSensingofEnvironment,Vol.66,No.1,pp.1–16,Elsevier Publishing. ManagoN.,andH.Kuze(2012).Determinationoftroposphericaerosolcharacteristicsbyspectral measurements of solar radiation using a compact, stand‐alone spectroradiometer, Appl. Opt.,Vol.49,No.8,pp.1446‐1458,TheOpticalSocietyofAmerica. Nakajima,T.,etal.(1996),Useofskybrightnessmeasurementsfromgroundforremotesensingof particulatepolydispersions,Appl.Opt.,Vol.35,No.15,pp.2672‐2686,TheOpticalSocietyof America. Takara,Y.,etal.(2012),RemotesensingapplicationswithNHhyperspectralportablevideocamera. In Proceedings of SPIE 8527, Multispectral, Hyperspectral, and Ultraspectral Remote SensingTechnology,TechniquesandApplicationsIV. Yamamoto, Y., et al. (2002), Development of a calibration standard of the spectral radiance for opticalsensors,In:Proc.ofthe41stSICEAnnualConferenceVol.3,pp.1885‐1890.
© Copyright 2024 ExpyDoc