Radiance Calibration of DMSP-OLS Low-Light

Radiance Calibration of DMSP-OLS Low-Light Imaging Data of Human SettlementsChristopher D.Elvidge,*Kimberly E.Baugh,†John B.Dietz,‡Theodore Bland,§Paul C.Sutton, and Herbert W.Kroehl¶N octurnal lighting is a primary method for enabling hu-tering the global environment include greenhouse gas man activity.Outdoor lighting is used extensively world-emissions from fossil fuel consumption,air and waterpollution,and land cover/land use change.wide in residential,commercial,industrial,public facili-Far from being evenly distributed across the land ties,and roadways.A radiance calibrated nighttime lightssurface,to a great extent human activities with environ-image of the United States has been assembled from De-mental consequences are concentrated in or near human fense Meteorological Satellite Program(DMSP)Opera-population centers.One approach to modeling the spatial tional Linescan System(OLS).The satellite observationdistribution of human activities is to use population den-of the location and intensity of nocturnal lighting providesity as an indicator for the phenomenon of interest(e.g., a unique view of humanities presence and can be used asthe percent coverage of impermeable surfaces).There a spatial indicator for other variables that are more diffi-are two primary disadvantages to using population den-cult to observe at a global scale.Examples include thesity as an indicator for human activities with environmen-modeling of population density and energy related green-tal consequences.At a global level,population density is house gas emissions.Published by Elsevier Science Inc.not well characterized.Currently available global popula-tion density grids cover approximately60sq km per datacell(Tobler et al.,1995),far too coarse for many envi-INTRODUCTIONronmental applications.In areas where high spatial reso-Much of global change research is dedicated to discern-lution population density data sets are available,the envi-ing and documenting the impacts of human activities on ronmental applicability of the data suffer due to the fact natural systems.Human population numbers have ex-that population density is defined as a residential param-panded fromف750million in the mid-1700s,will reach eter.As a result transportation corridors,public,com-6billion in1999,and could double in the next49years mercial,and industrial zones have very low population if current growth continues(Haub and Cornelius,1998).density.In many cases these areas have much higher Human activities which are known to be cumulatively al-densities of people present8–12h/day than the associ-ated residential zones.Thus the use of population den-sity as an indicator for percent of land area covered by *NOAA National Geophysical Data Center,Boulderimpermeable surface(roads,roofs,parking lots)would †Cooperative Institute for Research in Environmental Sciences,University of Colorado,Boulder result in a substantial skewing of the results towards resi-‡Cooperative Institute for Research on Atmosphere,Colorado dential areas.State University,Fort CollinsHaving a capability for direct global observation of a §Electronic Sensors and Systems Division,Northrop GrummanCorporation,Baltimore widespread and distinctly human activity that varies in Department of Geography,University of California,Santa intensity could substantially improve understanding of Barbarathe magnitude of humanities presence and modeling hu-¶Solar-Terrestrial Physics Division,NOAA National GeophysicalData Center,Boulder man impacts on the environment.Available satellites Address correspondence to C.D.Elvidge,NOAA National Geo-sensors with global data acquisition capabilities have all physical Data Center,325Broadway,Boulder,CO80303.E-mail:focused on the observation of natural systems as their cde@Received9May1998;revised2October1998.design criteria.If global observation of human activity REMOTE SENS.ENVIRON.68:77–88(1999)Published by Elsevier Science Inc.0034-4257/99/$–see front matter 655Avenue of the Americas,New York,NY10010PII S0034-4257(98)00098-478Elvidge et al.1976.In1999,DMSP operates three OLS(satellitesF-12,F-13,and F-14).Many details of the OLS instru-ment and the onboard data processing are described byLieske(1981).The OLS is an oscillating scan radiometer whichgenerates images with a swath width ofف3000km.With14orbits per day,each OLS is capable of generatingglobal daytime and nighttime coverages of the Earth ev-ery24h.The full resolution data,having a ground sam-ple distance(GSD)of0.56km,is referred to as“fine.”Onboard averaging of five by five blocks of fine data pro-duces“smoothed”data with a GSD of2.7km.The“visi-ble”bandpass straddles the visible and near-infrared(VNIR)portion of the spectrum(Fig.1).The thermal Figure1.Bandpass for the visible band for the OLS instru-infrared channel has a bandpass covering10–13l m.The ment aboard DMSP F-12.visible band has6-bit quantization,producing digitalnumber(digital number)values ranging from0to63. was set as a design criteria,what wavelength(s)would beThe thermal infrared band has8-bit quantization(digital investigated?Radio frequencies emitted by power lines,number range from0to255).A constellation of two sat-electrical devices,and cellular telephones might be aellites provides for global coverage four times a day: good starting point!As an alternative,we have been in-dawn,day,dusk,and night.Satellite attitude is stabilized vestigating the nocturnal observation of artificial lightingusing four gyroscopes(three-axis stabilization),a star as a measure or indicator of human activity using data mapper,Earth limb sensor,and a solar detector.The potential use of nighttime OLS data for the ob-collected by the U.S.Air Force Defense MeteorologicalSatellite Program(DMSP)Operational Linescan System servation of city lights and other VNIR emission sources (OLS).This instrument has a low-light imaging capabil-was first noted in the1970s by Croft(1973;1978;1979). ity,which was designed for the observation of clouds illu-Welch(1980)and Foster(1983)speculated that OLS minated by moonlight.In addition to moonlit clouds,thedata could be used to map the distribution human settle-data can be used to detect light sources present at the ments and inventory the spatial distribution of human ac-Earth’s surface.tivities,such as energy consumption.Sullivan(1989)pro-We have explored the gain control of the OLS and duced a10km resolution global image of OLS observed discovered that it is possible to adjust the gain to coverVNIR emission sources using film data.The global map the range of radiances encountered from brightly lit published by Sullivan(1989)was derived from single commercial centers in the largest cities down to residen-dates of OLS imagery,selected based on the presence tial areas in urban,suburban,and many rural settings.of large number cloud-free VNIR emission sources and The OLS low-light imaging data can be converted tomosaicked into a global product.As a result,many of the brightness values based on the preflight calibration of the features presented in areas such as Africa are ephemeralVNIR emissions from fires.These early studies with OLS OLS sensor.The objective of this article is to review thelow-light imaging features of the OLS,to describe meth-data relied on the analysis of film strips,which limited ods used for the assembly of radiance calibrated night-the scope of the studies.time lights products of the United States,and to com-In1992the U.S.Air Force and NOAA established pare this product to several other data types.a digital archive for DMSP data at the NOAA NationalGeophysical Data Center.Subsequently,algorithms havebeen developed to identify and geolocate VNIR emission BACKGROUNDsources in nighttime OLS imagery(Elvidge et al.,1997a). Since1970the DMSP has operated polar orbiting satel-NGDC has produced a global inventory of fires,gas lite sensors capable of low-light imaging the Earth at flares,fishing lightboats,and human settlements using ing light intensifying photomultiplier tube tech-DMSP-OLS observations acquired between1October nology,these sensors were designed to observe clouds il-1994through31March1995.The thermal infrared band luminated by moonlight using a single broad spectralOLS data are used to identify clouds,and the number band(Fig.1).With sunlight eliminated,it possible to of times lights are detected in cloud-free areas are tal-also detect city lights,gas flares,and fires.From1970tolied.By dividing this tally by the total number of cloud-1975DMSP Blocks5A,B,and C flew equipped with the free observation,it is possible to normalize the frequency SAP(S ensor A erospace Vehicle Electronics P ackage).of the lights for differences in the number of usable ob-The Operational Linescan System(OLS)first flew on servations.The percent frequency values are then DMSP Block5D-1satellite F-1,launched in Septemberthresholded to eliminate ephemeral events like noise,Calibration of Low-Light,Nighttime Imagery79Figure2.Relationship between the F-12OLS visible band gain when operated in PMT mode,dig-ital numbers,and observed radiances derived from the preflight sensor calibration.lightning,and fires.Our experience indicates that most OLS LOW-LIGHT IMAGING GAIN CONTROL ephemeral events can be eliminated by deleting detec-An optical instrument’s gain can be thought of as the tions which occurred in less than6–10%of the cloud-amplification of the incoming signal,from the front end free observations for a time series spanning a6-monthof the telescope to the output of the digital number data time period.The finished stable lights data record thestream.The full system contains both gains and losses; percent frequency with which lights were observed withinhowever,the overall system amplifies the original signal. the set of cloud-free observations.The OLS has analog preamplifiers and postamplifier with Stable lights for the1994–1995time period havefixed gains as well as VDGA(Variable Digital Gain Am-been produced for most of the Americas,Europe,Asia,plifier)gain.The gain of the photomultiplier tube(PMT) and northern Africa(Elvidge et al.,1997b).Sutton et al.also contributes to the overall system gain for night scenes. (1997)examined the potential use of the stable lights dataPrimary control of the low-light imaging gain is achieved to spatially apportion population.Imhoff et al.(1997a,b)through ground command of the VDGA,which has values used the stable lights to estimate the extent of land areasfrom0to63decibels(dB is the log of the amplification). withdrawn from agricultural production.Gallo et al.Prior to launch,the OLS is calibrated under condi-(1995)used nighttime lights to assess urban heat islandtions which simulate the space environment.During the impacts on meteorological records.Elvidge et al.(1997b,c)calibration the OLS views light sources having known ir-found that the area lit(km2)from the stable lights of in-radiance,and digital numbers are recorded over the full dividual counties to be highly correlated to Gross Do-range of VDGA.The calibration data equate telescope mestic Product.input illuminance to digital numbers at specific VDGA The stable lights product indicates the percent fre-gain settings.quency with which lights were detected within the set ofFigure2shows the relationship between OLS cloud-free observations,with no indication of the bright-VDGA gain settings and the observed range of radiances, ness of the lights.The stable lights products were assem-based on the preflight sensor calibration.This calibration bled utilizing data acquired under low lunar illuminationdata makes it possible to relate OLS digital numbers conditions,making it is possible to exclude moonlit cloudsback to the laboratory observed radiances known in and reservoirs from the analysis.The high gain settingsterms of W/cm2/sr/l m.The solid diagonal line indicates (or amplification)used during the dark half of the monththe saturation radiance(digital numberϭ63).Observed make it possible to detect faint sources of VNIR emis-pixels with radiances greater than this radiance yield dig-sion,but result in saturated data for many urban areas.ital number values of63.The dashed diagonal line indi-This prevented implementation of a radiance calibrationin our earliest nighttime lights products.cates the radiance for a digital number of1,the lowest80Elvidge et al.detectable radiance.At a VDGA gain setting of63,thefull system gain,including the fixed gain amplifiers,is136dB.This is a light intensification of1013.6times theoriginal input signal,permitting measurement of radi-ances down to10Ϫ10W/cm2/sr/l m.This is approximately6orders of magnitude lower than the OLS daytime visi-ble band or the VNIR bands of other sensors,such asthe NOAA AVHRR or the Landsat Thematic Mapper.The primary objective of the on-orbit OLS gain con-trol is the generation of consistent imagery of clouds atall scan angles for visual interpretation by Air Force me-teorologists.In normal operations the VDGA is modifiedto track scene illumination predicted from lunar phaseand elevation.The resulting base gains are modified ev-ery0.4ms by an onboard along scan gain control(ASGC)algorithm.In addition,a BRDF(bidirectionalreflectance distribution function)algorithm further ad-justs the gain in the scan segment where the illuminationangle equals the observation angle.The lowest gain settings occur under full moon con-Figure3.F-12OLS visible band data acquired with a range ditions,producing imagery which looks strikingly similarof gain settings on16March1996.to daytime visible band data.Gain settings gradually riseas lunar illumination declines,reaching levels at or near60dB(Fig.2)between the last quarter and first quartersaturation.This examination lead to the selection of a lunar phases.During the darkest10nights of each lunarVDGA gain setting of26dB for use in the second24h. cycle,illumination is too low to detect clouds in the OLSIt was found that this gain setting continued to produce visible band data,even with high gain settings.Undersaturated data in the bright centers of major cities and these conditions the effects of the along scan gain andthe gain level was reduced to24dB for the third24-h BRDF algorithms are minimized,and the gain rises toperiod.Very few urban centers continued to have satu-its maximum monthly level.rated pixels in the data acquired at24dB.For example,a set of three saturated OLS pixels were found in the DATA ACQUISITION center of Las Vegas,Nevada.However,at this low gainsetting,only the bright cores of urban centers were de-In early1996,NGDC requested and received permissiontected.To detect the larger number of the smaller and from the DMSP program office to acquire OLS data atdimmer sites,it was decided to acquire data with gain reduced gain settings,with onboard ASGC(along scansettings of40dB on the fourth24-h period and to then gain control)and BDRF functions turned off.The“low-alternate between24dB and40dB every24h(Fig.2). gain”data acquisition was requested for an8-night pe-Initial examination of the1996low-gain revealed riod,corresponding to the darkest nights in March1996,that it is possible to observe brightness variations within with a start date of16March.In order to avoid interfer-urban centers.A global radiance calibrated product could ence with ongoing observations of the leading edge ofnot yet be assembled due the presence of solar glare at the aurora at high northern latitudes,the NGDC requestlatitudes greater than45Њand the relatively small num-was for the acquisition of reduced gain nighttime visibleber of orbits available at each gain setting.It was also band data for theϪ55Њto55Њlatitude range.The north-concluded that a gain setting higher than40dB would ern limit of the usable data was reduced to latitudes lessbe required to detect the large numbers of small towns than45Њdue to the presence of solar glare(see Fig.3).detected in the1994–1995stable lights products.As a Because the OLS gain had never been operated withresult,NGDC requested and received additional low-the objective of observing the brightness patterns of lightgain data acquisitions for5–14January and3–12Febru-sources present at the Earth’s surface,there was no aary1997fromϪ65Њto75Њlatitude,using alternating gain priori information available on what gain setting wouldsettings of24dB and50dB(Fig.2).These correspond be required to avoid saturation on major urban centers.to the10darkest nights of January and February1997. To address this question,OLS data was acquired at aseries of VDGA gain settings,by switching the gain in asteplike fashion for the first24h of the8-day period.AnDATA PROCESSINGexample of this data is shown in Figure3.As a first step in the processing,sections(suborbits)of On17March the step-gain OLS data values of ma-jor urban centers were examined to determine extent of usable nighttime lights data were extracted from theCalibration of Low-Light,Nighttime Imagery81Figure5.Total number of cloud-free data coverages for ob-Figure4.Total number of data coverages for observations servations made at gains greater than30dB(top)andmade at gains greater than30dB(top)and for obser-observations made at gains less than30dB(bottom). vations made at gains less than30dB(bottom).which draws a polygon at the edges of the scan and along original full orbit files.The data processing then involvedthe edges of data gaps.These polygons are ultimately the following primary steps:1)establishment of a refer-used to track the total number of cloud-free observations ence grid;2)identification and geolocation of lights,within the reference grid.clouds,and coverage areas;3)establishing a digital num-In our previous nighttime lights products,lights ber to radiance scale for the final product;4)cloud-freewere detected using an automatic algorithm that estab-compositing within two overlapping gain ranges(highlished a local threshold based on background brightness and low);5)threshold to eliminate isolated detections;levels(Elvidge et al.,1997a).This methodology was de-and6)combine the calibrated images from the two gainveloped to reduce human involvement in the identifica-ranges by averaging based on the number of detections.tion of the lights,to contend with the high noise levels1.Reference Grid:Development of a spatially coher-typically encountered(see bottom section of imagery in ent temporal image composite requires the use of a ref-Fig.3),and to permit processing at all lunar illumina-erence grid with finer spatial resolution than the inputtion conditions.imagery.We have used the1km equal area InterruptedThe local thresholding algorithm operates very Goode Homolosine Projection(Goode,1925;Steinwand,quickly and identifies lights from major cities and towns 1993)developed for the NASA-USGS Global1km(points of light)under a wide range of conditions.How-AVHRR project(Eidenshink and Faundeen,1995).Theever,the algorithm does not generally detect“diffuse IGHP is optimized to provide a uniform grid cell size atlighting,”which is dim and scattered across the landscape. all latitudes and contiguous land masses(except Antarc-These areas can be observed in OLS data in locations like tica).These are favorable characteristics for the productionthe northeastern United States and surrounding major of global land datasets from raster imagery(Steinwand,metropolitan areas.1993).We found that noise was greatly reduced in the data2.Identification and Geolocation of Lights,Clouds,that were acquired for our study.In an attempt to iden-and Coverages:Data gaps within suborbits are filled withtify areas of diffuse lighting as well as lights in focused digital number values of zero.The outline of each blockof valid data was identified using an automatic program centers such as cities and towns,we developed a soft-82Elvidge et al.Figure6.Digital number versus ra-diance diagram for the compositedoutput product.ware tool that an image analyst can use to select athreshold a single digital number above the noise levelFigure7.Radiance calibrated OLS image composites of encountered in a suborbit.An image analyst then re-Minneapolis,Minnesota for two overlapping gain ranges, viewed each suborbit,selecting thresholds,which were indicating the use of compositing to overcome the dynamicused to extract pixels with detectable levels of VNIR range limitations of the OLS sensor.emission.The analyst also identified defective scan linesand scan lines with lightning.In addition,the analyst cre-ated image masks for image areas which contained exces-sive noise or phenomena such as solar glare.Data withinthe marked scan lines of under the image masks wereeliminated from the compositing procedure describedbelow.Because of the low level of lunar illumination pres-ent in the orbital segment,it was not possible to use thevisible band to identify clouds.The cloud identificationwas based entirely on thresholds set on the thermal in-frared band.Clouds are generally colder than the Earth’ssurface.However,the separation of cloud pixels fromEarth surface pixels using thermal infrared thresholdingis complicated by seasonal,latitudinal,and altitudinalvariations in the background Earth surface temperature.The separation of clouds from Earth surface pixels is rel-atively easy at low latitudes,where there is generally alarge temperature difference between pixels of cloudtops and pixels containing land or ocean.Because of thestrong latitudinal effects on the thermal infrared thresh-old for cloud screening in our data,we segment each su-borbit into200-pixel-wide latitudinal bands and visuallyselect a thermal infrared threshold for identification ofthe cloud pixels.Calibration of Low-Light,Nighttime Imagery83Figure8.Radiance calibrated nighttime lights of the United States.Lights,clouds,and the coverage polygons were then calibrated radiance observed was3.17ϫ10Ϫ7W/cm2/sr/l m.In establishing a digital number to radiance conver-geolocated.Our geolocation algorithm operates in theforward mode,projecting the center point of each pixel sion scale,we had two criteria:1)The digital number onto the Earth’s surface.The geolocation algorithm esti-data in byte format(0–255),and2)the scale should ex-mates the latitude and longitude of pixel centers based ceed the range of current data to accommodate later ex-on the geodetic subtrack of the satellite orbit,satellitetensions made at lower or higher OLS gain settings.We altitude,OLS scan angle equations,an Earth sea level developed a logarithmic scaling using the function:Radi-anceϭ(digital number)3/2ϫ10Ϫ10W/cm2/sr/l m(Fig.6). model,and digital terrain data.The geodetic subtrack ofeach orbit is modeled using daily radar bevel vector positing within High and Low Gain Ranges: sightings of the satellite(provided by Naval Space Com-Suborbits were divided into two,roughly equal data vol-mand)as input into an Air Force orbital mechanics ume groups:A)data acquired at gains less than30dB model that calculates the satellite position every0.4208s.and B)data acquired at gains greater than30dB.The satellite heading is estimated by computing the tan-To determine the total number of“cloud-free”ob-gent to the orbital subtrack.We have used an oblate el-servations,the data coverage polygons and cloud detec-lipsoid model of sea level and have used30arc second tions are tallied within the reference grid.Subtraction of digital terrain data provided by the U.S.Geological Sur-these two images yields the number of cloud-free obser-vey,EROS Data Center.vations in the time series.This was performed to both3.Digital Number to Radiance Scale:The thresholdsthe less than and greater than30dB data sets(Figs.4 for the light detection established the range of radiances and5)to ensure that each land surface area had multiple which would be encompassed in the output product.Thevalid observations.Inspection of the images in Figure5 lowest radiances,detected in data acquired with a gain revealed that land areas had at least three cloud-free ob-of50dB,was1.54ϫ10Ϫ9W/cm2/sr/l m.The maximumservations in both gain ranges.84Elvidge et al.Figure9.Stable lights of the United States derived from a1994–1995DMSP-OLS time series(Elvidge et al.,1997a).The second step in the preliminary compositing stage to the stable lights:1)The radiance calibrated lights pro-is to tally the number of light detections and to producevide spatial detail related to brightness variations within an average radiance for each reference grid pixel with urban centers(Fig.10)that was not present in the stable light detection.lights.2)The radiance calibrated lights show low levels5.Thresholding:The high and low gain composite of“diffuse lighting”(shown as purple on Fig.8)indensely populated rural environments in the Washington, images were filtered to eliminate detections which oc-curred only once.This step removes noise and many DC to Boston corridor,western Pennsylvania,Indiana, ephemeral events such as fires or lightning.Michigan,Wisconsin,and surrounding major metropoli-bining High and Low Gain Composites:The tan areas like Minneapolis-St.Paul,Minnesota.high and low gain cloud-free composites are averaged.The radiance calibrated nighttime lights are substan-The radiance calibrated digital number average from tially different in character than other available geos-each image is weighted by the total number of detec-patial data sources.Figure10a shows a Landsat MSS tions.Images demonstrating this final step in the product scene of the St.Louis,Missouri region in a traditional generation are shown in Figure7.“false color”rendition.The information content is pri-marily land cover and geomorphology.Major roads canbe observed,but there appears to be substantial spectral RESULTS AND DISCUSSIONoverlap between human settlements and other land cover Figure8is a color-coded image of the radiance cali-types.Figure10b shows a color-coded version of popula-brated nighttime lights of the United States from1996–tion density,processed from1990U.S.Census Bureau 1997.For comparison,the1994–1995stable lights prod-data by CIESIN(1996).Note that there is substantial uct of the United States(Elvidge et al.,1997a),is shown detail present in the St.Louis area but that outside of in Figure9.As with the1994–1995stable lights data,thethe metropolitan area the data appear to be in large ho-radiance calibrated product shows the cities and many of mogeneous blocks.This is due to the fact the reporting the smaller towns of the United States.Two major differ-units used in the census data are postal ZIP codes,which ences have been identified in the spatial information can be quite large in rural areas.Figure10c shows the content of the radiance calibrated lights when comparedradiance calibrated nighttime lights data for the sameCalibration of Low-Light,Nighttime Imagery85Figure10.a)Landsat MSS data of the St.Louis,Missouri area,Path24Row33,Bands4,2,1as red,green,blue,3Octo-ber1992.b)Population density of the St.Louis,Missouri area in a1km grid(CIESIN,1996).Derived from1990U.S. Census Bureau data.c)Radiance calibrated nighttime lights of the St.Louis,Missouri area derived from DMSP-OLS data acquired during March1996and January–February1997.d)Stable lights of the St.Louis,Missouri area derived froma1994–1995DMSP-OLS time series.area.Note that radiance calibrated data is more similar out the ability of the OLS data to inventory human activ-to the population density data than the Landsat data.ity that is not associated with residences.The1994–1995 Close examination of the data in Figures10b and10c OLS stable lights data are presented in Figure10d.The and more detailed maps indicates that the St.Louis air-units for the stable lights data are percent frequency of port and surrounding commercial and industrial zones detection rather than brightness.The stable light data are quite bright in the radiance calibrated nighttimeprovide clear indications of the locations of small towns lights but have low population density.This is one of the in rural settings,but do not provide the detailed zonation busiest areas within the metropolitan region and pointsof the radiance calibrated lights within the metropolitan。