Web-enabled Landsat Data (WELD): Landsat ETM+ composited mosaics

Remote Sensing of Environment 114 (2010) 35–49

Contents lists available at ScienceDirect

Remote Sensing of Environment

j ourna l homepage: www.e lsev ie r.com/ locate / rse

Web-enabled Landsat Data (WELD): Landsat ETM+ composited mosaics of theconterminous United States

David P. Roy a,⁎, Junchang Ju a, Kristi Kline b, Pasquale L. Scaramuzza b, Valeriy Kovalskyy a,Matthew Hansen a, Thomas R. Loveland b, Eric Vermote c, Chunsun Zhang a

a Geographic Information Science Center of Excellence, South Dakota State University, Brookings, SD 57007, USAb U.S. Geological Survey Center for Earth Resources Observation and Science, 47914 252nd Street, Sioux Falls, SD 57198, USAc Department of Geography, University of Maryland, 1113 LeFrak Hall, College Park, MD 20742, USA

⁎ Corresponding author.E-mail address: [email protected] (D.P. Roy).

0034-4257/$ – see front matter © 2009 Elsevier Inc. Aldoi:10.1016/j.rse.2009.08.011

a b s t r a c t
a r t i c l e i n f o
Article history:Received 21 May 2009Received in revised form 1 August 2009Accepted 7 August 2009

Keywords:Landsat ETM+Time seriesFree dataContinentalLong-term data recordPhenologyCompositeMosaic

Since January 2008, the U.S. Department of Interior / U.S. Geological Survey have been providing free terrain-corrected (Level 1T) Landsat Enhanced Thematic Mapper Plus (ETM+) data via the Internet, currently foracquisitions with less than 40% cloud cover. With this rich dataset, temporally composited, mosaics of theconterminous United States (CONUS) were generated on a monthly, seasonal, and annual basis using 6521ETM+ acquisitions from December 2007 to November 2008. The composited mosaics are designed toprovide consistent Landsat data that can be used to derive land cover and geo-physical and bio-physicalproducts for detailed regional assessments of land-cover dynamics and to study Earth system functioning.The data layers in the composited mosaics are defined at 30 m and include top of atmosphere (TOA)reflectance, TOA brightness temperature, TOA normalized difference vegetation index (NDVI), the date eachcomposited pixel was acquired on, per-band radiometric saturation status, cloud mask values, and thenumber of acquisitions considered in the compositing period. Reduced spatial resolution browse imagery,and top of atmosphere 30 m reflectance time series extracted from the monthly composites, capture theexpected land surface phenological change, and illustrate the potential of the composited mosaic data forterrestrial monitoring at high spatial resolution. The composited mosaics are available in 501 tiles of5000×5000 30 m pixels in the Albers equal area projection and are downloadable at http://landsat.usgs.gov/WELD.php. The research described in this paper demonstrates the potential of Landsat data processing toprovide a consistent, long-term, large-area, data record.

© 2009 Elsevier Inc. All rights reserved.

1. Introduction

The Landsat satellite series, operated by the U.S. Department ofInterior / U.S. Geological Survey (USGS) Landsat project, with satellitedevelopment and launches supported by the National Aeronauticsand Space Administration (NASA), represents the longest temporalrecord of space-based land observations (Williams et al., 2006). Untilrecently the primary limitations to using Landsat data have been thecost and availability of data, which have precluded continental toglobal scale Landsat studies (Hansen et al., 2008). In January 2008,NASA and the USGS implemented a free Landsat Data DistributionPolicy that provides Level 1 terrain corrected data for the entire U.S.Landsat archive, over 2.2 million globally distributed Landsat acquisi-tions, at no cost via the Internet. Landsat acquisitions with a cloudcover of less than or equal to 40% are processed and made freelyavailable as they are acquired, and users may request any other scene

l rights reserved.

in the U.S. Landsat archive to be processed and made available via theInternet at no cost. Free Landsat data will enable reconstruction of thehistory of the Earth's land surface back to 1972, with appropriatespatial resolution to enable chronicling of both anthropogenic andnatural changes (Townshend & Justice, 1988), during a time when thehuman population has doubled and the impacts of climate changehave become noticeable (Woodcock et al., 2008).

The Landsat 7 Enhanced Thematic Mapper Plus (ETM+) is themostrecent in a series of Landsat sensors that acquire high spatial resolutionmulti-spectral data over an approximately 183 km×170 km extent,defined inaWorldwideReferenceSystemof path (groundtrackparallel)and row (latitude parallel) coordinates, with a 16 day revisit capability(Williams et al., 2006). Every Landsat overpass of the conterminousUnited States (CONUS) is acquired by theU.S. Landsat Project, providing22 or 23 acquisitions per year per path/row (Ju & Roy, 2008). TheLandsat project does not acquire every acquisition globally due toground system processing and archiving constraints (Arvidson et al.,2006). Cloud cover reduces the number of Landsat surface observations;for example, the annualmeanLandsat ETM+cloud cover for theCONUSand global scenes stored in theU.S. Landsat archive is about 40% and35%

http://landsat.usgs.gov/WELD.php


mailto:[email protected]

http://dx.doi.org/10.1016/j.rse.2009.08.011

http://www.sciencedirect.com/science/journal/00344257

36 D.P. Roy et al. / Remote Sensing of Environment 114 (2010) 35–49

respectively (Ju & Roy, 2008). In addition, in May 2003 the ETM+scan line corrector (SLC) failed, reducing the usable data in each LandsatETM+ SLC-off scene by 22% (Storey et al., 2005; Maxwell et al., 2007).

Arguably, the utility of Landsat data for long-termand/or large-areamonitoring has not been fully assessed; to date, the majority of thedata in the U.S. Landsat archive have not been used in applicationsscience. A number of regional, continental and global Landsat data setshave been generated however. Regional mosaics of Landsat imageryare increasingly being developed to meet national monitoring andreporting needs across land-use and resource sectors, for example,in Canada (Wulder et al., 2002) and the Congo basin (Hansen et al.,2008). Large volume Landsat processingwas developed by the LandsatEcosystem Disturbance Adaptive Processing System (LEDAPS) thatprocessed over 2100 Landsat Thematic Mapper and ETM+ acquisi-tions to provide wall-to-wall surface reflectance coverage for NorthAmerica for the 1990s and 2000s (Masek et al., 2006). More recently, aLandsat mosaic of Antarctica was generated from nearly 1100 LandsatETM+ austral summer acquisitions (Bindschadler et al., 2008). Atglobal scale, the Global Land Survey decadal Landsat data set providesrelatively cloud-free acquisitions selected for each path/row fromthe 1970s, 1990s and 2000s (Tucker et al., 2004). These data sets arecomposed of single date manually selected Landsat acquisitions. Withthe advent of free Landsat data it becomes feasible to apply temporalcompositing approaches tomulti-temporal Landsat acquisitions of thesamepath/row. Compositingprocedures are applied independently ona per-pixel basis to gridded satellite time series data and providea practical way to reduce cloud and aerosol contamination, fill missingvalues, and reduce the data volume of moderate resolution globalnear-daily coverage satellite data (Holben, 1986; Cihlar et al., 1994).Thus, instead of spatiallymosaicing select relatively cloud-free Landsatacquisitions together (Zobrist et al., 1983), all the available multi-temporal acquisitions may be considered, and at each gridded pixelthe acquisition that satisfies some compositing criteria selected. Inthis way, the Global Land Survey (GLS) 2005 Landsat ETM+ data setis generated by compositing up to three circa 2005 low cloud coveracquisitions per path/row (Gutman et al., 2008). Recently, Lindquistet al. (2008) examined the suitability of the GLS data sets compared tomore data intensive Landsat compositing methods (Hansen et al.,2008) and showed that over the Congo Basin compositing an in-creasing number of acquisitions reduced the percentage of SLC-offgaps and pixels with a high likelihood of cloud, haze or shadow.Similar observations have been observed for compositing moderateand coarse spatial resolution satellite data (Holben, 1986, Cihlar et al.,1994, Roy, 2000).

In this paper we describe the generation of Landsat ETM+composited mosaics of the CONUS for December 2007 to November2008. The composited mosaics are designed to provide consistentLandsat data that can be used to derive land cover and geo-physical andbio-physical products needed for detailed regional assessments of land-cover dynamics and to study Earth system functioning (Gutman et al.,2008; Wulder et al., 2008). The composited mosaics are generated on amonthly, seasonal, and annual basis to provide data that capturetemporal surface variations. The compositing approach is designed topreferentially select valid land surface observationswithminimal cloud,snow, and atmospheric contamination; consequently the compositedmosaics are not appropriate for Landsat studies of cloud, snow or theatmosphere. The processing steps and data products are informed byour MODIS Land processing, quality assessment and validationexperience (Justice et al., 2002). The processing approach is intention-ally designed to facilitate automated processing with minimal humanintervention, including the requirement for composited mosaics to beupdated regardless of the chronological order of the Landsat acquisitionand processing dates, and to provide processing in near-real time i.e.,updating composited mosaics shortly after the Landsat data areacquired. Informationonhow toobtain the compositedmosaic productsand further research is described.

2. Input web-enabled level 1T Landsat ETM+ data

The Landsat data made freely available by the U.S. Landsat projectare sometimes called “web-enabled” data as they are made availablevia the Internet. These data are nominally processed as Level 1 terrain-corrected (L1T) data. The L1T data are available in GeoTIFF format inthe Universal Transverse Mercator (UTM) map projection with WorldGeodetic System 84 (WGS84) datum which is compatible withheritage GLS and Landsat MSS data sets (Tucker et al., 2004). TheLevel 1T processing includes radiometric correction, systematicgeometric correction, precision correction using ground controlchips, and the use of a digital elevation model to correct parallaxerror due to local topographic relief. The L1T geolocation error in theCONUS is less than 30 m even in areas with substantial terrain relief(Lee et al., 2004).Whilemost web-enabled Landsat data are processedas L1T (i.e., precision and terrain-corrected), certain acquisitions donot have sufficient ground control or elevation data necessary forprecision or terrain correction, respectively. In these cases, the bestlevel of correction is applied and, over the CONUS, the data areprocessed to Level 1G systematic (L1G) (WWW1). The product filemetadata notes if the acquisition was processed to L1T or L1G.

Landsat acquisitions with cloud cover less than or equal to 40% areprocessed and made freely available as they are acquired, and usersmay request any other scene in the U.S. Landsat archive to beprocessed and made available at no cost via the Internet. Initially, theU.S. Landsat project implemented a 20% cloud cover threshold but thiswas increased to 40% in October 2008 after a processing load analysisand in response to requests from the user community including fromthis research project. In, addition, the U.S. Landsat project bulkprocessed all the CONUS ETM+ acquisitions with less than 40% cloudcover acquired from May 2007 to October 2008 in support of thisproject. The cloud cover of each ETM+ acquisition is estimatedoperationally by the automatic cloud cover assessment (ACCA)algorithm during the archiving process (Irish et al., 2006).

Fig. 1 shows the CONUS study area, defined by 459 Landsat path/row coordinates, covering about 11,000,000,000 30 m land pixels. Allthe web-enabled Landsat ETM+ data acquired over the study area fora 12 month period from December 2007 to November 2008 wereobtained. A total of 6959 acquisitions were copied by dedicated filetransfer protocol from the U.S Landsat project to computers at theGeographic Information Science Center of Excellence, South DakotaState University. Of the 6959 acquisitions, 438 (6.3%) were found to beprocessed as L1G and not as L1T. Fig. 2 shows histograms of the ACCAcloud cover percentage for the 6521 L1T (left) and the 438 L1G (right)acquisitions. The acquisitions that were processed as L1G havemarkedly higher cloud cover (mean 61%) compared to thoseprocessed as L1T (mean 18%). This is expected as higher cloud coverreduces the number of available ground control chips needed in theL1T processing, although a minority of acquisitions processed as L1Talso have high cloud cover. The locations of acquisitions that had lessthan 20% ACCA cloud cover but that were processed to L1G areillustrated in Fig. 1 (black filled circles). For all 105 of these theavailability of ground control chips and/or the reliability of the groundcontrol matching algorithm was compromised for reasons other thancloud cover: 100 acquisitions were of 10 coastal water path/rowlocations where land ground control was less available, another (path30, row 27) was extensively contaminated by snow reducing theavailability of ground control, and four were at the same location inKansas (path 30, row 34) where there is extensive center-pivotirrigation (Wardlow et al., 2007) that are thought to confuse thematching algorithm.

Table 1 summarizes the number, ACCA cloud cover, and seasonaldistribution of acquisitions in the U.S. Landsat project archive. Thenumber of acquisitions in the archive is about the same in each season(Table 1, column 2), but because of the higher cloud cover in winterand spring, fewer acquisitions were available with less than 40% ACCA

Fig. 1. The Landsat ETM+ path and row coordinates (circles) defining the conterminous United States (CONUS) study area. The black filled circles show path/row coordinates where105 web-enabled ETM+ acquisitions over the 12 month study period had less than 20% ACCA cloud cover but could not be processed as L1T and so were processed as L1G by the U.SLandsat project.

37D.P. Roy et al. / Remote Sensing of Environment 114 (2010) 35–49

cloud cover than in summer and autumn (Table 1, column 3). All theavailableweb-enabled acquisitions processed to L1Twere used for theresearch (Table 1, column 4). The 438 acquisitions processed to L1Gwere discarded and only the remaining 6521 acquisitions processed

Fig. 2. Histograms of the ACCA cloud cover percentage for the 6521 L1T (left) and the 438 L1from the U.S. Landsat project.

to L1T were used in order to maximize the Landsat geolocation and socompositing accuracy (Roy, 2000). A total of 734 L1T acquisitions hadhigher than 40% ACCA cloud cover reflecting high cloud coveracquisitions ordered by the user community. The high degree of

G (right) ETM+ December 2007 to November 2008 web-enabled acquisitions obtained

Table 1The availability of Landsat ETM+ L1T data over the CONUS study area (Fig. 1) for winter(December 2007, January, February 2008), spring (March, April, May, 2008), summer(June, July, August, 2008), and autumn (September, October, November, 2008).

Number ofacquisitions in theU.S. Landsat Projectarchive. Mean cloudcover percentageshown inparenthesis

Number of acquisitionswith ≤40% cloudcover. Mean cloudcover percentageshown in parenthesis

Number of acquisitionsused for the researchreported in this paper.Mean cloud coverpercentage shown inparenthesis

Winter 2569 (53%) 1067 (13%) 1379 (25%)Spring 2605 (40%) 1467 (11%) 1639 (17%)Summer 2614 (32%) 1736 (11%) 1851 (15%)Autumn 2531 (36%) 1517 (10%) 1652 (15%)Annual 10,319 (40%) 5787 (11%) 6521 (18%)

The cloud cover summary statistics are derived from the ACCA reported cloud coverpercentages.


cloud contamination in some of the L1T acquisitions was notconsidered important however because it was expected that cloudypixels would be removed by the compositing process.

3. Top of atmosphere reflectance, brightness temperature,vegetation index, and band saturation computation

The spectral radiance sensed by each ETM+ detector is stored asan 8-bit digital number (Markham et al., 2006). The digital numbersshould be converted to radiance (units: W m−2sr−1μm−1), tominimize changes in the instrument radiometric calibration, andthen converted to top of atmosphere reflectance to minimize remotesensing variations introduced by variations in the sun–earth distance,the solar geometry, and exoatmospheric solar irradiance arising fromspectral band differences (Chander et al., 2009). This is particularlyimportant for applications that use Landsat data acquired over largeareas or long time periods. The conversion of the radiance sensed atthe Landsat reflective and thermal wavelengths to reflectance(unitless) and brightness temperature (units: Kelvin) respectivelyprovides data that has physical meaning and, for example, can becompared with laboratory and ground based measurements, modeloutputs, and data from other satellite sensors (Masek et al., 2006), andimportantly provides data that can be used to derive geo-physical andbio-physical products (Justice et al., 2002).

In this research all the Landsat ETM+ bands were used, except thepanchromatic band, i.e., the bands used were the 30 m blue (0.45–0.52 μm), green (0.53–0.61 μm), red (0.63–0.69 μm), near-infrared(0.78–0.90 μm), and the two mid-infrared (1.55–1.75 μm and 2.09–2.35 μm) bands, and the 60 m thermal (10.40–12.50 μm) low and highgain bands. The L1T 8-bit digital numbers were converted to spectralradiance using sensor calibration gain and bias coefficients derivedfrom the ETM+ L1T file metadata (Chander et al., 2009). The radiancesensed in the Landsat reflective wavelength bands, i.e., the blue,green, red, near-infrared, and the two mid-infrared bands, wereconverted to top of atmosphere reflectance using the standardformula as:

ρλ =π⋅Lλ⋅d2

ESUNλ⋅cosθsð1Þ

where ρλ is the top of atmosphere (TOA) reflectance (unitless), Lλ isthe TOA spectral radiance (Wm−2 sr−1 μm−1), d is the Earth–Sundistance (astronomical units), ESUNλ is the mean TOA solar spectralirradiance (Wm−2 μm−1), and θs is solar zenith angle at the center ofthe Landsat acquisition (radians). The quantities ESUNλ and d aretabulated by Chander et al. (2009) and θs is calculated from the solarelevation angle stored in the ETM+ L1T file metadata. The TOAreflectance computed as [1] is the TOA bi-directional reflectance

factor and can be greater than 1, for example, due to specularreflectance over snow or water under certain solar and viewinggeometries (Schaepman-Strub et al., 2006). In addition, due toinstrument artifacts not accommodated for by the calibration, theretrieved TOA reflectance can be negative, for example, over waterbodies. The 30 m TOA reflectance for each reflective band were storedas signed 16-bit integers after being scaled by 10,000, in the samemanner as the MODIS surface reflectance product (Vermote et al.,2002).

The normalized difference vegetation index (NDVI) is a commonlyused vegetation index, derived as the near-infrared minus the redreflectance divided by their sum (Tucker, 1979), and is used inmaximum NDVI compositing to preferentially select pixels withreduced cloud and atmospheric contamination (Holben, 1986). The30 m TOA NDVI was computed from the TOA red and near-infraredLandsat reflectance and stored as signed 16-bit integers after beingscaled by 10,000, in the same manner as the MODIS NDVI product(Huete et al., 2002).

The radiance sensed in the Landsat low and high gain thermalbands were converted to TOA brightness temperature (i.e., assumingunit surface emissivity) using the standard formula as:

T =K2

logðK1 = Lλ + 1Þ ð2Þ

where T is the 10.40–12.50 μm TOA brightness temperature(Kelvin), K1 and K2 are thermal calibration constants set as 666.09(Wm−2 sr−1 μm−1) and 1282.71 (Kelvin) respectively (Chander etal., 2009), and Lλ is the TOA spectral radiance. This equation is aninverted Planck function simplified for the ETM+ sensor consider-ing the thermal band spectral responses. For convenience for thesubsequent processing, and subsequent product use, the two 60 mbrightness temperature thermal bands were resampled to 30 m toprovide the same image spatial dimensions as the 30 m TOAreflectance bands. The upper-left 60 m L1T pixels are co-registeredat the pixel centers (WWW2). Consequently, the even-numberedrow and column 30 m pixels fall within the 60 m pixels and so wereassigned the 60 m pixel brightness temperature values. The odd-numbered 30 m pixels have centers on the boundaries of twoneighboring 60 m pixels, and their pixel values were systematicallyobtained from the nearest 60 m pixel with a smaller row and/orcolumn number. The 30 m low and high gain TOA brightnesstemperature data were stored as signed 16-bit integers with units ofdegrees Celsius by subtracting 273.15 from the brightness temper-ature and then scaling by 100.

The Landsat ETM+ calibration coefficients are configured in anattempt to globally maximize the range of land surface spectralradiance in each spectral band (Markham et al., 2006). However,highly reflective surfaces, such as snow and clouds, may over-saturatethe reflective wavelength bands, with saturation varying spectrallyand with the illumination geometry (solar zenith and surface slope)(Cahalan et al., 2001; Dowdeswell & Mcintyre, 1986; Bindschadleret al., 2008). Similarly, hot surfaces may over-saturate the thermalbands (Flynn & Mouginis-Mark, 1995), and cold surfaces may under-saturate the high gain thermal band (WWW3). Over and under-saturated pixels are designated by digital numbers of 255 and 1respectively in the L1T data. As the radiance values of saturated pixelsare unreliable, a 30 m 8-bit saturationmask was generated, storing bitpacked band saturation (1) or unsaturated (0) values for the eightLandsat bands used in this study.

4. Cloud masking

It is well established that optically thick clouds preclude opticaland thermal wavelength remote sensing of the land surface but thatautomated and reliable satellite data cloud detection is not trivial


(Kaufman, 1987; Platnick et al., 2003). Recognizing that clouddetection errors, both of omission and commission, will alwaysoccur in large data sets, both the Landsat automatic cloud coverassessment (ACCA) algorithm and a classification tree based clouddetection approach were implemented.

4.1. Automatic cloud cover assessment (ACCA) cloud detection

The U.S. Landsat project uses the ACCA algorithm to estimate thecloud content of each acquisition (Irish, 2000; Irish et al., 2006). Thepercentage of pixels that are flagged as cloudy are stored as metadata(e.g., Fig. 2). The ACCA takes advantage of known spectral propertiesof clouds, snow, bright soil, vegetation, and water, and consists oftwenty-six filters/rules applied to five of the ETM+ bands (Irish et al.,2006). The primary goal of the algorithm is to quickly produce scene-average cloud cover metadata values, that can be used in futureacquisition planning (Arvidson et al., 2006), and that users may queryas part of the Landsat browse and order process. The ACCA was notdeveloped to produce a “per-pixel” cloudmask; despite this, the ACCAhas an estimated 5% error for 98% of the global 2001 ETM+acquisitions archived by the U.S. Landsat project (Irish et al, 2006).

A copy of the ACCA code was obtained and applied to everyLandsat ETM+ acquisition to produce a 30 m per-pixel cloud datalayer, stored as an unsigned 8-bit integer where 0=not cloudy,1=cloudy, and 255=fill value denoting missing data in any of theLandsat bands.

4.2. Classification tree cloud detection

The state of the practice for automated satellite land-coverclassification is to adopt a supervised classification approach wherea sample of locations of known land-cover classes (training data) iscollected. The optical and thermal wavelength values sensed at thelocations of the training pixels are used to develop statisticalclassification rules, which are then used to map the land-cover classof every pixel. Supervised classification tree approaches have beenapplied to Landsat data to identify cloud contaminated pixels (Hansenet al., 2008). Classification trees are hierarchical classifiers that predictcategorical class membership by recursively partitioning data intomore homogeneous subsets, referred to as nodes (Breiman et al.,1984). They accommodate abrupt, non-monotonic and non-linearrelationships between the independent and dependent variables, andmake no assumptions concerning the statistical distribution of thedata (Prasad et al., 2006). Bagging tree approaches use a statisticalbootstrapping methodology, whereby a large number of trees aregrown using random subsets of the training data, to reduce over-fitting and improve predictive ability (Breiman, 1996). Conventionallymultiple bagged trees are used to independently classify the satellitedata and the multiple classifications are combined using some votingprocedure. In this studywe developed a single parsimonious tree frommultiple bagged trees so that only one tree was used to classify theLandsat data, reducing the computational overhead.

Supervised classification approaches require training data. A globaldatabase of Landsat Level 1G and corresponding spatially explicitcloud masks generated by photo-interpretation of the reflective andthermal bands were used. This database was developed to prototypethe cloud mask algorithm for the future Landsat Data ContinuityMission (Irons & Masek, 2006). The Landsat interpreted cloud maskdefines each pixel as thick cloud, thin cloud, cloud shadow or notcloudy. We reconciled these interpreted cloud states into cloud (i.e.,thick and thin cloud) and non-cloud (i.e. cloud shadow and not cloudystates) classes. In addition, to avoid mixed pixel cloud edge problems,the cloud labeled regions were morphologically eroded by one 30 mpixel and not used (Serra, 1982). A 0.5% sample of training pixels wasextracted randomly from each Landsat scene, where data werepresent and not including the cloud boundary regions. A total of 88

northern hemisphere Landsat scenes acquired in polar (19 acquisi-tions), boreal (22 acquisitions), mid-latitude (24 acquisitions) andsub-tropical latitudinal zones (23 acquisitions) were sampled. TheLandsat Level 1G data were processed to TOA reflectance, brightnesstemperature and the band saturation flag computed as described inSection 3. Only pixels with reflectance greater than 0.0 were used. Atotal of 12,979,302 unsaturated training pixels and 5,374,157saturated training pixels were extracted.

After some experimentation it became apparent that it was neces-sary to develop two classification trees; one for saturated training dataand the other for the unsaturated training data. Saturation occurs inany of the reflective ETM+ bands, and also for the high gain thermalwavelength band which may over- or under-saturate typically whenpixels contain illuminated thick cumulonimbus cloud, illuminatedsnow, or sunglint over water. The saturated TOA reflectance andbrightness temperature values were unreliable but still providedinformation that could be classified. Consequently, better cloud non-cloud discrimination was afforded by classifying the saturated andunsaturated pixels independently. For both the saturated andunsaturated classification trees, all the 30 m TOA reflective bandswere used, except the shortest wavelength blue band which is highlysensitive to atmospheric scattering (Ouaidrari & Vermote, 1999). Forboth trees, the low gain thermal band was also used. The high gainthermal band was not used because it under-saturated or over-saturated frequently over the wide surface brightness temperaturerange of the CONUS e.g., from winter snow covered surfaces to hotsummer desert surfaces. The unsaturated classification tree also usedreflective band simple ratios similar to those used by ACCA. Thesaturated classification tree did not use band ratios as they could notbe computed when one or both bands in the ratio formulation weresaturated.

Twenty five bagged classification trees were generated, runningthe Splus tree code on a 64 bit computer, each time, 20% of the trainingdata were sampled at randomwith replacement and used to generatea tree. Each tree was used to classify the remaining (“out-of-bag”) 80%of the training data, deriving a vector of predicted classes for each out-of-bag pixel. In this way, each training pixel was classified 25 or fewertimes. The most frequent predicted class (cloud or non-cloud) foreach training pixel was derived; and used with the correspondingtraining data to generate a single final tree, i.e. the final tree wasgenerated using approximately 25×0.8×n training pixels, where nwas either the 12,979,302 unsaturated training pixels or the 5,374,157saturated training pixels. To limit over-fitting, all the trees wereterminated using a deviance threshold, whereby additional splits inthe tree had to exceed 0.02% of the root node deviance or tree growthwas terminated. The final unsaturated and saturated classificationtree was then used to classify every Landsat pixel according to itssaturation status.

The 30 m cloud classification result was stored as an unsigned 8-bitintegerwhere0=not cloudy, 1=cloudy, 2=not cloudy but adjacent toa cloudy pixel, 200=could not be classified because of negative reflec-tance, and 255=fill value denoting missing data in any of the Landsatbands.

5. Reprojection, resampling and tiling

The L1T data are defined in the UTM projection which is defined inzones, each 6° of longitude in width and centered over a meridian oflongitude, with ten zones encompassing the CONUS (Snyder, 1993).Consequently, after each Landsat ETM+ L1T acquisition was pro-cessed (Sections 3 to 4), the 30 m TOA reflective bands, TOA NDVI,TOA brightness temperature bands, band saturation mask, and thetwo cloudmasks were reprojected from the L1T UTM coordinates intoa single continental map projection. The high Landsat L1T data volumerestricts provision of multiple product instances in different mapprojections, even though users will inevitably prefer different

Table 2Compositing criteria used to compare two acquisitions of a pixel, the criteria areconsidered in decreasing priority until a selection is made (see text for further details).

Priority Compositing criteria

1 If either fill: select non-fill2 If either saturated: select unsaturated3 If both saturated: select the one with greatest brightness temperature4 If one cloudy and one non-cloudy: select non-cloudy5 If one cloudy and one uncertain cloud: select uncertain cloud if it has

greater brightness temperature or greater NDVI, else select cloudy6 If one non-cloudy and one uncertain cloud: select non-cloud if it has greater

brightness temperature or greater NDVI, else select uncertain cloud7 If either “unvegetated”: select the one with greatest brightness temperature8 Select the one with greatest NDVI


projections (Teillet et al., 2000). The Albers Equal Area projection wasselected as it is suitable for large areas that are mainly east–westoriented such as the CONUS (Snyder, 1993), and was defined withstandard parallels 29.5°N, 45.5°N and central Meridian 96°W toprovide heritage with the USGS EROS National Land Cover Database(Homer et al., 2004; Chander et al., 2009a). As there are about11,000,000,000 30 m Landsat pixels covering the CONUS study area itis not physically possible to store the reprojected Landsat 30 m data ina single file. The largest file size achievable is limited by the amount ofaddressable memory on a user's personal computer, usually conser-vatively considered to be a 32 bit computer with a maximum file sizeof 2 GB. To ensure manageable file sizes, the 30 m Landsat data werereprojected into 501 fixed Albers tiles where each tile was composedof 5000×5000 30 m Landsat pixels. This tile dimension (number ofrows and columns) is smaller than the dimensions of individual L1TETM+ acquisitions.

The Landsat ETM+ pixels were allocated to the Albers coordinatesystem using the inverse gridding approach, sometimes known as theindirect method (Konecny, 1976). In this approach the centercoordinates of each Albers 30 m pixel are mapped to the nearestpixel center in the Landsat data, and the ETM+processed data for thatpixel are allocated to the Albers output grid. This processing approachis computationally efficient and geometrically is the equivalent ofnearest neighbor resampling (Wolfe et al., 1998). The GeneralCartographic Transformation Package (GCTP) developed by theUSGS and used to develop a number of applications including theMODIS global browse imagery (Roy et al., 2002) and the MODISReprojection Tool (WWW4) was used to transform coordinatesbetween the UTM and Albers map projections. The GCTP iscomputationally expensive. Consequently, a sparse triangulationmethodology was used where the GCTP was invoked to projectAlbers 30 m pixels to UTM coordinates only at the vertices of triangles,and Albers 30 m pixel locations falling within the triangles wereprojected to UTM coordinates using a simplicial coordinate transfor-mation (Saalfeld, 1985). In this approach, any point (px, py) in atriangle with vertices (x1, y1), (x2, y2), (x3, y3) can be represented bythree simplicial coordinates (s1, s2, s3) defined as:

s1 = a1py + b1px + c1s2 = a2py + b2px + c2s3 = 1−s1−s2

ð3Þ

where:

a1 = ðx3−x2Þ= t a2 = ðx1−x3Þ= tb1 = ðy2−y3Þ= t b2 = ðy3−y1Þ= tc1 = ðx2y3−x3y2Þ = t c2 = ðx3y1−x1y3Þ= tt = x1y2 + x2y3 + x3y1−x3y2−x2y1–x1y3

Given apoint (px,py) defined inAlbers coordinates the correspondinglocation in UTM coordinates is:

p0x = s1x01 + s2x

02 + s3x

03

p0y = s1y01 + s2y

02 + s3y

03

ð4Þ

where (x1′, y1′), (x2′, y2′), (x3′, y3′) are the coordinates of the trianglevertices in UTM calculated by projecting the corresponding Alberstriangle vertices (x1, y1), (x2, y2), (x3, y3) using the GCTP. A regularlattice of triangles was defined by bisecting squares with side lengthsof 450 m (i.e., fifteen 30 m pixels) defined from the northwest originof the Albers coordinate system so there in each square there weretwo triangles with different topologies. This approach is computa-tionally efficient as the GCTP is only called for each triangle vertex andthe coefficients a, b, c and t are computed only once for each triangle.The maximum coordinate difference between the simplical interpo-lation and GCTP projected coordinates, occurs at the triangle centers,

and for the CONUS study area is less than 1 cm east–west and north–south.

6. Compositing and composited mosaic format

Compositing was developed originally to reduce residual cloudand aerosol contamination in AVHRR time series to producerepresentative n-day data sets (Holben, 1986). Compositing proce-dures either select from colocated pixels in different orbits ofgeometrically registered data the pixel that best satisfies somecompositing criteria, or combine the different pixel values together.Compositing criteria have included the maximum NDVI, maximumbrightness temperature, maximum apparent surface temperature,maximum difference in red and near-infrared reflectance, minimumscan angle, and combinations of these (Roy, 2000). Ideally, the criteriashould select from the time series only near-nadir observations thathave reduced cloud and atmospheric contamination. Compositesgenerated from wide field of view satellite data, such as AVHRR orMODIS, often contain significant bi-directional reflectance effectscaused by angular sensing and illumination variations combined withthe anisotropy of reflectance of most natural surfaces and theatmosphere (Gao et al., 2002; Roy et al., 2006). Compositingalgorithms that model the bi-directional reflectance have beendeveloped to compensate for this problem and combine all validobservations to estimate the reflectance at nadir view zenith for someconsistent solar zenith angle (Roujean et al., 1992; Schaaf et al., 2002).This approach does not provide a solution for compositing thermalwavelength satellite data, and is not appropriate for application toLandsat data as the comparatively infrequent 16 day Landsat repeatcycle and the narrow 15º Landsat sensor field of view do not provide asufficient number or angular sampling of the surface to invert bi-directional reflectance models (Danaher et al., 2001; Roy et al., 2008).Consequently, in this research, compositing based on the selection of a“best” pixel over the compositing period was implemented.

Composited mosaics of the CONUS were generated on a monthly,seasonal and annual basis for December 2007 to November 2008. Thecompositing criteria were focused on selecting valid surface observa-tions with minimal cloud, snow and atmospheric contamination.Initially, a simple selection over the compositing period of interest(month, season, year) of pixels that were not missing due to temporalLandsat acquisition gaps or SLC-off gaps, and that were not flagged ascloudy by the ACCA or classification tree cloud masking algorithmswas attempted. This did not provide visually coherent composites,due to the incidence of saturated pixels, pixels that were contam-inated by atmospheric aerosols, pixels that contained sub-pixel cloudand cloud edges, and the unresolved problem of how to select fromseveral observations of the same pixel that were not missing and notflagged as cloudy. Consequently, the compositing approach wasrefined and incorporated the heritage maximum NDVI and maximumbrightness temperature compositing criteria as clouds, snow and

Table 3Composited mosaic product format.

SDS name Datatype

Valid range Scalefactor

Units Fill value Notes

Band1_TOA_REF int16 −32,767–32,767 10,000 Unitless −32,768 Top of atmosphere (TOA) reflectance, computed using standard formulae and calibrationcoefficients associated with sensed L1T dataBand2_TOA_REF int16 −32,767–32,767 10,000 Unitless −32,768

Band3_TOA_REF int16 −32,767–32,767 10,000 Unitless −32,768Band4_TOA_REF int16 −32,767–32,767 10,000 Unitless −32,768Band5_TOA_REF int16 −32,767–32,767 10,000 Unitless −32,768Band61_TOA_BT int16 −32,767–32,767 100 Celsius −32,768Band62_TOA_BT int16 −32,767–32,767 100 Celsius −32,768Band7_TOA_REF int16 −32,767–32,767 10,000 Unitless −32,768NDVI_TOA int16 −10,000–10,000 10,000 Unitless −32,768 NDVI value.Day_Of_Year int16 1–366 1 Unitless −32,768 Day of year the selected ETM+ pixel was sensed on.Saturation_Flag uint8 0–255 1 Unitless NA Bit packed for the 8 ETM+ bands used. The least significant bit to the most significant bit

corresponds to bands 1, 2, 3, 4, 5, 61, 62, and 7 with a bit set to signifying saturation in that band.Classification TreeCloud_State

uint8 0, 1, 2, 200 1 Unitless 255 Classification Tree Cloud Classification, 0=not cloudy, 1=cloudy, 2=not cloudy but adjacent to acloudy pixel, 200=could not be classified reliably due to negative reflectance.

ACCA_State uint8 0, 1 1 Unitless 255 ACCA Cloud Classification, 0=not cloudy, 1=cloudyNum_Of_Obs uint8 0–255 1 Unitless NA Number of ETM+ observations considered over the compositing period.


aerosols typically depress NDVI and brightness temperature over landsurfaces (Holben, 1986; Cihlar et al., 1994; Roy, 1997).

ThemaximumNDVI compositingcriterionwas selectedas theprimarycompositing criterion, rather than the maximum brightness temperaturecriterion, because among cloud-free observations it preferentially selectsvegetated observations and arguably the majority of terrestrial Landsatapplications are concerned with vegetation. We observed however thatfor certain low vegetation covers, including certain dark and bright soils,water and snow, that the top of atmosphere (TOA) NDVI of a cloud couldbe higher than the TOA NDVI of the cloud-free surface. The sensitivity ofNDVI to the brightness of soil beneath vegetation canopies (Huete, 1988)and to atmospheric effects (Liu & Huete, 1995) is well established.Consequently, a pixel was considered as “unvegetated” if NDVI<0.09AND 2.09–2.35 μm TOA reflectance<0.048. These two thresholds werederivedempirically.When therewere “unvegetated”pixels themaximumbrightness temperature compositing criterion was used, as cloudbrightness temperatures tended to be lower than the cloud-freebrightness temperature. The low gain 10.40–12.50 μm TOA brightnesstemperature was used as it has a wider range than the high gain TOAbrightness temperature and does not saturate (Chander et al., 2009). Forcompositing purposes only, a pixel was considered saturated if either thered or near-infrared TOA reflectance was saturated as the NDVI isunreliable when one or both of these bands are saturated.

The ACCA and classification tree cloud masks were used tocomplement the maximum NDVI and maximum brightness temper-ature criteria and to provide a more reliable differentiation between

Fig. 3. Scatterplots of Classification Tree and corresponding ACCA cloud percentages derivedlinear regression lines are shown superimposed; Left: y=4.94+0.7328 x (R2=0.66), Right:to extract the example subsets illustrated in Fig. 4.

clouds and the land surface. These two cloud masks did not alwaysagree, but it was not possible to quantitatively evaluate their relativeomission and commission errors as a function of different clouds andbackground reflectance and brightness temperature. Consequently, apixel was considered as cloudy if both the ACCA and the ClassificationTree algorithms detected it as cloud, as uncertain cloud if only onealgorithm detected it as cloud, or as non-cloudy if both algorithmsdetected it as non-cloud.

Table 2 summaries the compositing logic; each row reflects acomparison of two acquisitions of the same pixel. If the criterion in arow is not met then the criterion in the row beneath is used and thisprocess is repeated until the last row. This implementation enablesthe composites to be updated on a per-pixel basis shortly after theinput ETM+ data are processed and regardless of the chronologicalprocessing order. For example, after 16 days the same Albers pixellocationmay be sensed again, and the compositing criteria are used todecide if the more recent ETM+ pixel data should be allocated tooverwrite the previous data. For each composited Albers pixel, the30 m TOA reflective bands, TOA NDVI, TOA brightness temperaturebands, band saturation mask, and the two cloud masks are stored. Inaddition, the day of the year that the selected pixel was acquired on,and the number of different valid acquisitions considered at that pixelover the compositing period, are stored.

Table 3 summarizes the file format of the monthly, seasonal andannual compositedmosaics. These data are stored in Hierarchical DataFormat (HDF). HDF is a data file format designed by the National

from 548 January 2008 (left) and 654 July 2008 (right) L1T ETM+ acquisitions. Simpley=−0.53+0.9025 x (R2=0.88). The large black dots denote the L1T acquisitions used


Center for Supercomputing Applications to assist users in the storageand manipulation of scientific data across diverse operating systemsand machines. For example, it is used to store the standard MODISLand products (Justice et al., 2002). Each composited mosaic pixel has14 bands (termed HDF science data sets) storing the information

Fig. 4. Landsat L1T 450×600 30 m spatial subsets (defined in UTM coordinates). The left coluand blue (0.45–0.52 µm), the middle column illustrates the ACCA and the right column illuscloud detected, black: no cloud detected, gray: fill value due to SLC-off gaps, yellow (Classificloudclassified. Top rowsubset: January132008, path20, row38 (center longitude−86.7900°, latlatitude 33.1332°), bottom row subset: January 24, path 33, row 27 (scene center longitude−10

described in Table 3 and in the text above. The different HDF sciencedata sets are stored with appropriate data types, and HDF internalcompression, to reduce the file size. A 5000×5000 tile is typically240 MB and the total size of the 501 tiles defining a monthly, seasonalor annual composited mosaic is about 110 GB.

mn illustrates the true color TOA reflectance: red (0.63–0.69 µm), green (0.53–0.61 µm)trates the Classification Tree cloud mask. Colors illustrating the cloud masks are: blue:cation Tree only): pixels with negative reflectance in any of the reflective bands and noitude31.6983°),middle rowsubset: January 7, path42, row37 (center longitude−120.3933°,1.8852°, latitude 47.4056°).


7. Browse generation

Browse images with reduced spatial resolution were generatedfrom the composited mosaics to enable synoptic product evaluationwith reduced data volume (Roy et al., 2002), and with the expectationthat the browse imagery could be used for Internet data ordering.CONUS browse images were generated in the JPEG format with fixedcontrast stretching and color look-up tables to enable consistenttemporal comparison. The browse images were generated using themedian pixel values falling in a given window size (Boschetti et al.,2008) defined with dimensions set as an integer multiple of the 30 mAlbers pixels. For example, a 17×17 30 m pixel window size producesa browse with approximately 500 m resolution. Browses weregenerated from the monthly, seasonal and annual composites usingthe same window size. Single band browses of a science data set ofinterest (Table 3) were generated using the median value of the pixelvalues falling in a given window. In addition, true color reflectancebrowseswere generated from the TOA red, green, and blue reflectancedata, selecting within each window the pixel with the median redreflectance and the corresponding blue and green reflectance valuesfor that pixel. In this way, the reflectance for the same pixel isobtained which produces a more coherent true color reflectancebrowse than selecting the median reflectance values for eachwavelength independently. The red reflectance was used as the“master” since it is less sensitive to atmospheric contamination thanshorter wavelength blue and green reflectance (Ouaidrari & Vermote,1999). In the results shown in the next section browse images werecomputed ignoring pixels flagged as cloud.

8. Results and preliminary assessment

8.1. Cloud

It is difficult to definitively validate cloud detection algorithms (Irishet al., 2006; Platnick et al., 2003). Consequently, the ACCA andClassification Tree cloud masks were compared statistically and alsoby visual inspection. The Classification Tree cloud mask was generatedusing unsaturated and saturated ETM+ L1T data; 12,979,302 unsatu-rated trainingpixelswere used to generate a single tree defined by 1595nodes that explained 98% of the tree variance, and 5,374,157 saturatedtraining pixels generated a tree of 188 nodes that explained 99.9% of thetree variance. Despite these high levels of explanation, it is recognized

Fig. 5. The locations and frequency of Landsat ETM+ L1T acquisitions used to produce the Cnumber of unique Landsat observations over the 12 month period reported in 17×17 30 m

that supervised classification approaches are only as good as the qualityof the training data and the degree to which the training data arerepresentative. Although every attempt was made to use a large welldistributed high quality training data set (Section 4.2) the suitability ofthe training data used to develop the saturated and unsaturatedclassification trees is unknown, especially considering the diverse andtemporally variable landscapes and clouds of the CONUS.

Fig. 3 shows scatterplots of the percentage of each Landsat L1Tacquisition flagged as cloudy by the classification tree and by ACCA.Each point corresponds to the cloud percentage derived by countingthe number of cloud labeled 30 m pixels in the L1T acquisition dividedby the number of non-missing 30 m pixels. The figure illustratesscatterplots for the 548 and 654 L1T CONUS acquisitions sensed inJanuary 2008 (left) and July 2008 (right) respectively. Simple linearregression lines are shown superimposed; the two cloudmasks at thisreporting scale have better correspondence in July (R2=0.88) than inJanuary (R2=0.66). For both January and July the classification treeunderestimates the CONUS cloud cover relative to ACCA (regressionline slopes of 0.73 and 0.90 respectively). It is established that ACCAmay misclassify snow as cloud, particularly under low solarillumination (Irish et al., 2006), and this may explain the poorercorrespondence between the cloud masks in January (Fig. 4, left)when a substantial proportion of the CONUS is snow covered andwhen the Landsat solar zenith angles are higher than in summer.

Scatterplots of the sort illustrated in Fig. 3 may be misleading inthat similar cloud proportions may be reported by both algorithmsbut with the clouds mapped at different locations within the LandsatL1T acquisition. As a confidence check, Fig. 4 shows three 450×60030 m spatial subsets of L1T January 2008 acquisitions. The left columnillustrates the 30 m true color TOA reflectance and the middle andright columns show the ACCA and Classification Tree cloud classifi-cation results respectively. These three subsets were extracted fromscenes with cloud cover percentages highlighted by the large blackdots in Fig. 3 (left). The top row of Fig. 4 shows a subset of anacquisition in which the ACCA and the Classification Tree cloud coveragree at the scene level (12%). The middle and bottom rows of Fig. 4illustrate subsets where the ACCA and Classification Tree scene levelcloud cover percentages are the most different among all the Januaryacquisitions.

The top row of Fig. 4 illustrates a subset in Alabama acquiredJanuary 13th 2008 with no apparent snow cover. Locally, both cloudmasks capture the general pattern of cloud evident in the TOA true

ONUS mosaics, shown in the Albers projection. The color scheme illustrates the medianpixel bins.


color reflectance and it is not possible to state which cloud mask isperforming better. The middle row illustrates a subset in Californiaacquired January 7th 2008 with isolated thick clouds and with thinclouds which completely obscure the land surface. The ACCA detectedthe thick clouds but failed to detect the thin cloud, while theClassification Tree more correctly labeled the majority of the scene ascloud. The bottom row illustrates a subset acquired over a snow-covered landscape in North Dakota acquired January 24th 2008. TheACCA misclassified the snow as cloud while the Classification Treecloud mask did not. Despite this evidence that the Classification Treemay be more reliable than the ACCA, we combined both cloud masksin the compositing procedure to provide a conservative cloudassessment used by the compositing procedure (Section 6).

8.2. Monthly, season and annual composited mosaics

Monthly, seasonal and annual composites were generated fromthe 6521 L1T acquisitions processed in the manner described in the

Fig. 6.Monthly true color TOA reflectance red (0.63–0.69 µm), green (0.53–0.61 µm) and blu30 m Landsat pixels to provide an approximate spatial resolution of 500 m (see text for de

previous sections. The standard seasonal definition adopted by theclimate modeling community is used where winter is defined by themonths: December, January and February (Ju & Roy, 2008). Fig. 5illustrates for the CONUS study area, the median temporal frequencyof Landsat L1T acquisitions in 17×17 30 m pixel bins (i.e. ~500 m)over the 12 month study period. The majority of the CONUS path/rowlocations have 10 or more acquisitions processed to L1T over the year;however, in cloudy regions including the Pacific Northwest, and theGreat Lakes, less acquisitions are available; conversely, in the arid, lesscloudy, southwest up to 55 acquisitions are available. These patternshave been observed previously (Ju & Roy, 2008). The maximumnumber of overpasses of a path/row per year is 23 but the number ofacquisitions may be greater because of overlapping acquisitions alongthe same path and between adjacent paths. Adjacent Landsat pathsare sensed with a seven day time difference and the L1T acquisitionsacquired in the same row overlap in the across-track direction.Overlapping acquisitions occur along the same path because of theway the L1T data are generated and not because the same location is

e (0.45–0.52 µm), CONUS browse images, each pixel is shown generalized from 17×17tails).

Fig. 7. Seasonal true color TOA reflectance CONUS browse images. Winter (December 2007, January, February 2008), spring (March, April, May, 2008), summer (June, July, August,2008), and autumn (September, October, November, 2008).


sensed twice on orbit. We retain these overlapping pixels however asthe calibration coefficients and the scene center solar geometry can bedifferent between consecutive L1T acquisitions along the same path.

Figs. 6–8 show the true color CONUS browse images for the twelvemonthly, four seasonal, and the annual composited mosaics respec-tively. The browse images are generated in the Albers Equal Areaprojection. A 17×17 pixel window size, that approximates a 500 mproduct resolution, was used to generate the browse images becausethe browse dimensions of 9700×6500 pixels are sufficient to enablevisualization of all the CONUS study area in a single JPEG file.

The true color browse images of the monthly composited mosaicshave evident holes (Fig. 6) that are due to the unavailability of L1T

Fig. 8. Annual (December 2007 to November, 200

acquisitions at path/row coordinates with less than 40% ACCA cloudcover. Geographically, the incidence of holes tends to be higher inmore cloudy regions near the Great Lakes and the U.S.–Canada borderand lower in the Southwest, which is related to cloud distributions atthe time of Landsat overpass (Ju & Roy, 2008). Some path/rowcoordinates have greater than 40% ACCA cloud cover reflectingadditional acquisitions ordered by the user community, confusingthis pattern. Monthly variations in the greenness of the vegetation areevident, with a characteristic wave of growth moving northwardacross the CONUS in the early spring and with differences in thetiming of growth between agricultural and natural vegetation, forexample, in and around the Mississippi valley. Vegetation senescence,

8) true color TOA reflectance CONUS browse.

Table 4The percentage of 30 m CONUS composited mosaic pixels that are missing and flagged as cloudy by the ACCA and Classification (C. Tree) cloud masks.

Period ACCA C. Tree Missing Period ACCA C. Tree Missing Period ACCA C. Tree

December 26.1 23.8 28.3January 28.1 24.0 20.1 Winter 12.95 11.79 1.75February 21.8 20.3 19.7March 14.9 15.1 17.1April 10.1 10.1 16.3 Spring 3.83 3.54 0.74May 12.0 11.1 18.5 Annual 0.19 0.21June 9.4 7.8 15.8July 8.0 6.8 10.6 Summer 1.41 1.29 0.43August 7.7 6.2 11.4September 7.1 6.0 13.5October 6.6 6.5 14.9 Autumn 1.79 1.60 1.43November 11.2 11.9 26.5

The missing percentages are expressed as the percentage of 30 m pixels in each CONUS monthly and seasonal composited mosaic that are missing relative to the CONUS annualcomposited mosaic.


particularly in the deciduous forests on the Eastern sea-board mayalso be apparent, where the leaf chlorophyll responsible for the greenreflectance of the leaves throughout the spring and summer reducesin the autumn and the yellow-red reflectance of other leaf pigmentsbegin to dominate in November. These vegetation patterns have beenobserved with coarser spatial resolution MODIS true color reflectancecomposites (Schaaf, 2007). The spatio-temporal distribution of snowis also evident in the monthly composited mosaics, with snow pre-dominating in the Northern states and over the Rocky mountains inthe winter and spring. Cloud contamination looks the same as snow inthe visible bands illustrated and is most obviously present when therewas only one cloudy L1T acquisition, for example, December, oversouthern Texas.

The seasonal and annual composited mosaics (Figs. 7 and 8) haveconsiderably fewer holes and clouds than the corresponding monthlycomposites. The seasonal composited mosaics still retain some clouds,primarily in the winter, spring and autumn where there were only asmall number of cloud-free L1T observations over the 3 month period.In addition, errors in the ACCA and Classification Tree cloud maskingalgorithms lead to incorrect selection of cloudy pixels in the seasonalmosaics although these are not apparent in the browse imagery. Thereare a small number of holes in the winter, spring and autumn seasonalcomposited mosaics where there were no L1T acquisitions over the3 months (Fig. 7). The annual composited mosaic has no missingacquisitions and less apparent cloud contamination over the land(Fig. 8).

Table 4 quantifies the percentage of 30 m pixels in each CONUScomposited mosaic that are missing and flagged as cloudy by the twocloud masks. There are a greater percentage of compositedmissing andcloudy pixels in winter than in the other seasons which is due to thegreaterwinter cloudiness at the Landsat overpass time (Ju & Roy, 2008).The Landsat ETM+L1Tprocessing and compositing isworking correctlyas increasing the compositing period from monthly, to seasonal, toannual periods, reduces the percentage of cloudy and missing pixels(Table 4). The percent cloud was computed by dividing the number ofcloud flagged pixels by the number of valid pixels without “Fill” values(Table 3). The ACCA cloud cover in the winter, spring, summer andautumn seasonal CONUS composited mosaics were 13.0%, 3.8%, 1.4%,and 1.8% respectively (Table 4), i.e., lower than the equivalent 25%, 17%,15% and 15% mean seasonal ACCA cloud cover percentages of the inputL1T data used to generate the composited mosaics (Table 1, column 4).The Classification Tree underestimated the CONUS cloud cover relativeto ACCA in the input L1T January and July data (Fig. 3) and this dif-ferencewas generallymaintained for the compositedmosaics (Table 4).The annual composited mosaic Classification Tree cloud cover (0.21%)and the annual composited mosaic ACCA cloud cover (0.19%) weresignificantly lower than the 18% mean annual ACCA cloud cover of theinput L1T data (Table 1, column 4).

As the number of L1T acquisitions composited is increased, thepercentage of missing data should decrease monotonically becausethe spatial phase of the SLC-off gaps is not constant betweenacquisitions of the same path/row (Lindquist et al., 2008). Thepercent missing statistics reported in Table 4 were computed for eachmonthly and seasonal composited mosaic by counting the number of“Fill” values in the composited mosaic occurring where there was no“Fill” value in the annual mosaic, divided by 10,838,923,633 (thenumber of annual mosaic pixels without “Fill” values). The monthlycomposites had a variable percentage of missing data from 11% (July)to 28% (December) reflecting the seasonal variability of L1T acquisi-tions with 40% or less cloud cover, and, as expected, the percentage ofmissing pixels in the seasonal composites was much less than in themonthly composites, at most 1.8% in the Winter seasonal compositedmosaic (Table 4).

The illustrated monthly, seasonal and annual composites exhibitacross-track reflectance variations and discontinuities betweenadjacent paths. It is unclear what proportion of this effect is due tothe surface or the atmosphere, although we note that even if theLandsat data were atmospherically corrected, solar zenith variationswill remain (Roy et al., 2008). It is established that for wide field ofview sensor data, such asMODIS or AVHRR, reflectance anisotropy canresult in the maximum NDVI compositing selection of a pixel withhigh NDVI caused by directional rather than reduced atmosphericeffects (Cihlar et al., 1994; Meyer et al., 1995; Gao et al., 2002). TheLandsat ETM+ sensor, with its comparatively narrow field of view, isnot as severely affected, but Landsat directional reflectance effectshave been documented and research on methods to minimize theseeffects undertaken (Danaher et al., 2001; Toivonen et al., 2006;Hansen et al., 2008; Roy et al., 2008). In addition, all compositingprocedures implicitly assume that the surface state remains static, andso increasing the compositing period increases the incidence oftemporal biases associated with violating this assumption (Roy et al.,2006). Thismay cause adjacent pixels to be sensed from the beginningor the end of the compositing period. The maximum NDVIcompositing criterion, leads to the preferential selection of the greenestdates over the compositing period (Roy, 1997, Roy, 2000). This isapparent in Fig. 9 which shows the day of the year that the 12 monthannual composited mosaic pixels were selected on. The majority of theCONUS annual composited mosaic pixels were selected in the summerJune–August months when NDVI is maximal. However, some path/rowcoordinates in the annual composite have fewer acquisitions (Fig. 5)confusing this pattern, and temporal variations driven by phenologicalvariations are also apparent, for example, the Southwest and certainhigh altitude mountains have later September–November peak NDVI,and regions in the southern states have peakNDVI in thewintermonths(Fig. 9). In the monthly composites this compositing timing bias is lessapparent and is illustrated in the next section.

Fig. 9. Day of year pixel selected from the annual composited mosaic. Each pixel shows the day of the year that the reflectance illustrated in Fig. 8 was sensed on.


8.3. Phenology

To further check that the processing was performing correctly, thetemporal coherence of the top of atmosphere NDVI stored in the 12monthly composites was examined. Fig. 10 shows the monthly top ofatmosphere NDVI values from December 2007 to November 2008extracted at four 30 m pixel locations. These four sites representdifferent surface phenologies and three of them arewell studied in theliterature. The NDVI values in successive months are not necessarilymultiples of 16 days apart because of overlapping acquisitions (Fig. 5).The absence of NDVI values in some months, occurring at all the sitesexcept the White Sands site, is due to the unavailability of L1T datawith less than 40% cloud cover. The phenological curves illustrated in

Fig. 10. Monthly top of atmosphere NDVI values from December 2007 to November 2008 atdashed lines illustrate the last day of each month.

Fig. 9 are consistent with the broad patterns of phenology reported inthe literature. The Harvard Forest site is in Massachusetts (42.5360°N,72.1729°W) in a temperate deciduous broad-leaf forest with highNDVI (~0.8) in May–September and rapid phenological change in thespring (March–April) and autumn (October–November) (Schwartzet al., 2002; Fisher & Mustard, 2007). The FIFE (First InternationalSatellite Land Surface Climatology Project (ISLSCP) Field Experiment)site is a grassland site (39.1097°N, 96.5225°W) near Manhattan,Kansas (Sellers et al., 1992). The grassland NDVI trend is similar tothat of Harvard Forest, but with a shorter growing season and lowerNDVI. The Los Angeles site (34.5000°N, 118.6000°W) is a shrublandsite in Los Angeles county, California (Dennison et al., 2005). The NDVIclearly shows the typical wet winter–spring and dry summer–autumn

four 30 m pixel locations extracted from the monthly composited mosaics. The vertical


seasons of the Mediterranean Californian climate. The White SandsMissile Range site (32.9190°N, 106.3510°W) is an arid desert sitedevoid of vegetation and commonly used for satellite sensorradiometric calibration (Thome, 2001). Consequently, the NDVI isconsistently low all year around. Only four sites are shown here due tospace constraints and they by no means represent the full range ofCONUS surface phenology (Morisette et al., 2009) but illustrate thepotential of the monthly composited mosaic data for time seriesmonitoring at 30 m resolution.

9. Summary

The 2008 free Landsat Data Distribution Policy opens a new era forutilizing Landsat data; users are no longer restricted by data costs toselection of acquisitions with low cloud cover, rather, with the adventof free Landsat data it becomes feasible to analyze and process long-term and/or large-area Landsat data sets. For example, a total of 6521Landsat acquisitions were used to generate the monthly, seasonal andannual composited mosaics of the CONUS described in this paper; thisdata volumewould have cost nearly four million US dollars when eachLandsat acquisition was priced at six hundred dollars (Reichhardt,2002).

The research described in this paper represents preliminary resultsof a project with the goal of providing consistent continental scale30 m Landsat long-term data records. Such data records are needed tomonitor land-cover change and study Earth system functioning(Gutman et al., 2008; Wulder et al., 2008) and may provide a highspatial resolution analogue to the moderate and coarse spatialresolution land products generated from the MODIS and AVHRRdata streams (Justice et al., 2002; Tucker et al., 2005). The generationof long-term Landsat data records, for example from 1972 to present,and at continental to global scale, is technically feasible using theapproach described in this paper but will be constrained primarily byLandsat data access and the ability to calibrate and geolocate the datain a consistent manner. The challenges to obtaining such consistentlyprocessed data are not inconsiderable, as the Landsat data reside indifferent international archives with different formats, documenta-tion, and processing. The planned Landsat Data Continuity Mission(LDCM) will provide an opportunity to continue the development ofsuch a long-term large-area data record.

The Landsat ETM+ processing described in this paper wasnecessarily designed to enable large volume processing. The proces-sing steps included conversion of digital numbers to calibratedradiance to top of atmosphere reflectance and brightness tempera-ture, per-band radiometric saturation identification, cloud screening,reprojection, and compositing. Increasing the compositing periodfrom monthly, to seasonal, to annual periods, reduced the percentageof cloudy and missing pixels. Using more Landsat 7 ETM+ acquisi-tions, and also acquisitions from the Landsat 5 Thematic Mapper,would further reduce the incidence of missing and cloudy pixels. Themonthly, seasonal and annual composited mosaic products qualita-tively capture CONUS spatio-temporal surface variations andexpected land surface phenology. Reflectance discontinuities remainin the composited mosaics due to factors including the impact of theatmosphere and residual clouds, bi-directional reflectance effectsassociated with variations in solar illumination and observationangles, the unavailability of L1T acquisitions at path/row coordinateswith greater than 40% ACCA cloud cover, and the sensitivity of thecompositing procedure to these effects. Future research is needed tomore fully characterize these effects and to implement systematicatmospheric correction and radiometric/BRDF normalization, forexample, by a fusion of information from contemporaneous MODISTerra data (Roy et al., 2008).

The composited mosaic data products described in this paper, andsoftware tools to utilize them, are being made available via theInternet from the U.S. Landsat Project at http://landsat.usgs.gov/

WELD.php. This web site is evolving, and more years of compositedmosaic products and future versions with refined processing will beprovided.

Acknowledgements

This Web-enabled Landsat Data (WELD) project is funded byNASA's Making Earth System Data Records for Use in ResearchEnvironments (MEaSUREs) program, grant number NNX08AL93A.The LEDAPS team led by Dr. Masek is thanked for their feedback andprovision of the ACCA cloud masking code. The U.S. Landsat projectmanagement and staff are thanked for provision of the Landsat ETM+data.

References

Arvidson, T., Goward, S. N., Gasch, J., & Williams, D. (2006). Landsat-7 long-termacquisition plan: Development and validation. Photogrammetric Engineering &Remote Sensing, 72, 1137−1146.

Bindschadler, R., Vornberger, P., Fleming, A., Fox, A., Mullins, J., Binnie, D., et al. (2008).The Landsat image mosaic of Antarctica. Remote Sensing of Environment, 112,4214−4226.

Boschetti, L., Roy, D. P., & Justice, C. O. (2008). Using NASA's world wind virtual globe forinteractive Internet visualisation of the global MODIS burned area product. Inter-national Journal of Remote Sensing, 29, 3067−3072.

Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification andregression trees. Monterey: Wadsworth & Brooks/Cole.

Breiman, L. (1996). Bagging predictors. Machine Learning, 26, 123−140.Cahalan, R. F., Oreopoulos, L., Wen, G. Y., Marshak, A., Tsay, S. C., & DeFelice, T. (2001).

Cloud characterization and clear-sky correction from Landsat-7. Remote Sensing ofEnvironment, 78, 83−98.

Chander, G., Huang, C., Yang, L., Homer, C., & Larson, C. (2009a). Developing consistentLandsat data sets for large area applications: The MRLC 2001 protocol. IEEETransactions on Geoscience and Remote Sensing, 6, 777−781.

Chander, G., Markham, B. L., & Helder, D. L. (2009b). Summary of current radiometriccalibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors. RemoteSensing of Environment, 113, 893−903.

Cihlar, J., Manak, D., & D'Iorio, M. (1994). Evaluation of compositing algorithms forAVHRR data over land. IEEE Transactions on Geoscience and Remote Sensing, 32,427−437.

Danaher, T., Wu, X., & Campbell, N. (2001). Bi-directional reflectance distributionfunction approaches to radiometric calibration of Landsat TM imagery. Proceed-ings of the IEEE Geoscience and Remote Sensing Symposium (IGARSS 2001), vol. 6.(pp. 2654−2657).

Dennison, P. E., Roberts, D. A., Peterson, S. H., & Rechel, J. (2005). Use of normalizeddifference water index for monitoring live fuel moisture. International Journal ofRemote Sensing, 26, 1035−1042.

Dowdeswell, J. A., & Mcintyre, N. F. (1986). The saturation of Landsat MSS detectorsover large ice masses. International Journal Of Remote Sensing, 7, 151−164.

Fisher, J. I., & Mustard, J. F. (2007). Cross-scalar satellite phenology from ground Landsatand MODIS data. Remote Sensing of Environment, 109, 261−273.

Flynn, L. P., & Mouginis-Mark, P. J. (1995). A comparison of the thermal characteristicsof active lava flows and forest fires. Geophysical Research Letters, 22, 2577−2590.

Gao, F., Jin, Y., Li, X., Schaaf, C. B., & Strahler, A. H. (2002). Bidirectional NDVI andatmospherically resistant BRDF inversion for vegetation canopy. IEEE Transactionson Geoscience and Remote Sensing, 40, 1269−1278.

Gutman, G., Byrnes, R., Masek, J., Covington, S., Justice, C., Franks, S., et al. (2008).Towards monitoring land-cover and land-use changes at a global scale: The GlobalLand Survey 2005. Photogrammetric Engineering and Remote Sensing, 74, 6−10.

Hansen, M. C., Roy, D. P., Lindquist, E., Adusei, B., Justice, C. O., & Altstaat, A. (2008). Amethod for integrating MODIS and Landsat data for systematic monitoring of forestcover and change and preliminary results for Central Africa. Remote Sensing ofEnvironment, 112, 2495−2513.

Holben, B. (1986). Characteristics of maximum-value composite images from temporalAVHRR data. International Journal of Remote Sensing, 7, 1417−1434.

Homer, C., Huang, C., Yang, L., Wylie, B., & Coan, M. (2004). Development of a 2001national land-cover database for the United States. Photogrammetric Engineeringand Remote Sensing, 70, 829−840.

Huete, A. (1988). A soil-adjusted vegetation index (SAVI). Remote Sensing of Environment,25, 295−309.

Huete, A., Didan, K., Miura, T., Rodriguez, E. P., Gao, X., & Ferreira, L. G. (2002). Overviewof the radiometric and biophysical performance of the MODIS vegetation indices.Remote Sensing of Environment, 83, 195−213.

Irish, R. I. (2000). Landsat7 automatic cloud cover assessment, in algorithms formultispectral,hyperspectral, and ultraspectral imagery. Proc. SPIE, vol. 4049. (pp. 2000).

Irish, R. I., Barker, J. L., Goward, S. N., & Arvidson, T. (2006). Characterization of theLandsat-7 ETM+ automated cloud-cover assessment (ACCA) algorithm. Photo-grammetric Engineering & Remote Sensing, 72, 1179−1188.

Irons, J. R., & Masek, J. G. (2006). Requirements for a Landsat data continuity mission.Photogrammetric Engineering and Remote Sensing, 72, 1102−1108.




Ju, J., & Roy, D. P. (2008). The availability of cloud-free Landsat ETM+ data over theconterminous United States and globally. Remote Sensing of Environment, 112,1196−1211.

Justice, C., Townshend, J., Vermote, E., Masuoka, E., Wolfe, R., Saleous, N., et al. (2002).An overview of MODIS Land data processing and product status. Remote Sensing ofEnvironment, 83, 3−15.

Kaufman, Y. J. (1987). The effect of subpixel clouds on remote sensing. InternationalJournal of Remote Sensing, 8, 839−857.

Konecny, G. (1976). Mathematical models and procedures for the geometricregistration of remote sensing imagery. International Archives of Photogrammetryand Remote Sensing, 21, 1−33.

Lee, D. S., Storey, J. C., Choate, M. J., & Hayes, R. (2004). Four years of Landsat-7 on-orbitgeometric calibration and performance. IEEE Transactions on Geoscience and RemoteSensing, 42, 2786−2795.

Lindquist, E., Hansen, M., Roy, D. P., & Justice, C. O. (2008). The suitability of decadalimage data sets for mapping tropical forest cover change in the DemocraticRepublic of Congo: Implications for the mid-decadal global land survey. Interna-tional Journal of Remote Sensing, 29, 7269−7275.

Liu, H. Q., & Huete, A. R. (1995). A feedback based modification of the NDVI to minimizecanopy background and atmospheric noise. IEEE Transactions on Geoscience andRemote Sensing, 33, 457−465.

Markham, D., Goward, G., Arvidson, T., Barsi, J., & Scaramuzza, P. (2006). Landsat-7long-term acquisition plan radiometry—Evolution over time. PhotogrammetricEngineering & Remote Sensing, 72, 1129−1135.

Masek, J. G., Vermote, E. F., Saleous, N. E., Wolfe, R., Hall, F. G., Huemmrich, K. F., et al.(2006). A Landsat surface reflectance dataset for North America, 1990–2000. IEEETransactions on Geoscience and Remote Sensing Letters, 3, 68−72.

Maxwell, S. K., Schmidt, G. L., & Storey, J. C. (2007). A multi-scale segmentationapproach to filling gaps in Landsat ETM+ SLC-off images. International Journal ofRemote Sensing, 28, 5339−5356.

Meyer, D., Verstraete, M., & Pinty, B. (1995). The effect of surface anisotropy and viewinggeometry on the estimation of NDVI from AVHRR. Remote Sensing Reviews, 12, 3−27.

Morisette, J. T., Richardson, A. D., Knapp, A. K., Fisher, J. I., Graham, E. A., Abatzoglou, J.,et al. (2009). Tracking the rhythm of the seasons in the face of global change:Phenological research in the 21st century. Frontiers in Ecology and theEnvironment, 7, 253−260.

Ouaidrari, H., & Vermote, E. F. (1999). Operational atmospheric correction of LandsatTM data. Remote Sensing of Environment, 70, 4−15.

Platnick, S., King, M. D., Ackerman, S. A., Menzel, W. P., Baum, B. A., Riédi, J. C., et al.(2003). The MODIS cloud products: Algorithms and examples from Terra. IEEETransactions on Geoscience and Remote Sensing, 41, 459−473.

Prasad, A. M., Iverson, L. R., & Liaw, A. (2006). Newer classification and regression treetechniques: Bagging and random forests for ecological prediction. Ecosystems, 9,181−199.

Reichhardt, T. (2002). Satellite-image users fear private price hike. Nature, 419, 328.Roujean, J. -L., Leroy, M., & Deschamps, P. -Y. (1992). A bidirectional reflectance model

of the Earth's surface for the correction of remote sensing data. Journal OfGeophysical Research, 97, 20455−20468.

Roy, D. P. (1997). Investigation of the maximum normalised difference vegetationindex (NDVI) and the maximum surface temperature (Ts) AVHRR compositingprocedures for the extraction of NDVI and Ts over forest. International Journal ofRemote Sensing, 18, 2383−2401.

Roy, D. P. (2000). The impact of misregistration upon composited wide field of viewsatellite data and implications for change detection. IEEE Transactions on Geoscienceand Remote Sensing, 38, 2017−2032.

Roy, D. P., Borak, J., Devadiga, J., Wolfe, R. E., Zheng, M., & Descloitres, J. (2002). TheMODIS land product quality assessment approach. Remote Sensing of Environment,83, 62−76.

Roy, D. P., Lewis, P., Schaaf, C., Devadiga, S., & Boschetti, L. (2006). The global impact ofcloud on the production of MODIS bi-directional reflectancemodel based compositesfor terrestrial monitoring. IEEE Geoscience and Remote Sensing Letters, 3, 452−456.

Roy, D. P., Ju, J., Lewis, P., Schaaf, C., Gao, F., Hansen, M., et al. (2008). Multi-temporalMODIS–Landsat data fusion for relative radiometric normalization, gap filling, andprediction of Landsat data. Remote Sensing of Environment, 112, 3112−3130.doi:10.1016/j.rse.2008.03.009.

Saalfeld, A. (1985). A fast rubber-sheeting transformation using simplicial coordinates.The American Cartographer, 12, 169−173.

Schaaf, C. (2007). The sunlit Earth. In Michael D. King, Claire L. Parkinson, Kim C.Partington, & Robin G. Williams (Eds.), Our changing planet: A view fromspaceSeptember 30th 2007. (pp. 129−135) : Cambridge University Press.

Schaaf, C., Gao, F., Strahler, A., Lucht, W., Li, X., Tsang, T., et al. (2002). First operationalBRDF, albedo and nadir reflectance products from MODIS. Remote Sensing ofEnvironment, 83, 135−148.

Schaepman-Strub, G., Schaepman, M. E., Painter, T. H., Dangel, S., & Martonchik, J. V.(2006). Reflectance quantities in optical remote sensing—Definitions and casestudies. Remote Sensing of Environment, 103, 27−42.

Schwartz, M. D., Reed, B. C., & White, M. A. (2002). Assessing satellite-derived start-of-season (SOS) measures in the conterminous USA. International Journal of Climatology,22, 1793−1805.

Sellers, P. J., Hall, F. G., Asrar, G., Strebel, D. E., & Murphy, R. E (1992). An overview of theFirst International Satellite Land Surface Climatology Project (ISLSCP) FieldExperiment (FIFE). Journal of Geophysical Research, 97(D17), 18345−18371.

Serra, J. (1982). Image analysis and mathematical morphology. London Academic.Snyder, J. P. (1993). Flattening the Earth: Two thousand years of map projections. Chicago

and London: The University of Chicago Press.Storey, J., Scaramuzza, P., Schmidt, G., & Barsi, J. (2005). Landsat 7 scan line corrector-off

gap filled product development. Proceedings of Pecora 16 Global Priorities in LandRemote Sensing (pp. 23−27). Sioux Falls, South Dakota: American Society forPhotogrammetry and Remote Sensing.

Teillet, P. M., Saleous, N. El., Hansen,M. C., Eidenshink, J. C., Justice, C. O., & Townshend,J. R. G. (2000). An evaluation of the global 1-kmAVHRR land data set. InternationalJournal of Remote Sensing, 21, 1987−2021.

Thome, K. J. (2001). Absolute radiometric calibration of Landsat 7 ETM+ using thereflectance-based method. Remote Sensing of Environment, 78, 27−38.

Townshend, J. R. G., & Justice, C. O. (1988). Selecting the spatial-resolution of satellitesensors required for global monitoring of land transformations. InternationalJournal of Remote Sensing, 9, 187−236.

Toivonen, T., Kalliola, R., Ruokolainen, K., &Malik, R. N. (2006). Across-path DN gradientin Landsat TM imagery of Amazonian forests: A challenge for image interpretationand mosaicing. Remote Sensing of Environment, 100, 550−562.

Tucker, C. J. (1979). Red and photographic infrared linear combinations for monitoringvegetation. Remote Sensing of Environment, 8, 127−150.

Tucker, C. J., Grant, D. M., & Dykstra, J. D. (2004). NASA's global orthorectified Landsatdata set. Photogrammetric Engineering and Remote Sensing, 70, 313−322.

Tucker, C. J., Pinzon, J. E., Brown, M. E., Slayback, D. A., Pak, E. W., Mahoney, R., et al.(2005). An extended AVHRR 8-km NDVI data set compatible with MODIS and SPOTvegetation NDVI data. International Journal of Remote Sensing, 26(20), 4485−4498.

Vermote, E. F., El Saleous, N., & Justice, C. (2002). Atmospheric correction of the MODISdata in the visible to middle infrared: First results. Remote Sensing of Environment,83, 97−111.

Wardlow, B. D., Egbert, S. L., & Kastens, J. H. (2007). Analysis of time-series MODIS250 m vegetation index data for crop classification in the U.S. Central Great Plains.Remote Sensing of Environment, 108, 290−310.

Williams, D. L., Goward, S., & Arvidson, T. (2006). Landsat: yesterday, today, andtomorrow. Photogrammetric Engineering and Remote Sensing, 72, 1171−1178.

Wolfe, R., Roy, D., & Vermote, E. (1998). The MODIS land data storage, gridding andcompositing methodology: L2 grid. IEEE Transactions on Geoscience and RemoteSensing, 36, 1324−1338.

Woodcock, C. E., Allen, A. A., Anderson, M., Belward, A. S., Bindschadler, R., Cohen, W. B.,et al. (2008). Free access to Landsat imagery. Science, 320, 1011.

Wulder, M. A., Loubier, E., & Richardson, D. (2002). A Landsat-7 ETM+ orthoimagecoverage of Canada. Canadian Journal of Remote Sensing, 28, 667−671.

Wulder, M. A., White, J. C., Goward, S. N., Masek, J. G., Irons, J. R., Herold, M., et al. (2008).Landsat continuity: Issues and opportunities for land cover monitoring. RemoteSensing of Environment, 112, 955−969.

WWW1, http://edcsns17.cr.usgs.gov/helpdocs/landsat/product_descriptions.html,Landsat Product Type Descriptions, Last accessed May 15th 2009.

WWW2, http://landsathandbook.gsfc.nasa.gov/handbook/handbook_htmls/chapter12/htmls/level1_differences.html, Landsat Handbook, Chapter 12, Level 1 ProductDifferences, Last accessed May 15th 2009.

WWW3, http://landsathandbook.gsfc.nasa.gov/handbook/handbook_htmls/chapter6/chapter6.html, Landsat Handbook, Chapter 6—Data Properties, Last accessed May15th 2009.

WWW4, https://lpdaac.usgs.gov/lpdaac/tools/modis_reprojection_tool, MODIS Repro-jection Tool, Last accessed May 15th 2009.

Zobrist, A. L., Bryant, N. A., & McLeod, R. G. (1983). Technology for large digital mosaicsof Landsat data. Photogrammetric Engineering and Remote Sensing, 49, 1325−1335.

http://dx.doi.org/10.1016/j.rse.2008.03.009

http://edcsns17.cr.usgs.gov/helpdocs/landsat/product_descriptions.html

http://landsathandbook.gsfc.nasa.gov/handbook/handbook_htmls/chapter12/htmls/level1_differences.html

http://landsathandbook.gsfc.nasa.gov/handbook/handbook_htmls/chapter12/htmls/level1_differences.html

http://landsathandbook.gsfc.nasa.gov/handbook/handbook_htmls/chapter6/chapter6.html

http://landsathandbook.gsfc.nasa.gov/handbook/handbook_htmls/chapter6/chapter6.html

https://lpdaac.usgs.gov/lpdaac/tools/modis_reprojection_tool

Web-enabled Landsat Data (WELD): Landsat ETM+ composited mosaics

Documents

Transcript of Web-enabled Landsat Data (WELD): Landsat ETM+ composited mosaics