ELSEVI ER

Development of Methods for Mapping GlobalSnow Cover Using Moderate ResolutionImaging Spectroradiometer Data

Dorothy K. Hall,* George A. Riggs,'

An algorithm is being developed to map global snowcover using Earth Observing System (EOS) ModerateResolution Imaging Spectroradiometer ('MODIS) data be-ginning at launch in 1998. As currently planned, digitalmaps will be produced that will provide daily, and per-haps maximum weekly, global snow cover at 500-m spa-tial resolution. Statistics will be generated on the extentand persistence of snow cover in each pixelfor each weeklymap, cloud cover permitting. It will also be possible togenerate snow-cover maps at 250-ni spatial resolutionusing MODIS data, and to study snow-cover characteris-tics. Preliminary validation activities of the prototypeversion of our snow-mapping algorithm, SNOMAP, havebeen undertaken. SNOMAP will use criteria tests and adecision rule to identify snow in each 500-m MODISpixel. fUse of SNOMAP on a previously mapped LandsatThematic Mapper (TAl) scene of the Sierra ]Nevadas hasshown that SNOMAP is 98% accurate in identifying snowin pixels that are snow covered bn 60% or more. Resultsof a comparison of a SNOMAP classification with a super-vised-classification technique on six other TM scenes showthat SNOMAP and supervised-classification techniquesagree to within about I I % or less for nearly cloud-freescenes and that SNOMAP provided motre consistent re-sults. About 10 % of the snow cover, known to be presenton the 14 March 1991 TM scene covering Glacier NationalPark in northern Montana, is obscured by dense forestcover. Mapping snow cover in areas of dense forests is alimitation in the use of this procedure for global snow-cover mapping. This limitation, and sources of error willbe assessed globally as SNO1MAP is refined and testedbefore and following the launch of MOD)IS.

Ilvdrologieal Sciecuces Branch. L.doirator0 fko dI\ drosplheriProcesses, NASA/ Goddard Space Flight (Xenter, (Grelibelt, Ml)

Research and Data Svstenis (Coirporation, (Greenb ielt, Ml)'Earth Sciences Directorate, NAASA /Goddard Space Flight (Cen-

ter, Greenbelt, MD

Address correspondence to I)r. l)orotbN K. Hlall, N ASA GIoddlladFlight (enter, Code 974, Greenibelt. Ml) 20771.

Received .9 1'ebruawnj 1995. acce'pted 1/3 Matil 1995

*and Vincent V. Salomonsont

INTRODUCTION

The highly reflective nature of snow combined with itslarge surface coverage (snow can cover up to 40% of theEarth's land surface during the Northern Hemispherewinter), make snow an important determinant of theEarth's radiation balance (Foster and Chang, 1993).Snow on the ground influences biological, chemical, andgeological processes (Walsh et a]., 1985; Robinson andKukla, 1985; Allen and Walsh, 1993; Robinson et al.,1993). Many areas of the world rely on snowmelt forirrigation and drinking water and must monitor snow-packs closely throughout the winter and spring for as-sessment of water supply (Carroll et al., 1989).

Snow cover is a key component of regional andglobal climate, and it is vital to have an accurate andlong-term database established on snow-extent variabil-ity. General circulation models (GCMs) do not simulatethe present Arctic climate very well (Bromwich andTzeng, 1994), and thus improvements in the measure-ment of global snow cover and other cryospheric ele-ments are key to improving the GCMs.

The Moderate Resolution Imaging Spectroroadio-meter (MODIS) will be launched as part of the firstEarth Observing System (EOS) platform in 1998 witha capability to study geophysical features globally (Salo-monson and Toll, 1991), including mapping the arealextent and reflectance of global snow cover on a dailybasis. A prototype algorithm, called SNOMAP, is cur-rently being developed using TM data to enable auto-mated mapping of areal extent of snow cover usingfuture MOI)IS data.

Efforts to develop and refine SNOMAP, and toassess its accuracy, are discussed in this paper. Examplesof work on algorithm development from Landsat the-matic mapper (TM) scenes acquired over northern Mon-tana (including Glacier National Park); the ChugachMountains, Alaska; Glacier Bay, Alaska; northern Minne-sota; and the Sierra Nevada Mountains, California in

REMOTE SENS. ENVIRON. 54:127-,1-1 (19951)EIsevier Scienice inc., 1995

655 Avenue of the Am ericas. N;\Ns Y,,ik NY 1011)00034-4257 / 95 / $9.50)

SSI)I 0034-4257(95)00137-P

128 Hall et al.

the United States, and an area in southeastern Icelandare presented.

BACKGROUND

Satellite sensors have been emplo'ed to map snowcover, and to measure (or estimate) snow depth andreflectance. Using satellite data, available since 1966, ithas been shown that there is an inverse relationshipbetween hemisphere-averaged, monthly mean snow cm-er, and temperature fluctuations (Robinson and Dewed,1990; Robinson et al., 1991; G;utzler and Rosen, 1992;NOAA, 1994). National Oceanic and Atmospheric Ad-ministration (NOAA) satellite data including the Ad-vanced Very High Resolution Radiometer (AVHRR) en-able the measurement of snow extent using visible.near-infrared and thermal-infrared sensors at a resolt-tion of about I kin (Matson et al., 1986; Matson, 1991).NOAA snow charts are digitized weekly using the Na-tional Meteorological (Center's standard-analysis grid, an89 x 89 cell Northern hemisphere grid with polar-stereographic projection. Cell resolution ranges from16,000-42,000 km2 . Only cells with a least 50% snowcover are mapped as snow (Robinson et al., 1993).

Other snow-mapping studies are performed on rt-gional and local scales using ground-based measure-ments, NOAA AVHRR, Landsat inultispectral scaninev(MSS) and TM data, and aircraft data (for example, seeCarroll, 1990; Rango, 1993). RegionIl snow productsare produced in > 4000 drainage basins in the westernUnited States and Canada on a weekly basis during thesnow season using NOAA AVHRR 1-km data (Carroll,1990; Rango, 1993). Use of passive-microwave sensorson board the Nimbnts -5, -6, and -7 satellites and theDefense Meteorological Satellite Program (DMSP) satel-lite has allowed successful measurement of snow extentat a 25 to 30-km resolution through cloud cover anddarkness since 1978.

The Landsat MSS and TM sensors may be used formeasurement of snow-covered area over drainage basins(Rango and Martinec, 1982). AVHRR data have alsobeen used successfully to measure snow at the drainage-basin scale. The 16-day repeat cycle of the Landsat-4and -5 satellites, however, coupled with the potentialfor cloud cover, precludes the use of Landsat data foroperational snow mapping. Additionally, Landsat TMdata are useful for the quantitative measurement ofsnow reflectance (Dozier et al., 1981; Dozier, 1984,1989; flall et al., 1992; Winther, 1992).

Various techniques, ranging from visual interpreta-tion, multispectral image classification, decision trees,change detection, and ratios (Kyle et al., 1978; Buntingand d'Entremont, 1982), have been used to map snowcover with remotely sensed data (see, for example,Rango, 1975). Additionally, spectral-mixture modelinghas been demonstrated as an important new techniquefor su1)pixel classification of snow in a scene (Nolin etal., 1993; Rosenthal, 1993).

DESCRIPTION OF THE MODIS SENSOR

The Moderate Resolution Imaging Spectroradiometeris an imaging radiometer that uses a cross-track scanmirror and collecting optics, and a set of individualdetector elements to acquire imagery of the Earth'ssurface and clouds in 36 discrete spectral bands. MODISis scheduled for launch as a facility instrument of theEOS polar-orbiting platform in 1998. The primary pur-pose of MODIS data is to permit the regional to globalstH(1y of the land, atmosphere, and ocean on a daily ornear-daily basis (Salomonson et al., 1992). Key landscience objectives are to study global vegetation andground cover, global land surface change, vegetationproperties, surface albedo, surface temperature, and

Tablec 1. MO)DIS Band Nuimber (#). and Bandwkidtlis in rni

MODIS band # bandwidth baln(l # bandwidth

92:3

4

6

S

'I(1

I I

12

1:3

l 4'.51 31 71s

0. 6200-() .60.841 -0.S-T60.459-(.47

o 545-0.5651 2:30-1 2501.628-1 .i522.105-2.155().405-(.42()0.438-0(44Xo 483-(.4930.526-0.5:36( 546- (55f60(662-).6720.673-0.63:0 743-(,53o 862-0(.7,,o 890-0.9200 9:31 -0(941

19

2(0

21

_222;324252 627i

2S29:30

:31

;3'3:34:3;5:3(6

0.915-0.9653.66-3.843.93-3.993.93-3.994.02-4.084.43-4.504.48-4.551.36-1.39

6.54-6.907.18-7.488.40-8.7,09.58-9.88

10.78-11.2811.77 12.2,13.19-13.4913.49-1 3., 7913.79-14.09

14.09-14.39

Mfethodsfor Mapping Snow Cover 129

Table 2. TM Band Number (#), and Bandwidths in I±ni

TM hand # bandwidthI 0.45-0.52

2 )0.52-0.603 (0.63-0.694 0.76-.90)5 1.55-1.75

6 10.40(-12.507 2.08-2.35

snow and ice cover and characteristics (Salomonson et

al., 1992; Running et al., 1994).Spatial resolution of the MODIS sensor at nadir

varies with spectral band and is 250, 500, or 1000 in.The spectral bands cover parts of the electromagneticspectrum from about 0.4-14.0 am (Table 1), thus spaii-ning the visible and thermal-infrared parts of the spec-trum. MODIS bands covering visil)le, near-infrared andshort-wave infrared parts of the spectrum will be used inthe snow-mapping algorithm. On the first FOS platform.MODIS is scheduled to have a morning (10:30 AA1.

+ 15 min) overpass (descending inode platform). Fur-ther details about the MODIS instrument characteristicscan be found in Salomonson and Barker (1992) andKing et al. (1992).

The wide swath ( + 55) of the MO)lIS sensor willbe suitable for large-area coverage. Only data from+ 450 will be used for production of the snow mapsbecause the distortions in pixel geometry and the in-creases in snow anistropy at angles greater than + 45'are likely to adversely affect our ability to calculatesnow-covered area using SNOMAP. Even with this re-striction, at a scan angle of + 450, virtually all seasonallxsnow-covered areas cal be imaged daily (Fleig et al.. inpress).

MODIS band selection for SNOMAP has beenlargely determined by research donen with comparablewavelength data from the TM sensors. TM bands artlisted in Table 2. MODIS Airborne Simulator (MAS)data will increasingly be used to refine( SNOMAP. AsMAS are acquired and analvze(l. selection of optimumMODIS bands to use in SNOMAI3 max! change in theprelaunch time frame.

Snow typically has ver\ high \ isible reflectance, andbased on MODIS specifications. NMODIS band 4 shouldnot saturate if snow is present. thus it is a good bandfor snow measurement and identification. This is animportant advance as saturation in some AV7HRR chan-nels and TM bands has been a p roblem over snow-covered areas.

RESULTS

Description of the SNOMAP Algorithm for MappingSnow Cover

SNOMAP (Riggs et al., 1994) is an) algorithm designedto identify snow, if present, in each MODIS pixel each

MODIS PRODUCT INPUTS

ANCILLARY

ALGORITHMPRODUCTS

tigvi, 1. A comiceptual flow diagram for SNOMAP.

dax. As cunrrently planned, if snow is present in any

pixel on any day during the compositing period, thatpixel will be considered to be snow covered. There willbe a daily and perhaps a weekly snow-cover product.

In the context of the Earth Observing System DataInformation System (EOSDIS), it is expected that sur-tace reflectance and a cloud mask produced by otherMOD)IS investigators will be used as input to SNOMAP(Fig. 1). When SNOMAP is implemented globally, usingMODIS data, a land / water mask will also be used.Seasonal and geographic adjustments will be made tothe algorithm, and mapping / gridding programs will beemuployedl to generate the global snow-cover data prod-uet. A daily and weekly snow-cover data product, andstatistics on snow-cover persistence for the weekly prod-uet are the outputs of SNOMAP. Areas that are coveredby persistent cloud cover will be identified as such.(:loid-cover considerations will be discussed further ina later Section.

Unique aspects of the MODIS-derived global snowmaps inc(llde: fully automated production, anticipatedimproved spectral discrimination between snow andother features, relative to what is available today, andstatistics (lescribing snow-cover persistence in eachpixel of the weekly product.

Landsat TM data are used as a surrogate for someof the spectral bands of MODIS in order to develop analgorithm to map snow using future MODIS data. For

130 Halletal.

initial algorithm-development efforts, TM data are bet-ter suited to represent MODIS data than are any othersatellite data for simulation of spectral characteristicsthat apply to mapping snow cover. There are importantdifferences between the TM and MODIS data that makethe use of TM data alone inadequate to simulate MODISdata, however. First, the TM data have a pixel resolutionof 30 m whereas the MODIS pixel resolution will be250( m to 1 km. The finer resolution of the TM data issuitable for mapping snow in drainage basins, whereasthe MODIS 250-m-1-km resolution is suitable for re-gional- and global-scale snow mapping as well as formapping snow' in drainage basins. There are fewer TMbands than there are MODIS bands, and the MOI)ISand TM band widths differ. Also, because the TM sensorhas only one band in the thermal-infrared part of thespectrum (TM band 6), it is difficult to determine thepotential utility of the MODIS thermal-infrared bands,using TM data alone. Data from other sensors, such asthe MAS, and the ANVHRR, are also l)eing used indevelopment efforts.

At-satellite reflectance is calculated using the fol-lowing equiation (Markham and Barker, 1986) in ourprototype algorithm,

P, = (mrL;d)/(ESU T Ncosd) ( l

where:

p = unitless effective at-satellite planetaryreflectance:

1, = spectral radiance at sensor aperture(Inl2l1cmn -2sr 'prn ).

d = Earth-Simn distance (Astronomical Units):ESI CA = mean solar exoatmnosph eric irradiance

(in X'V cnm 2- rn - );0, = solar zenith angle (degrees).

Because the above formulation assumes isotropy ofthe surface, and snow is an anisotropic reflecting surface.there is an inherent error in the reflectance value calcim-lated using this technique. The error will be smaller forfreshlv fallen snow than for ol(ler, metamorphosed snow.because fresh snow can be nearlv an isotropic reflectingmedium (Steffen. 1987). This error is expected to besmall enough so that SNONMAP results should not bealtered significantly, if at all. bw this source of error.

Use of Reflectances versus Digital NumbersBecause the same D)N values on diffrent TM scenesare likely to correspond to different reflectances, theuse of reflectances improves the identification of snowbecause reflectance is based on fraction of incomingsolar radiation, anml the cosine eflect of' sun angle onincident radiation is accoulnte(l for. Thus, optimum de-tection of snow cover requires that data be expressedin physical units, for example. reflectance. MODIS cali-

1.

wIU)z0-LJ

U-

0zM2

0.8-

0.6

0.4

0.2

LANDSAT 5 TM BANDS 2,4,5G1scoeNP 14 Mo,91 solszenih=58.0

0 50 1o0 150DN

200 250

TM 2

TM 4

TM5

300

Figure 2. Relationship between digital number (DN) and re-flectance on the 14 March 1991 TM scene covering GlacierNational Park, Montana. Reflectance was calculated usingformulation by Markham and Barker (1986).

brated, geolocated, atmospherically corrected radiancesare the planned input data for SNOMAP.

In order to test the assumption that the use ofreflectances is better than the use of DNs for snow-covermapping, a study was conducted using two TM scenesin northern Montana: 14 March 1991 and 06 March1994, to identify snow cover. The relationship betweenthe DN and reflectance values for the 14 March 1991TM scene covering Glacier National Park is shown for3 TM hands in Figure 2. Sensor saturation is commonin TM bands over snow and no further informationcan l)e derived from reflectance once saturation hasoccurred. Both DNs and reflectances were used to cal-cimlate the Normalized Difference Snow Index (NDSI).The NDSI is an integral part of the SNOMAP algorithmfor the identification of snow:

N DSI = (TM band 2 - TM band 5)(TM band 2 + TM band 5).

(2)

Having its heritage with the the Normalized DifferenceVegetation Index (NDVI) (Tucker, 1979, 1986), andhand-rationimig techniques (Kyle et al., 1978; Buntingand d'Entremont; Dozier, 1984), the NDSI is used toidentify snow in an automated-algorithm environment.The utility of the NDSI is based on the fact that snowreflects visible radiation more strongly than it reflectsradiation in the middle-infrared part of the spectrum.Because the reflectance of clouds remains high in theregion of the spectrum in which TM band 5 is located,an(d the reflectance of snow drops to near-zero values,the NDSI also functions as a snow / cloud discriminator.

A greater snow cover was calculated when reflec-tances were used to calculate the NDSI versus whenl)Ns were used, and ground-truth measurements simul-taneous with the 14 March 1991 data acquisition indi-

I I __ /d

X �., 71�1

Methods for Mapping Snow Cover 131

Table 3. Comparison of the Number ol Snow-CoveredPixels (and Percentage of Scene that Was SnowCovered) Calculated Using DNs and Reflectances (R)to Calculate the NDSI

14 March 1991 06 March 1994

DN method 1(),272,700) (25.7%) 9.733,780 (23.4%)R method 13,483,976 (32.4%) 13,088,391 (31.5%)

cated that the larger amount of snow cover was moreaccurate (Table 3).

In testing snow-covered areas using TM scenes inthe United States (Alaska, California, Montana, and Min-nesota), and Iceland, NDSI values greater than or equalto approximately 0.4 were found to represent snowcover well, and to separate snow from most clouds. Thiswas also found to be an effective threshold for snowmapping in the Sierra Nevadas by Dozier (1989). Sensi-tivity of individual criteria tests can be studied by chang-ing the threshold value incrementally and analyzing theeffect on results. These results reveal that there is notan exact NDSI threshold for snow, but that a crediblethreshold for snow mapping can be established.

For six TM scenes, the NDSI threshold value forsnow was increased incrementally from 0 to 1.0 in stepsof 0.05 as shown in Figures 3a and 3b. If a pixel hadan NDSI value equal to or greater than the threshold,it was identified as snow. Results are expressed as totalnumber of snow pixels for a threshold as shown in

Figure 3a, and as change in snow cover between succes-sive NDSI thresholds as shown in Figure 3b.

A consistent observation is that the areal extent(number of snow pixels) decreases as the NDSI thresh-old was increased from 0 to 1.0 (Figure 3a). The amountof change between thresholds varied from relativelyconstant to rapid depending on the particular scene andthreshold. Those results indicate that a range of NDSIthreshold values between 0.10 and 0.50 produces consis-tent results among images, with only small changes insnow cover between adjacent thresholds as shown inFigure 31). Generally a 10% or less change in snow coverwas observed between successive thresholds over theNDSI range of 0.10 to 0.50 for all scenes studied. NDSIthreshold values greater than 0.50 resulted in largechanges in snow extent; visual analysis indicated thatactual snow cover was eliminated at these higher thresh-olds. At low NDS1 thresholds, that is 0.20 and lower,many non-snow pixels are identified as snow. Acceptablesnow cover results have been found with NDSI thresh-olds in the range of 0.25 to 0.45.

Water bodies may have NDSI values in the range ofthose for snow, however, water has a lower near-infraredreflectance and can thus can be distinguished fromsnow. In SNOMAP, as in previous work (Dozier, 1989),an additional test for snow must be conducted after theNDSI is calculated. In order to be mapped as snow inSNOMAP, the TM band 4 reflectance must be>11%.A pixel is classified as snow covered if results of both theNDSI and reflectance tests lie within the intersection of

Figure 3. (a) In this figure, the total number of snow pixels at each NDSI threshold is mapped in different TM scenes.GNP refers to a 14 March 1991 scene of northern Montana, including Glacier National Park; Glacier Bay refers to a 6October 1992 scene covering Glacier Bay in southeastern Alaska; Sierra refers to a 10 May 1992 scene covering the SierraNevada Mts. in California; Minn refers to a 9 March 1985 scene covering part of northern Minnesota; Vatna refers to a 19October 1992 scene of southeastern Iceland including Vatnajdkull, an ice cap; Chugach refers to a 29 September 1992 sceneof the Chugach Nits., in southern Alaska. (b) In this figure, the change in snow cover mapped is shown at each NSDIthreshold value. Note that small changes in amount of snow cover mapped occur until a Normalized Difference Snow Index(NDSI) value of about 0.45 is reached.

NDSI

- GNP - SIERRA - VATNA

|-- GLACIER BAY ---- MINN - CHUGACH

z

0z

Z.

I

(b)

-100 ____ _

-50- __ _ e \ .

-80- _ \ ,

-100

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9NDSI THRESHOLD

GNP - SIERRA VATNA

GLACIER BAY - MINN CHUGACH

C'lLU -

C,

(a)

132 Hall et al.

1.0

0.5

Ma 0.0z

-0.5

-1 f

0.0 0.2 0.4 0.6 0.8TM BAND 4 REFLECTANCE

Figure 4. Snow decision region for SNOMAP criteria tests.Thresholds for TM band 4 reflectance and NDSI tests arethe solid lines. Pixels falling in quadrant one of the planeformed by the intersection of the threshold lines (shadedarea) are identified as snow. Data are from TM scenes ofGlacier National Park, Montana: 14 March 1991 (for snow,n = 2000; for cloud, n = 10,000; for snow-covered forest,n = 1750); and 03 September 1990 (for forest with no under-lying snow, n = 1050, for water, n = 1750). Scene identifica-tion numbers for the 14 March 1991 and 03 September1990 TM scenes, respectively, are # 5256917454 and# 5237717441.

the acceptance regions of these criteria tests (seehatched area on Figure 4). Pure pixels of water, forest,cloud, and snow were selected to show separation offeatures. Also shown are pixels from a snow-coveredforest in Glacier National Park, Montana from the 14March 1991 TM scene. The snow-covered forest se-lected was the same area (west of Lake Mc Donald) asselected for pure forest cover from the 03 September1990 Glacier National Park scene. Most of the snow-covered forest is not classified as snow by SNOMAP asseen in Figure 4. Adjusting the NDSI threshold down-ward would permit more snow in forests to be mappedas snow, but, in doing so, non-snow pixels may also bemapped. This illustrates the problem inherent in thethresholding techniques, and will be addressed furtheras the snow-cover algorithm evolves.

Figures 5a and b show a TM band 5, 4, 2 false-colorcomposite of northern Montana including Glacier Na-tional Park (14 March 1991), and the results of SNOMAPon that scene, respectively. Note that in addition tomapping bright, sunlit snow, SNOMAP identifies some

snow in mountain and cloud shadows and in forestedareas. SNOMAP correctly identifies most or all cloudsas non-snow features. Also, the non-snow-covered plainsin the northeastern part of the image, which were largelysnow-free at the time of image acquisition, are identifiedcorrectly as non-snow covered. The snow-covered fea-tures in that part of the image are snow- and ice-coveredlakes. Similar results have been obtained on all the TMdata we have studied so far, including other scenes inthe United States (northern Montana, and scenes innorthern Minnesota, the Chugach Mountains, Alaska;Sierra Nevada Mountains, California), Antarctica, south-eastern Iceland, and the east coast of Greenland. Weare continually seeking to identify scenes on whichSNOMAP does not accurately represent snow cover inorder to improve the algorithm in the pre-launch timeframe.

Cloud Cover ConsiderationsSNOMAP is capable of separating most clouds and snow.Cumulus clouds are generally readily distinguished fromsnow because the reflectance of cumulus clouds remainshigh in the region of the spectrum from 1.55-1.75 pm(TM band 5), whereas the reflectance of snow drops.The reflectance of cirrus clouds, however, may not beas high in this region of the spectrum if they are thinand if the reflectance from the ground cover beneaththe cloud is visible through the cloud. SNOMAP hasbeen run on TM scenes with cirrus clouds and resultsare mixed; sometimes the cirrus clouds are identifiedas clouds and other times they are confused with snow.We do not yet know at what optical thickness the cloudsbecome opaque in TM band 5.

There may be areas in which cloud cover is soprevalent that no useful snow data will be acquired ina given week due to cloud cover. Rossow (1993) hasmapped global cloud amount by latitude and season. Inhis analysis, very few areas are either completely cloudyor completely clear over a 30-day period. He empha-sizes, however, that the results in the polar regionsprobably miss significant amounts of cloudiness becauseof the difficulty in separating snow and clouds. Duringthe EOS era, with the development of global cloud-cover maps from MODIS, and the global snow and seaice-cover maps, we will gain a better understanding ofthe cloud persistence in the polar regions.

ERRORS

Assessment of the inherent errors in hemispheric-scalesnow-cover mapping has been difficult (Robinson et al.,1993). Also, the inability to map snow through denseforests is an important limitation in the use of satellitedata for mapping snow. Errors are inherent in the calcu-lation of snow-covered area due to several factors. Errors

Yw )qi/ X X I g/

We' if' ,' o~~~~X / X,

ORESTWINTER

_ ~~~~~~CLOUDFREST

SUMMER

Methods for Mapping Snow Cover 133

Figure 5. (a) TM band 5, 4, 2 color composite of northern Montana including Glacier National Park, 14 March 1991(i.d.# 5256917454).

may be caused by the presence of clouds using satellitedata from the optical part of the electromagnetic spec-trum; smaller errors due to cloud and mountain shadowsare also often present. Additionally, errors due to theeffects of topography can be large because satellite datadepict a flat surface and do not permit areal snow coverto be mapped on mountain slopes; without the use ofa digital elevation model, snow-covered area can be

greatly underestimated in mountainous areas (Hall etal., in press).

Other sources of error in mapping snow cover usingSNOMAP are to be expected due to the inability tomap snow that covers less than 60% of a pixel withoutmapping bright, non-snow features elsewhere in theimage. Furthermore, if there are any inaccuracies inthe input products, that is, inaccuracy in the calculation

134 Hall et al.

Figure 5. (b) Result of applying the SNOMAP algorithm to image in 5a: white represents snow cover and black representsnon-snow-coveredl areas in cluding (louids.

of the atmospheric correction, this may cause errors inthe amount of snowo that is mapped by SNOMAP.

Limitations Association with Analysis of Snow Coverin Forested Areas

The presence of dense forests has long been a sourceof difficulty in snoxw mapping from aircraft and satellites(Tiuri and Hallikainen, 1981; Foster et al., 1991). Ratherthan considering this an error, it is prohably better

deemed a limitation to automated, global snow-covermapping because the presence of dense forests maskssnow cover on the forest floor. Using passive-microwavedata, it is also difficult to map snow under dense forests(for example, see Hallikainen et al., 1988; Goodison etal., 1986; Foster et al., 1994). Various models have beendeveloped to permit more snow to be detected undertrees using passive-microwave data (Hall et al., 1982;(Chaiig et al., 1992a), however, the problem remains.

Mlfethods for Mapping Snow Cover 135

CNP 14 MAR 1991* SNOW AND HIGH NDVI L CLOUDS

* HIGH NDVI U SHADOWS

* SNOW [3 OTHER

Figure 6. Results of applying the Normalized Difference \ egetation Index (NDVI) to a subscene of the 14 March 1991 TMscene covering part of Glacier National Park, Montana (icd. # 5256917454). Blue depicts snow cover, dark green depictsareas of high NDVI for which snow coxer was not mapped. red depicts snow cover mapped in areas of high NDVI, andwhite, grey and yellow depict cloud cover. Light-green areas represent noni-snow-covered areas. Note the high NormalizedDifference Vegetation Index values around L1ake Mc Donald.

Analysts using NOAA data to derive Northern Hemi- Because of the uncertainty of assuming snow cover insphere weekly snow-cover maps assume a complete an automated-algorithm environment at certain timessnow cover beneath dense forests if the surrounding of the year, the MODIS snow-mapping algorithm willareas are observed to be snow covered. Using an auto- not, at least initially, contain code that checks for snowmated-algorithm approach. it cannot be determined cover in forested areas adjacent to snow-covered areaswhether snow persists (and for how long) underneath for the daily snow-cover product, because this couldthe forest canopy once snow in adjacent areas melts, introduce errors and inconsistent results between years.

136 Ihall et al.

The weekly snow-cover product will be used to developa maximum snow-extent map that will show the maxi-mum position of the continental snowline for that week.In doing so, snow cover under forested areas will bemapped.

Using the 14 March 1991 TM scene of northernMontana (covering Glacier National Park), dense forestsare defined as places where the NDVI is high (Fig. 6).The NDVI has a close relationship with the photosyn-thetic capacity of specific vegetation types (Tucker,1979; Townshend et al., 1993). The theoretical rangeof the NDVI is - I to + 1, but in practice it varies from0.03 for water to 0.05 for deserts, to 0.6 for areas withthe highest levels of photosynthetic activity (Townshendet al., 1993). The NDVI may be calculated using TMdata as follows:

NDVI = (TM band 4 - TM band 3)(TM band 4 + TM band 3)

An area of dense forest just west of Lake Mc Donaldhas an average NDVI of about 0.54 in the 03 September1990 TM image. The same area has an average NDVIof 0.38 in the 14 March 1991 TM image when snowwas present under the forest canopy. In areas of denseforests such as the area near Lake Me Donald, however,snow under the canopy is not consistently identified bySNOMAP in the March scene. In fact, 10% of areas ofdense forests were not identified as snow covered bySNOMAP even though simultaneous field measurementsrevealed a complete snow cover. Those areas where theaverage NDVI was high (about 0.38), and no snow ismapped, are shown in green on Figure 6; areas wherethe average NDVI is high and snow is mapped, areshown in red. Thus SNOMAP permits the mapping ofsome, but not all snow cover in densely forested areas.The circumstances under which snow is mapped utnderdense forests will be investigated further.

VALIDATION OF SNOWMAP

Comparison with a Spectral-MixtureModeling TechniqueAs part of the validation activities and to determine itsaccuracy for individual TM scenes, SNOMAP resultshave been compared with the results from a subpixelsnow-cover mapping technique (Rosenthal, 1993). Ro-senthal used spectral-mixture modeling and decisiontrees to map fractional snow cover on a 10 May 1992TM scene covering part of the Sierra Nevada Mountains,California; ground-trulth measurements and aerial pho-tographs were used for validation. Fractional snow-coverclasses derived from the spectral-mixture modeling rangefrom 100% snow covers, to 0% sIIow covers, with inter-mediate classes containing mixtures of snow, vegetationand rock (Rosenthal, 1993).

Based on comparison with Rosenthal's results,SNOMAP is 98% accurate in identifying pixels coveredat least 60% by snow in the 10 May 1992 TM scene ofthe Sierra Nevada Mountains. Rosenthal's results show-ing pixels covered 60% or more by snow, and theSNOMAP result, are shown in Figure 7.

Comparison with a Supervised-ClassificationTechniqueAdditionally, supervised classification was performed on6 TM scenes. The results of the supervised classificationswere then compared with the results of the SNOMAPclassification. Results of each were also later comparedinteractively with a TM band 5, 4, 2 digital reflectanceimage of each scene.

Detailed analysis of each scene indicated that, over-all, a better classification was performed using SNOMAPthan with the supervised classification, Also, SNOMAPdid a more consistent job in the snow classificationsthan was done with supervised classification. In the fournearly cloud-free images, supervised versus SNOMAPresults compared well; results of the two classificationtechniques compared less well in the two scenes (GNP14Mar91 and Ch 29Sep92) where cloud cover was asignificant factor as discussed below (Table 4).

In the case of the 14 March 1991 Glacier NationalPark scene comparison, because of extensive cloudcover, the supervised classification provided poor re-sults. The supervised classification found 39.3% lesssnow cover than did the SNOMAP classification (Table4). Also, using supervised classification, it was difficult todefine pixels in cloud shadows that were snow-coveredwithout inadvertently mapping non-snow pixels. In thecase of the 29 September 1992 Chugach Mts. scene, acirrus cloud in the northeastern part of the image wasmapped erroneously as snow by SNOMAP, but not bythe supervised-classification technique. The supervisedclassification technique found 19.8% less snow coverin this scene as compared to the SNOMAP classificationtechnique (Table 4). This is because in the supervisedclassification, the cirrus cloud was selected as a uniquefeature and mapped accordingly.

SNOMAP mapped areas of shadowed snow muchbetter than did the supervised classification, althoughin some cases (e.g., the Glacier National Park sceneacquired on 09 May 1994), the supervised classificationmapped more snow at the edges of snow-covered areas.Both techniques mapped a few, stray, non-snow pixelsoutside of the snow-covered areas. SNOMAP mappedmore snow in dense forests (e.g., around Lake Mc Don-ald on the 14 March 1991 Glacier National Park scene)than did the supervised classification. Interestingly,SNOMAP did not map very dark glacier ice as snow onthe Iceland scene covering Vantnaj6kull ice cap, whereasthe supervised-classification technique did. Using either

Metbods lor MIapping Snow Cover 137

Figure 7. Comparison of results of mapping snow on the 1() \tlax 1992 lINI scene (id. # .5299217572) of the Sierra Neva-das in California using results from Rosenthal's (1993) spectral-inixtore inodeling techni(pue, wvith SNOMAP results.

technique, when there is insufficient signal froln tiesnow, as in the case of completely shadowed pixels. (itpixels under a dense tree caanopx, there is no meinas ofdetermining ground cover using optical sensors.

DISCUSSION

Analysis of the prototype \ersion of our sn0-miappinLgalgorithm, SNOMAP, has shown it to be effective inmapping snow cover on TM scenes in wNhich a \arictNof surface covers is represented. SN(O)MAP provided(Imore consistent results than were derixed from a super-vised-classification technique. (Compared with a T\Iscene of the Sierra Nevada Moutains. mnapped I spe( tral-mixture modeling, SNOMAP was 98% accurate mc

ma~ppilng slio\s in pixels that e-erc at least 60% snowco veredl.

An estimnate of the limitation, globally, in mappingsno\o-cov\red area (hie to dense forests will be devel-oped in the pre-lauinch time framne. As land-cover mapsi1m0prove, we will be al)le to determine how much of

the land is covered by forests that are too dense topenrrlit s51ow to be adequately mapped below using thecemirent algorithm. It will be possible to apply a differentalgorithm to densely forested areas that will Inore accu-rate1', identify snow underneath forests.

Work is ongoing to determine if' the thresholds cur-

eiti tised iii SNOMAP will apply universally. If, forexamplte. images acquired when the solar elevation iskXm requimire a (lifferent threshold setting, this can be

138 Hall et at.

Table 4. Snow-Covered Area (SCA) in km2 and Percentageof Full TM Scene Determined Using Supervised VersusSNOMAP Classification Techniques

Supervised SNOMAP Percentage of

km2 (percentage) km2 (percentage) change in SCA

GNP 14Mar91 6,450 (19.1) 10,631 (31.5) -39.3CNP 06Mar94 10,253 (30.3) 10,953 (32.4) - 6.4CNP 09Mav94 4,126 (12.2) 4,006 (11.9) 2.9Ch 29Sep92 12,841 (38.0) 16,021 (47.5) - 19.8vat 19Oct92 12,020 (35.6) 13,033 (38.6) - 7.8MN 09Mar85 19,443 (57.6) 21,534 (63.8) -9.7

CNP refers to Clacier National Park, Montana; Ch refers to the Chugach Mts., Alaska; vatrefers to Vatnajbkull, Iceland; and MN refers to northern Minnesota. Percentage of changerefers to the change in the amount of SCA mapped using the two different approaches formappimsg snow cover, relative to the SNOMAP result.

accommodated in the algorithm. The next challenge inthis ongoing research is to locate and analyze TM scenesthat were acquired tinder conditions of low solar ele-vation.

The MODIS snow maps will complement the suiteof planned MODIS products (Salomonson and Toll,1991; Running et al., 1994). Land cover, surface albedo,sea ice cover, net-primary productivity, surface tempera-ture, and other products will be available for inter-comparison. SNOMAP can be applied using MODIS1-km and 500-m resolution data, thus enabling bothglobal- and regional-scale measurements to be made,though only 500--m daily and 7-day composite global mapswill be produced as MODIS products. Two 250-m reso-lution bands will permit more detailed, but limited,mapping of regional snow-cover characteristics.

MODIS-derived global snow-cover data productswill be available to use in general circulation models(GCMs) and will represent a consistent snow-cover dataset for long-term climatology studies. Snow-cover datawill also be available to use as input into regional-scalehydrological models to permit improved estimates ofrunoff, and for hydrological- ancl energy-balance mod-eling.

In September 1995 a MODIS snow and ice work-shop was held at Goddard. An objective of the workshopwas to seek advice from the snow and ice communityon the MODIS snow product. There was considerablediscussion and many excellent ideas were put forth. Asa result of the workshop, the snow product will bemodified in the pre-launch time frame so that it willbetter meet the needs of potential users. Thus, thereare likely to be enhancements to the product in thenext couple of years that are not discussed in this paper.

Advanced-classification techniques. currently beingdeveloped (Nolin et al., 1993), permit sunpixel classifi-cation and improve the identification of fractional snowcover. It is likely that spectral-mixture modeling, oranother advanced-classification technique that does notrely on thresholcling, will be implemented after ]asmuch

when we can "train" on the entire globe to select suitableendmembers for analysis. If a better snow-cover productcan be obtained using advanced-classification techniques,then all data will be reprocessed in the post-launch timeframe to provide an optimum product.

Outlook for the FutureThe combined use of visible, near-infrared, and micro-wave sensors is likely to offer a better way to map snowextent and water equivalent in the future. Much workis already ongoing in this area (Chang et al., 1992b;Salomonson et al., 1995; Grody and Basist, in press).Because passive-microwave sensors are generally un-affected by cloud cover over snow-covered areas, it willbe advantageous to use MODIS data in conjunction withAdvanced Microwave Sounding Radiometer (AMSR)data to map daily snow extent and water equivalent.The AMSR may be flown on the EOS PM satellitealong with MODIS early in the next century (personalcommunication, Claire Parkinson/GSFC). The spatialresolution of the passive-microwave data on the AMSRmay be as good as 5 km, thus AMSR will be far moreuseful for mapping snow parameters than are the satel-lite-borne passive-microwave sensors of today, whichhave a maximum spatial resolution of about 15 km.

CONCLUSION

It is concluded that the planned MODIS snow-coverproducts will represent an advance over products pres-ently available because the daily and weekly maps willbe derived automatically using the unique capabilitiesof the MODIS at 500-m resolution (e.g., snow/clouddiscrimination using spectral bands that are not availabletoday and ability to derive spectral reflectance). Addi-tionally, there will be snow-cover persistence statisticsderived for each MODIS 500-m pixel globally from theweekly maps. Results also show that an automated-classification technique gives more consistent results

Alethods for Mapping Snow Cover 139

than does a supervised-classification technique. TheSNOMAP algorithm will continue to evolve before andafter the planned 1998 launch of MODIS and if more-advanced snow-cover mapping techniques emerge, thendata will be reprocessed after launch. In short, the MODISsnow-cover products as part of a suite of MODIS prod-ucts, will allow us to improve our ability to monitor andmodel key geophysical processes.

The authors would like to thanrk Walter Rosenthal / Universitof California at Santa Barbara, for discussious concerning re-sults of fractional-snowr mapping weork; Janet Chien / GeneralScience Corporation, Laurel, MD, for image processing of theTAI data; and Tom Carroll of N(O-A /INational HydrologicRemote Sensing Center, Minneapolis, Minnesota; Jim FosterNASA /GSFC,; and Anne Nolin/ Univ(ersity of Colorado, Borl-der, Colorado, for their reviews of the paper.

REFERENCES

Allen, T. R. arid Walsh, S. J. (1993). (:liaracterizirg mirltitemi-poral alpine snowvielt patterrts for ecological inferenees.Photogram. EIrg. Renmote Seis. 59(10):1521-1529.

Bromwich, D. H., and Tzeig, R.-Y. (1994), Simulation ofl thmodern arctic climate by the NCAR CCMs-IL. J. Climate7:1050-1069.

Bunting, J. T. an(ridEitreriorit, R. P (1982), Improved clouddetection ritilizing dtele nse meteorological satellite progrannear infraredr measurremrserits. Aim Foree Geophssics Lahora-torN, Hanscori AFB. MA, AF(.I.-TR-X2-01(27. Encirolioen-tal Research Papers No. 76.5 91p1-

Carroll, T, i s(1990), O)peratiomal ai rror-ne and satellite silo\s

tover prodrcts of the Nationral ()peratiorral lIvdrologi -

Remote Sensing Center, Proteediings of the Forty-seenthbAn nual Eia stern Snot ('onJerenrcc. Jllte 7-8. 1990, Barigoin-Maine, (.RREl. Special Report 9(0-4-4.

Carroll, T. R.. Blaglio. Jr..1. \.. NVerdir. J. P. arid Ilolro d, 111.E. W. (1989). Operational tiiapp)inig oi snowN mose in tlnUnited States anrd ( Carada disirg airn-bonie aid satellite data.Proceedings of the 121/i Ca nadianii Str ipo.sioni t oti ReiornoSensing, V.3, I(ARSS '9, 1)--14 Jl ' 1989. Varicoimet.

(Cariada.(:harg, A. 1. ( ., lFiostci. J I . atih hiaiigo. \ (I 992a!) TlI(

role of' paSsix n ncitlo\sN seas ito -11ar.lcterizirng Sno\\ (t0\('l ill

tlre (Colorado Ii iver Basin. co jurnaul 26A.3:31I-3XX.Clhaig, A. 1T. C.. Foster. 1. L.. liall. I). K. Rtiolisorr. 1). \.

Pieji, L., anrd Mlishring. t . J 9921. olin tse of mini. ro\\a\s

rad1iomnete r data for cli a r a-tn rizir g sniow stitra.ge ill we str 1-

(hIirna, :Xtiri. (;lritiol. 16:215 219

Dozier, J. (1984). Sneis\ relleet ari(e tiomnI .andsla.t-4 tdernratirrinapper. I. It. F. - I. Tran Gco-0 v- tri. Bc i tnic Si -us. 22:323-32S

D:)ozier'. (I9')X9), Spectral Sigruitullt-(i tillpiril Slut\S' i.l\l- tro

thle L ,arndsat I'lienrati- NIl apper. ietrotc Sinus E tirotll. 2S

Dozier. J., Schneideir. S B .B inI \lai.-iliis. D). I.. Jr. 19SIlEffect of' grain sizt arid slu'\\p.ack \\.altt' e(ii'.aleit-( (\iSitle r- a (r uel ar-inulraredl ,it ellit- obisrl\.atiolls of snno

Uatcr Ri'soturrr-s R,. I 7: 1 )-- 121IFleig \. J . Iltubanks. I' .\- Stres,\ I I - a ( .urprnta-ilt. 1.

(in press), An analysis of MODIS earth location errors,Version 2.0.

Foster, J. L. and Chang, A. T. C. (1993), Snow cover, In Atlasof Satellite Observations Related to Global Change, R. J.Gurney, C. L. Parkinson, and J. L. Foster (Eds.), CambridgeUniversity Press, Cambridge, pp. 361-370.

Foster, J. L., Chang, A. T. C., Hall, D. K., and Rango, A. (1991),Derivation of snow water equivalent in boreal forests usingmicrowvave radiometry, Arctic 44 (Suppl. 1):147-152.

Foster, J. L., Chang, A. T. C., and Hall, D. K. (1994), Snowmass ii bhoreal forests derived from a modified passivemicrow-ave algorithm, Proceedings of the European Sympo-Sinn o0n Satellite Remote Sensing, Conference on MicrowaveSensing for Forestry and Hydrology, Rome, Italy, Septem-her 1994.

(;oodison, B. E., Rubinstein, I., Thirkettle F. W., and Lan-ghalum E. J. (1986), Determination of snow water equivalenton the Canadian Prairies using microwave radiometry,IAHS Publication No. 155, 163-173.

(;rodv, N. C. and Basist, A. (in press), Global identification ofsnoweover using SSM / I measurements.

(;utler. I). S., and Rosen, R. D. (1992), Interannual variabilityof wintertime snow' cov er across the Northern Hemisphere,J. (limuate. 5:1441-1447.

Ilall. D. K., Foster, J. L., and Chang, A. T. C. (1982), Measure-ment anrd modeling of microwave emission from forestedsrowvbelds in Michigan, Nordic Hydrol. 13:129-138.

llall, 1). K., Foster, J. L., and Chang, A. T. C. (1992), Reflec-tance of snow as measured in situ and from space insubrlarctic areas in Canada arid Alaska, IEEE Trans. Geosci.Reoiotc Snlls. V.30(3):634-637.

llall, I). K., Foster, J. L., Chien, J. Y. L., and Riggs, G. A. (in

press), D)eterminationr of actual snow cover using LandsatFNM and digital elevation model data in Glacier NationalPark, Montana, Polar Rec.

Iflallikairreri, M, T., Jolma, P. A., ard Hyvppa, J. M. (1988),

Satellite microwave radiometry off forest and surface typesin Finland, IEEE Trans. Geosci. Reorote Sens. 26:622-628.

King, MI. 1)., Kaufman, Y. K., Menzel, W. P., and Tanre, D.i 1992), Renmote sensing of cloud, aerosol, and water vaporproperties from the Moderate Resolhtion Imaging Spectro-radiometer (MOI)I8), IEEE Trans. Ceosci. Remiote Sens. 30:2-27.

Ks I(. If. L.., (:rrarr, R. J., Bal-rnes, W. L., and Escoe, D.1l1S, N) clouid physics radliorneter, Third Conference on

Atonosplhcric Radiation, IaXis, CA, pp. 107-109.Mlarkinari. 1i. L., antI Barke. J. L. (1986), Landsat MSS anrd

*lM\1 post-calibration dvrna]unc ranges, exoatmospheric re-tlectarc-es arid at-satellite terniperatirres, in EOSAT Techni-,al Notes, No. 1, 3-8.

\latsori, NM. Roeplewski, C. F., and Varriadore, M. S. (1986),Aol -lt/o.s of Satellite-l)erived .Voetthern Hemisphere Snow(Co (r lre(reneny, Nationral Weathier Service, Washington,

\latsori. \., (1991), NOAA satellite snow cover data, Palaeo-rog, rinld Palcoe-ol. 90:213-218.

Nt)A:V\ (Natiorral Oceanic and Atmospheric Administration),1l994). Iiftl A;Xutal Climate Assessmeurt 1993, Climate

Baksi,, Center, ( anp Springs, NIarylarnd, March 1994,

Nol ii. V. D)ozier, J., antI NMertcs, I. -. K. (1993), Mapping

140 llill et al.

alpinc m snosw Usilmg a spectral mnixtimre mmoldclimg techmiqmme.

Ann. Glaciol. 17:121-124.Rango, A. (Ed.). (1975). ()pcratiomnal Applications o/ Satellite

Snouwcover Observatiomms, NASA SP-391, 43() pp.

Rango, A. (1993), Simow hyldrolog ploc(tsses anmd remote semns-

ing, Hy~drol. Proc. 7:121-138.Rango, A., anld Martintcm , J. (1982). SnoiN aci( umlm dation deriv ed

from modified depletion cur\es tf fsnowx coverage, ,Sym)po)m -siuni orm lIty1drological Aspe(t.s ol Alpilme and High Monntaimm

Areas. in Exeter, IAIIS Pubmlicatiomm No. 1.38, Plp 8:3-90.

Riggs, G. A., llall, D. K., and Salomor ison. \ . V. (1994), A sno\\index for the LIanmsdsat Themnatic XI fapper amid Moderate

Resol tioh i Imaging Spectroradiom ne ter, Proceedin gs of theInternational 01Gcoscience and( Reolote Semn.snmg Srynmpo~sinin.I(ARSS'94. 8-12 Augmist 1994. Pasadema. CA. pp. 1942-

1944.Robinson. 1). A. n a Km Kukla. C. (1985), Nlaximiunu smirLace

alhedo of seasonally sno" covm lretd lamids ill the Nortl meromI lemisphere. 1. Cl mniate A.ppl. Ale teoroml. 2:4402-411.

Rolsinsostm, D. A. andl 1)eNwet, K. F ( 199(1). Pie cent see lai

variatiomms iil the extent of NOIrthermm Ilemiispiere snIO\Sc oer. Geopli s. Rles. Le tt. 17 :155, 156f(

Rohinsom, 1). A. Keimnig, F. T., amlnd I)ewtv , K. F. (1991).

Recemit Variatiolms ill Northern H ilem isphere smiIoS (ON er

Procm'ed ings of time I .t/b A.m ii0l C/ia mm ti 1 I )iagmm sm ic s W ork-

smop. Asheville, NC:. N()AA. p. 219 221.Ro]imIsmn. 1). A., I)e Ktv. . F. amid HI timi. R. R., Jr. (199t3).

Glohal simmi\ civem IlmollitOrlimlg Aim idllilitc. 13m/ .. AI i. Mute-oiol. So(. 74:1689-169(.

Rosemnthal (C. WN. (1993). \Mappinmg mimimitalit smios c ostr iat

smmhpixel resomlution 1rom the 1 .diltlSnd t thi'imattic Immppll)tl.

NI.A. thesis, I )(epart mitmmt of (.tmog rapl In U is m rsih ivo ( Cali-lonlia at Samtia Bamrhara. 70 pp

Rossosw, \\'. B. (1s99:3) (. ( 1lm It/os ohm i/l Ohcsvriatiiln.s

Rlletated o (Glob)al (liammgm . :3 l (; rminlt . I. l.. Foster.

atld (C. L. Pai-kimisomm. Etlds., ( alm mritlgu lVnusitrsit\ Prmss.

Cambridge. Fnmglam d. pp. 141 1 6:3.

RIsmmimig, S. WN.. JtSmtieC. (: . SalOMmiosoui, V., it al. 1994).

Terrestrial rmemmote sm'nmsimmg scimime .imd aligtimithsins p)lamsitled

for ()S NI()DIS, hit. J. Riomotc Setmi I 517 ):3587-:362(Salommmonisom \V. V.. autd Toll, 1) I. 1991 ). Exe(mititii phiimsi

( I 1)) spitrlal handt helsaiettristits s l liii 11)8 lluOdiratti

resolution imaging spectrometer-nadir (MODIS-N) facilityinstrument, Adv. Space Res. 11 :231-236.

Salomnimson, V. V. and Barker, J. L. (1992), EOS ModerateResolution Imaging Spectroradiometer: Phase C / D statusantI comments on calibration and georeferencing ap-proaches, Proceedings of the 15th Annual AAS Guidanceand Control Conference. 8-12 February 1992, Keystone,(Colorado, 1-14.

Salonmonson. V. V., Toll, 1). ., anmd Lawrence W. L. (1992),Moderate Resolution Imaging Spectroradiometer (MODIS)ancl observations of the land surtace, Proceedings of theInternational Geoscience and Remote Sensing Society AnnualMeeting, 26-29 May 1992, Houston, TX, pp. 549-551.

Salon1ommsomm, V. V., Hall l). K., and Chien, J. Y. L. (1995),Ulse of passive microwave anld optical data for large-scalesn1ow-coxer mapping, Proceedings of the Combined Optical-Microwave Earth an0( Atmnospheric Sensing, 3-6 April 1995,Atlanta. GA.

Stefftn, K. (1987), Bidirectional reflectance of snow, In LargeScale Effrets of Seasonal Snoic Cover, B. E. Goodison, R. G.Barry, ani( J . Dozier, Eds. Proceedings of the IAHS Sympo-siun held in Vancouver on 19-22 August 1987, IAHS,Vancouver, Canada., pp. 415-425.

T'iuli. M. and 1lallikaine, Ml. (1981). Microwave emissionclharacteristics of snow covered Earth surfaces measuredbN the Nimbus-7 satellite, 11th European Microwave Con-jf'renec, Amsterllami. Septemher 1981. 233-238.

tok nshemmd, J. R. G., Tucker, C. J. and Coward, S. N., (1993),(Johal vegetatiomi mapp)imig, Iu Atlas of Satellite Obs)rea-timom.s Related to Global Chanmge R. J. Gurney, J. L. Fosteramdl (C. 1. Parkinsomm. (Eds.). Cambridge University Press,I.onmon, England, pp. 301-311.

I'ucker, (i J. (1979). Red ancl plotographic infrared linearco mbhinmatiomms f'or mo iitoring v (egetation, Remote Sens. Envi-omm. S: I27-15(0.

Ilicker. (:. .1. (19'86) \laximmmm miormalized difference vegeta-timm imdex images for sub-Saharan Africa for 1983-1985,

lot. 1. lRinmmotv Sens. 7:1383-1384.\Valsh. J. V.. Jasperson, X. It., ammd Ross, B. (1985), Influences

if smoN% cover anmd soil mioisture on monthly air tempera-tmre, Mlonthl/y Wcatlethe let. 11:3:7,56-768.

A inother, l.-G. (1992), Lamndsat thematic mapper (TNf) derivedrefleetamme from a lin m m taillonms \\ atemsh ed doming the snowmelt Stcasoiim .NVordic llymdrol. 23 :27.3-290.