AMSR-E algorithm for snowmelt onset detection in sub-arctic heterogeneous terrain

HYDROLOGICAL PROCESSESHydrol. Process. 21, 1587–1596 (2007)Published online in Wiley InterScience(www.interscience.wiley.com) DOI: 10.1002/hyp.6721

AMSR-E algorithm for snowmelt onset detection in sub-arcticheterogeneous terrain

Jeremy D. Apgar,1* Joan M. Ramage,1 Rose A. McKenney2 and Patrick Maltais3

1 Earth and Environmental Sciences, Lehigh University, Bethlehem, PA 18015, USA2 Geosciences/Environmental Studies, Pacific Lutheran University, Tacoma, WA 98447, USA

3 Water Survey of Canada, Environment Canada, 91782 Alaska Highway, Whitehorse, Yukon Territory Y1A 5B7, Canada

Abstract:

The onset of snowmelt in the upper Yukon River basin, Canada, can be derived from brightness temperatures (Tb) obtained bythe Advanced Microwave Scanning Radiometer for EOS (AMSR-E) on NASA’s Aqua satellite. This sensor, with a resolutionof 14 ð 8 km2 for the 36Ð5 GHz frequency, and two to four observations per day, improves upon the twice-daily coverage and37 ð 28 km2 spatial resolution of the Special Sensor Microwave Imager (SSM/I). The onset of melt within a snowpack causesan increase in the average daily 36Ð5 GHz vertically polarized Tb as well as a shift to high diurnal amplitude variations (DAV)as the snow melts during the day and re-freezes at night. The higher temporal and spatial resolution makes AMSR-E moresensitive to sub-daily Tb oscillations, resulting in DAV that often show a greater daily range compared to SSM/I. Therefore,thresholds of Tb > 246 K and DAV > š10 K developed for use with SSM/I have been adjusted for detecting the onset ofsnowmelt with AMSR-E using ground-based surface temperature and snowpack wetness relationships. Using newly developedthresholds of Tb > 252 K and DAV > š18 K, AMSR-E derived snowmelt onset correlates well with SSM/I observations inthe small subarctic Wheaton River basin through the 2004 and 2005 winter/spring transition. In addition, the onset of snowmeltderived from AMSR-E data gridded at a higher resolution than the SSM/I data indicates that finer-scale differences in elevationand land cover affect the onset of snowmelt and are detectable with the AMSR-E sensor. On the basis of these observations,the enhanced resolution of AMSR-E is more effective than SSM/I at delineating spatial and temporal snowmelt dynamics inthe heterogeneous terrain of the upper Yukon River basin. Copyright 2007 John Wiley & Sons, Ltd.

KEY WORDS passive microwave; SSM/I; AMSR-E; Wheaton River; Yukon River; snowmelt

Received 17 June 2006; Accepted 18 October 2006

INTRODUCTION

Hydrologic and atmospheric processes are closely tiedwith snow characteristics throughout the arctic and sub-arctic regions of the Northern Hemisphere. The extentof snow cover, snow water equivalent (SWE), and thetiming of spring snowmelt influence local hydrologyas well as regional and global water budgets, and thehigh albedo of snow has an important influence onthe global energy budget (Vorosmarty et al., 2001). Thespring snowmelt is an annual event in snow-covered,high-latitude drainage basins that sends significant quan-tities of meltwater downstream during a relatively shortperiod of time. The timing and magnitude of these eventsare variable, but they can create a large impact onthe basin hydrology and influence geomorphic changewithin the basin. The use of passive microwave datato examine the characteristics associated with this sea-sonal snowmelt may provide a better understanding ofthe sensitivity of snowmelt timing to climatic conditionsas well as the effects of the snowmelt upon basin hydrol-ogy.

* Correspondence to: Jeremy D. Apgar, Earth and Environmental Sci-ences, Lehigh University, 31 Williams Drive, Bethlehem, PA 18015,USA. E-mail: [email protected]

The Yukon River extends from the Juneau Icefield,British Columbia, to the Bering Sea, passing through theYukon Territory and Alaska. It receives drainage from abasin more than 850 000 km2 in area. The majority of thisdrainage flows from the Bering Sea northward to provide8% of the total fluvial freshwater input to the ArcticOcean (Aagaard and Carmack, 1989). The basin includesa variety of ecosystems as well as a range of elevationsand local relief. This region has a heterogeneous terrainmarked by boreal forests, mountain ranges, low-lyingvalleys, and a variety of lakes. Much of the upper basinis seasonally snow covered aside from some glaciatedregions near the headwaters.

This paper focuses on the timing of the springsnowmelt transition in the Wheaton River sub-basin ofthe Yukon River (Figure 1), from which the findings maythen be applied to larger sub-basins and eventually theupper Yukon River basin as a whole. The Wheaton Riverbasin is a small (875 km2) upland basin that has hetero-geneous topography and terrain with only a small glacialinput.

Two passive microwave sensors, the Special SensorMicrowave Imager (SSM/I) and Advanced MicrowaveScanning Radiometer for EOS (AMSR-E), provide atleast twice-daily observations in the high-latitude regions,and the record of SSM/I observations goes back nearly

Copyright 2007 John Wiley & Sons, Ltd.

1588 J. D. APGAR ET AL.

Figure 1. Wheaton River basin EASE-Grid (a) 25 ð 25 km2 pixels and(b) 12Ð5 ð 12Ð5 km2 pixels superimposed on a 1 : 250 000 (90 m) digitalelevation model (Environment Yukon, 2000). While the actual sensorfootprints are more elliptical in shape, these square pixels provide agood representation for the land area covered. The inset map showsthe location of the basin within the Yukon Territory. The basin (blackoutline) is represented by parts of six 25 ð 25 km2 pixels and fourteen12Ð5 ð 12Ð5 km2 pixels. The Wheaton River near the Carcross streamgauge is located near the river’s mouth at an elevation of 668 m (white

dot in pixel B02 and B02B)

two decades. Due to differences in brightness temper-ature emitted from dry versus wet snow, these pas-sive microwave sensors are effective at differentiatingbetween snow that is melting and snow that is not melt-ing, or frozen. In addition, the ability of these sensorsto observe snow characteristics through clouds and mostother atmospheric conditions, as well as during the night,proves useful for examining snowmelt in high-latitudeareas that may have increased cloud cover and darknessduring the winter and spring months.

The purpose of this study is to derive the timingof spring snowmelt in the Wheaton River basin withrecently acquired AMSR-E observations and to test thesensitivity of this sensor to the dynamic and variedregional snowpack characteristics. A snowmelt onsetalgorithm developed by Ramage et al. (2006) for usewith SSM/I data is modified for use with AMSR-E datato allow for a direct comparison of snowmelt onsettiming from both of these passive microwave sensors. Inaddition, higher-resolution AMSR-E data and elevationdata are used to show improvements of this sensorover the SSM/I sensor in detecting snowmelt within

heterogeneous terrain. The algorithm is fine-tuned usinghourly near-surface air temperature data and ground-based snow wetness measurements from within theWheaton River basin. Due to the greater spatial andtemporal resolution of the AMSR-E sensor comparedto SSM/I, snowmelt dynamics can be more accuratelydescribed with AMSR-E data in the mixed terrain of theupper Yukon River basin.

DATA

Passive microwave satellites provide a means by whichto differentiate between melting and frozen snow. BothAMSR-E and SSM/I brightness temperature data wereused for this analysis. In addition, 90 m digital elevationmodels of the Yukon Territory were utilized (Environ-ment Yukon, 2000), as well as hourly temperature mea-surements and snow wetness measurements from withinthe Wheaton River basin.

AMSR-E swath data, in the form of Level-2A globalswath spatially re-sampled brightness temperatures (Tb)V001, are supplied by the National Snow and Ice DataCenter (NSIDC) through the EOS Data Gateway witha record of observations that extend back to 2002(Ashcroft and Wentz, 2004). The AMSR-E sensor pro-vides two to four observations of the study area perday and has a mean pixel resolution at 36Ð5 GHz of12 ð 12 km2 (actual footprint size is 14 ð 8 km2). Twice-daily SSM/I data extending back to 1987 come fromthe Defense Meteorological Satellite Program (DMSP)and have a nominal resolution for the 37 GHz chan-nel of 37 ð 28 km2. SSM/I images gridded to the EqualArea Scalable Earth-Grid (EASE-Grid) 25 ð 25 km2 res-olution are provided by the NSIDC, and this prod-uct separates the ascending and descending passes.The AMSR-E data were gridded first at the EASE-Grid 25 ð 25 km2 resolution to compare directly withthe SSM/I data and establish to what degree they aresimilar with a minimum of complicating factors. TheAMSR-E data then were gridded at the finer EASE-Grid12Ð5 ð 12Ð5 km2 resolution to examine improvements insnowmelt detection of AMSR-E upon the SSM/I sen-sor.

Passive microwave sensors have been used to examinecharacteristics of snow such as snow extent (e.g. Walkerand Goodison, 1993; Abdalati and Steffen, 1997; Wanget al., 2005), snow depth (e.g. Josberger and Mognard,2002; Kelly et al., 2003), and SWE (e.g. Goita et al.,2003; Derksen et al., 2005; Foster et al., 2005). TheSSM/I sensor has been used in previous studies to estab-lish the timing of the spring melt transition in the upperYukon River basin (Ramage et al., 2006) as well as in theJuneau Icefield (Ramage and Isacks, 2002, 2003). Eventhough the SSM/I sensor provides twice-daily observa-tions and has been shown to correlate well with ground-based brightness temperature measurements over fairlyhomogeneous terrain such as the Alaskan North Slope(Kim and England, 2003), the pixel resolution of greater

Copyright 2007 John Wiley & Sons, Ltd. Hydrol. Process. 21, 1587–1596 (2007)DOI: 10.1002/hyp

AMSR-E ALGORITHM FOR SNOWMELT ONSET DETECTION 1589

than 25 ð 25 km2 that results from the passive nature ofthe sensor is a problematic issue in monitoring dynamicchanges over heterogeneous terrain. The AMSR-E sensor,recently launched aboard NASA’s Aqua satellite in 2002,provides more observations of the study area per day and,with an improved pixel resolution over SSM/I, can pro-vide a more accurate examination of snow characteristicsover mixed terrain.

A significant difference in brightness temperature (Tb)between dry and wet snow occurs at frequencies greaterthan 10 GHz. The Tb of a material is related to its surfacetemperature (Ts) and emissivity (E),

Tb D ETs. �1�

A rapid increase in emissivity occurs as a result ofa small amount (¾1–2%) of liquid water within thesnowpack, causing the Tb to increase for wet snow(Ulaby et al., 1986). The Tb in the 19 and 37 GHzfrequencies associated with the SSM/I sensor are usefulin detecting melt on glaciers (Ramage and Isacks, 2002,2003) and on heterogeneous terrain (Ramage et al., 2006)since the Tb transition from dry to wet snow occurs assurface temperatures approach 0 °C. From the AMSR-Esensor, the Tb from the vertically polarized 36Ð5 GHzfrequency (wavelength of 0Ð82 cm) is comparable tothe SSM/I sensor for the detection of snowmelt. Here,we compare and contrast Tb from the two sensors andcompare their ability to detect the onset of snowmelt.

During the spring snowmelt, the snowpack experiencescyclical daytime melt and nighttime freeze, changes thatare manifested as diurnal differences in Tb. As a result ofthe at least twice-daily observations by the sensors, thesehigh diurnal amplitude variations (DAV) of Tb can bedetected and are useful in identification of these dynamictransition periods.

The Wheaton River basin is covered by six EASE-Grid25 ð 25 km2 pixels (Figure 1(a) and Table I). Due to thecoarse pixel size in relation to the size of the basin, somepixels only cover a small percentage of the basin (seeTable I). The Wheaton basin has a range of elevationsand land cover that include bare mountain slopes, denseboreal forest, and slopes of varying aspect (Brabets et al.,2000; Ramage et al., 2006). Pixels A02 and A03 coverthe upland headwaters of the Wheaton River and havehigh mean elevations. The rest of the basin is mostlycovered by pixels B03 and B02, representing the middleand lower parts of the basin respectively. Pixels C02

and C03 cover a small fraction of the basin but can beused to examine the lowest elevations of the basin. Inrelation to the AMSR-E data gridded to the EASE-Grid12Ð5 ð 12Ð5 km2 resolution, the Wheaton River basin iscovered by parts of 14 pixels, improving the spatialscale at which differences in snowmelt characteristicscan be investigated (Figure 1(b)). Adjacent pixels inthis area appear to have similar Tb signatures thatdiffer only slightly due to differences in land cover aswell as elevation and topography. Seasonal variationsrelated to frost, vegetation, and lakes are also factorsin this complex landscape, but snowmelt has the largestinfluence on Tb.

Near-surface air temperatures for the Wheaton Riverbasin were acquired in 2004 and 2005 using aHOBO automatic temperature logger located within pixelB02 at 668 m, approximately 10 m from EnvironmentCanada’s ‘Wheaton River near Carcross’ stream gauge(60°0800500N, 134°5304500W). The HOBO recorded hourlyair temperature and relative humidity and was situatedabout 1Ð5 m above the forest floor on the north-facingside of a spruce tree, and thus was shaded duringmost daylight hours. It was installed under a protectiveplatform in August 2004, and data were downloadedperiodically by Pat Maltais of the Water Survey ofCanada.

A field campaign was carried out during the spring of2005 to collect a variety of in situ snowpack measure-ments from snowpits, including snow density and thedielectric properties of the snow at various depths, attimes that coincided with AMSR-E satellite overpasses.The dielectric properties of the snow, and ultimately thesnow liquid-water content as a percentage of the vol-ume, were derived from these measurements to relateto AMSR-E brightness temperature data. Other observa-tions of the snowpack in the upper Yukon River basinprovided support to show that the response in Tb couldbe attributed to snowmelt instead of liquid precipitationevents.

METHODS

Ramage et al. (2006) developed a snowmelt onset algo-rithm for use with SSM/I 37V GHz data to detect the tim-ing of snowmelt onset within the Wheaton River basin.This algorithm determines the onset of snowmelt basedon the day of the year when certain thresholds for both

Table I. Wheaton River basin EASE-Grid 25 ð 25 km2 pixel characteristics (from Ramage et al., 2006)

PixelID

EASEX

EASEY

Latitude(°)

Longitude(°)

Meanelevation (m)

Minimumelevation (m)

Maximumelevation (m)

Relief(m)

% in WheatonBasin

A02 267 266 59Ð84 �135Ð30 1435 720 2192 1472 8A03 268 266 60Ð00 �135Ð61 1576 802 2479 1677 30B02 267 267 60Ð00 �135Ð30 1177 672 2193 1521 42B03 268 267 60Ð17 �135Ð30 1373 798 2193 1395 48C02 267 268 60Ð17 �134Ð69 1084 691 1998 1307 4C03 268 268 60Ð33 �135Ð00 1051 717 1883 1166 6



Table II. Snowmelt onset thresholds

Sensor Tb (K) DAV (šK)

SSM/I 246 10AMSR-E 252 18

brightness temperature (Tb) and DAV are met (Table II).For the SSM/I data, ‘DAV Melt’ occurs when DAV isgreater than š10 K, and ‘Melt Onset’ occurs on the daywhen Tb is greater than 246 K and DAV is greater thanš10 K, indicating the snow is both melting periodicallyand experiencing strong melt/re-freeze cycles (Ramageand Isacks, 2002, 2003). In addition, ‘End DAV’ repre-sents the end of the transition when DAV falls belowš10 K, and the duration of the transition period is cal-culated as the number of days between ‘Melt Onset’ and‘End DAV’.

A number of assumptions must be made in detect-ing the onset of snowmelt using this passive microwavedata. The heterogeneous and mountainous terrain withinthe upper Yukon River basin most likely creates a com-plex Tb signal. However, based on field observations andlocal weather records, we assumed that the greater partof the land is snow covered during the spring transitionperiod. In addition, since the methods focus on snowmeltevents during the spring transition (March through May),intermittent snowmelt and thaw events during the wintermonths are largely ignored. Field observations also sup-port that the events identified with the passive microwavedata as melting snow are representative of snowmelt andnot liquid precipitation events.

AMSR-E swath data were processed using the AMSR-E Swath-to-Grid Toolkit (AS2GT) in NSIDC’s PassiveMicrowave Swath Data Tools (PMSDT). The data werethen gridded to the EASE-Grid Northern Hemisphereprojection with a nominal resolution of 25 ð 25 km2,the same format and pixel resolution as the SSM/I dataprovided by the NSIDC. The Tb were extracted from eachchannel using Interactive Data Language (IDL) code,creating a chronological time series of all Tb observationsfor 2004 and 2005.

Since there are two to four AMSR-E observations perday, DAV cannot simply be calculated as the differencefrom one observation to the next, as is done with thetwice-daily SSM/I observations. Extra observations in agiven day are often less than 2Ð5 h from the previousone and have fairly similar Tb (Figure 2). These extravalues cause the DAV to artificially approach 0 K asa result of similar Tb values even if the actual dailyoscillation is much greater. An alternative to the raw datais to average the observations that are less than 2Ð5 hapart (Figure 2). This creates a total of two observationsfor each day and eliminates extraneous DAV valuesthat approach 0 K, allowing for more representativeDAV values to be calculated and used for examiningsnowmelt characteristics in a similar way to SSM/I DAV.Other techniques used to calculate DAV, including usingdaily minimum and maximum values, did not provide

Figure 2. Calculation of DAV from AMSR-E 36Ð5V GHz Tb. In order toexamine daily fluctuations, Tb less than 2Ð5 h apart (medium gray circles)were averaged. The running differences of these observations with theother observations produces a more accurate DAV representation (solidblack line) than if all observations are used to calculate daily variations

(gray line)

representative diurnal variations as well as the abovemethod. The absolute value of DAV is used since DAVcan be either positive or negative. For the purposesof examining snowmelt, only the 36Ð5 GHz verticallypolarized observations were used.

The SSM/I snowmelt onset algorithm was modified foruse with AMSR-E 36Ð5 GHz vertically polarized Tb data.An initial examination of observations from the AMSR-E sensor for 2005 indicated that the sensor was moresensitive to daily fluctuations and that the previous Tb

snowmelt threshold for SSM/I of 246 K was too low. Todevelop a new threshold for the snowmelt onset algo-rithm, the AMSR-E 36Ð5V GHz Tb for the six WheatonRiver basin pixels were plotted against temporally cor-responding near-surface air temperature data from theWheaton River for the period of August 2004 to Decem-ber 2005 (Figure 3). When the air temperature is near0 °C, the Tb rapidly increases due to the presence ofliquid water in the snowpack. The Tb at which this tran-sition from frozen to melting snow occurs is consideredthe threshold. A histogram of annual AMSR-E 36Ð5VGHz Tb data was also produced to identify the new Tb

threshold (Figure 4). A typical year of observations inthis seasonally snow covered region has a bimodal distri-bution of Tb related to frozen, dry snow during the coldmonths and wet snow or no snow during the warmermonths, and the Tb that separates these two conditions isthe Tb threshold (Figure 4, Table II).

Measurements of snowpack properties were obtainedin the Wheaton River basin during the spring of 2005.More than fifty snowpits were dug in a variety of loca-tions, encompassing a range of elevations, terrains, andforest cover in an attempt to represent the heterogeneousnature of the region. Characteristics of the snowpack weredescribed and measured, including the general stratigra-phy as well as grain size, air and snow temperature, andsnow structure at various depths. A comprehensive analy-sis of these parameters is beyond the scope of this paper.



Figure 3. Scatterplots of Wheaton River near-surface air temperatures compared to AMSR-E 36Ð5 GHz vertically polarized Tb for the six 25 ð 25 km2

Wheaton River basin pixels for 2004–2005. The transition in Tb from frozen (1, 2) to melting (3) snow, which appears as an inflection in the slopenear a surface temperature of 0 °C, helps to establish the Tb threshold of snowmelt as 252 K. The two unique distributions that exist at temperaturesbelow 0 °C in most of the pixels (1, 2) may be related to snowpack grain size growth or the presence of ice on vegetation. The area of high

temperatures (4) represents snow-free surfaces during the summer months

A snow surface dielectric capacitance probe and measure-ments of snow density were used to derive snow wetnessas a percentage of the volume at different depths (Denoth,1989). Surface and near-surface wetness values wereaveraged from multiple measurements at each site, andthese wetness values were later compared with AMSR-E 36Ð5V GHz Tb from overpasses obtained less than1 h before or after the ground observations (Figure 5(a)and (b), Table III). In addition to supporting the newTb threshold, these observations helped show that theobserved Tb response is linked to melting snow and isnot the result of liquid precipitation or other events.

Once the new Tb threshold was determined, theAMSR-E 36Ð5V GHz data, along with melt onset andspring transition data, were compared with the melt onset

results of the SSM/I 37V GHz data for all six of theWheaton River basin 25 ð 25 km2 pixels (Figure 6(a)and (b), Table IV). This allowed for a new DAV thresholdto be created for the AMSR-E data (Table II), sincethe daily fluctuations appear to have increased with theAMSR-E sensor, and also helped to determine if the newthresholds are effective at detecting snowmelt as relatedto previous snowmelt onset detection studies using SSM/ITb data.

After determining that similar snowmelt onset resultscould be obtained from AMSR-E and SSM/I Tb data grid-ded at the same 25 ð 25 km2 resolution, the effectivenessof the AMSR-E sensor’s improved spatial resolution wasinvestigated. The AMSR-E swath data were gridded tothe EASE-Grid with a resolution of 12Ð5 ð 12Ð5 km2,



Figure 4. Frequency histogram of all AMSR-E 36Ð5V GHz Tb for thesix Wheaton River basin 25 ð 25 km2 pixels during the 2004–2005period. The Tb are grouped in bins of 2 K. The bimodal distributionsuggests a threshold of 252 K is more suitable for distinguishing betweenfrozen (left distribution) and wet (right distribution) snow using AMSR-Ein the Wheaton River basin than the SSM/I threshold of 246 K. Theindividual pixel sets, while not as clearly defined, also have minima near

the threshold of 252 K

Figure 5. (a) Scatterplot of snowpack wetness compared to AMSR-E36Ð5V GHz Tb within the Wheaton River basin. Below 2% wetness,Tb are clustered between 230 and 245 K, but an increase in liquid-watercontent beyond 2% creates an abrupt rise in Tb around the threshold of252 K. (b) The snowpack wetness (gray) relates well to the AMSR-E Tb

(black) during the period prior to the spring transition

which is more representative of the 12 km mean spa-tial resolution of the 36Ð5V GHz channel (Ashcroft andWentz, 2004). This gridded resolution allowed for thecreation of four finer-resolution pixels of informationfor each 25 ð 25 km2 pixel (Figure 1(b)). Elevation data

Figure 6. (a) Comparison of pixel B02 AMSR-E and SSM/I Tb for2005. The AMSR-E Tb (black) match closely with the SSM/I Tb(gray), although they appear to show greater variability during the springtransition and summer months. The melt onset date and duration forthe two records are also similar. (b) Comparison of pixel B02 AMSR-Eand SSM/I DAV for 2005. The AMSR-E DAV (black) is consistentlyhigher than the SSM/I DAV (gray) throughout the spring and summer.The AMSR-E DAV threshold of š18 K appears to match with the SSM/I

DAV threshold of š10 K and the timing of SSM/I melt onset

were used to identify the fine-resolution pixels withinany coarse-resolution pixel with the highest and lowestmean elevations (Table V). The AMSR-E 36Ð5V GHzTb for these pixels, as well as the snowmelt onset tim-ing, were then compared with the coarser 25 ð 25 km2

data to examine the effects of the AMSR-E sensor’sincreased spatial resolution on improving the ability todetect snowmelt in heterogeneous terrain (Figure 7(a) and(b), Table V).

RESULTS

The two main techniques used to determine a newsnowmelt onset Tb threshold for the AMSR-E 36Ð5VGHz channel both found the threshold between non-melting and melting snow to be 252 K. The first method,in which the Tb for the six Wheaton basin pixels werecompared with air temperature data for 2004 and 2005(refer to Figure 3), illustrates that a Tb threshold ofaround 252 K is an effective boundary between non-melting (�1 and �2 ) and melting snow (�3 ) near 0 °C.In these non-linear plots, it can be noted that two distinct



Table III. Spring 2005 Wheaton River basin snow wetness measurements and AMSR-E 25 km2 Tb

PitID

Date &Time (PST)

Latitude(°)

Longitude(°)

Depth(cm)

# ofSamples

AverageWetness (%)

AMSR-E 25 km2

Pixel Tb (K)

1 3/24/05 12Ð48 60Ð13 �134Ð88 0 0 n/a B02 245Ð20Ð5–2 2 3Ð12

2 3/24/05 13Ð68 60Ð13 �134Ð88 0 1 0Ð47 B02 244Ð00Ð5–2 3 1Ð56

3 3/24/05 14Ð16 60Ð23 �135Ð17 0 0 n/a B02 244Ð00Ð5–2 4 1Ð45

4 3/27/05 13Ð20 60Ð21 �135Ð27 0 6 0Ð97 B03 232Ð90Ð5–2 7 1Ð40

5 3/28/05 14Ð16 60Ð35 �135Ð03 0 3 1Ð55 C03 238Ð00Ð5–2 6 1Ð96

6 3/28/05 14Ð64 60Ð35 �135Ð03 0 3 1Ð24 C03 238Ð00Ð5–2 4 1Ð42

7 4/1/05 12Ð96 60Ð35 �135Ð07 0 6 1Ð57 B03 230Ð80Ð5–2 6 1Ð75

8 4/1/05 14Ð88 60Ð36 �135Ð04 0 5 1Ð30 C03 242Ð10Ð5–2 5 1Ð49

9 4/4/05 11Ð76 60Ð29 �135Ð00 0 8 2Ð72 B02 251Ð10Ð5–2 8 2Ð60

10 4/4/05 13Ð44 60Ð30 �134Ð99 0 5 2Ð71 B02 252Ð00Ð5–2 5 2Ð25

11 4/4/05 15Ð36 60Ð29 �134Ð98 0 6 3Ð23 B02 252Ð00Ð5–2 6 2Ð58

12 4/6/05 13Ð44 60Ð23 �135Ð17 0 5 4Ð40 B02 254Ð90Ð5–2 5 2Ð84

13 4/6/05 15Ð12 60Ð34 �135Ð00 0 5 3Ð79 C03 259Ð40Ð5–2 5 2Ð97

14 4/11/05 13Ð68 60Ð13 �134Ð88 0 5 3Ð30 B02 248Ð60Ð5–2 5 2Ð58

Table IV. AMSR-E and SSM/I melt dates (day of year) for theWheaton River Basin

2004 2005

PixelID

AMSR-E meltonset (day)

SSM/I meltonset (day)

AMSR-E meltonset (day)

SSM/I meltonset (day)

A02 100 100 112 110A03 116 115 111 110B02 100 99 110 109B03 116 116 110 112C02 95 96 94 94C03 96 97 106 105

relationships occur on either side of the 0 °C mark. Below0 °C, the Tb are below 252 K and increase very slightly

with an increase in temperature. As the temperatureapproaches and exceeds 0 °C, the Tb rapidly increasesabove 252 K (Figure 3). This rapid change in the slope ofTb when the air temperature is near 0 °C at Tb D 252 Kis due to the high sensitivity of microwaves to liquidwater within the melting snowpack. The transition fromfrozen to melting snow is especially clear for pixel B02,which contains the site of the near-surface air temperatureobservations, as the change in slope between these twoconditions at 0 °C occurs at the Tb threshold of 252 K(Figure 3).

It is also worth noting that two distinct distributionsof Tb occur when compared to air temperatures below0 °C (Figure 3). The upper Tb response (Figure 3, �1 )is from periods in November from both 2004 and 2005,while the lower response (Figure 3, �2 ) is from periods

Table V. EASE-Grid 12Ð5 ð 12Ð5 km2 pixel characteristics and melt onset dates for 2005

PixelID

Melt onset(day)

Transition end(day)

Transitionduration (days)

Meanelevation (m)

Minimumelevation (m)

Maximumelevation (m)

Relief(m)

B02a 110 131 21 1177 672 2193 1521B02A 110 131 21 976 672 1875 1203B02B 96 131 35 940 696 2160 1464B02C 110 154 44 1493 711 2193 1482B02D 109 132 23 1170 719 1842 1123

a This EASE-Grid 25 ð 25 km2 pixel is compared with the four finer-resolution 12Ð5 ð 12Ð5 km2 pixels contained within it.



Figure 7. (a) Comparison of high-resolution AMSR-E Tb from 2005 forpixels B02B and B02C. The lower elevation B02B (black) has higheroverall Tb than B02C (gray) that relate to warmer temperatures. Thisresults in an earlier melt onset for B02B in relation to B02C and theother pixels (Table V). (b) A detailed look at days 90 to 100 reveals thatthe Tb threshold is met consistently in the lower elevations of B02B(black) but not in the higher elevations of B02C (thick gray). Pixel B02(thin gray) is more closely related to B02B, but the Tb threshold is not

met as early as in B02B

in December of the same years. The cause of the upperresponse may be related to ice accumulation on treesduring the earlier winter months, a condition that wasobserved in the 2006 winter season. During the winter,especially during unusually warm years, open lakes thathave not yet frozen can create an icy fog that creepsup valleys and allows for ice to accumulate on vegeta-tion. These distinct distributions may also be an effect ofgrain size growth during the early winter months, influ-encing the scattering properties of the snowpack. At the36Ð5 GHz frequency, emissivity decreases proportionallyto increases in grain size (Matzler, 1994), and so, asthe grains within the snowpack grow early in the win-ter, Tb decreases from the upper distribution (�1 ) to thelower distribution (�2 ) (Figure 3). As these events bothoccur within the same time frame, further investigationis required to identify the source of these differing rela-tionships.

The second technique, examining a histogram ofAMSR-E 36Ð5V GHz Tb distribution from the Wheatonbasin for all 2004 and 2005 data, also produced a Tb

threshold of 252 K (Figure 4). The frequency histogram,with a bin size of 2 K, displays a bimodal distribution

from which it is interpreted that brightness temperaturesless than 252 K relate to frozen, non-melting snowand brightness temperatures greater than 252 K relateto wet, melting snow as well as snow-free conditionsduring the summer months. The sets of individual pixels,representing the upper (pixels A02 and A03), middle(pixels B02 and B03), and lower (pixels C02 and C03)portions of the Wheaton River basin, do not have as cleara bimodal distribution, but do have comparable minimumvalues around 252 K (Figure 4). This, along with theplots of Tb against air temperature, suggests that the Tb

threshold is effective in differentiating between dry andwet snow at the drainage basin and sub-drainage basinscales in this heterogeneous terrain.

The DAV calculation, which involves taking the differ-ence between night and day observations, shows the dailycontrast in brightness temperatures at this time of year.The timing of the AMSR-E overpasses for this regionare around 03 : 30 and 13 : 30 local time (PST), whichare likely near the times of minimum and maximumdaily melt, and so the difference in the Tb from thesetimes provides a good indicator of melting and re-freezingsnow. The DAV threshold increased due to the appar-ent increased sensitivity of the AMSR-E sensor to dailyTb fluctuations (Figure 6(a)). A new DAV threshold ofš18 K, which was determined by comparing the AMSR-E DAV with SSM/I DAV for the Wheaton River pixels in2005 (Figure 6(b)), represents times when strong melt/re-freeze cycles are occurring during the spring snowmelttransition.

The change in Tb and DAV thresholds from SSM/I toAMSR-E (Table II) is most likely the result of differencesin the times of data acquisition for the two sensors.Daily AMSR-E overpasses for the area occur around03 : 30 and 13 : 30 PST, whereas SSM/I overpasses occuraround 08 : 30 and 18 : 30 PST. The AMSR-E sensormakes observations during the early morning, whentemperatures and corresponding Tb would be near theirdaily minimum, and again in the early afternoon, whenmuch of the mountainous terrain would be near thedaily maximum air and brightness temperatures. Thiscreates larger DAV in relation to the DAV from SSM/I,since the timing of SSM/I overpasses is further fromthese daily minimum and maximum temperatures, andtherefore requires a larger DAV threshold. In addition,the DAV threshold must increase so as to not indicatethat melting is occurring when large DAV are observedduring the summer, when much of the land is snow-free(Figure 6(b)).

Field-observed snow liquid-water content measure-ments from the Wheaton River basin during the springof 2005 support the Tb threshold of 252 K and alsoshow that the assumed response in Tb to melting snowcan be directly attributed to snowmelt instead of othereffects, such as liquid precipitation events. Surface (0 cm)and near surface (0Ð5–2 cm) wetness measurementsrelate very well to AMSR-E 36Ð5V GHz Tb observa-tions obtained at coincident times (Figure 5(a) and (b),



Table III). At wetness values below 2%, the Tb are clus-tered between 230 K and 245 K, but as the liquid-watercontent of the snowpack increases above 2% by volume,the Tb appear to increase near and above the threshold of252 K (Figure 5(a)). For the period prior to and includ-ing the beginning of the spring transition, the changes inthe snowpack wetness through time closely follow thatof the AMSR-E Tb, and at wetness values above 2%,the Tb are at or above the 252 K threshold (Figure 5(b)).This relationship verifies that the Tb response is due tothe presence of liquid water within melting snow.

With these new AMSR-E snowmelt onset thresholds(Tb > 252 K and DAV > š18 K), the AMSR-E springtransition in 2004 and 2005 was characterized for the sixWheaton River 25 ð 25 km2 pixels and compared withSSM/I data for the same pixels and years (Figure 6(a)and (b), Table IV). The snowmelt onset algorithm devel-oped by Ramage et al. (2006) was used for the SSM/Idata, using thresholds of Tb > 246 K and DAV > š10 K(Table II). For pixel B02, the onset of snowmelt comparesvery well between AMSR-E and SSM/I, as does the endof the spring transition and the total duration (Figure 6(a)and (b)). It should be noted that although the AMSR-Ethresholds were met for this pixel around day 95, theywere not met for a consistent length of time, and so thespring transition is interpreted as beginning about 10 dayslater (Figure 6(a) and (b)). The DAV of the AMSR-Eobservations are higher than the corresponding SSM/Iobservations throughout the spring transition as well asfor much of the summer. Preliminary investigations ofAMSR-E derived snowmelt in the larger Pelly Riverbasin, approximately 250 km northwest of the WheatonRiver basin, indicate the AMSR-E Tb and DAV thresh-olds are effective for detecting snowmelt on a regionalscale (Ramage et al., 2007).

For both 2004 and 2005, the AMSR-E and SSM/I meltonset dates for the six Wheaton River 25 ð 25 km2 pixelsmatch up very well, with the dates differing by less than2 days (Table IV). Pixels A02 and B03 in 2004 werefound to have the same melt onset date for both sensors,and the same was true for pixel C02 in 2005. These well-correlated melt onset dates indicate that the new AMSR-Edata and thresholds are effective in detecting snowmelt inthe Wheaton River basin at the same resolution as SSM/I.In addition, this relationship may eventually allow for theobservations from both sensors to be combined to createan even more robust satellite record of observation.

Since the AMSR-E sensor provides enhanced resolu-tions over the SSM/I sensor, especially in the 36Ð5 GHzchannel, the timing of snowmelt within the WheatonRiver basin was also determined using Tb data gridded tothe EASE-Grid 12Ð5 ð 12Ð5 km2 resolution (Figure 1(b)).Pixel B02 (25 ð 25 km2 resolution) and the four finer-resolution pixels within it were examined in an attemptto extract the effects of the heterogeneous terrain on theonset of snowmelt (Table V). Differences between theplots of AMSR-E Tb in the first half of 2005 for pix-els B02C and B02B, which represent the highest andlowest mean elevations within pixel B02 respectively,

indicate that the spatial resolution of the AMSR-E sen-sor improves upon the SSM/I sensor’s ability to detectsnowmelt in mixed terrain, especially in areas with arange of elevations (Figure 7(a) and (b), Table V).

In relation to the 2005 melt onset timing and springtransition duration for pixel B02, the timing of theseevents for the finer-resolution pixels is similar (Table V).Two notable exceptions are pixel B02B, the lowest inmean elevation, and pixel B02C, the highest in mean ele-vation. Melt onset occurs much earlier in B02B than inthe other pixels, most likely as a result of warmer temper-atures within its lower elevations. For B02C, the springtransition period lasts more than 20 days longer than theother pixels, and this is due to colder temperatures inthe higher elevations keeping the nighttime snowpackfrozen and DAV high well into the spring. Examiningthe time during which snowmelt onset has begun forB02B (Figure 7(b)), it can be noted that the Tb for B02Bare consistently above the threshold from day 94 to 100,whereas the Tb for B02C never reach the threshold dur-ing this period. Furthermore, Tb for pixel B02 indicatemelting during the day from days 96 to 100, and whilethis concurs with B02B, it disagrees with B02C, and sothe higher-resolution AMSR-E data is more effective atdifferentiating melting and non-melting areas in this het-erogeneous terrain than the lower resolution AMSR-Eand SSM/I data.

CONCLUSIONS

AMSR-E passive microwave data can be used to inves-tigate snowmelt at the basin and regional scales. In addi-tion, the finer resolution of AMSR-E over that of SSM/Iallows for the snowpack characteristics of areas with het-erogeneous terrain to be more accurately examined. Forthe Wheaton River dataset, AMSR-E is more sensitive todaily brightness temperature fluctuations than SSM/I atthe same spatial scale. Relating the two datasets at thesame spatial scale provides the foundation for a longer-term study of snowmelt dynamics combining the histor-ical record of both sensors. Plots of AMSR-E Tb andnear-surface air temperatures, as well as a bimodal dis-tribution of Tb and comparisons with snow liquid-watercontent, suggest a Tb threshold of 252 K can be usedto make the distinction between melting and non-meltingsnow in combination with a DAV threshold of š18 K thatreflects the spring transition when large melt/re-freezecycles are occurring. The larger oscillations in AMSR-E daily Tb are most likely due to the timing of dataacquisition being closer to the actual maximum and min-imum daily Tb. The snowmelt onset algorithm will alsobe applicable in most arctic and subarctic landscapes.

Snowmelt onset dates from AMSR-E, as determinedusing the new Tb and DAV thresholds, compare verywell with SSM/I for the Wheaton River basin, differingby a maximum of 2 days. The enhanced spatial resolutionof the AMSR-E data provides an improvement overSSM/I in detecting the onset of snowmelt in areas of



mixed terrain. Using this AMSR-E melt onset algorithm,the snowpack dynamics of the Wheaton River basin,as well as other areas of the upper Yukon River basin,can be further investigated spatially and temporally forthe record of AMSR-E observations. The improvementsby the AMSR-E sensor upon SSM/I allow for a moreeffective spatial and temporal examination of snowmeltdynamics in the heterogeneous terrain of the upper YukonRiver basin. It will be a significant improvement to ourability to understand the melt processes and timing in animportant, remote, high-latitude drainage basin.

ACKNOWLEDGEMENTS

We would like to extend our thanks to Mary Jo Brodzik,Amanda Leon, and the National Snow and Ice DataCenter for EASE-grid satellite data and continued assis-tance. Thanks to Environment Canada for ongoing sup-port with field logistics and additional data. We thankSarah Kopczynski and Shannon Haight for their assis-tance collecting field data. We are also thankful for help-ful comments provided by Andrew Klein and two anony-mous reviewers. Additional thanks to Michael Chupa.Funding was provided by Lehigh University and theNational Aeronautics and Space Administration’s Terres-trial Hydrology Program (grant # NNG04GR31G).

REFERENCES

Aagaard K, Carmack EC. 1989. The role of sea ice and other freshwaterin the Arctic circulation. Journal of Geophysical Research 94:14485–14498.

Abdalati W, Steffen K. 1997. Snowmelt on the Greenland Ice Sheet asderived from passive microwave satellite data. Journal of Climate 10:165–175.

Ashcroft P, Wentz F. 2004. updated daily. AMSR-E/Aqua L2A Globalswath spatially-resampled brightness temperatures (Tb) V001 . January2004 to December 2005. Boulder, CO,: National Snow and Ice DataCenter. Digital media.

Brabets TP, Wang B, Meade RH. 2000. Environmental and hydrologicoverview of the Yukon River basin, Alaska and Canada, Water-Resources Investigations Report 99–4204. US Dept. of the Interior,US Geological Survey: Anchorage, AK.

Denoth A. 1989. Snow dielectric measurements. Advances in SpaceResearch 9: 233–243.

Derksen C, Walker AE, Goodison BE, Strapp JW. 2005. IntegratingIn Situ and multiscale passive microwave data for estimation ofsubgrid scale snow water equivalent distribution and variability. IEEETransactions on Geoscience and Remote Sensing 43: 960–972.

Environment Yukon. 2000. 90 Meter Yukon Digital Elevation Model,http://www.environmentyukon.gov.yk.ca/geomatics/data/90m dem.html, [13 March 2005].

Foster JL, Chaojiao S, Walker JP, Kelly RE, Chang A, Dong J,Powell H. 2005. Quantifying the uncertainty in passive microwavesnow water equivalent observations. Remote Sensing of Environment94: 187–203.

Goita K, Walker AE, Goodison BE. 2003. Algorithm development forthe estimation of snow water equivalent in the boreal forest usingpassive microwave data. International Journal of Remote Sensing 24:1097–1102.

Josberger EG, Mognard NM. 2002. A passive microwave snow depthalgorithm with a proxy for snow metamorphism. HydrologicalProcesses 16: 1557–1568.

Kelly RE, Chang AT, Tsang L, Foster JL. 2003. A prototype AMSR-Eglobal snow area and snow depth algorithm. IEEE Transactions onGeoscience and Remote Sensing 41: 230–242.

Kim EJ, England AW. 2003. A yearlong comparison of plot-scale andsatellite footprint-scale 19 and 37 GHz brightness of the Alaskan NorthSlope. Journal of Geophysical Research 108: ACL 7.

Matzler C. 1994. Passive microwave signatures of landscapes in winter.Meteorology and Atmospheric Physics 54: 241–260.

Ramage JM, Apgar JD, McKenney RA, Hanna W. 2007. Spatialvariability of snowmelt timing from AMSR-E and SSM/Ipassive microwave sensors, Pelly River, Yukon Territory, Canada.Hydrological Processes 21(12): 1548–1560.

Ramage JM, Isacks BL. 2002. Determination of melt-onset and refreezetiming on southeast Alaskan icefields using SSM/I diurnal amplitudevariations. Annals of Glaciology 34: 391–398.

Ramage JM, Isacks BL. 2003. Interannual variations of snowmelt andrefreeze timing on southeast-Alaskan icefields, U.S.A. Journal ofGlaciology 49: 102–116.

Ramage JM, McKenney RA, Thorson B, Maltais P, Kopczynski SE.2006. Relationship between passive microwave-derived snowmelt andsurface-measured discharge, Wheaton River, Yukon Territory, Canada.Hydrological Processes 20: 689–704.

Ulaby FT, Moore RK, Fung AK. 1986. Microwave Remote Sensing:Active and Passive, Vol. III: From Theory to Applications . ArtechHouse: Dedham, MA.

Vorosmarty CJ, Hinzman LD, Peterson BJ, Bromwich DH, Hamil-ton LC, Morison J, Romanovsky VE, Sturm M, Webb RS. 2001. TheHydrologic Cycle and its Role in Arctic and Global EnvironmentalChange: A Rationale and Strategy for Synthesis Study . Arctic ResearchConsortium of the U.S.: Fairbanks, AK; 84.

Walker AE, Goodison BE. 1993. Discrimination of a wet snowcoverusing passive microwave satellite data. Annals of Glaciology 17:307–311.

Wang L, Sharp M, Brown R, Derksen C, Rivard B. 2005. Evaluation ofspring snow covered area depletion in the Canadian Arctic from NOAAsnow charts. Remote Sensing of Environment 95: 453–463.


AMSR-E algorithm for snowmelt onset detection in sub-arctic heterogeneous terrain

Documents

Transcript of AMSR-E algorithm for snowmelt onset detection in sub-arctic heterogeneous terrain