Remote Sensing of Environment 114 (2010) 16991709

Contents lists available at ScienceDirect

Remote Sensing of Environment

j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / r s e

Development of a tundra-specic snow water equivalent retrieval algorithm for satellite passive microwave data
C. Derksen a,, P. Toose a, A. Rees b, L. Wang a, M. English b, A. Walker a, M. Sturm c
a b c

Climate Research Division, Environment Canada, Toronto, Ontario, Canada Department of Geography and Environmental Studies, Wilfrid Laurier University, Waterloo, Ontario, Canada Cold Regions Research and Engineering Laboratory, U.S. Army Engineer Research and Development Center, Fort Wainwright, Alaska, United States

a r t i c l e

i n f o

a b s t r a c t
Airborne and satellite brightness temperature (TB) measurements were combined with intensive eld observations of sub-Arctic tundra snow cover to develop the framework for a new tundra-specic passive microwave snow water equivalent (SWE) retrieval algorithm. The dense snowpack and high sub-grid lake fraction across the tundra mean that conventional brightness temperature difference approaches (such as the commonly used 37 GHz19 GHz) are not appropriate across the sub-Arctic. Airborne radiometer measurements (with footprint dimensions of approximately 70 120 m) acquired across sub-Arctic Canada during three eld campaigns during the 2008 winter season were utilized to illustrate a slope reversal in the 37 GHz TB versus SWE relationship. Scattering by the tundra snowpack drives a negative relationship until a threshold SWE value is reached near 130 mm at which point emission from the snowpack creates a positive but noisier relationship between 37 GHz TB and SWE. The change from snowpack scattering to emission was also evident in the temporal evolution of 37 GHz TB observed from satellite measurements. AMSR-E brightness temperatures (2002/032006/07) consistently exhibited decreases through the winter before reaching a minimum in February or March, followed by an increase for weeks or months before melt. The cumulative absolute change (|37V|) in vertically polarized 37 GHz TB was computed at both monthly and pentad intervals from a January 1 start date and compared to ground measured SWE from intensive and regional snow survey campaigns, and climate station observations. A greater (lower) cumulative change in |37V| was signicantly related to greater (lower) ground measured SWE (r2 = 0.77 with monthly averages; r2 = 0.67 with pentad averages). |37V| was only weakly correlated with lake fraction: monthly r2 values calculated for January through April 20032007 were largely less than 0.2. These results indicate that this is a computationally straightforward and viable algorithmic framework for producing tundra-specic SWE datasets from the complete satellite passive microwave record (1979 to present). Crown Copyright 2010 Published by Elsevier Inc. All rights reserved.

Article history: Received 25 November 2009 Received in revised form 23 February 2010 Accepted 26 February 2010 Keywords: Snow water equivalent Passive microwave Tundra Sub-Arctic

1. Introduction A number of high priority science questions require snow water equivalent (SWE) information across high latitude regions, including: determining whether there is an increase in high latitude winter season precipitation to corroborate recent evidence from model simulations and the sparse network of conventional observations (Min et al., 2008; Zhang et al., 2007). identifying the role of snow cover variability and change in the potential intensication of the high latitude water cycle through increased precipitation, earlier melt, higher peak runoff, and greater freshwater input to the Arctic Ocean (see Dery et al., 2009).

Corresponding author. E-mail address: Chris.Derksen@ec.gc.ca (C. Derksen).

Conventional observations are not adequate to answer these questions because the station network is sparse and coastally biased, and the measurements themselves are uncertain. Snowfall gauge and shield combinations are not standard between countries (Yang et al., 1999), and the required auxiliary measurements for systematic undercatch correction (such as wind speed at gauge height) are often not available. Point snow depth measurements are subject to local scale wind drifting or scour and may not represent the prevailing regional conditions. Even when they do, the large distances between stations does not allow for meaningful spatial interpolation (i.e. kriging), and coastal stations do not represent vast inland areas. Satellite passive microwave measurements address the spatial limitations of conventional observations, but not necessarily the uncertainties in snow cover information. A large imaging footprint (25 km grid cell dimensions), wide swath, and general insensitivity to cloud cover produce spatially continuous daily brightness temperatures (TB) across latitudes north of approximately 60 N. The response

0034-4257/$ see front matter. Crown Copyright 2010 Published by Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2010.02.019

1700

C. Derksen et al. / Remote Sensing of Environment 114 (2010) 16991709

Table 1 Summary of sub-Arctic in situ snow cover measurements. Location Coordinates Time period Distance (km) 18 33 28 Measurements (n) Snow depth 12,000 4400 5900 SWE 369 668 757 Snowpits 10 23 27 SWE (mm) Mean 101 100 111 Min 21 3 15 Max 431 428 534 SD 66 64 61

Puvirnituq, QC Daring Lake, NT Trail Valley Creek, NT

59.83N76.39W 64.85N111.63W 68.70N133.62W

February 2008 April 2008 April 2008

of these TB measurements to seasonally evolving sub-Arctic snow and lake ice cover is poorly understood, however, and the performance of hemispheric SWE retrieval algorithms (for example, Biancamaria et al., 2008; Chang et al., 1990; Kelly et al., 2003) across northern regions is not well documented due to the lack of high latitude measurements for algorithm validation. Algorithms developed for regional application across the boreal forest and open prairies do not perform well over the tundra (Derksen et al., 2005; Koenig & Forster, 2004; Rees et al., 2006) because of the unique physical properties of tundra snow and the microwave contribution of the high fraction of sub-grid lakes. The extreme variability of tundra snow on meter-tometer length scales further complicates tundra algorithm development and validation at the scale of satellite microwave measurements due to the complexity in characterizing ground-truth SWE. Still, passive microwave retrievals remain an attractive option for snow cover applications because of the theoretical relationship between SWE and TB at 37 GHz (Matzler, 1994), and the considerable length of the data record (1978present). In this study, we present a new framework for a tundra-specic passive microwave SWE algorithm developed through analysis of high resolution airborne passive microwave measurements coupled with detailed in situ snow measurements from three eld campaigns across sub-Arctic Canada, and satellite data from the Advanced Microwave Scanning Radiometer (AMSR-E). Specically, we address two fundamental challenges to develop a tundra-specic SWE algorithm: 1. The unique radiometric properties of lake ice, which can constitute up to 40% of the sub-Arctic tundra land surface (primarily through small sub-grid sized lakes). Derksen et al. (2009b) showed that lake fraction is a primary control on tundra TB magnitude at all satellite measured frequencies and therefore must be considered as part of a tundra specic retrieval scheme. 2. The slope reversal in the SWE versus 37 GHz Tb relationship, that occurs between 120 and 180 mm water equivalent (Derksen, 2008; De Seve et al., 1997; Schanda et al., 1983). Below this threshold, increasing SWE is associated with lower TB due to volume scatter. Above this threshold, emission from the snowpack produces higher TB's with increasing SWE. Tundra snowpacks are typically shallow, dense, ne grained, and contain pronounced wind slab layers (Sturm et al., 1995). The inuence of this stratigraphy on scattering and emission behaviour can be complex, but there is documented observational evidence of the scattering to emission transition during late winter (Kim & England, 2003). Accounting for these two factors is essential to avoid the systematic SWE underestimation that is produced from contemporary brightness temperature difference algorithms in tundra environments (Koenig & Forster, 2004; Rees et al., 2006). 2. Data Interpreting TB response over tundra landscapes in winter requires in situ observations of snow cover properties and lake ice characteristics. Optimally, these measurements would be available at contin-

uous intervals through the season and adequately capture sub-grid variability below the scale of satellite passive microwave measurements (25 km grid cell dimensions). In short, datasets with these spatial and temporal characteristics are not available. In this study, we utilize datasets from a series of eld campaigns conducted during discrete time periods across the Canadian sub-Arctic tundra as part of International Polar Year activities between February and April 2008. As summarized in Table 1, these campaigns included the deployment of airborne passive microwave radiometers coupled with intensive in situ snow measurements. Collectively, these datasets provide the opportunity to determine relationships between snow cover properties and microwave TB at multiple scales (airborne and satellite) from multiple tundra sites. In turn, these relationships can be used to develop the framework for a new tundra-specic SWE retrieval algorithm that can be applied to the entire passive microwave satellite data record that dates back to 1978.

2.1. Airborne radiometer measurements and in situ snow surveys At all three tundra eld sites (see Table 1; Fig. 1), airborne passive microwave measurements were acquired covering a wide range of tundra terrain. The radiometer installation (dual polarization 6.9, 19, 37, and 89 GHz) on the National Research Council Twin Otter aircraft is described in MacPherson et al. (2001) and Walker et al. (2002). The airborne radiometers (parameters for the 19 and 37 GHz radiometers used in this study are provided in Table 2) were calibrated pre- and post each ight using warm (ambient temperature microwave absorber) and cold (liquid nitrogen) targets as described by Solheim (1993). Uncertainty in the measurement of the calibration target temperature was estimated at +/-2 K. The 19 and 37 GHz radiometers were calibrated simultaneously so the same target temperature uncertainties for a given calibration apply to both frequencies. Estimates of inter-calibration receiver drift were made by examining the pre- and post-ight calibration target brightness temperatures. Radiometer stability was dependant on the frequency and varied somewhat from campaign to campaign, but overall uncertainty was estimated at +/2 K at 19 GHz, and b 1 K at 37 GHz. Transects of snow depth and SWE measurements were acquired along segments of the radiometer ight lines, typically several kilometers in length. Snow depth was sampled using a self recording snow depth probe (Magna Probe U.S. Patent No. 5864059; cf. Sturm & Liston, 2003) linked to a GPS. At all sites, manual SWE cores (using an ESC-30 corer) were taken approximately every 250 m along the depth transects, in order to compute bulk density values used to convert the measured snow depth to estimated SWE. Average land cover specic snow densities were determined for each study site by relating each ESC-30 measurement to a landscape class determined from a Landsat classication (Natural Resources Canada, 2008). These land cover specic snow densities were used to estimate a SWE value for each snow depth measurement based on the land cover class it was measured in. Density is an inherently conservative variable across the tundra compared to snow depth, so this is an effective technique at converting a more straightforward measurement (depth) to a more

Fig. 1. (a) Study area overview showing the three IPY eld campaign sites and the Baker Lake climate station. (b) SnowSTAR snow survey route and measurement sites; grid shows extent of AMSR-E EASE-Grid data used for analysis.

C. Derksen et al. / Remote Sensing of Environment 114 (2010) 16991709

1701

1702 Table 2 Summary of radiometer characteristics.

C. Derksen et al. / Remote Sensing of Environment 114 (2010) 16991709

3. Results
Frequency

3.1. Radiometric measurements

37 2000 1.2 0.03 b1 6 53

Frequency [GHz] Bandwidth [MHz] Integration time [s] Sensitivity [K] Accuracy [K] 3 dB [] i [] Scan method Flight altitude [m] Airspeed [m/s] Footprint size (w l)

19 1000 1.2 0.04 b2 6 53 Sideways along track 916 60 70 120

70 120

desirable variable (SWE; see Derksen et al., 2005). A summary of the in situ snow sampling transects is provided in Table 1. Snowpits were also excavated at regular intervals along the ight lines. Stratigraphic observations including layer thickness were determined by visual and physical examination of the snowpit face. Average grain size in each layer was estimated from manual observations using a stereo-microscope and comparator card. Density proles were determined with 100 cm3 cutters.

2.2. Satellite passive microwave measurements AMSR-E TB's in both level 2 swath and EASE-Grid (Knowles et al., 2006) projections were acquired for October 1 through April 30, 2002 to 2008. The use of the swath data is advantageous because the original frequency dependant imaging characteristics are retained. Inter-orbit variability in swath level footprint locations, however, necessitated the use of the 25 km EASE-Grid dataset for the compilation of brightness temperature time series. A ve-day moving average was applied to the descending orbit (approximately 0130 local overpass time) AMSR-E data to reduce high frequency variability in TB driven by atmospheric and physical temperature effects (Markus et al., 2006).

2.3. Other snow measurements To address regional snow information, data are available from a coordinated series of snow measurements made across the Northwest Territories and Nunavut, Canada in April 2007 during a snowmobile traverse from Fairbanks, Alaska to Baker Lake, Nunavut (see Fig. 1b). Systematic snow measurements at the regional scale had not been previously collected across this region of northern Canada. A full description and analysis of the traverse measurements is provided in Derksen et al. (2009b). The most intensive period of sampling occurred during the Daring Lake, Northwest Territories to Baker Lake, Nunavut portion of the traverse (16 to 26 April, 2007). Sample sites were located at least every 1 C of longitude. At these sites, snow depth, density, snow water equivalent (SWE), stratigraphy, and grain size measurements are available. Whenever possible, measurements were made at paired sites: one on tundra (land) and one on ice (lake or river ice depending on what was available) in order to capture mean regional SWE in this lake rich environment. Temporally continuous snow observations are available from climate stations, but across the sub-Arctic tundra of Canada these stations are extremely sparse and coastally biased. The only inland observations across the sub-Arctic Canadian tundra are from the community of Baker Lake (64 18N 96 4.8 W; Fig. 1). Bi-weekly, manual observations of snow depth, density, and water equivalent measurements were acquired from the Environment Canada snow course at Baker Lake for algorithm development and validation.

Heterogeneity in land cover, terrain, vertical snowpack properties, and horizontal snow distribution are the reality within satellite passive microwave measurements (25 km grid cell dimensions). This heterogeneity is also often unavoidable at airborne scales (even with 80 100 m footprint dimensions) particularly in tundra environments due to local scale snowpack variability. Given the analysis requirement of linking ground measured SWE to airborne measured TB, it was essential to accurately co-locate surface snow measurements within the airborne footprints to ensure a comparison covering the same physical location on the ground. ArcGIS Geographic Information System (GIS) software was used to link surface snow measurements with airborne data by calculating the approximate extent of the instantaneous eld of view (IFOV) for each airborne TB and then overlaying these airborne footprints with the snow survey data to identify regions of overlap. The GPS Magna Probes geotag each snow depth observation with a latitude and longitude coordinate. These units were WAAS enabled allowing for positional accuracy of better than 3 m 95% of the time (Garmin, 2009; WAAS signal reception is ideal for open land and marine applications). The airborne radiometer's IFOV is dependent on the aircraft's ground speed and altitude, as well as the radiometer's beamwidth, view angle, and integration time. The radiometer's data acquisition system automatically calculates the latitude and longitude coordinates for each TB taking into account the height above ground and the radiometer's beamwidth and view angle (based on the aircraft's pitch, roll and yaw). The airborne radiometers had an integration time of approximately 1.2 s. During this 1.2 s sampling sequence, there are 20 unique integration cycles in which the radiometers measure emission in both horizontal and vertical polarization from the ground scene within the IFOV, as well as a noise diode and a load. Because of the forward momentum of the aircraft, the radiometer actually measures a slightly different scene during each of the 20 integration cycles, each adjacent and overlapping one another, resulting in a larger IFOV in the alongtrack axis than what would be sampled from a stationary tower mounted radiometer system. The 20 cycles are averaged together to produce a single TB measurement for the entire IFOV. The radiometer integration time is a xed parameter and was not changed during the IPY eld campaigns; however the aircraft speed did occasionally vary between ights and during the same ight, but overall was quite consistent at approximately 110 nautical miles per hour ( 60 m/s, resulting in a typical IFOV size uctuations in the along-track axis of less than 10 m). These parameters resulted in a typical IFOV extent of approximately 80 m in the across-track axis and 100 m in the alongtrack axis. ArcGIS was used to create overlapping buffers, with variable along-track axis lengths for each airborne TB measurement. These variable length buffers better represent the actual IFOV of the airborne radiometers, making it possible to determine which land cover characteristics and surface measurements were found within the footprint of each airborne radiometer. Fig. 2 shows a visual representation of the snow survey and airborne data collected during a short segment of data acquired near Puvirnituq, Quebec using ArcGIS software. The snow depth transect transitions from open tundra on to a small pond and then back on to land. The black dashed ellipses approximate the IFOV of the high resolution airborne radiometers ( 80 m 100 m). The two lines of points passing through the IFOVs represent two side-by-side snow depth transects, and each individual point is a snow depth measurement recorded using the GPS Magna Probes. The cross symbols represent the location of the manual ESC-30 SWE measurements recorded to obtain the bulk snowpack density and SWE. In reality, the precise co-location of surface snow measurements within an airborne TB footprint does not always occur due to errors in

C. Derksen et al. / Remote Sensing of Environment 114 (2010) 16991709

1703

Fig. 2. Approach of linking the transect snow measurements to the airborne footprints. Large ellipses illustrate airborne microwave footprints, red circles indicate Magna Probe snow depths, and yellow crosses indicate ESC-30 snow cores used to convert depth measurements to SWE.

the aircraft ight path and/or the ground location of the snow surveyors. In total 995 TB's from all three trans-tundra study site locations had 11,909 coincident SWE measurements within the IFOV's of the radiometers. However, this includes microwave footprints which potentially only had a single snow measurement and/or had measurements at the very edge of an IFOV boundary. Therefore, to ensure that each TB was compared to an adequate and representative number of SWE measurements a simple ltering process was used. The rst stage of ltering eliminated all TB's with snow measurements at the margins of the IFOV that might not have been representative of the snowpack within the IFOV itself. This was accomplished by using ArcGIS software to select only those TB's with snow measurements at least 10 m inside the boundary of the IFOV, resulting in a total of 803 TB's with 9253 coincident SWE measurements. The second step of the ltering process involved selecting only those TB's with a large enough number of snow measurements to capture the local scale variability in SWE that is common for the heterogeneous tundra snowpack. Therefore, only those airborne TB's with at least 10 SWE measurements within the IFOV were selected for this analysis. A sample size of at least 10 snow measurements was considered large enough to not be overly affected by one or two non-representative measurements within the IFOV. This selection process resulted in a total of 714 TB's with 9028 coincident SWE measurements. The expected TB response at 37 GHz to a range of SWE between 0 and 600 mm is shown in Fig. 3a (modied from Matzler et al., 1982). Two distinct slopes are evident: a negative relationship driven by snowpack volume scatter, and a positive relationship driven by

snowpack emission. The slope reversal in this relationship (at approximately 150 mm in the case of these data) illustrates the SWE saturation problem with empirical brightness temperature difference algorithms that rely on 37 GHz measurements: the same TB difference is associated with two SWE estimates that can differ by over 200 mm. With the full range of SWE encountered across all three of the eld campaigns (as summarized in Table 1), we sought to reproduce the relationship in Fig. 3a with our much larger IPY eld dataset. A comparison of in situ measured SWE and 37V TB from the 714 airborne measurements retained as described above is shown in Fig. 3b. The eld observations do capture the same slope reversal phenomenon, albeit with some noise (both relationships identied in Fig. 3b are statistically signicant at 99%). The scattering inuence of the snowpack is evident only to a threshold just below 130 mm after which increasing SWE is coincident with an increase in 37 GHz TB due to snowpack emission. This threshold is slightly lower than reported in some studies (i.e. De Seve et al., 1997; Matzler, 1994) but very similar to direct measurements (with ground based microwave radiometers) of the transition of a tundra snowpack from volume scatter to emission provided in Sturm et al. (1993). Their measurements identied volume scatter to a critical snow depth of 31 cm; depths above this threshold produced an increase in the effective emissivity at 37 GHz. These measurements were made by removing layers from the snowpack near the end of the accumulation season, as opposed to measuring the natural accumulation and metamorphosis of the snowpack. Regardless, presuming a basal depth hoar layer density of 350 kg m 3, the SWE at the 31 cm threshold in the Sturm et

1704

C. Derksen et al. / Remote Sensing of Environment 114 (2010) 16991709

Fig. 3. (a) Expected relationships between 37 GHz TB and SWE (adapted from Matzler et al., 1982); (b) airborne TB versus in situ SWE from all three eld campaigns; (c) monthly averaged 36.5 GHz (V-pol) AMSR-E TB's from 2004/05 for the domain in Fig. 1b.

al. (1993) measurements was approximately 110 mm, close to the 130 mm scattering to emission inection identied with the airborne data in Fig. 3b. The degree of noise in the ground measured SWE versus airborne TB relationship was anticipated given the inherent uncertainty in the ability of the eld observations to accurately characterize mean footprint SWE. The challenge of acquiring snow cover measurement transects within a 100 m wide airborne measurement swath across the tundra should not be underestimated. In addition, variability driven by factors such as underlying vegetation properties and snowpack properties such as depth hoar layer thickness were not considered. A detailed comparison of vertical snowpack properties and measured TB was addressed by Rees et al. (2010). An important scaling question is whether the expected TB versus SWE slope reversal (Fig. 3a) captured by the airborne measurements (Fig. 3b) was also observed at the satellite scale. Snow course measurements from Baker Lake were compared to ve-day (pentad) averaged AMSR-E 37 GHz TB's from the EASE-Grid cell in which the snow course is located (Fig. 3c) over the 2002 to 2008 time period. The transition from snowpack scatter to emission at 37 GHz is evident at the satellite scale, but with a TB minima near 70 mm SWE, followed by increased TB with higher SWE. While the slope reversal is therefore evident at both airborne and satellite scales, the inection point in the TB versus SWE relationship occurs at a much lower SWE threshold at the satellite scale than would be expected given the ground based and airborne radiometer results in Fig. 3a and b. The difference between the airborne and satellite inection points is likely related to scale issues. Unlike the relatively homogeneous airborne measurements, the AMSR-E footprint that covers Baker Lake is composed of approximately 30% lake fraction. This sub-grid heterogeneity will have a seasonally evolving inuence on TB throughout the winter that does not inuence the airborne data. In addition, the Baker Lake snow course covers a relatively localized area, which may underestimate the surrounding tundra SWE. 3.2. Tundra-specic algorithm concept The multi-scale behaviour of 37 GHz TB's over snow covered tundra illustrated in Fig. 3 led to the development of a single frequency solution for tundra SWE retrievals based solely on vertically polarized 37 GHz measurements. The primary advantages of this single frequency approach for tundra regions are: 1. 37 GHz measurements exhibit consistent correlation with lake fraction across the sub-Arctic tundra from January onwards (Derksen et al.,

2009b). While low frequency measurements (6.9; 10.7 GHz) are controlled by emission from below the ice layer through the entire winter season, higher frequency measurements including those commonly used to retrieve SWE ( 19 and 37 GHz) have penetration depths that are strongly inuenced by water beneath the ice for part of the season, but are also inuenced by the ice and overlying snowpack by the end of the season. The timing of this shift in the primary emission source is different at 19 versus 37 GHz, so the traditional brightness temperature difference approach (37 GHz19 GHz) is not appropriate for lake rich tundra environments. By January, the combined thickness of lake ice and snow cover exceeds the 37 GHz penetration depth, and so is sufcient to mask the liquid water signal beneath the ice, creating consistent correlations with lake fraction through the remainder of the snow cover season. 2. Unlike complications that can arise across the boreal forest (see Derksen, 2008; De Roo et al., 2007), vegetation inuences on 37 GHz Tb's are minimal across the tundra. 3. Vertical polarization 37 GHz measurements exhibit low sensitivity (compared to horizontal polarization measurements) to vertical snowpack structure, including ice lenses (Derksen et al., 2009b; Rees et al., 2010). The primary challenge with a single frequency approach, however, is the fact that multiple values of SWE are related to the same TB as illustrated in Fig. 3 and discussed in the previous section. To address this, the month to month cumulative absolute change for vertically polarized 37 GHz TB (|37V|) was calculated on a grid cell by grid cell basis for the seasons spanning 2002/03 through 2006/07. Boxplots were produced that show both the monthly averaged 37 GHz TB and the absolute cumulative change in 37 GHz TB (Fig. 4). The use of absolute change at 37 GHz accounts for the decrease in TB during the rst portion of the winter season (primary mechanism = snowpack scatter), and the subsequent increase in TB as SWE exceeds a critical threshold (primary mechanism = snowpack emission). The hypothesis is that any change in TB observed during the January through April period in the tundra environment is largely due to an increase in SWE. Therefore, the greater (lower) the cumulative change in 37 GHz TB calculated between January and April, the greater (lower) the SWE. During heavy snow years, the minimum TB corresponding to the SWE threshold is reached early, and increases in SWE act to increase TB's resulting in higher absolute cumulative change in 37 GHz TB. During light snow years, the minimum TB corresponding to the 100 mm SWE threshold is reached later or not at all, and continued changes in TB are minimal resulting in lower

Fig. 4. Boxplots of monthly averaged AMSR-E 36.5 GHz TB (left column) and cumulative absolute change (|37V|) over the domain shown in Fig. 1b for 2002/03 (top) through 2006/ 07 (bottom).

C. Derksen et al. / Remote Sensing of Environment 114 (2010) 16991709

1705

1706

C. Derksen et al. / Remote Sensing of Environment 114 (2010) 16991709

cumulative change in 37 GHz TB. A source of uncertainty will be TB changes associated with snow metamorphism and not changes in SWE. A quantitative examination of this issue depends on seasonally continuous physical snowpack observations coupled with high resolution TB measurements; eld experiments during the 2009/10 winter in support of the The COld REgions Hydrology High-resolution Observatory (CoRe-H2O; Rott et al., 2008) will provide this opportunity. We would expect the single frequency change in 37 GHz to be insensitive to lake fraction from January onwards based on the correlation results between TB and lake fraction presented in Derksen et al. (2009b). To conrm this assumption, correlation analysis was performed on monthly averaged change |37V| TB (calculated from the EASE-Grid AMSR-E brightness temperatures) and lake fraction (produced by the U.S. Geological Survey and re-gridded to the EASEGrid). Five winter seasons (January through April; 2002/032006/07) of AMSR-E data were analyzed, producing ve correlation values for each month. Boxplots of these correlation values over the ve years are shown in Fig. 5, with the relationships between monthly |37V| TB and lake fraction interannually consistent and weak. 3.3. Prototype algorithm and evaluation AMSR-E data, averaged over monthly periods to remove high frequency noise, were initially used to investigate relationships between cumulative absolute change in 37 GHz TB and SWE. Five winter seasons (2002/032006/07) were investigated. On a grid cell by grid cell basis, mean monthly 37V TB was calculated for December through April. The absolute month to month change in 37V was then calculated: 37Vmonth i = j37Vmonth i 37Vmonth i1 j 1

for each individual December relative to the monthly mean over the full AMSR-E period of record (20022008). For example, the starting point for the 2002/03 calculations is shown in Eq. (2).
37VSTART = 37V 37V

Dec2002

Dec20022008 :

This value can be either positive or negative. For January, 37V is calculated as: 37VJan = j37VSTART 37VJan j: 3

For each subsequent month, it follows Eq. (1). The monthly 37V values were then compared to tundra SWE datasets derived from surface measurements taken from three sources: 1. Environment Canada snow course measurements from Baker Lake (see Fig. 1a) available bi-weekly for all six seasons. 2. Mean SWE for an AMSR-E grid cell covering the Daring Lake study area (see Fig. 1b), determined from a terrain weighted average of all in situ SWE measurements acquired during eld campaigns in April between 2003 and 2008. 3. Regional SWE measurements acquired between Daring Lake and Baker Lake during the April 2007 SnowSTAR traverse (see Fig. 1b) described in Derksen et al. (2009b). Results of this assessment were very encouraging when the datasets were compared at a monthly time step (r2 = 0.77; signicant at 95%), as illustrated in Fig. 6a. The measured SWE values cover a reasonable range in SWE, encompass different snowpack ages (thin tundra snow in January; a mature multi-layer snowpack near peak SWE in April) and lake rich and lake poor areas. The analysis was repeated at the pentad (ve-day) average time step to see if the monthly results were transferable to a ner temporal scale. All calculations follow Eqs. (1)(3), but

This step produces an absolute 37V term for each month, January through April. These values were then summed (|37V|) for each EASE-Grid cell in a cumulative fashion from month to month through the season. For example, the value for March is the sum of the monthly change in absolute 37V TB from December to January, January to February, and February to March. This process shifts the seasonal Ushaped evolution of 37V TB (as illustrated in Fig. 3 and the left column of Fig. 4) to an approximately linear seasonal evolution (as illustrated in the right column of Fig. 4). To account for variability in December 37V TB caused by a variety of factors including early season snow accumulation, lake ice thickness, and physical temperature, the starting point for the calculation of Eq. (1) for January was determined by calculating the 37V TB difference

Fig. 5. Box plots of correlation between monthly averaged change in 37 GHz TB and lake fraction, 2002/032007/08, for the domain shown in Fig. 1. Shading denotes insignicant correlations at 95%.

Fig. 6. Relationship between AMSR-E cumulative 37V and ground measured SWE at monthly (a) and pentad (b) time steps.

C. Derksen et al. / Remote Sensing of Environment 114 (2010) 16991709

1707

with monthly TB data replaced with pentad data. Pentad averages were calculated from 1 January through 30 April (24 pentads). These values were then summed (|37V|) for each pixel in a cumulative fashion from pentad to pentad through the season. For example, by 30 January you have summed 6 pentads; by 30 April you have summed 24 pentads. As in Eq. (3), the starting point for the calculation of Eq. (1) for pentad 01 (January 1 to 5) is determined by computing the 37V December difference as dened previously. The pentad resolution relationship between |37V| and SWE is shown in Fig. 6b. While somewhat weaker than the monthly average results, a signicant relationship (r2 = 0.67; signicant at 95%) is still evident.

4. Discussion and conclusions There is growing evidence that the traditional use of the 37 and 19 GHz TB difference to retrieve SWE from satellite passive microwave data is not appropriate across sub-Arctic tundra and taiga. While not discussed in this study, the emission signal from vegetation, and deep snow volume scatter at 19 GHz are problematic across the boreal forest (Derksen, 2008). In the tundra, high sub-grid lake fraction and microwave emission from the snowpack at SWE values as low as 70 mm (see Fig. 3a) hamper the ability to retrieve SWE with these conventional TB difference algorithms. While the use of multiple frequencies is preferred from the perspective of mitigating problematic physical temperature and atmospheric effects (see Markus et al., 2006; Wang & Tedesco, 2007), airborne and satellite measurements coupled with intensive snow surveys across the Canadian sub-Arctic tundra were used in this study to present the framework for a tundra-specic SWE retrieval scheme that relies solely on 37 GHz satellite measurements. Relying on a single frequency mitigates the inuence of sub-grid lake fraction, and accounts for the slope reversal in the 37 GHz TB versus SWE relationship. The theoretical relationship between 37 GHz TB and SWE, illustrated clearly with point scale surface based radiometer measurements at an alpine site in Matzler et al. (1982) was observed in ground snow surveys and airborne radiometer measurements from three eld sites that span the Canadian sub-Arctic tundra. This relationship, in turn, was evident with coarse resolution AMSR-E satellite measurements over the same region. The consistent nature of these airborne and satellite datasets suggests that tundra-specic algorithmic approaches that rely on 37 GHz measurements should be transferable across the sub-Arctic at continental scales. To investigate algorithm transferability further, a comparison of AMSR-E retrievals with eld measurements of SWE acquired in March

3.4. Comparison with a previous algorithm The spatial pattern of tundra SWE produced with the new 37V GHz algorithm for pentad 19 in 2007 (1 through 5 April) is shown in Fig. 7a. For comparison, SWE derived using a TB difference (37V19V) approach developed for open environments (Goodison & Walker, 1995) is shown in Fig. 7b. The modal SWE retrieval for the 37V19V algorithm is between 40 and 80 mm (Fig. 7d). By comparison, the modal retrieval for the new algorithm described in this study is between 100 and 140 mm (Fig. 7c). While it is difcult to construct a meaningful SWE histogram for observations across this region due to the sparse observing network, the 37V GHz algorithm produces a more appropriate distribution relative to other published eld observations (i.e. Sturm & Liston, 2003), estimates from other independent products such as the global daily gridded snow depth product from the Canadian Meteorological Centre (Brasnett, 1999), and the observations cited previously in this study. For instance, note in Fig. 3b that eld measurements acquired across the sub-Arctic during the airborne passive microwave campaigns are predominantly greater than 80 mm.

Fig. 7. Regional sub-Arctic tundra SWE maps, 1 April 2007, from (a) the relationship shown in Fig. 6b, and (b) the open environments algorithm of Goodison and Walker (1995). SWE retrieval histograms are shown below each map in (c) and (d).

1708

C. Derksen et al. / Remote Sensing of Environment 114 (2010) 16991709 Brasnett, B. (1999). A global analysis of snow depth for numerical weather prediction. Journal of Applied Meteorology, 38, 726740. Chang, A., Foster, J., & Hall, D. (1990). Satellite sensor estimates of northern hemisphere snow volume. International Journal of Remote Sensing, 11(1), 167171. Derksen, C. (2008). The contribution of AMSR-E 18.7 and 10.7 GHz measurements to improved boreal forest snow water equivalent retrievals. Remote Sensing of Environment, 112, 27002709. Derksen, C., Brown, R., & Wang, L. (2009). Terrestrial snow (Arctic). State of the Climate in 2008. Bulletin of the American Meteorological Society, 90. (pp. S1S196). Derksen, C., Sturm, M., Liston, G., Holmgren, J., Huntington, H., Silis, A., et al. (2009). Northwest Territories and Nunavut snow characteristics from a sub-Arctic traverse: Implications for passive microwave remote sensing. Journal of Hydrometeorology, 10 (2), 448463. Derksen, C., Walker, A., & Goodison, B. (2005). Evaluation of passive microwave snow water equivalent retrievals across the boreal forest/tundra transition of western Canada. Remote Sensing of Environment, 96(3/4), 315327. De Roo, R., Chang, A., & England, A. (2007). Radiobrightness at 6.7-, 19-, and 37-GHz downwelling from mature evergreen trees observed during the Cold Land Processes Experiment in Colorado. IEEE Transactions on Geoscience and Remote Sensing, 45(10), 32243229. Dery, S., Hernandez-Henrquez, M., Burford, J., & Wood, E. (2009). Observational evidence of an intensifying hydrological cycle in northern Canada. Geophysical Research Letters, 36, L13402. doi:10.1029/2009GL038852. De Seve, D., Bernier, M., Fortin, J. -P., & Walker, A. (1997). Preliminary analysis of snow microwave radiometry using the SSM/I passive-microwave data: The case of La Grande River watershed (Quebec). Annals of Glaciology, 25, 353361. Garmin, 2009. About GPSWhat is WAAS? Available at: http://www8.garmin.com/ aboutGPS/waas.html [Accessed July 29th 2009] Goodison, B., & Walker, A. (1995). Canadian development and use of snow cover information from passive microwave satellite data. In B. Choudhury, Y. Kerr, E. Njoku, & P. Pampaloni (Eds.), Passive Microwave Remote Sensing of LandAtmosphere Interactions (pp. 245262). Utrecht, Netherlands: VSP BV. Kelly, R., Chang, A., Tsang, L., & Foster, J. (2003). A prototype AMSR-E global snow area and snow depth algorithm. IEEE Transactions on Geoscience and Remote Sensing, 41 (2), 230242. Kim, E., & England, A. (2003). A yearlong comparison of plot-scale and satellite footprintscale 19 and 37 GHz brightness temperature of the Alaskan North Slope. Journal of Geophysical Research, 108(D13), 43884403. doi:10.1029/2002JD002393. Knowles, K., Savoie, M., Armstrong, R., Brodzik, M. J. (2006, updated 2008). AMSR-E/ Aqua daily EASE-Grid brightness temperatures, 20032008. Boulder, Colorado USA: National Snow and Ice Data Center. Digital media. Koenig, L., & Forster, R. (2004). Evaluation of passive microwave snow water equivalent algorithms in the depth hoar-dominated snowpack of the Kuparuk River watershed, Alaska, USA. Remote Sensing of Environment, 93, 511527. MacPherson, I., Marcotte, D., & Jordan, J. (2001). The NRC atmospheric research aircraft. Canadian Aeronautics and Space Journal, 47(3), 111. Markus, T., Powell, D., & Wang, J. (2006). Sensitivity of passive microwave snow depth retrievals to weather effects and snow evolution. IEEE Transactions on Geoscience and Remote Sensing, 44(1), 6877. Matzler, C. (1994). Passive microwave signatures of landscapes in winter. Meteorology and Atmospheric Physics, 54, 241260. Matzler, C., Schanda, E., & Good, W. (1982). Towards the denition of optimum sensor specications for microwave remote sensing of snow. IEEE Transactions on Geoscience and Remote Sensing, GE-20, 5766. Min, S. -K., Zhang, X., & Zwiers, F. (2008). Human-induced Arctic moistening. Science, 320, 518520. Natural Resources Canada. (2008). Circa-2000 Northern Land Cover of Canada. [Digital Raster Data]. 2008.Ottawa, Ontario: Canada Centre for Remote Sensing, GeoAccess Division Available at: http://geogratis.cgdi.gc.ca/download/landsat_7/Northern_Land_Cover [Accessed June 2008]. Rees, A., Derksen, C., English, M., Walker, A., & Duguay, C. (2006). Uncertainty in snow mass retrievals from satellite passive microwave data in lake-rich high-latitude environments. Hydrological Processes, 20, 10191022. Rees, A., Lemmetyinen, J., Derksen, C., Pulliainen, J., & English, M. (2010). Observed and modelled effects of ice lens formation on passive microwave brightness temperatures over snow covered tundra. Remote Sensing of Environment, 114, 116126. Rott, H., Cline, D., Duguay, C., Essery, R., Haas, C., Kern, M., et al. (2008). Scientic preparations for CoReH20, a dual frequency SAR mission for snow and ice observations. CD-ROM Proceedings, 2008 International Geoscience and Remote Sensing Symposium, Boston MA. Schanda, E., Matzler, C., & Kunzi, K. (1983). Microwave remote sensing of snow cover. International Journal of Remote Sensing, 4(1), 149158. Solheim, F. (1993). Use of pointed water vapor radiometer observations to improve vertical GPS surveying accuracy. Unpublished Ph.D. Dissertation, Dept. Physics, Univ. Colorado, Boulder. Sturm, M., Grenfell, T., & Perovich, D. (1993). Passive microwave measurements of taiga and tundra snow covers in Alaska, U.S.A.. Annals of Glaciology, 17, 125130. Sturm, M., Holmgren, J., & Liston, G. (1995). A seasonal snow cover classication system for local to global applications. Journal of Climate, 8, 12611283. Sturm, M., & Liston, G. (2003). Snow cover on lakes of the Arctic Coastal Plain of Alaska, U.S.A. Journal of Glaciology, 49(166), 370380. Takala, M., Pulliainen, J., Metsamaki, S., & Koskinen, J. (2009). Detection of snowmelt using spaceborne microwave radiometer data in Eurasia from 1979 to 2007. IEEE Transactions on Geoscience and Remote Sensing, 47(9), 29963007. Walker, A., Strapp, J. W., & MacPherson, I. (2002). A Canadian Twin Otter microwave radiometer installation for airborne remote sensing of snow, ice and soil moisture.

Table 3 Comparison of algorithm performance assessed with Alaskan north slope measurements from Sturm and Liston (2003). Year Dates Mean SWE (mm) 100 84 37V19V RMSE 49 29 Mean bias 45 12 37V RMSE 22 39 Mean bias 5 37

2007 2008

29 March2 April 2024 April

2000 and April 2002 from the north slope of Alaska (described in Sturm & Liston, 2003) is summarized in Table 3. Signicant underestimation with a TB difference algorithm (37V19V) is apparent for both years, consistent with the results of Koenig and Forster (2004). Retrievals with the new 37V GHz algorithm agree strongly with the eld measurements from 2000 with an RMSE of 22 mm and a mean bias of 5 mm. In 2002, when the eld measurements were made relatively late in the season (April 20 to 24), the results highlight a challenge with the new 37 GHz single frequency approach. Late in the winter, large changes in TB can occur over short time periods as the snowpack begins to melt and refreeze. Because the SWE retrievals will continue to increase as TB's evolve through the season, these diurnal changes create an overestimation of SWE, as illustrated by the high positive bias with the new algorithm (Table 3). It will be necessary to identify an end date after which large TB uctuations make the SWE retrievals unreliable. This should be relatively straightforward through the use of melt onset detection algorithms available for QuikSCAT (Wang et al., 2008) and passive microwave data (Takala et al., 2009). The next phase in this work is to examine the data record prior to AMSR-E, and apply this prototype algorithm to Scanning Multichannel Microwave Radiometer (SMMR; 19781987) and Special Sensor Microwave/Imager (SSM/I) measurements. Because only a single frequency is utilized, ensuring inter-sensor TB homogeneity will be relatively straightforward. This 30 year data record will characterize variability in SWE across the tundra, and address priority research issues that are difcult to answer with sparse conventional observations. Analysis of snow extent and snow cover duration from the NOAA snow chart record (1966present) shows a decrease in the length of the Arctic snow cover season, driven by earlier spring snow melt (Derksen et al., 2009a). Linking trends in SWE from the satellite passive microwave record with snow extent and melt timing (for example, Wang et al., 2008) datasets will contribute to characterizing the role of terrestrial snow cover in the high latitude water cycle over the past three decades. Acknowledgments Thanks to Ken Asmus, Walter Strapp and Arvids Silis of Environment Canada for support with radiometer measurements and eld activities. AMSR-E brightness temperatures were acquired from the National Snow and Ice Data Center, Boulder, Colorado. Field activities during 2008 near Puvirnituq, Daring Lake, and Inuvik were supported by the Government of Canada International Polar Year funding to the Variability and Change in the Canadian Cryospere network project. SnowSTAR 2007 was supported by the National Science Foundation, the Cold Regions Research and Engineering Laboratory, and Environment Canada. Field measurements at Daring Lake prior to 2008 were supported by the Canadian Foundation for Climate and Atmospheric Science. References
Biancamaria, S., Mognard, N., Boone, A., Grippa, M., & Josberger, E. (2008). A satellite snow depth multi-year average derived from SSM/I for the high latitude regions. Remote Sensing of Environment, 112, 25572568.

C. Derksen et al. / Remote Sensing of Environment 114 (2010) 16991709 CD-ROM Proceedings, International Geoscience and Remote Sensing Symposium, Toronto, June, 2002. Wang, L., Derksen, C., & Brown, R. (2008). Detection of Pan-Arctic Terrestrial Snowmelt from QuikSCAT, 20002005. Remote Sensing of Environment, 112(10), 37943805. Wang, J., & Tedesco, M. (2007). Identication of atmospheric inuences on the estimation of snow water equivalent from AMSR-E measurements. Remote Sensing of Environment, 111, 398408.

1709

Yang, D., Goodison, B., Metcalfe, J., Louie, P., Leavesley, G., Emerson, D., et al. (1999). Quantication of precipitation measurement discontinuity induced by wind shields on national gauges. Water Resources Research, 35(2), 491508. Zhang, X., Zwiers, F., Hegerl, G., Lambert, H., Gillett, N., Solomon, S., et al. (2007). Detection of human inuence on twentieth-century precipitation trends. Nature, 448, 461465.