Journal of Geophysical Research: Atmospheres

A multi-data set analysis of variability and change in Arctic spring snow cover extent, 1967–2008



[1] A new multi-data set estimate of Arctic monthly snow cover extent (SCE) in the May–June melt period is derived from 10 data sources covering different time periods from 1967 to 2008. The data sources include visible and microwave satellite observations, objective analyses of surface snow depth observations, reconstructed snow cover from daily temperature and precipitation, and proxy information derived from thaw dates. The new estimates show a more linear reduction in spring SCE than previously characterized by the National Oceanic and Atmospheric Administration weekly snow chart data set, with air temperature explaining 49% of the variability in Arctic SCE in May and 56% of the variability in June. The Arctic Oscillation is only significantly linked to Arctic SCE in May where it explains 25% of the variance in Eurasian sector SCE. Trend analysis of the multi-data set series (including an annually varying estimate of error) reveals that May and June SCE have decreased 14% and 46%, respectively, over the pan-Arctic region over the 1967–2008 period in response to earlier snow melt. These results are confirmed with in situ data from Canada, Alaska and Russia that show significant reductions in spring snow cover duration over the last 30 years. The spring snow cover temperature sensitivity over the pan-Arctic region during this period is estimated to be in the range −0.8 to −1.00 × 106 km2 °C−1. The observed reductions in June SCE over the 1979–2008 period are found to be of the same magnitude as reductions in June sea ice extent with both series significantly correlated to air temperature changes over the Arctic region and to each other. This result underscores the close relationship between the cryosphere and surface air temperatures over the Arctic region in June when albedo feedback potential is at a maximum.

1. Introduction

[2] The Arctic region has experienced the largest and most rapid warming in the Northern Hemisphere over the past 30 years [Trenberth et al., 2007] with predominantly surface-based warming in the spring period consistent with positive feedbacks from decreasing snow and ice cover [Graversen et al., 2008]. Arctic summer minimum sea ice extent has experienced a well-documented rapid decrease since satellite observations began in 1979 [Stroeve et al., 2007; Serreze et al., 2007; Kwok et al., 2009; Wang and Overland, 2009], but Arctic spring snow cover response appears to be less consistent with the Arctic warming trend. Foster et al. [2008] document a step change in Arctic snow cover during the mid to late 1980s they link to a regime shift in the Arctic Oscillation (AO).

[3] Published estimates of change in Arctic snow cover extent (SCE) are for the most part based on the National Oceanic and Atmospheric Administration (NOAA) weekly data set based on manual interpretation of visible satellite imagery [e.g., Dye, 2002; Foster et al., 2008] or the passive microwave record [e.g., Takala et al., 2009; Tedesco et al., 2009] with assessment of uncertainty usually limited to local to regional-scale comparisons with in situ observations. Monitoring spring snow cover changes over the Arctic region is a major challenge for many reasons including strong local controls on snow cover, frequent cloud cover, large gaps and biases in surface observing networks, and possible confusion of lake ice and snow cover during the melt season. The latter problem and inconsistencies in mapping of patchy snow contribute to the poor performance of the NOAA weekly data set at capturing interannual variability in spring snow cover over northern Canada [Wang et al., 2005a; Brown et al., 2007].

[4] Reliable information on spring snow cover change over the Arctic region is needed for climate monitoring, for understanding the Arctic climate system, and for the evaluation of the representation of snow cover and snow cover feedbacks in climate models [e.g., Barry, 2002; Frei et al., 2005; Roesch, 2006; Key et al., 2007; Brown and Frei, 2007; Räisänen, 2007; Fernandes et al., 2009; Fletcher et al., 2009]. A multi-data set approach to mapping snow cover over the Arctic is a logical response to the challenges mentioned above as it exploits the strengths of the various platforms and methodologies (e.g., the all-weather capability of microwave remote sensing, the high spatial resolution of optical sensors, and the spatial and temporal continuity of operational analysis products) as well as providing an estimate of the uncertainty in SCE.

[5] The main objective of this paper is to develop a multi-data set estimate of spring (May and June) snow cover extent over the pan-Arctic region for the 1967–2008 period to provide new insights into the response of Arctic SCE to recent warming, to examine the consistency of Arctic snow cover products, and to provide estimates of the uncertainty in monthly SCE. The multi-data set SCE series is also used to examine the relative importance of air temperature and the AO in observed SCE variability (i.e., the importance of the AO regime shift noted by Foster et al. [2008]). The paper focuses on the spring period primarily because this is the season when snow cover temperature feedbacks are the strongest [Groisman et al., 1994; Déry and Brown, 2007] but where potential observational errors are also the largest [Wang et al., 2005a; Brown et al., 2007]. There is also a practical consideration to focusing on the spring period as several of the available data sets only provide information on snowmelt onset or snow-off dates.

2. Data Sets

[6] For the purpose of this paper, the Arctic was defined as the area north of 60°N with the North American and Eurasian sectors of the Arctic defined as NA and EUR, respectively. Greenland was excluded from the analysis as there are few products providing reliable seasonal snow cover information due to the complex coastal topography and difficulty of discriminating between snow and ice. A summary of the data sets used to examine Arctic spring snow cover extent is provided in Table 1 with more detailed descriptions provided below. Gridded (5° × 5°) surface air temperature data for the Arctic land areas were obtained from the Climatic Research Unit (CRU) CRUtem3v data set [Brohan et al., 2006], and the AO index used to examine its influence on Arctic spring snow cover variability was obtained from the NOAA Climate Prediction Center ( Winter (January, February, March, JFM) values of the AO index were used following the findings of Rigor et al. [2002] and Bamzai [2003] that spring temperatures and snow cover are most closely linked to AO over the previous winter.

Table 1. Summary of Data Sources Used in the Analysis
DescriptionAcronymPeriodResolutionData Source
Snow-off date from Arctic Polar Pathfinder advanced very high resolution radiometer (AVHRR) data setCCRS (Canada Centre for Remote Sensing)1982−20045 kmCCRS, Zhao and Fernandes [2009]
Daily snow depth analysis from in situ observationsCMC (Canadian Meteorological Centre)1998−2008∼35 kmNational Snow and Ice Data Center (NSIDC), Brasnett [1999]
ERA-40 reanalysis daily snow depthsERA-401957−2002∼275 kmEuropean Centre for Medium-Range Weather Forecasts (ECMWF), Uppala et al. [2005]
ERA-40 reconstructed snow cover duration with temperature index snow model of Brown et al. [2003]ERA-40rec1957−2002∼275 km (with 5 km empirical elevation adjustment)Brown et al. (this study) with 6 hourly temperature and precipitation from ERA-40
NOAA IMS daily 24 km snow/no-snow productInteractive Multisensor Snow and Ice Mapping System (IMS), IMS-24 (interactive multisensor 24 km)1997−200824 km(NSIDC, Ramsay [1998]
NOAA IMS daily 4 km snow/no-snow productIMS-4 (interactive multisensor 4 km)2004−20084 kmNSIDC, Helfrich et al. [2007]
MODIS 0.05° monthly mean snow cover fraction (MOD10CM Version 5) productMODIS (Moderate Resolution Imaging Spectroradiometer)2000−2008∼5 kmNational Aeronautics and Space Administration (NASA), Hall et al. [2006]
NCEP thaw index: snow-off date estimated from 0°C crossing date with NCEP1 Reanalysis daily temperaturesNCEP (National Centers for Environmental Prediction)1948−2008∼275 kmEarth System Research Laboratory, NOAA, Kalnay et al. [1996]
NOAA weekly snow/no snowNOAA (National Oceanic and Atmospheric Administration)1966−2008190.5 kmRutgers U., Robinson et al. [1993]
Snow water equivalent from scanning multichannel microwave radiometer (SMMR, 1978–1987) and the Special Sensor Microwave/Imager (SSM/I, 1987–2008)PMW (passive microwave)1978−200824 kmNSIDC, Savoie et al. [2009]
QuikSCAT snow-off dateQSCAT (SeaWinds-on-QuikSCAT scatterometer)2000−2008∼5 kmWang et al. [2008]

[7] In situ daily snow depth observations from Alaska (P. Groisman, personal communication, 2009) and Canada [Meteorological Service of Canada (MSC), 2000] were used to construct regionally averaged time series of the spring snow cover duration over the Arctic sector of North America to corroborate the trend observed in the multi-data set SCE series over NA. Spring snow cover was defined following Brown and Goodison [1996] as the number of days with snow depth ≥ 2 cm in the February–July period. Annual series were converted to anomalies with respect to a 1971–2000 reference period before regional averaging. Snow depth data from the Eurasian sector of the Arctic were unavailable to carry out the same analysis, but recent insights into snow cover trends in this region are provided by Radionov et al. [2004], Kitaev et al. [2005], and Bulygina et al. [2009]. Monthly sea ice extent data for the Arctic were obtained from the National Snow and Ice Data Center (NSIDC) sea ice index [Fetterer et al., 2002] derived from passive microwave data using the NASA Team algorithm [Cavalieri et al., 1984].

2.1. Description of Snow Cover Data Sets

[8] Canada Centre for Remote Sensing Snow-off Dates (CCRS): Snow-off dates over the pan-Arctic region were supplied by H. Zhao and R. Fernandes (personal communication, 2009) based on daily snow cover fraction (SCF) information derived by Zhao and Fernandes [2009] from the 5 km resolution Arctic Polar Pathfinder advanced very high resolution radiometer (AVHRR) data set over the 1982–2004 period. The snow mapping methodology was based on an adaptive temporal filtering technique developed by Fernandes and Zhao [2008] to take account of cloud cover and provides complete spatial and temporal coverage over the Arctic for the 23 year period. Evaluation with surface observations gave performance results that were comparable to Moderate Resolution Imaging Spectroradiometer (MODIS) [Zhao and Fernandes, 2009]. Snow-off date was defined as the day when each grid cell had no snow for at least three continuous days during the period 1 April to 31 August. Results for 1985 were excluded from the analysis as the SCF values were determined to be anomalously low because of problems with the AVHRR radiances (R. Fernandes, personal communication, 2009). Snow cover was assumed to be continuous up to the snow-off date.

[9] Canada Meteorological Service Global Daily Snow Depth Analysis (CMC): The CMC global daily snow depth analysis is based on optimal interpolation of in situ daily snow depth observations with a first-guess field generated from a simple snow accumulation and melt model driven with analyzed temperatures and forecast precipitation from the Canadian forecast model [Brasnett, 1999]. The analysis has been run essentially unchanged over a 1/3° Gaussian global grid since 12 March 1998. An error in the correction of air temperatures to model topography was introduced into the operational implementation of the analysis in October 2006 that affects snow depths at higher elevations. However, this has been corrected in offline runs, and a corrected offline version of the analysis is being run daily to maintain data continuity. There are few in situ observations over the Arctic so the analysis is based to a large extent on estimated snow depths from the first-guess field. In addition, the available snow depth observations tend to be made at open areas where the snow melts out earlier than the surrounding terrain [Brown et al., 2003]. Grid cells were considered completely snow covered for snow depth values > 1 cm.

[10] European Centre for Medium-Range Weather Forecasts Reanalysis (ERA-40): Monthly SCE information was derived from the daily snow water equivalent (SWE) values in the ERA-40 2.5° reanalysis [Uppala et al., 2005] by assuming grid cells were completely snow covered for SWE values > 1 mm. The SWE values in ERA-40 are obtained from a blend of snow depth observations and model-derived information with the snow depth observations covering different regions and time periods. SWE is estimated from the analyzed snow depth and the corresponding snow density simulated by the forecast model [Drusch et al., 2004]. There is a documented bug in the snow analysis for 1973–1974 and 1990–1994 that caused SWE values to be too low by about a factor of 0.25 during these periods [Betts et al., 2003]. SCE should be relatively insensitive to these errors as well as errors in precipitation because of the low SWE threshold used to estimate snow cover and because SCE variability is usually more closely linked to air temperature. Betts et al. [2003] found that ERA-40 melted snow too quickly over the Mackenzie Basin, and Simmons et al. [2004] showed that ERA-40 had a warm bias compared to the CRU temperature data set before about 1975. The later bias may affect conclusions about trends in SCE derived from ERA-40. Khan et al. [2008] found that ERA-40 provided the best agreement with station data over the former Soviet Union in a comparison with the National Center for Environment Prediction (NCEP) and Japanese 25 year Reanalysis Project reanalyses.

[11] ERA-40 Reconstruction (ERA-40rec): To remove the impact of spatial and temporal variability in the assimilation of snow observations and the previously mentioned errors, daily snow depth was reconstructed using the simplified temperature index model described in the study by Brown et al. [2003] with 6 hourly temperature and precipitation fields from the 2.5° ERA-40 reanalysis described above. The snowpack model includes most of the temperature-dependent processes included in detailed physical models (e.g., partitioning of precipitation into solid and liquid fractions, melt from rain-on-snow events, specification of new snowfall density, snow aging, and snowmelt). Blowing snow and sublimation processes are not included in the simplified snow model. The latter is taken into account through a 20% reduction of precipitation, which falls within the range of estimated snow loss from sublimation for open and forested terrain provided by Pomeroy and Gray [1995]. Temperature index snow models have been shown to be capable of providing realistic simulations of key snowpack properties such as snow cover onset date, maximum accumulation, snow cover disappearance date, and runoff with only temperature and precipitation data as input [Vehviläinen, 1992; Ferguson, 1999; Hock, 2003; Turcotte et al., 2007]. Monthly snow cover duration (SCD) was derived from daily snow depth (threshold > 0) with a 5 km resolution topographic downscaling for SCD of 3.37 days.100 m−1 obtained by Brown et al. [2007] over the Canadian Arctic. The downscaling was applied to the difference in elevation between the observed 5 km mean elevation and the corresponding elevation interpolated from the ERA-40 topography to provide a more realistic spatial distribution of snow cover in mountainous regions. The elevation adjustment is comparable to the 4.3 days.100 m−1 obtained by Tong et al. [2009] for spring snow cover in the Quesnel River Basin of British Columbia using MODIS. The reconstruction will be sensitive to the biases in ERA-40 temperatures noted above.

[12] Moderate Resolution Imaging Spectroradiometer (MODIS): Mean monthly fractional snow cover was obtained from the MODIS/Terra Snow Cover Monthly L3 Global 0.05° MOD10CM version 5 product [Hall et al., 2006] available from September 2000 (missing data for June 2001, March 2002, and December 2003). This data set provides mean monthly SCF based on all the available daily snow cover fractional information computed from a Normalized Difference Snow Index at a nominal pixel spatial resolution of 500 m. A confidence index is provided for each grid cell based on the amount and quality of information available over a month. For this analysis, all valid monthly snow cover fraction values were used irrespective of confidence index to maximize the spatial coverage. According to D. Hall (personal communication, 2009), the monthly SCF is more likely to represent the maximum monthly SCF than the mean. This was investigated by comparing MODIS monthly snow cover with the comparable resolution NOAA IMS 4 km product (defined below) over the Arctic for the 2004–2008 period. The results (not shown) revealed that both products map the same areas as snow covered but with contrasting seasonal relationships in snow cover fraction. The two products are virtually identical in the fall when snow cover fraction increases rapidly from 0% to 100%; however, during the spring melt period, MODIS consistently maps lower snow cover fractions than IMS by an average 28% over NA and 21% over EUR in June. IMS includes daily snow cover information from MODIS so the observed spring differences may be partly procedural in origin.

[13] National Centers for Environmental Prediction Reanalysis (NCEP): Daily SWE values are provided on T62 Gaussian grid output of the NCEP/NCAR (NCEP1) reanalysis [Kalnay et al., 1996] from 1948 but were not analyzed because of well documented errors in the NCEP1 snow analysis that included repeated use of the 1973 snow cover data for the entire 1974−1994 period as well as a 100 mm limit on snow accumulation [Kanamitsu et al., 2002]. Daily SWE values are also available from 1979 for the NCEP/Department of Energy (NCEP2) Reanalysis 2 product (U.S. Department of Energy) [Kanamitsu et al., 2002], but SCE in this product is constrained by the NOAA weekly snow cover extent analysis. Several studies have shown that NCEP1 temperatures have some ability to capture melt dynamics over northern latitudes [e.g., Atkinson et al., 2006; Brown et al., 2007]. Wang et al. [2008] showed that snow clearance was associated with a well-defined transition to above-freezing temperatures in the Arctic region so a proxy spring SCE series was developed from the NCEP1 daily surface air temperatures assuming that snow was present on the ground until the 7 day running mean of daily mean surface air temperature exceeded 0°C. Kanamitsu et al. [2002] show that the snow cover analysis error in NCEP1 generated a warm bias in transition months (their Figure 1) that may impact this kind of proxy series. However, comparison of the proxy SCE series with corresponding monthly air temperature anomalies from CRU (not shown) did not show any evidence of systematic shifts in the relationship between snow cover and temperature during the 1974–1994 period of the snow analysis error.

Figure 1.

Average % error in Arctic monthly SCE for NOAA weekly and IMS-24 km products compared to the higher resolution IMS-4 km product over the 2004–2008 period.

[14] National Oceanic and Atmospheric Administration (NOAA) Weekly Snow Cover Product (NOAA): The NOAA weekly product is described in the study by Robinson et al. [1993] and consists of weekly charts of snow cover extent derived from manual interpretation of visible satellite imagery. Before 1999, the presence/absence of snow was manually digitized onto a 190.5 km polar stereographic grid over the Northern Hemisphere with presence/absence determined based on a 50% snow cover threshold in each grid cell. In practice, Brasnett [1999] found that a lower 30% threshold was required to emulate the snow-covered area in the NOAA analysis so there is built-in conservatism in the product particularly in mountain regions. The charting method changed in May 1999 with the introduction of the higher resolution 24 km daily IMS snow cover product [Ramsay, 1998]. A pseudoweekly product is automatically derived from the IMS 24 km daily product assuming the analysis for Sunday is representative of the previous week. The data used in this study were monthly snow cover duration fraction information derived by Dr. David Robinson at Rutgers University. It should be noted that the NOAA product includes changes in the volume and resolution of satellite information over time; procedural changes in the way patchy snow was mapped by analysts, as well as the shift to automated generation of the weekly chart from the IMS daily product in 1999. This has resulted in inconsistencies in the mapping of snow that are particularly apparent over the Arctic [Wang et al., 2005a; Brown et al., 2007; Déry and Brown, 2007]. A major effort has just been completed to create a climate date record (CDR) with the NOAA data set (D. Robinson, personal communication, 2009), but the CDR version of the NOAA weekly data set was not available for analysis at the time this study was undertaken.

[15] NOAA IMS daily 24 and 4 km snow cover products (IMS-24, IMS-4): The IMS snow cover analysis system is described by Ramsay [1998] and Helfrich et al. [2007] and consists of an interactive workstation for snow cover mapping by a snow analyst with tools for overlaying and interpolating snow cover information from a variety of data streams. The system relies mainly on visible satellite imagery (including MODIS data) but is augmented by station observations and passive microwave data. The IMS system started mapping snow cover at a 24 km resolution in early 1997 with the data officially coming on-stream in May 1999. The analysis was run at a higher 4 km resolution from early 2004. Daily data from both the 24 and 4 km versions of IMS snow cover analysis are archived at NSIDC. There were insufficient data for the IMS-4 km to be included in the development of a multi-data set SCE series but they allowed an estimate to be made of the influence of spatial resolution on Arctic SCE estimates by comparing it to the derived IMS-24 km daily and NOAA 190.5 km weekly products over the common 2004–2008 period of data. The comparison (Figure 1) indicated that error remained below 10% throughout the whole year for the 24 km product but exceeded 10% for the coarser 190.5 km NOAA weekly product from July to October. The NOAA weekly product resolution error is below 10% in May and June during the main period of snow cover depletion over the Arctic. A comparison of the IMS 4 km and MODIS products mentioned previously showed that IMS mapped higher snow cover fractions over the Arctic during the spring melt period. This tendency was previously documented with the NOAA weekly product by Wang et al. [2005a] and Brown et al. [2007] and attributed to cloud cover and less frequent satellite information over higher latitudes. The fact that the IMS product now includes MODIS information suggests there may be more of a procedural origin for the differences between the two products in the spring period, in agreement with the comment by Wang et al. [2005a] that snow analysts were trained to be very “aggressive” in their classification of snow cover.

[16] Passive microwave (PMW): Passive microwave data from the scanning multichannel microwave radiometer (SMMR, 1978–1987) and the Special Sensor Microwave/Imager (SSM/I, 1987–2008) offer the potential for all-weather monitoring of Arctic snow cover over a 31 year period. There are well-documented uncertainties in using PMW measurements to retrieve hemispheric SWE information due to regional differences in land cover and snow cover properties that influence microwave emission and scatter [Derksen, 2008; Kelly et al., 2003; Biancamaria et al., 2008]. There is better skill, however, at capturing the snowmelt signal [Takala et al., 2009; Tedesco et al., 2009] because of the relatively straightforward and strong influence of liquid water in the snowpack on microwave emission. The daily coverage and insensitivity to cloud cover makes PMW data attractive for mapping changes in spring snow cover over the pan-Arctic.

[17] The data obtained for this study were daily SWE estimates from SMMR and SSM/I provided by NSIDC (version NSIDC_S11_SWE) [Savoie et al., 2009] from a baseline version of the NSIDC microwave algorithm [Armstrong and Brodzik, 2001] where SWE is derived as a function of the difference in 19 and 37 GHz horizontally polarized microwave brightness temperatures following Chang et al. [1987]. Previous experience has shown that although the SWE estimates can be highly uncertain in Arctic regions, standard 19–37 GHz SWE retrievals are able to provide realistic estimates of spring snow extent variability when used only as a proxy for snow cover [Brown et al., 2007]. The SWE estimates provided by NSIDC had a minimum threshold value of 7.5 mm (values less than this were set to zero in the data set) and a complete snow cover was assumed to exist if the SWE was greater than or equal to this minimum value. A constant melt duration of 9 days was added to the PWM-derived SCD to take account of the time to melt wet snow that is not seen by the SWE retrieval algorithm. The 9 day value is based on an evaluation of QuikSCAT melt onset dates and surface snow cover observations over the Canadian Arctic [Brown et al., 2007]. An additional 3 days was added to the SMMR SCD values to take account of the shift in brightness temperatures between SMMR and SSM/I [Jezek et al., 1991] following the recommendation of Takala et al. [2009]. Comparison of the SMMR and SSM/I monthly SCE climatologies over the Arctic with corresponding values from NOAA (Figure 2) show the two platforms have similar patterns of seasonal differences with virtually identical results in the May–June period analyzed in this study that confirms the appropriateness of the SMMR SCD adjustment. The PMW data set consistently mapped anomalous small amounts of snow over the southern region of the Central Siberian Plateau in June compared to the other data sets. The exact cause of this is unknown, but the snow cover amounts are small and do not appear to have a major impact on the performance results presented in section 4.

Figure 2.

Difference in monthly mean SCE between SMMR and NOAA (1978–1987) and SSM/I and NOAA (1988–2008) for the Arctic region with SMMR adjusted for brightness temperature offset versus SSM/I.

[18] QuikSCAT (QSCAT): Enhanced resolution (4.45 km) Ku-band backscatter measurements from the SeaWinds-on-QuikSCAT scatterometer are able to provide all-weather, high-resolution information on melt onset and melt end dates for land ice and snow cover over the pan-Arctic region [Wang et al., 2005b, 2008]. This study used QSCAT-derived melt end dates (Q) from Wang et al. [2008] over the 2000–2008 period to estimate snow-off dates (S) with the following empirical relationship presented in the study by Wang et al. [2008] where rmse is the root mean squared error:

equation image

Wang et al. [2008] found their QSCAT melt detection algorithm underestimated melt end date in heavily forested regions by about 1 week so an additional 7 days was added in areas with more than 40% tree cover using the 500 m resolution MODIS Vegetation Continuous Field product [Hansen et al., 2003]. Monthly SCE in the spring period was estimated by assuming continuous snow cover at each grid point until the snow-off date (the same procedure as the CCRS snow-off date).

3. Methods

[19] With the exception of MODIS, all the data set s were converted to gridded monthly snow cover duration fraction (SCF = total number of days with snow cover in a month divided by the number of days in the month) assuming 100% snow cover above the various SWE or depth thresholds appropriate to each data set or by assuming a complete snow cover up to the end of the snow cover period in the case of data sets with snow-off dates. The monthly SCF was then multiplied by the grid point area to arrive at the total SCE with Greenland excluded. MODIS differs from all the other data sets in that the monthly SCF is based on the mean daily snow cover area fraction.

[20] The development of a combined data set is not straightforward as the individual sources of information cover different time periods and “see” different amounts of SCE over the Arctic depending on spatial resolution, the method (or algorithm) used to detect snow cover, as well as different definitions of snow cover. This is clearly illustrated in Figure 3 that shows the amount of snow cover seen by the different data sets in May and June of the common year 2002. In May, the differences mainly occur over Scandinavia and western Russia, but differences between data sets are more apparent in June including spurious PMW-derived snow cover in Eurasia over areas south of the snowline and more extensive snow cover over Siberia and Alaska in the NOAA and IMS products.

Figure 3.

Comparison of Arctic monthly snow cover fractions (%) seen by the various data sets in May 2002.

Figure 3.


[21] A summary of mean Arctic SCE estimates from the various data sets (Table 2) for two periods of overlapping data (1982–2002 and 2004–2008) show that May SCE averages around 10–11 million km2 with an interdata set standard deviation of 0.5–0.8 million km2. In June, SCE decreases to around 4 million km2, but the interdataset standard deviation increases to close to 1 million km2. The larger noise in June is related to a more patchy snow cover during the end of the melt period, to a greater potential for wet snow conditions that cause problems for microwave-based snow mapping systems, and to the concentration of a large fraction of the Arctic snow cover outside Greenland over the Canadian Arctic Archipelago that is characterized by small islands and complex terrain.

Table 2. Mean Values of May and June SCE (million km2) Seen by Various Sensors and Data Sets Over the Northern Hemisphere North of 60° (Excluding Greenland) for Two Common Periods 1982–2002 and 2004–2008
1982–2002CCRSERA-40ERA-40recNCEPNOAAPMWAverage ± 1 Standard Deviation  
May average SCE10.611.510.410.611.711.111.0 ± 0.53  
June average SCE3. ± 0.96  
2004–2008CMCIMS-24IMS-4MODISNCEPNOAAPMWQSCATAverage ± 1 Standard Deviation
May average SCE9. ± 0.80
June average SCE3. ± 1.06

[22] To compare the data sets on a consistent basis, monthly SCE was converted to standardized anomalies using the mean and standard deviation for two reference periods chosen to maximize data set overlap: Six data sets used a 1982–2002 reference period (NOAA, NCEP, ERA-40, ERA-40rec, CCRS, and PMW) and a 2001–2008 (NCEP, PMW, MODIS, QSCAT, IMS-24, CMC). The NOAA and IMS-24 data sets are essentially the same after 1999 so NOAA was not included in the 2001–2008 analysis. The impact of converting the data to standardized anomalies is demonstrated in Figure 4 for the pan-Arctic region in June. The consistency of each data set was evaluated by computing the correlation and rmse to the multi-data set mean excluding the data set being verified. The CCRS and MODIS data sets were considered benchmarks for each evaluation period as they both represent consistent, objectively derived snow cover estimates from high resolution visible satellite data. The evaluation was carried out for the North American and Eurasian sectors of the Arctic as well as for the pan-Arctic region.

Figure 4.

Example of the process of converting SCE estimates to standardized anomalies for the pan-Arctic region in June. Raw SCE values are shown in the top two frames for each of the common periods of data overlap, with standardized anomalies in the bottom two frames.

[23] The final multi-data set SCE series is obtained by averaging the anomaly series from each reference period, converting the average anomaly series to first differences (i.e., year 2 minus year 1, etc) then joining the difference series in 2001. Use of a first-difference series minimizes the impact of differences in the means and standard deviations between the reference periods. The final SCE anomaly series is then constructed by providing a starting value in 1967 (there were at least four data sets per year available for averaging from 1967). An estimate of the uncertainty in SCE in each year is obtained from the standard error (SE)

equation image

which depends on the standard deviation s of the n data sets included in the average anomaly. The uncertainty estimates were included in linear trend analysis of Arctic SCE using subroutine FITEXY (Numerical Recipes, Press et al. [1992]), which takes into account of errors in the dependent variable.

4. Results

4.1. Data Set Evaluation

[24] The results of the data set consistency evaluation are summarized in Tables 3 and 4. The evaluation results indicate generally consistent agreement between data sets with correlations greater than 0.7 and rmse values less than 0.8 × 106 km2 in most cases. NOAA, IMS-24, and CMC are exceptions with noticeable departures from the multi-data set average in June for reasons previously mentioned. The multi-data set average was observed to give consistently good agreement with the CCRS and MODIS benchmarks for May and June over both sectors of the Arctic with an average correlation of 0.87.

Table 3. Correlation of Monthly SCE Anomaly Series With the Average Anomaly Series From the Five Other Data Setsa
  • a

    Correlations with the benchmark series for each evaluation period are shown in parentheses. Bold values indicate where the correlation to the average anomaly series was below 0.7

1982–2002 Period (CCRS Benchmark)
NA-Arctic0.950.94 (0.89)0.94 (0.93)0.86 (0.93)0.75 (0.77)0.81 (0.82)
EUR-Arctic0.820.88 (0.68)0.90 (0.89)0.89 (0.75)0.75 (0.66)0.89 (0.74)
Pan-Arctic0.830.88 (0.73)0.89 (0.92)0.83 (0.75)0.60 (0.58)0.74 (0.64)
NA-Arctic0.890.73 (0.69)0.89 (0.87)0.78 (0.81)0.11 (0.48)0.86 (0.89)
EUR-Arctic0.830.77 (0.71)0.94 (0.81)0.82 (0.86)0.55 (0.69)0.80 (0.71)
Pan-Arctic0.850.76 (0.73)0.88 (0.81)0.80 (0.88)0.29 (0.59)0.73 (0.75)
2001–2008 period (MODIS Benchmark)
NA-Arctic0.74 (0.74)0.84 (0.82)0.890.94 (0.80)0.80 (0.70)0.77 (0.84)
EUR-Arctic0.89 (0.94)0.88 (0.92)0.950.74 (0.81)0.63 (0.57)0.92 (0.89)
Pan-Arctic0.74 (0.86)0.83 (0.83)0.930.87 (0.89)0.45 (0.37)0.92 (0.94)
NA-Arctic0.40 (0.16)0.56 (0.86)0.760.76 (0.51)0.92 (0.80)0.93 (0.74)
EUR-Arctic0.64 (0.82)0.47 (0.20)0.830.83 (0.75)0.97 (0.87)0.85 (0.81)
Pan-Arctic0.45 (0.48)0.42 (0.51)0.850.85 (0.73)0.96 (0.84)0.88 (0.86)
Table 4. Root Mean Squared Error for Standardized Monthly SCE Anomalies From Each Data Series Versus the Average Standardized Anomaly Series From the Five Other Data Sets for Both Evaluation Periodsa
1982–2002 Period
  • a

    Bold values indicate where rmse exceeded 0.8 to highlight outliers. The units are dimensionless.

2001–2008 Period

[25] The rmse results show a general increase in uncertainty levels from May to June for most of the data sets related to the smaller snow-covered area and greater potential for errors. The microwave-based measurements (PMW and QSCAT) were an exception, particularly over NA, where agreement with the average anomaly series improved in June. The added value of microwave-based observations during the melt season provides additional justification for using a multi-data set approach for mapping snow cover over the Arctic. It had been intended to remove the poorer performing data sets from the final average anomaly series. However, multiple regression analysis revealed that all of the data sets were statistically significant (0.05 level) variables in explaining the variance in the multi-data set anomaly series so there was no compelling reason to eliminate any of the data set s from the final anomaly series.

[26] Stratification of the evaluation results in Tables 3 and 4 by data set resolution reveals that the highest resolution data sets (CCRS, QSCAT, MODIS, ERA-40rec with resolution ∼5–10 km) have the highest correlations and lowest rmse compared to the multi-data set average in both May and June. There is no clear difference in performance between the medium resolution (IMS-24, CMC, PMW with resolution ∼25–35 km) and coarser resolution data sets (ERA-40, NCEP, NOAA with resolution ∼200–300 km).

4.2. Analysis of Multi-Data Set SCE Series

[27] The combined monthly anomaly series were converted back to snow-covered area in million km2 using the 1982–2002 reference period mean and standard deviation from the CCRS benchmark data set. The series are shown in Figure 5 along with error bars representing 2 times the standard error defined in equation (2). Linear trend analysis including the annual error term is summarized in Table 5 along with trend values obtained using the NOAA data set. The multi-data set shows statistically significant (0.05 level) decreases in SCE in both May and June over both sectors of the Arctic with a 46% decrease in pan-Arctic SCE in June over the 1967–2008 period in response to earlier snowmelt. This value is almost identical to that obtained with the NOAA data set except that the NOAA data set shows greater losses over Eurasia. Pan-Arctic May SCE was estimated to have decreased by 14% over the 1967–2008 period that is slightly higher than the 9% reduction estimated from NOAA.

Figure 5.

Variability and trend in (top) May and (bottom) June SCE over the Arctic region with error bars showing ±2 times the standard error of the multi-data set average. Trend estimates were computed from least squares and are summarized in Table 5.

Table 5. 1967–2008 Least Squares Trend Analysis Results for the Multi-Data Set Series With Annual Values of the Standard Error in SCE Included in the Regressiona
MonthRegion1967–2008 Mean SCE (Standard Deviation) 106 km2Multi-Data Set Trend Over 1967–2008 106 km2 10 yr−1 (Slope Standard Error)% Change in Mean SCE Over 1967–2008 (NOAA % Change)
  • a

    Corresponding trends from the NOAA data set are shown in parentheses in the last column. All trends are significant at the 0.05 level with the exception of NOAA trend in NA May SCE. The mean and standard deviation of the multi-data set SCE over the 1967–2008 period are included in the first column for reference.

MayNA-Arctic3.86 (0.28)−0.11 (8.61 × 10−3)−11.6 (−4.8)
EUR-Arctic6.85 (0.53)−0.19 (1.58 × 10−2)−12.1 (−11.1)
Pan-Arctic10.71 (0.62)−0.36 (1.89 × 10−2)−14.2 (−8.6)
JuneNA-Arctic1.82 (0.34)−0.13 (1.17 × 10−2)−31.1 (−23.8)
EUR-Arctic1.96 (0.49)−0.23 (1.65 × 10−2)−50.3 (−68.1)
Pan-Arctic3.81 (0.67)−0.40 (2.67 × 10−2)−45.7 (−45.6)

[28] The new multi-data set estimate of change in June SCE over the pan-Arctic region from 1979 to 2008 period corresponds closely with observed decreases in Arctic sea ice extent from passive microwave data [Fetterer et al., 2002] (Figure 6): June sea ice extent decreased at a rate of 0.41 × 106 km2 10 yr−1, whereas June SCE decreased at a rate of 0.39 × 106 km2 10 yr−1 with the two detrended series significantly correlated (r = 0.57). The decreases in both cases are consistent with observed strong warming over the Arctic in June with detrended series of snow cover and sea ice extent exhibiting significant negative correlations with June air temperature anomalies of −0.75 and −0.66, respectively. In May, only SCE was significantly correlated with Arctic air temperature anomalies and snow cover and sea ice extent were not significantly correlated with each other.

Figure 6.

Variation in Arctic June snow cover extent, sea ice extent, and air temperature (plotted as negative anomaly) over the 1979–2008 period. Least squares trends shown for snow cover and sea ice extent.

[29] Analysis of snow cover duration data from in situ daily snow depth measuring sites in Canada and Alaska identified 23 stations (14 in Canada and 9 in Alaska) with more-or-less complete data over the 1967–2007 period for verifying the multi-data set result (there were too few stations with data in 2008 to include this year in the analysis). The regionally averaged anomaly series (Figure 7) exhibited a significant (0.05 level) reduction over the 1967–2007 period of 2.2 days per decade with CRU May temperature anomalies over NA Arctic land areas explaining 32% of the variance (winter AO did not explain a significant fraction of the variance). Data were unavailable to carry out the same analysis over northern Eurasia, but a review of recent Russian publications [Radionov et al., 2004; Kitaev et al., 2005; Bulygina et al., 2009] show that annual SCD increased over much of northern Eurasia from the mid-1930s when measurements began until the late 1970s when a rapid decline set in of about 4 days per decade based on data presented in the study by Kitaev et al. [2005]. According to Radionov et al. [2004] the largest decreases have occurred in coastal regions.

Figure 7.

Variability and linear (least squares) trend in North American Arctic spring (February–July) snow cover duration (SCD) over the 1967–2007 period for 23 stations (14 in Canada and 9 in Alaska) with more-or-less complete data.

[30] The Russian papers unfortunately do not provide any information on changes to start and end dates of snow cover so it is unclear how much of the reported decreases in annual SCD are due to earlier snowmelt or a later start to the snow season. Analysis of fall (August–January) and spring (February–July) snow cover duration trends with the NOAA gridded data set over the 1972–2008 period of continuous data (Figure 8) show a marked difference in seasonal snow cover trends over Eurasia with many regions experiencing earlier starts to the snow cover season, so most of the recent documented decreases in annual SCD must be occurring in the spring period. The exceptions to this are the Canadian Arctic Archipelago, Alaska, northern Scandinavia, and northeastern Siberia where SCD has decreased in both the fall and spring. Both continents show clear evidence of stronger reductions in spring snow cover in northern coastal regions in agreement with the findings of Radionov et al. [2004]. This coastal response is likely linked to enhanced local warming related to thinning sea ice [Lindsay et al., 2009] and earlier sea ice retreat [Howell et al., 2009] that generate positive feedbacks from increased heat transfer to the atmosphere and increased solar heating of open water [Perovich et al., 2007].

Figure 8.

Trend in (left) fall and (right) spring snow cover duration (days/decade) over the 1972–2008 period from the NOAA weekly data set maintained at Rutgers University. Trends were computed using least squares method.

[31] These results underscore the close connection between the cryosphere and surface air temperatures over the Arctic region in June when albedo feedback potential is at a maximum [Déry and Brown, 2007]. Previous estimates of SCE change over the Arctic with the NOAA data set [Foster et al., 2008] emphasized a step change in the mid-1980s linked to a regime change of the Arctic Oscillation (AO). Multiple regression analysis of the multi-data set Arctic SCE in May and June with winter (JFM) AO, monthly temperature anomalies, and linear trend revealed that air temperature was the dominant influence explaining 49% of the variability in Arctic SCE in May and 56% of the variability in June. AO was only significantly linked to Arctic SCE in May where it explained 25% of the variance in Eurasian SCE. There was no significant link between AO and NA SCE which supports the findings of Tedesco et al. [2009] of a weaker AO influence on NA SCE. The presence of the late 1980s shift in AO was detectable in the May and June SCE series over both continents using a number of standard homogeneity tests in the AnClim software package [Štěpánek, 2005]. However, only the June SCE series for Eurasia showed visible evidence of this change with SCE relatively stable before 1988 but entering a decreasing trend after 1988.

[32] The temperature sensitivity of Arctic SCE (dSCE/dT) was obtained from regression analysis of the multi-data set SCE series against monthly temperature anomalies over Arctic land areas obtained from the CRU data set. This kind of information can be used to evaluate the climate sensitivity of Arctic snow cover in climate models. For example, Brown and Mote [2009] noted that most of the CMIP3 climate models were unable to reproduce observations of significant reductions in high-latitude snow cover over the last three decades of the 20th century. Results are summarized in Table 6 along with sensitivities obtained from the NOAA data set. Snow cover temperature sensitivity was estimated to be ∼0.8−1.0 × 106 km2 °C−1 over the pan-Arctic region during the May–June melt period. The NOAA data set yielded a larger range in dSCE/dT (0.5−1.5 × 106 km2 °C−1) over the melt period, but both data sets agreed on higher temperature sensitivities in June versus May and higher temperature sensitivities over the Eurasian sector of the Arctic versus the NA sector. The higher temperature sensitivity in June and the higher temperature sensitivity over Eurasia are both consistent with positive albedo feedback potential as outlined by Déry and Brown [2007].

Table 6. SCE Temperature Sensitivity (106 km2 °C−1) for the Multi-Data Set Series and NOAA Over the 1967–2008 Period From Linear Regression With Climatic Research Unit Monthly Temperature Anomalies Averaged Over Land Areas North of 60°Na
Multi-Data SetNOAAMulti-Data SetNOAA
  • a

    All values were statistically significant at the 0.05 level. r2 values are shown in parentheses.

NA-Arctic−0.22 (0.69)−0.14 (0.21)−0.33 (0.59)−0.37 (0.32)
EUR-Arctic−0.54 (0.74)−0.44 (0.36)−0.57 (0.68)−0.77 (0.30)
Pan-Arctic−0.76 (0.69)−0.53 (0.21)−0.99 (0.77)−1.45 (0.38)

[33] The multi-data set exhibited much stronger correlations to air temperature anomalies by reducing the impact of a number of anomalous years in the NOAA series. It could also be argued the stronger correlation is related to the inclusion of sources of snow cover information that are strongly controlled by air temperature (e.g., the NCEP thaw index and the ERA-40 reconstruction based on a temperature index snowpack model). Comparison of the average anomaly series with and without these two sources of information (not shown) gave virtually identical results confirming the multi-data set anomaly series is not being unduly influenced by air temperature.

5. Summary and Conclusions

[34] To date, assessments of spring snow cover change over the Arctic region have relied on single sources of information that can be affected by changes in mapping procedures, biases in snow retrievals, and changes in satellite sensors. The multi-data set approach taken in this analysis reduces the impact of such inconsistencies and produces robust estimates of Arctic SCE that agree closely with the best available independent information over both the North American and Eurasian Sectors of the Arctic and provide an estimate of the uncertainty in SCE in each year. The robustness of the approach stems from the fact that despite seeing very different amounts of snow cover over the Arctic, most of the data sources exhibit a relatively high degree of coherence at capturing snow cover anomalies. The multi-data set evaluation process also showed that the highest resolution data sets (MODIS, QSCAT, CCRS, ERA-40rec) had the best performance and that microwave-based sources of snow cover information provided added value in June.

[35] The multi-data set SCE series show a 46% reduction in Arctic SCE in June and a 14% reduction in May over the 1967–2008 period. These numbers are not that different from estimates derived from the NOAA weekly data set traditionally used to document snow cover variability and change over the Northern Hemisphere most recently by Lemke et al. [2007] in the 4th Intergovernmental Panel on Climate Change Assessment. However, the multi-data set series shows a much closer link to Arctic temperature anomalies than the NOAA data set by reducing the impact of anomalous values in the NOAA record. While a reanalyzed version of the NOAA data set has recently been developed to correct some of these problems (D. Robinson, personal communication, 2009), the authors recommend a multi-data set approach for snow cover monitoring over the Arctic region in light of the poorer performance of the NOAA IMS-24 km product in June where it holds onto snow too long over large areas of the Arctic. An added benefit of the multi-data set approach is the straightforward determination of standard error in the time series.

[36] The multi-data set estimate of June SCE exhibits a downward trend over the 1979–2008 period that closely follows observed trends in warming and reductions in June sea ice extent. The significant correlations between Arctic snow and ice cover and air temperature in June is consistent with the albedo feedback mechanism that has the strongest feedback potential in June [Déry and Brown, 2007]. The enhanced spring snowmelt in coastal regions observed in the NOAA data set and by Radionov et al. [2004] is also consistent with earlier sea ice retreat and positive albedo feedbacks. The winter AO is only significantly correlated to pan-Arctic SCE variability in May and then only explains a relatively small amount of the variance (21%) compared to temperature (49%). This finding differs from previous studies [e.g., Foster et al., 2008] that characterized a step change in Arctic SCE related to the 1988 shift in AO and provides some support for the conclusions of Cohen and Barlow [2005] that the large-scale features of the global warming trend over the last 30 years are unrelated to AO.

[37] Finally, one encouraging point that emerged from the data set evaluation process was the strong performance of the ERA-40 snow cover reconstruction with topographic adjustment. This data set exhibited the strongest agreement with the CCRS benchmark over the 1982–2002 period and had the lowest rmse values compared to the multi-data set average in both May and June. This finding suggests that high-resolution downscaling of snow cover information from relatively unsophisticated snow models can generate useful information for monitoring snow cover extent variations over the Arctic region.


[38] This study was carried out as part of the Canadian International Polar Year project “Variability and change in the Canadian Cryosphere.” The authors thank Hongxu Zhao and Richard Fernandes (CCRS) for supplying their AVHRR-derived snow-off date data set, Mary-Jo Brodzik (NSIDC) for supplying the PMW data set, David Robinson (Rutgers U.) for providing the NOAA weekly data set, and Bruce Brasnett (CMC) for the CMC global daily snow depth analysis. Thanks are also extended to NSIDC for online access to the MODIS and IMS data sets. European Centre for Medium-Range Weather Forecasts (ECMWF) are acknowledged for providing the ERA-40 data used in this study through the ECMWF data server ( NCEP data were provided by the Physical Sciences Division, Earth System Research Laboratory, NOAA, Boulder, CO ( The helpful comments from four anonymous referees are gratefully acknowledged.