2.1. Snowmelt Detection Algorithms for Satellite Data
 Time series of both active and passive microwave measurements have been widely used for melt detection in the Arctic. During the transition from dry to wet snow, the radar backscattering coefficient (σ0) decreases dramatically as surface scattering begins to dominate over volume scattering [Ulaby et al., 1982]. The exception is over first-year sea ice, where increased σ0 is observed with melt onset because of an accompanying increase in brine volume, which increases the dielectric permittivity and contributes to enhanced volume scattering [Barber and Nghiem, 1999]. For passive microwave observations, the microwave emissivity (ɛ) increases distinctly from dry to wet snow because of the much higher dielectric constant of water, resulting in an increase in the brightness temperature in accordance with the Rayleigh-Jeans approximation [Ulaby et al., 1982, 1986]. These responses to melt onset provide the basis for melt detection using active and passive microwave data.
 Temporal variations in backscatter or brightness temperature allow identification of the date of melt onset. Threshold techniques have been applied to QS data to retrieve melt onset dates for terrestrial snow cover, lake ice, ice caps across the Arctic, and sea ice in the CAA. Multiple indicators have been used to determine the timing of Arctic sea ice melt onset from passive microwave data. Melt detection algorithms used for each surface type in this study are described briefly below and summarized in Table 1.
Table 1. Melt Detection Algorithms and Data for Each Element of the Cryosphere Included in This Study
|Cryosphere Element||Melt Detection Algorithm||Data Set, Resolution, and Reference||Name of Algorithm|
|Terrestrial snow||Multiple melt events; σ0 < 1.7 dB of the previous 5 day average for three or more consecutive days; main melt event has the longest melt duration||QSCAT, 4.45 km [Wang et al., 2008]||qs_terrestrial|
|Ice cap/sheet||σ0 < 3.0 dB of winter mean for 3 or more consecutive days, or σ0 < 3.5 dB of winter mean for 1 day||QSCAT, 2.225 km [Sharp and Wang, 2009]||–|
|Lake ice||σ0 < 4 dB of winter mean for 2 or more consecutive days||QSCAT, 4.45 km [Howell et al., 2009b]||–|
|CAA sea ice||absolute change in σ0 > 2 dB of winter mean; a kriging interpolation method is applied to estimate spatially continuous onset dates||QSCAT, 4.45 km [Howell et al., 2006]||qs_caa|
|Arctic sea ice||Multiple indicators: daily change in 37V, spectral gradient ratio for 37V and 19, P = Tb(19V) + 0.8Tb(37V)||SMM/I, 25 km [Markus et al., 2009]||ssmi_early ssmi_melt|
 1. Terrestrial snow: Melt onset was detected if the daily QS σ0 was 1.7 dB lower than the previous 5 day average for three or more consecutive days. This algorithm, described by Wang et al. , is capable of distinguishing “preliminary” melt events from the “main” melt event and identifies and excludes areas for which a snowmelt event cannot be identified (typically because of dense forest cover or very shallow snow) to prevent erroneous melt retrievals. The main melt onset date was used in this study; this date identifies when snow is wet but still fully covering the ground.
 2. Ice caps/Greenland: On the basis of previous melt detection efforts [Sharp and Wang, 2009], an optimized single set of thresholds was developed to estimate melt onset for all the ice caps in the Arctic from QS data, including the Greenland Ice Sheet (GrIS). Onset was determined when σ0 was either (1) 3.0 dB lower than the winter mean σ0 for 3 or more consecutive days or (2) 3.5 dB lower than the winter mean for 1 day.
 3. Lake ice: Melt onset over Great Bear Lake and Great Slave Lake was detected when QS σ0 was 4 dB lower than the winter mean for 2 or more consecutive days [Howell et al., 2009b]. We include only these two lakes as they are the only large lakes north of 60°N.
 4. Canadian Arctic Archipelago (CAA) sea ice: Melt onset was estimated from QS data using a threshold of absolute change in σ0 (increase for first year ice, decrease for multiyear ice) of more than 2 dB from winter conditions. The algorithm was applied to pixels identified as homogeneous first-year ice (σ0 < −18 dB) or multiyear ice (σ0 < −11 dB) during the winter season. A kriging interpolation method (following Howell et al. ) was used to estimate spatially continuous patterns of onset dates in the CAA.
 5. Sea ice outside the CAA: Melt onset on sea ice outside the CAA was estimated using SSM/I passive microwave data [Markus et al., 2009]. The strength of the melt signal was determined by summing the normalized magnitudes of multiple melt indicators. The day with the greatest sum was taken as the “melt” onset day (hereafter referred to as “ssmi_melt”), and the second largest peak was taken as the “early melt” day (hereafter referred to as “ssmi_early”). “Early melt” is related to the transition period when transformation of the snowpack due to melt-freeze cycles begins. “Melt” is defined as the day after which free water is continuously present in the snowpack. “Early melt” was found to be closely related to melt onset from QS data [Markus et al., 2009] and was therefore used in this study.
 All the algorithms described above have been validated using in situ observations and assessed against other available melt onset data sets within the studies listed in Table 1. We consider these algorithms to be the optimal methods for melt detection for each element of the cryosphere. An attempt was made to develop a melt detection threshold algorithm for the entire cryosphere using only QS data. For sea ice outside the CAA, QS-derived melt onset dates (not shown) were in close agreement with those from the passive microwave data for areas with landfast ice and multiyear pack ice. However, for some areas between the central pack ice and the coastal landfast ice, patterns of QS-derived melt onset dates were noisy: melt onset dates were often either abnormally early or abnormally late compared to those in neighboring grids. The Ku-band QS data are very sensitive to changes in the surface properties of sea ice resulting from lead development, ice motion, and ridging. Cases where extremely early melt onset was thought to be related to lead opening were confirmed by means of Moderate Resolution Imaging Spectroradiometer (MODIS) clear-sky composite images [Luo et al., 2008]. Although similar issues were also seen in the results derived from passive microwave data, it was much less significant in terms of the area and frequency of grids affected. We therefore chose to use the passive microwave data for melt detection on sea ice outside the CAA.
 Considering the differences in the physical basis for melt detection using active and passive microwave data, an important question is whether the melt onset results derived from passive (e.g., SSM/I) and active (e.g., QS) microwave data are consistent. Markus et al.  showed that the 2000–2007 mean “early melt” onset dates from passive microwave data agreed well with melt onset dates derived from QS data on sea ice in the CAA. We also found that the melt onset dates (MODs) over sea ice in the CAA derived from passive microwave [Markus et al., 2009] and QS data [Howell et al., 2006] were consistent in the 2000–2009 period (Figure 1a). The mean difference between the two data sets for the period was 3.6 days, with a standard deviation of 4.0 days. To test this correspondence further, we applied the QS melt detection algorithm developed in Wang et al.  for the terrestrial Arctic to central Arctic sea ice and compared estimates derived from QS and passive microwave data for a region of the central Arctic (see location in Figure 3a). The annual melt onset dates derived from the two data sets are closely covariant, with a mean difference of 3.6 days and a standard deviation of 2.2 days (Figure 1b). This agreement justifies our use of the combined active and passive microwave data set to generate an integrated melt onset data set for the entire Arctic cryosphere.
Figure 1. Annual MOD (day of year) from SSM/I and QS for sea ice regions in the (a) CAA and (b) central Arctic. See location of central Arctic region in Figure 3a.
Download figure to PowerPoint
 Using the algorithms shown in Table 1, MODs were determined for each component of the cryosphere and combined into a single data set for each year of the 2000–2009 period. This study uses two data sources: (1) enhanced-resolution products produced from QS L1B data with the scatterometer image reconstruction algorithm [Early and Long, 2001], available from the Brigham Young University Centre for Remote Sensing [Long and Hicks, 2005]; and (2) daily averaged brightness temperatures from SSM/I [Maslanik and Stroeve, 1990], available from the National Snow and Ice Data Center in Boulder, Colorado. For visualization purposes, we downscaled the 25 km passive-microwave-derived sea ice data set and scaled up the 2.225 km QS-derived data set for ice caps to the 4.45 km QS polar stereographic grid using a nearest neighbor method. The mean melt onset dates (MMOD), along with the standard deviation, were computed for the most recent decade. The standard deviation was computed only for areas where melt occurred in at least 5 years in the 10 year period. All statistics were calculated from the original-resolution data to avoid artifacts due to data resampling.
 Our study area lies north of 60°N and is assumed to be completely snow covered prior to the spring melt period. The active- and passive-microwave-derived MOD is primarily associated with the early stage of spring snowmelt. However, snowmelt onset is sometimes not detected in tundra regions with shallow snow cover or in regions with dense forest cover [Wang et al., 2008]. For areas of relatively thin ice cover, such as the marginal sea ice zone, the melt signal is not always clear. In these areas, MOD was defined following Markus et al.  as the day on which ice concentration dropped below 80% for the last time before becoming ice-free. Hereafter in this paper we use “MMOD” to refer to the 10 year mean melt onset date during the period 2000–2009; we use “MOD” to refer to melt onset date in all other situations, such as the MOD at a specific year for a specific component of the cryosphere.
2.2. Simulated Snowmelt Onset in CGCM3
 The pan-Arctic melt onset data set described above was used to evaluate the simulated melt onset in the third version of the Canadian Coupled Global Climate Model (CGCM3) [Flato and Boer, 2001; Scinocca et al., 2008], which used a T63 horizontal resolution (approximately 330 km or 2.81° × 2.81°) and 31 vertical atmospheric layers. The particular CGCM3 simulation used in this study was integrated from 1850 to 2000 using historical forcings and then continued using forcings from the Special Report on Emission Scenarios A1B scenario [Nakicenovic et al., 2000] beyond year 2000. We then compared CGCM3 mean Arctic snow cover extent over the April-June melt period with that yielded by other Coupled Model Inter-comparison Project (CMIP3) GCMs [see Brown and Mote, 2009, Table 2] and the NOAA satellite data set (not shown); the comparison reveals that the seasonal ablation timing and rates in CGCM3 are similar to that from the satellite data and the GCM ensemble average. The annual melt onset date at each grid point north of 60°N was determined from daily values of snowmelt runoff simulated by CGCM3 [Flato and Boer, 2001; Scinocca et al., 2008]; these values were provided by the Canadian Centre for Climate Modeling and Analysis (CCCma), as this variable is not available from the CMIP3 archive. Melt onset was assumed when snowmelt runoff was greater than zero for 3 or more consecutive days. Mean values of melt onset date and standard deviation were computed over the 2000–2009 period to match the satellite record.
 In CGCM3, snow cover and snowmelt processes are treated in the Canadian Land-surface Scheme (CLASS V2.7) over land areas [Verseghy, 1991; Verseghy et al., 1993] and in a thermodynamic model over sea ice, as outlined by Flato and Brown . CLASS has three soil layers, a snow layer, and a vegetation canopy with physically based calculations of heat and moisture transfers at the surface and across layer boundaries. Snow is treated as a variable depth fourth “soil” layer and snow-covered and snow-free areas are treated separately [Brown et al., 2006]. On ice caps and GrIS, glacier ice at the bottom of the snowpack is treated as “soil” layers without pores. The energy balance of the snowpack is solved iteratively for surface temperature, taking into account incoming shortwave and longwave radiation, snow albedo and density, sensible and latent heat exchanges, and the ground heat flux. If the simulated surface temperature is above freezing, the surface temperature is reset to 0°C, and the excess energy is applied to melting snow. The amount of liquid water percolation through the snowpack (snowmelt or liquid precipitation) is determined from the heat balance of the snowpack. If the snowpack temperature is below freezing, liquid water refreezes and the heat released warms the snowpack. Once the snowpack is isothermal, all meltwater is assumed to percolate through the snowpack and contribute to snowmelt runoff.
 To assess simulated snowmelt onset patterns from CGCM3 (hereafter referred to as “model” or “simulated” MMOD), the combined active- and passive-microwave-derived data set (see section 2.1, hereafter referred to as “satellite” or “observed” MMOD) was regridded to the model grid (∼2.81° × 2.81° latitude-longitude grid) by averaging all the grid cells whose centers fall within the model grid cell. In the following sections, we present the satellite results on the model grid side by side with the model results. We compare the spatial and temporal distribution patterns of the simulated and the observed MMOD and attempt to explain the observed differences between them. We investigate whether or not the observed differences exhibit any latitudinal or longitudinal dependencies and, if they do, whether they are related to vegetation or topography, or both, in the terrestrial Arctic.
2.3. Analysis Methods and Other Data Sets
 Forest cover and topography can exert large effects on MMOD across the terrestrial Arctic [Wang et al., 2008]. Quantitative information about forest cover and topography is therefore helpful in interpreting the spatial distribution of MMOD. Tree fraction estimates (percentage of forest in each grid) for each QS grid were aggregated from the 500 m resolution MODIS vegetation continuous field product [Hansen et al., 2006]. Surface elevations were obtained from the Global 30 Arc-second Elevation Data set (GTOPO30), developed by the U.S. Geological Survey (http://www.webgis.wr.usgs.gov/globalgis/gtopo30/gtopo30.htm), which uses the same grid as the QS data. Multiple regression of MMOD against tree fraction, elevation, and latitude was conducted to investigate the combined effects of these geographic variables on melt onset in the North American and Eurasian Arctic mainland. Simple linear regression between MMOD and each individual variable was also performed to provide an indication of the relative importance of the different variables as influences on MMOD.
 We also investigated the effects of elevation on MMOD at the regional scale. The mean vertical melt progression rate, dz/dMMOD (N, m d−1) was obtained through linear regression analysis between elevation and MMOD over a 21 × 21 moving window. The size of the moving window (∼100 km) is about the same as that used by Brown et al.  to document the elevation dependence of spring snow cover duration (SCD) analysis. Only statistically significant regression results are reported.
 The mean horizontal melt progression rate (M, km d−1) was computed from the average absolute gradient in MMOD between each grid point and the eight adjacent points on the satellite grid or the four adjacent points on the model grid. This variable provides information on melt dynamics and identifies regions with strong feedback potential. To remove noise, a 5 × 5 spatial averaging was applied to MMOD on the satellite grid before the calculation. A cutoff was applied for M > 330 km d−1 as it approaches infinity as the spatial gradient in MMOD approaches zero.
 To examine whether or not there are latitudinal or longitudinal dependencies in the simulated and observed MMOD and their difference, zonal and meridional averages of MMOD were computed on the model grid for every grid in the latitude direction (∼2.81°) and for every three grids in the longitude direction (∼8.44°).
 Monthly sea ice extent data were obtained from the National Snow and Ice Data Center sea ice index [Fetterer et al., 2009], which is derived from passive microwave data using the NASA Team algorithm [Cavalieri et al., 1996]. The relationship between melt onset timing and monthly sea ice extent north of 60°N was investigated through correlation analysis. This analysis was performed between sea ice extent and MMOD over the terrestrial Arctic mainland (excluding Arctic islands and Greenland) and on the Arctic sea ice (excluding the Bering Sea). The linear trend in each data set was removed prior to the analysis.
 Monthly air temperature (April to June) at 850 hPa pressure level was obtained from the National Centers for Environmental Prediction/National Center for Atmospheric Research Reanalysis 1 data set [Kalnay et al., 1996]. The 850 hPa level was chosen because it is generally above the main topographic features across the continental Arctic. The significance levels of the regressions and correlations were determined using a two-tailed Student's t test. All reported correlations are statistically significant at a significance level of 0.05.