We show the first results of a large-scale survey of snow depth on Arctic sea ice from NASA's Operation IceBridge snow radar system for the 2009 season and compare the data to climatological snow depth values established over the 1954–1991 time period. For multiyear ice, the mean radar derived snow depth is 33.1 cm and the corresponding mean climatological snow depth is 33.4 cm. The small mean difference suggests consistency between contemporary estimates of snow depth with the historical climatology for the multiyear ice region of the Arctic. A 16.5 cm mean difference (climatology minus radar) is observed for first year ice areas suggesting that the increasingly seasonal sea ice cover of the Arctic Ocean has led to an overall loss of snow as the region has transitioned away from a dominantly multiyear ice cover.
 Knowledge of snow depth on sea ice is critical for understanding the precipitation, heat, and radiation budgets of the polar regions. The snow cover on sea ice has a high surface albedo which limits the amount of shortwave heating during the spring season when the shortwave flux is high and melt ponds have not yet formed [e.g., Lindsay, 1998]. The low thermal conductivity of snow makes it an effective insulator thereby impacting the growth and decay of the underlying sea ice layer, as well as reducing the transfer of heat between the ocean and atmosphere [Maykut, 1978; Kurtz et al., 2011]. Furthermore, knowledge of snow loading (snow depth and density) on sea ice is required for retrievals of sea ice thickness from airborne and spaceborne altimeters [e.g., Kurtz et al., 2009; Kwok and Cunningham, 2008; Giles et al., 2008]. For these reasons, the determination of snow depth using the University of Kansas' frequency modulated continuous-wave (FMCW) snow radar is one of the sea ice mission objectives for NASA's ongoing Operation IceBridge airborne campaign. The IceBridge mission has flown more than twenty trans-oceanic surveys over Arctic sea ice since its inception in 2009.
 To date, climatological [Warren et al., 1999], observational [e.g., Markus and Cavalieri, 1998; Gerland and Haas, 2011], and model approaches [e.g., Kwok and Cunningham, 2008] have all been used to investigate the snow cover on sea ice. These snow data sets have been combined with satellite altimetry data to determine sea ice thickness in the Arctic, but errors in the snow depth data are presently not well constrained. The climatology of snow depth on sea ice developed by Warren et al.  (hereafter referred to as W99) was derived from in-situ data gathered over multiyear ice during the 1954–91 time period. Due to the large observed loss of multiyear ice in recent years, the climatology may no longer provide an accurate representation of current snow conditions. Despite these potential differences, the W99 climatology has been combined with data from past and current satellite altimetry missions for the retrieval of sea ice thickness [e.g., Laxon et al., 2003; Giles et al., 2008; Kwok and Cunningham, 2008; Kurtz et al., 2009]. Model approaches suffer from uncertainties in the magnitude of precipitation [Serreze and Hurst, 2000], difficulties in estimating snow loss to leads [Déry and Tremblay, 2004; Leonard and Maksym, 2011], compaction and densification issues, and many other factors. Observational approaches to measuring snow depth using passive microwave data have been developed [e.g., Markus and Cavalieri, 1998], but uncertainties due to surface roughness variations remain [Stroeve et al., 2006] and the methodology is unsuitable for the retrieval of snow depth over multiyear ice in the Arctic [Comiso et al., 2003]. The IceBridge snow radar has been designed specifically for the determination of snow depth from an airborne platform [Panzer et al., 2010]. The feasibility of using an airborne radar for the retrieval of snow depth was established during previous surveys from ground and helicopter platforms [e.g., Kanagaratnam et al., 2007; Galin et al., 2011].
 Initial comparisons with an in-situ data set have shown that the IceBridge snow radar is capable of accurate snow depth retrieval [Farrell et al., 2011], providing a valuable tool for altimetric sea ice thickness retrievals and future studies of the Arctic snow pack. IceBridge snow depths will thus provide a new source for quantifying errors in the different approaches used to estimate basin-wide Arctic snow depth on sea ice and for investigating the impact of such errors on sea ice thickness retrievals from satellite altimetry data including ICESat [Zwally et al., 2002], CryoSat-2 [Wingham et al., 2006], and ICESat-2 [Abdalati et al., 2010]. The data will also be useful for the development and validation of improved methods to estimate snow depth on sea ice from existing model and observational data.
 In this study, we use snow depth retrievals from the snow radar system to quantify the large-scale snow depth properties of Arctic sea ice in April, 2009. A comparison with the climatology of W99 is conducted to determine the utility of climatological data for combination with current laser and radar altimeter data sets for ice thickness retrievals. In addition, the comparison with the historical climatology gives insight into the extent to which recently observed shifts in the Arctic sea ice regime towards younger and thinner sea ice have impacted the snow cover. This comparison provides a baseline for determining changes in the mean depth and spatial variability of the snow cover within the IceBridge survey area which have occurred since the 1954–1991 time period.
2. Data Sets and Methods
 The IceBridge snow radar data used in this study [Leuschen, 2009] consists of 4 flights between April 2nd and April 25th, 2009. An IceBridge flight over the Fram Strait was conducted on March 31st. However, our assessment of this data revealed that differences in the radar operating parameters may limit the quality of snow depth retrieved from the algorithm presented in this study. We therefore do not use data collected during this flight. The FMCW radar uses a nominal sweep from 2–8 GHz with a usable bandwidth of ∼4.5 GHz. This results in a maximum range resolution of ∼3 cm and an estimated 5 cm snow depth retrieval resolution based on the 3 dB width of returns observed over flat, specular targets. The radar has a pulse-limited footprint size of 16 m at the nominal IceBridge flight altitude (460 m) and a spatial sampling frequency of 1 m (see Panzer et al.  for full technical details on the radar design and operation). To boost the signal-to-noise ratio, the data are averaged to provide an along track resolution of 40 m for snow depth retrieval. The snow depth is found by identifying the snow-air and snow-ice interfaces in the time domain and multiplying the time separation of the two interfaces by the speed of light within the snow pack. The propagation velocity of light in the snow pack was taken to be constant (2.34e8 m/s) based on the W99 climatological mean snow density of 320 kg/m3 and the relation between snow density and dielectric constant given by Tiuri et al. . For a typical snow depth of 30 cm, an uncertainty in snow density of 100 (taken from W99) leads to a maximum snow depth retrieval error of 2 cm due to errors in the propagation velocity of light. The algorithm used for the retrieval of snow depth in this study follows from the technique described in detail by Farrell et al. . The conditions for the minimum return power level, Ps−a (in dB), for the snow-air interface are:
where is the mean noise level, σN, is the standard deviation of the noise level, xthresh = 2.3, and is the mean power in the 6 range bins that follow the estimated snow-air interface. These conditions were chosen to separate the beginning of the return (at the snow surface) from the noise level. For sea ice, the surface roughness is typically larger than the Rayleigh criterion over the first two Fresnel zones of the radar active area [Carsey, 1992]. Therefore, diffuse rather than specular reflections generally dominate the returns, but the manner in which the radar signal is backscattered from the snow-air interface is quite variable over the survey lines due to the snow property variations. To account for these variations, the snow-air interface was chosen to be the first local maxima in the radar signal beyond the minimum power level. But if the return power reached a point greater than Ps−a ≥ + xmaxσN (where xmax = 2.8) above the noise level before a local maxima was found then this point was set to be an initial estimate for the snow-air interface location. The threshold conditions restrict the return power from the snow-air interface such that it is less than or equal to the return power we expect from a more specularly reflecting snow surface, but still above the noise level. After determination of the estimated snow-air interface location, an initial estimate for the snow-ice interface was then taken to be the location of the largest peak in the radar return. Due to the larger dielectric contrast between ice and snow, the signal from the snow-ice interface is stronger and easier to identify than that from the snow-air interface. The final locations of the snow-air and snow-ice interfaces were then chosen using a second iteration of the previous method, but with adjustments to the values of xthresh and xmax to account for variations in the radar operating parameters which occurred during each flight (e.g. aircraft altitude and radar transmit power). The average power of the noise level, 〈N〉, and initial estimate snow-ice interface power, 〈Ps−i〉, were determined for ∼4 km segments, corresponding to the length of each radar echogram. The values of xthresh and xmax were then multiplied by the scale factor (〈Ps−i〉 − 〈N〉)/13.0, where 13.0 is the average power separation between the noise level and the snow-ice interface as observed in the study of Farrell et al. . A locally weighted robust linear regression was then applied at a 40 m length scale to reduce the impact of outliers in the final determination of the snow-air and snow-ice interface locations.
 A comparison between in-situ snow depth measurements and those retrieved from the airborne snow radar over fast ice north of Greenland [Farrell et al., 2011] proved the capability of the retrieval algorithm for deriving snow depth over a variety of ice types and snow thicknesses. On undeformed first year (FY) ice the mean snow depth measured in-situ was 9.5 cm with a mode at 4.5 cm, while mean snow depth retrieved from the radar was 9.6 cm with a mode at 5.7 cm, demonstrating the ability of the algorithm to accurately retrieve of snow depth over level ice with a thin snow cover. Similarly, over multiyear (MY) ice the mean difference between the in-situ and radar measurements was also small and the overall mean difference for the full 2 km survey was 0.8 cm [Farrell et al., 2011]. A comparison between coincident IceBridge laser altimetry and aerial photography data revealed two surface types where the retrieval algorithm could not accurately define the air-snow interface: 1) New or partially refrozen leads were found to cause a characteristic series of multiple bands within the radar signal rendering it unusable for interface detection. These areas were flagged when three bands four standard deviations above the noise level were present. The snow depth in these regions was set to zero corresponding to the negligible snow cover on new leads. 2) Data over the apexes of steep pressure ridges caused a loss in the return power such that the returns from the snow-air and snow-ice interfaces became indistinguishable from the noise level. Following Farrell et al.  we discard data where the maximum return power is less than six standard deviations above the mean noise level. Overall, 84% of the radar data had a sufficient return energy for the retrieval of snow thickness and was used in this study. Even though snow depth data directly over steep pressure ridges could not be found, comparisons with the distribution and mean snow depth from the in-situ and radar survey study suggests that the deep snow cover typical of snow drifts on the leeward side of ridges is largely captured by the snow radar system.
3. Snow Depth Results
 In 2009, the IceBridge flight lines were dedicated to surveying a variety of ice types and capturing the gradient in ice thickness on basin scales. Flights on April 2, 21, and 25 were primarily over areas of thick MY Arctic sea ice in the Canada Basin, while the flight on April 5 surveyed a mixture of FY and MY ice areas. A map of snow depth derived along IceBridge flight lines is presented in Figure 1. The 16 m by 40 m resolution snow depth derived from the snow radar data have been placed into the 12.5 km AMSR-E grid for comparison with the scale of current snow depth and gridded altimetric (e.g., ICESat) sea ice thickness retrieval methods and for ease of separating the FY and MY ice areas. The distinction between FY and MY ice areas was identified using passive microwave data from AMSR-E [Cavalieri et al., 2004] (updated daily), and is outlined in dark red in Figure 1. Also shown in Figure 1 is the W99 climatological snow depth for the month of April. Compared to the smooth gradient of the climatological snow depths, the observational data indicate higher variability in snow depth on scales of 12.5 km and greater. This higher variability has been seen in other observational studies as well [Gerland and Haas, 2011]. Therefore the combination of climatological snow depth values with higher-resolution altimeter data sets such as ICESat or CryoSat-2 may result in local ice thickness errors due to snow depth inaccuracies. Overall, the IceBridge radar data show similar gradients in the snow cover with the lowest snow depth generally occurring towards the Alaskan coast and the highest snow depth towards the northern coasts of Greenland and the Canadian Archipelago. This gradient is consistent with contemporary snow precipitation and ice circulation patterns from model data [e.g., Kwok and Cunningham, 2008] and ice type/age differences.
 Flight 3 shows an anomalous decrease in snow depth towards the western portion of the flight track that was not observed in the overlapping data set from Flight 1. Possible reasons for these anomalous snow depth values include dynamic redistribution of the snow pack or sea ice during the 3 week period between flights, or thermodynamic processes. The AMSR-E sea ice mask delineating FY and MY ice types did not shift significantly during the period (not shown) suggesting the flight survey was entirely over MY ice. To further investigate thermodynamic causes we calculated surface temperatures for each flight line using the thermodynamic sea ice model of Kurtz et al.  forced with ECMWF meteorological data. The mean surface temperature reached −4 °C in the region where the anomalously low snow depths occurred. In contrast, surface temperatures for all previous days and flights were typically less than −10 °C. Evidence of variations in radar penetration depth due to surface temperature effects have been observed in previous studies [e.g., Giles and Hvidegaard, 2006; Willatt et al., 2011]. Barber et al.  found that surface radiative forcing led to phase changes in the snow surface which enabled free water to percolate through the snow pack for temperatures greater than −5 °C, this free water greatly changed the dielectric properties of the snow pack. We speculate that wet layers such as this changed the dielectric properties of the snow pack causing errors in the snow depths retrieved at the western end of Flight 3. We therefore discarded data where surface temperatures were greater than −5 °C (i.e. westward of ∼235 W longitude for Flight 3) in the subsequent analysis.
 The distributions of the individual snow radar measurements (at 16 m by 40 m resolution) are shown in Figure 2 and demonstrate variability in the small-scale snow depth. The three cross-basin surveys (Flights 1–3) showed a broad distribution of snow depths with standard deviations of 15–19 cm for both the FY and MY ice regions (Figure 2). As expected, modal snow depth over FY ice was less than the modal snow depth over MY ice. Figure 1 shows a clear transition between the FY and MY snow depth covers for Flight 3 in particular. Overall, snow depth on sea ice in the central Arctic was observed to be highly variable, ranging from near zero (over refrozen leads) to 60 cm over thick MY ice.
 The snow depth climatology of W99 was developed using observations collected on MY ice only and is likely not applicable over FY ice areas since the FY ice areas may form after the heavy snowfall periods which typically occur in September and October. A comparison between the 12.5 km gridded FY and MY ice snow depths with W99 is shown in Table 1 for each flight line. Two values are shown: 1) the mean snow depth calculated for all data points i.e the relevant value for combination with satellite laser altimetry data, and 2) the mean snow depth excluding leads i.e. the value which is relevant for comparison with the W99 data set and for combination with satellite radar altimetry data. W99 estimate the upper bound of the interannual variability of snow depth for April to be 6.1 cm providing an estimate on the expected error in the climatological snow depths shown here. Using the mean of all available 12.5 km grid cells shown in Table 1, we find that the climatological snow depths of W99 are only 0.3 cm higher than those from the IceBridge radar data set implying consistency between the large-scale climatological mean snow depths and contemporary observations. However, on a flight-by-flight basis for the MY ice areas some of the results compare quite well while others do not. Flight 1 covered a large area of the MY ice cover with only a 0.7 cm difference between W99 and the observations (Table 1). In contrast, Flight 2 covered both FY and MY ice but showed a much more significant difference of 8.9 cm in the MY snow depth. Nonetheless, the results over MY ice show that over a very large scale (several hundred to thousands of km) the snow depth climatology continues to provide an accurate estimate of the average snow depth.
Table 1. Snow Depth Comparisons for the 12.5 km Gridded Dataa
 For FY ice areas, the climatological snow depth values compare poorly with the mean climatological value being 16.5 cm (191%) higher than the snow radar observations (Table 1). This is not unexpected, but illustrates that snow depths from climatologies such as W99 do not accurately represent snow depth on FY sea ice. While Flights 2 and 4 show large differences (>16 cm) between the climatology and snow radar, Flight 1 showed less significant differences of <7 cm. There is, however, little data in the FY ice area of Flight 1 (comprising 4% of the FY ice observations), but their proximity to the MY ice areas may mean that the ice formed relatively early in the growing season allowing the FY floes to accumulate more snow.
 Recently observed changes in the Arctic sea ice regime raise two questions which can be addressed with the current results, and future IceBridge surveys: 1) How has the snow cover changed in response to the changing MY sea ice cover in the Arctic? 2) Have changing precipitation patterns [e.g., Serreze et al., 2000] in recent years acted to change the spatial distribution and mean depth of the Arctic snow cover? The present study provides an answer to the first question, in that the observed April 2009 snow depths for FY ice areas were ∼52% of the 1954–1991 climatological snow depths. Snow accumulates primarily in the early part of the fall, so as the ice cover shifts to a more seasonal ice pack with later freeze up dates [Markus et al., 2009] less snow now collects on Arctic sea ice, precipitation which may have previously fallen onto the sea ice is now lost to the ocean. However, determining the magnitude of the precipitation loss and additional fresh water input to the ocean requires more surveys over a larger expanse of FY ice. This study also provides a baseline data set for answering the second question, but subsequent IceBridge surveys over the next several years will be needed to answer this question with more certainty. W99 found a trend of decreasing mean April snow depth of 0.1 cm/yr, representing a ∼4 cm loss in mean snow depth between the middle of the climatology period (1972) and the present time. We observe only a 0.3 cm difference between our surveyed snow depths and those from the climatology indicating that the mean precipitation (over the fall, winter, and spring time periods) and mean snow depth over MY ice may be similar to the 1954–1991 period. However, the maximum April interannual variability in snow depth was estimated to be 6.1 cm by W99, so more data over a longer time period is required to determine the long-term trend in the Arctic MY ice snow cover.
 The IceBridge surveys over Arctic sea ice have and will continue to provide much useful new information on the distribution of snow on sea ice. The snow depth surveys presented in this study reveal a snow cover that is more variable on regional scales than the W99 climatology, although on synoptic scales we observe a gradient similar to the W99 climatology. Thus, we find that the climatological snow depth is therefore useful for altimetric sea ice thickness retrievals on synoptic scales over MY ice, but may introduce errors when used for regional-scale studies. Additionally, due to the recent loss of MY ice across large areas of the Arctic basin the sea ice cover is becoming increasingly seasonal. Therefore, it may no longer be valid to use the W99 climatology on basin scales, where we have shown large differences with the observed snow depth over FY ice. Overall, an analysis of future IceBridge surveys will enable us to estimate the spatial and interannual variabilities of the April snow pack, determine trends during the observation period, and compare results to the historical climatology. The comparison of snow radar data from more recent IceBridge campaigns, including those conducted in 2010 and 2011, and the analysis of interannnual variability of the Arctic snow pack will be the subject of a follow-on study.
 This work was supported by the NASA Cryospheric Sciences Program under grant NNX10AV07G and the NOAA/STAR Ocean Remote Sensing Program. The authors would like to thank the editor and reviewers for their helpful suggestions on improving the manuscript.
 The Editor thanks Jan Lieser and an anonymous reviewer for their assistance in evaluating this paper.