Integrating snow science and wildlife ecology in Arctic-boreal North America

The ABR is a vast and geographically diverse region with sparsely distributed in situ observations of snow and other weather-related station observations; the spatial distribution of these measurements are generally inadequate to accurately represent environmental gradients that influence snow such as altitude, latitude, and the distance from the coast. During the snow cover season, the temporal resolution of in situ snow measurements in ABRs are typically hourly at automated meteorology and snow measurement stations and monthly at manual snow course sites. These measurements typically lack the spatial coverage required to quantify variability in snowscape properties across the range of spatial scales relevant to wildlife behavior. While interpolated data from reanalysis or physical and statistical models are widely used, the utility of such data in wildlife studies remains limited by the sparse in situ observations available for calibration and validation.

Relative to the snow cover of the western conterminous United States, which has <50% snow covered area for less than 6 months per year, Alaska has >95% snow covered for 7–10 months of per year (Liston and Hiemstra 2011, Robinson et al 2018). When compared to the Western states of the conterminous United States, Alaska has <1/8 the density of in situ snow-cover observations. Figure 1 highlights this, showing the number and density of active Natural Resource Conservation Service (NRCS) snow measurement sites in the combined area of Washington, Oregon, California, Arizona, Nevada, Utah, Idaho, Montana, Wyoming, and New Mexico (mean February snow covered area = 1.31 × 106 km²) compared with Alaska (mean February snow covered area = 1.72 × 106 km²) (wcc.nrcs.usda.gov/snow/). The limited periods of record and sparse spatial distribution of in situ data in Alaska further constrain calibration and validation of both remote-sensing observations and spatial models of snowscape properties (Bokhorst et al 2016). While many of the snow observing sites in the ABR measure SWE, snow depth, air temperature, and total precipitation, other variables critical to understanding wildlife behavior are rarely measured, including snow hardness and ice-layers, and local scale spatial heterogeneity in these and the other snow properties.

Zoom In Zoom Out Reset image size

In addition to readily accessible snow data, other public and private stakeholders (e.g. United States Department of Agriculture Forest Service, United States National Park Service, petroleum and mining companies) collect limited in situ snow observations in Alaska. These data are also sparsely distributed and may be reported using different formats or, in the case of private industry, unavailable to the public, making them more difficult to discover, access, and utilize. In recognition of the need for coordinated management and archiving of ABR data, some notable efforts have emerged to locate and consolidate these data such as the NSF Arctic Data Center (arcticdata.io/) and NOAA's Sustaining Arctic Observing networks (arcticobserving.org/).

The use of satellite data for global mapping of snow-cover extends over the modern satellite era. However, it was not until the launch of the daily-orbiting Moderate Resolution Imaging Spectroradiometer (MODIS) sensors on the NASA EOS Terra (1999) and Aqua (2002) satellites, and more recently the reanalysis of satellite passive microwave sensor records, that daily snow products describing the extent, albedo, grain size, and relative ice and water content have become available (Tedesco et al 2013). The spatially continuous nature of satellite information complements in situ observations that are spatially limited. Unfortunately, effective remote-sensing techniques to map and monitor snow properties at the spatial and temporal resolutions required to understand wildlife behavior are still lacking (Nolin, 2010, Dietz et al 2012, Bokhorst et al 2016). As shown in section 4, the spatial resolution of snow-related information required to characterize wildlife behavior can range from meters (e.g. polar bear (Ursus maritimus) denning habitat) to kilometers (e.g. caribou migration timing). Similarly, regular daily observations that capture transient snowscape conditions and anomalous events are also required (e.g. snow-free date and water or ice content of the snowpack). Another major challenge for remotely derived snow assessments is that wildlife respond to a broad range of ecosystem-specific snowscape properties that are not available from current remote-sensing observations and data products.

Existing satellite-derived snow products meet some snow-wildlife study requirements (table S1), but no single sensor or data record provides consistent and reliable observations of all of the relevant snow variables at the required spatial and temporal scales. Most available satellite records are limited by one or more factors, including the trade-off between spatial resolution and temporal frequency; where more frequent sensor scans (e.g. Passive Microwave and MODIS) have coarser footprints when compared to those scanning smaller areas that can resolve information at finer spatial resolution (e.g. Landsat) but are only available at much coarser temporal resolution (e.g. 16 d). Each sensor type also has limitations causing other data gaps and uncertainties. Optical and infrared sensors are limited by low solar illumination, polar night, and persistent cloud cover. The signal strength of microwave energy emitted from Earth's surface, complex terrain, vegetation cover, and snow-cover properties (e.g. stratigraphy, grain size) likewise complicate data retrieval by microwave sensors (Nolin 2010, Dietz et al 2012).

Airborne or satellite image retrievals are also used to create high resolution digital elevation or surface elevation models (DEM and SEM, respectively) that are in turn input into physical snowpack evolution models to study snow-wildlife interactions (see Model simulations of snowscape properties). Unfortunately, much of the digital elevation and surface mapping of the ABR is of relatively low spatial resolution, with two notable exceptions: the Alaska Interferometric Synthetic Aperture Radar DEM (5 m resolution), and the SEM called the ArcticDEM, created using stereo pairs of high resolution optical imagery (2 m resolution) (LPDAAC, 2017, Polar Spatial Center, 2017). Higher resolution airborne LiDAR and structure from motion products are also available in a few locations (as described in case study 1) and provide the opportunity to conduct wildlife relevant snowscape research by mapping snow depth through differencing (Deems et al 2013, Nolan et al 2015).

Temporally varying snowpack properties can be extrapolated over landscapes using process-based models that implement snow physics within the context of a numerical modeling system (e.g. Durand et al 1999, Bartelt and Lehning 2002, Lehning et al 2006, Liston and Elder 2006a). These models input time-invariant variables such as topography and land-cover, and time-evolving meteorological forcing such as air temperature, humidity, precipitation, wind, and radiation, and simulate the evolution of variables such as SWE, snow albedo, and snow density. Existing snow-evolution models faithfully represent key snow-related processes; including snow–water-equivalent accumulation of snowfall, redistribution by wind, snow-forest-canopy interactions, snow-vegetation interactions, and snowmelt. In addition, these snow modeling tools appropriately handle the spatial-distribution and temporal-evolution aspects of snowpack growth and decay. Often these models have been developed for a specific application (e.g. climate studies, water resource management, avalanche risk assessment), and the details of the model configuration (e.g. grid increment, time step, and represented physical processes) reflect that focus. To date, because the focus has rarely been on wildlife applications, model configurations have rarely met the needed criteria to produce snow products that are directly applicable to understanding wildlife-snow interactions (Liston et al 2007).

Modeling approaches for characterizing snowscape properties can be developed to produce a large suite of wildlife-relevant snow variables over large spatial and temporal extents, and over a wide range of spatial and temporal resolutions. In this way, modeled snowscape products can be tailored to fit the specific needs of wildlife studies (e.g. Vikhamar-Schuler et al 2013, Eaton and Businger 2014, Northrup et al 2016, Ouellet et al 2016, Rasmus et al 2016, Lendrum et al 2017, Reinking et al 2017). While current snow-evolution models can make valuable contributions to snow-wildlife interaction studies, the majority are not yet up to the task because they do not simulate the required snow variables that are often most critical to wildlife dynamics studies, although there are some notable exceptions (see table S1).

Other weaknesses associated with snow-evolution models limit their ability to simulate or estimate snowscape properties, which in turn limits their efficacy in studies of wildlife-snow interactions. For example, the models are only as good as the meteorological forcing inputs; these data can suffer from deficiencies in data quality or resolution. Currently the most problematic meteorological input is the precipitation forcing; the available data on water-equivalent precipitation (i.e. total precipitation of various forms falling from the sky and reaching the ground) are often of insufficient quality to exactly reproduce observed snow distributions (Goodison et al 1981, Yang et al 1998, Liston and Sturm 2002, Pan et al 2003, Liston et al 2008, 2016, Girotto et al 2014, Margulis et al 2015, Wrzesien et al 2017). Often times there are no validation data to compare the model outputs to, and available topography and vegetation datasets may be inadequate. In addition, it has been challenging to develop snow models that are general enough to consistently and accurately simulate snow properties throughout the entire ABR for the entire scope of processes found within the domain. This is because of the wide range of processes found in warm-wet, cold-dry, tundra, and forested environments. The existing models almost always produce qualitatively reasonable snow-property distributions when driven with realistic forcing data. However, their outputs may include biases when compared to in situ observations of snow properties made at a given location and time.

Pregnant polar bears typically den in snowdrifts found on sea ice or land (Amstrup 2003). As sea ice has become progressively thinner and unstable (Comiso et al 2008, Kwok and Rothrock 2009, Maslanik et al 2011, Polyakov et al 2012), the proportion of land-based dens has increased from 42% (in the 1980s) (Amstrup and Gardner 1994) to 63% (in the early 2000s) (Fischbach et al 2007). Coupled with concurrent increases in human activity in the coastal regions of Alaska's Beaufort Sea, the potential for human-bear interactions has heightened. To minimize interactions, as well as disturbance to denning bears, it is critical to identify active dens (MacGillivray et al 2003, Stirling and Derocher 2012). As such, Liston et al (2016) developed, applied, and tested a physically based numerical model (SnowDens-3D) to map land and barrier-island polar bear snowdrift den habitat along the Beaufort Sea coast.

Liston et al (2016) found that snowdrift dens are not randomly distributed, but are the direct result of wind blowing snow into topographic snowdrift traps. They used SnowDens-3D to simulate the year-specific physical interactions of concurrent snow, wind, and topographic dynamics. The model was run annually from 1995 through 2012, on a 2.5 m × 2.5 m horizontal grid; a grid sufficient to resolve the snowdrifts that polar bears den in. The minimum snow-depth required for a viable den was defined and applied to the SnowDens-3D simulated snow depths. This yielded annual maps of potential polar bear den habitat locations, which change every year depending on wind speed and direction, and snowfall amounts. They found that 97% of observed maternal den locations were correctly identified in the simulated snowdrifts modeled by SnowDens-3D (figure 2). The model's high level of accuracy is due to its use of high resolution (2.5 m) topographic data and its ability to account for the influences of year-specific snowfall and wind speed and direction. Given that the existing suite of widely available spatial snow products does not include snow depth data with comparable accuracy nor spatial resolution (see table S1), this case study illustrates the clear need for a highly specialized snow product for identifying polar bear denning habitat, and provides an example of a wildlife-relevant snow-related variable that is not available by field or remote sensing methods alone. Because of SnowDens-3D's basis in snow and weather physics, it can be incorporated into climate scenario simulations to explore how shifts in early-winter meteorological conditions are likely to impact the formation and availability of polar bear snowdrift den habitat.

Zoom In Zoom Out Reset image size

The Dall sheep (Ovis dalli dalli) is an iconic northern wildlife species endemic to the mountain ranges of Alaska and northwestern Canada. Range-wide, Dall sheep populations have declined by at least 20% during the past decade, with some areas declining up to 70% and causing emergency harvest closures (Koizumi et al 2011, Alaska Department of Fish and Game 2014). Dall sheep may be particularly sensitive to changing snow conditions, because sheep occur at both high elevations and high latitudes, and average temperature increases and the associated changes are likely amplified due to reductions in the cold regions' temperature inversions (Callaghan et al 2011).

Dall sheep exhibit seasonal movements that may be tied to seasonal changes in snow conditions, including use of wind-exposed patches of forage along ridgelines during peak snow coverage and short-distance elevational migrations (Simmons 1982, Nichols and Bunnell 1999). Mahoney et al (2018) evaluated the efficacy of MODIS snow-cover fraction and SnowModel (Liston and Elder 2006a) snow depth and density products in predicting Dall sheep movements at multiple spatial and temporal scales, finding that adding a snow covariate, regardless of type, substantially improved model predictions of sheep movements (figure 3(a)). However, the relative performances of snow products were scale-dependent. At higher resolutions (25 and 500 m), SnowModel products that included blowing-snow redistribution, snow-density evolution, orographic precipitation increases, and slope-aspect snowmelt relationships, outperformed MODIS at predicting the finest movement scales (i.e. the distances traveled over 7–56 h).

Zoom In Zoom Out Reset image size

However, the 500 m MODIS fractional snow cover outperformed SnowModel's grid-averaged snow products at coarser resolutions and movement scales (>112 h), likely because the coarser resolution SnowModel outputs did not include the sub-grid snow information reflected in the MODIS data (figure 3(a)). At fine movement scales (i.e. 7 h), Dall sheep generally selected low density, shallow snow (figures 3(b), (c)), likely to facilitate access to forage and reduce energy expenditure (Dailey and Hobbs 1989). However, they selected for higher snow density at or above the mean threshold of support (329 kg m⁻³, SE = 18; Sivy et al in press) when traveling through deep snow above mean chest height (54 cm; figures 3(b), (c)), likely to reduce snow hoof penetration and improve efficiency of movement (Parker et al 1984). At coarse scales, Dall sheep selected for areas with lower fractional snow cover (figure 3(d)), indicating sheep may be detrimentally affected if regional trends of increased winter precipitation continue (Olsen et al 2011, Winski et al 2017).

These scale-dependent differences in snow product performance indicate that use of a snowpack evolution model such as SnowModel may be necessary to obtain critical insights regarding fine-scale responses of animals to changing snow properties, whereas MODIS fractional snow cover data may be able to provide insight into the effects of snow on movements at broader scales. Although MODIS products are readily available and relatively easy to use, incorporating snow cover in this case required removing a substantial proportion of animal location data due to cloud cover inhibiting image acquisition (82.1% of used locations removed), which reduced the statistical power and the ability to characterize complex movements in response to fine-scale spatial processes.

During their spring migrations to calving grounds, barren-ground caribou (Rangifer tarandus spp.) travel up to one thousand kilometers within only weeks. There is considerable variability in the timing of the spring migrations across years, but relatively few large-scale studies have explored the environmental drivers of these variations (though see LeCorre 2017). The amount of spring snow is even more variable across years, but has shown a steady negative trend in high northern latitudes for several decades (Derksen and Brown 2012). In temperate zones, snowmelt precipitates spring green-up, which has been shown to be an important driver of ungulate migrations (Aikens et al 2017). It therefore seems reasonable to hypothesize that the timing of snowmelt might drive spring migration phenology of caribou, as well.

To explore these relationships, we analyzed caribou movement data collected during 15 consecutive spring migrations (2000–2014) from the Bathurst herd in central Canada that has historically been one of the largest migratory herds in the world. To quantify the timing of snowmelt, we computed snow disappearance day (SDD) for each 500 m MODIS pixel, averaging neighbors in space and time to fill gaps. Caribou generally started migrating more than three weeks prior to snowmelt, with high variability among years (mean DAS = 21.9 d, standard deviation 17.5), and arrived at the calving grounds one week before snowmelt (mean 7.3 d, s.d. 12).

Spring migration for collared caribou in 2011 overlaid on MODIS snow-cover data (MODIS/Terra Snow Cover Daily L3 Global 500 m Grid, Version 6—table S1, item 19) is presented in figure 4. The 2011 migration began well before snowmelt and continued until well ahead of the melting snow front, which was the case in all years of the migration. We investigated whether the inter-annual variation in migration timing was explained by timing of snowmelt, as indexed by the median SDD in the entirety of the wintering and calving ranges, respectively. These exploratory analyses indicate that caribou do not migrate in a consistent way with respect to MODIS-derived snowmelt timing. While caribou tended to move ahead of the snowmelt front, they were as likely to move anywhere from one day to three weeks ahead of the retreating snow front. Our findings suggest that if snow-related drivers are important to the timing of migration, snow properties other than cover—the most readily accessible of remotely sensed snow data—are more important. For example, local snow quality, including snow variables such as depth, hardness, and finer-scaled spatial patterns of snow cover and timing of ice cover melt on the tundra's many waterbodies, may be more important. To more fully explain the spring phenology of caribou migration a combination of more sophisticated snow products and more ecologically driven combinations of climatic variables (e.g. temperature and precipitation to assess snow quality) need to be taken into account.

Zoom In Zoom Out Reset image size

Given the sparseness of existing in situ snow and snow-related observations (see figure 1), improvement of the existing ground-based observing network should focus primarily on improving the spatial coverage and representativeness of the measurements. When selecting where to locate new stations, underrepresented landscape types and areas identified by snow models as particularly difficult to characterize should be prioritized. Areas that are most underrepresented include locations that are farthest from existing road networks and higher elevations. Unfortunately, the current trend in the total number of snow monitoring stations is flat and agencies operating them have tightening budget constraints. However, there may be opportunities to leverage existing infrastructure to meet the need for in situ data through greater collaboration across management agencies. For example, Alaska Transportable Array infrastructure (usarray.org/alaska#after-2018), located in underrepresented areas, could be identified, augmented with sensors, and operation supported to collect snowscape relevant data. Further, it is critical that augmented networks are equipped with sensors that measure the most critical variables required to make characterize snowscape properties: total precipitation, snow depth, snow water equivalence, and snow surface temperature.

Finally, we suggest a complimentary, ground-based observational approach to collecting information on snowscape properties—deployment of animal-borne sensors. The state of the art in movement ecology includes several miniaturized data logging sensors, including video, audio, biological, and atmospheric data (Kays et al 2015). When combined with GPS units, these sensors allow scientists to spatially and temporally reference a suite of co-registered data. For example, cameras have been used to record short videos on pre-programmed schedules, yielding qualitative data on ground conditions (Thompson et al 2012), with potential for remote documentation of a suite of vegetation and snowpack characteristics. In an analogous application, oceanographic sensors have been deployed on narwhals (Monodon monoceros) to make deep-water observations in Arctic waters that were otherwise unobtainable (Laidre and Heide-Jørgensen 2007). Animal-borne cameras could not only serve to augment traditional ground measurement networks, but also contribute to validation of remotely sensed snow products. Further, biotelemetry sensors that provide data on activity (e.g. accelerometers) and physiology (e.g. heart-rate or body temperature data loggers), can inform empirical relationships between snowscape properties and wildlife movement, activity, and physiological metrics. These empirical relationships can then, in turn, enable those fine-scale metrics to be used as indirect indicators of localized snow conditions throughout remote ABRs.

Developing more effective snow retrievals for wildlife studies will likely involve the use of different sensor combinations, including passive optical-IR, lidar, gamma radiation, and active and passive microwave remote sensing. Promising approaches include the fusion of satellite optical-IR and multi-frequency active and passive microwave sensor observations for delineating landscape freeze-thaw heterogeneity and wet snow conditions (Podest et al 2014, Kim et al 2015). In addition, the use of multi-frequency satellite radar data for landscape level SWE retrievals (Rott et al 2010) shows promise and, when combined with a measure of snow density, could be used to calculate the more wildlife-relevant snow depth. These products will offer the greatest benefit when combined with emerging high resolution land surface and digital elevation models that will provide better scaling and modeling opportunities (e.g. Noh and Howat 2015).

New developments in airborne remote sensing offer the potential for significant advances in the remote sensing of snowscapes using coordinated observations from different sensors to develop and test new algorithms for more effective snow retrievals (Deems et al 2013, Witz, 2016). For example, airborne observations of snow depth can be obtained through differencing snow-on and snow-off lidar observations (Bolton et al 2013, Kirchner et al 2014, Painter et al 2016, Zheng et al 2016), or less expensive structure from motion DSMs (Hopkinson et al 2004, Trujillo et al 2007, Mott et al 2011, Deems et al 2013, Nolan et al 2015). These methods are especially viable for small to medium scale (10–100 km²), un-forested areas where a few seasonal data collections (e.g. peak snow depth and bare earth) are sufficient (Nolan et al 2015, Harder et al 2016) Larger areas with tree cover, rugged topography, and the need for multiple observations quickly add to mission complexity and cost. Airborne activities such as these offer a path forward for developing next generation satellite-based methods to monitor snowscapes at wildlife-relevant spatial and temporal scales using data such as that expected from the Advanced Topographic Laser Altimeter System to be launched aboard the NASA ICESat-2 platform (see Treichler and Kaab 2017). Because snowscape properties in ABRs are inherently very different than those in other snow landscapes (Sturm et al 1995), an ABR-specific airborne snow campaign, conducted concurrent with ground observations—similar to NASA's SnowEx campaign focused on mapping SWE in Colorado's forested ecosystems (https://snow.nasa.gov/snowex)—will be essential to developing effective snow retrievals for the region.

To satisfy the needs of wildlife studies, existing snow-evolution models that were initially designed for climate, hydrologic, and/or avalanche applications should be modified and enhanced to include new wildlife-relevant snow variables such as snow depth, snow hardness, and ice-layer thickness, and vertical distribution within the snowpack. These models must also have the ability to be run at a wide range of spatial resolutions, depending on the wildlife application of interest; for example, as shown in our case studies, simulations of polar bear den habitat likely require much finer resolution than studies of caribou migration timing and routes.

While it is true that each meteorological input, each ground observation, each remote sensing instrument, and each snow modeling tool includes some valuable information on snow-property distribution and evolution, each of these contributions also comes with their own collection of weaknesses. We contend that a combined approach, using atmospheric forcing from meteorological stations (re)analyses, and/or climate scenario datasets; ground-based snow measurements; remote-sensing datasets; and modeling tools is required to answer the breadth of snow science and management questions being asked by the wildlife community. In such a synthesis, each meteorological dataset, ground measurement, remote sensing observation, and model result, contributes toward unveiling the 'snow puzzle'. We call this 'data-model fusion', as opposed to 'data assimilation', because it embodies a synergetic extracting and merging of information obtained from different sources, in an effort to produce an outcome that is better than the outcome obtained by using the different sources independently. By merging data from meteorological, ground in situ, remotely sensed, and/or modeled data sources, the data-model fusion approach—if carefully applied—dynamically offsets weakness in any individual component dataset at any given point in space and time, while including the strength of other component datasets. Under conditions where, or when, a given data source is known to have certain limitations or has missing data, another data source can fill in. For example, given that snow-evolution models suffer often from precipitation inputs that are inadequate to reproduce observed snow distributions (see section 3), integration of snow-on-the-ground observations using data-model fusion can correct those precipitation deficiencies (e.g. Liston and Sturm 2002, Liston et al 2008, Pedersen et al 2017). As another example, snow-depth datasets are often available instead of the more hydrologically important SWE data (Sturm et al 2010), this information can be used to add information to data-model fusion efforts. Collectively, the data-model fusion system should ultimately return a well-informed and valued product.

Data-model fusion is a multi-step process, where at each step there are interactions and feedbacks between (1) the observations, (2) the model configuration and representations, and (3) the modeling team (Williams et al 2009, Keenan et al 2011). Inherent in this process is the requirement that the model developers have intimate knowledge of the observational datasets, how they were collected, what features and processes they represent, what they contribute, and their limitations. The model developers must also have intimate knowledge of the modeling systems and how those systems relate to, and merge with, the various data inputs. This knowledge is particularly important to reduce the ultimate amount of error and uncertainty during product development.

The data-model fusion system that we are advocating includes three basic and practical requirements, which emerge from limitations in current (and likely future) snow-related datasets. All of these requirements can be met with modifications to existing snow-evolution modeling tools, and with inputs from ground and remote-sensing observations. The system should:

• 1.  
  Add physical properties to basic snow information and datasets (e.g. convert SWE data from a remote sensing instrument to snow depth, by adding model-simulated snow density data).
• 2.  
  Using our physical understanding and model representations of snow-evolution processes, fill in missing data between snow measurements sparsely distributed in space and/or time.
• 3.  
  Account for a wide range of snow-related observational datasets provided by both ground observations (e.g. SWE, snow depth, snow surface temperature), and remote sensing (SWE, snow presence or absence, fractional snow-covered area, grain size, albedo, snow surface temperature, snow wetness, and icing).

Figure 5 provides an example of the data-model fusion approach we used to simulate snow depth in the Arctic National Wildlife Refuge (ANWR) for 1 April 2015. We merged snow-on-the-ground observations (figure 5(a)) with the MicroMet (Liston and Elder 2006b), SnowModel (Liston and Elder 2006a), and SnowAssim (Liston and Hiemstra 2008) snow-evolution modeling tools. The models were run over the 400 km × 400 km domain using a 500 m grid increment and 3 h time step (figure 5(b)). MicroMet ingested atmospheric reanalysis data, SnowAssim assimilated end-of-winter (pre-melt) SWE observations, and SnowModel performed the spatial and temporal interpolations and evolutions and added the higher-resolution physics to the distributions, including conversion from SWE to snow depth (figure 5(b)).

Zoom In Zoom Out Reset image size

Figure 5(a) displays the spatially interpolated observations (no physically based adjustments have been applied). Figure 5(b) shows the value added when the physics in the models are used to distribute and convert the SWE to snow depth. The data-model fusion snow-depth distribution (figure 5(b)) includes features such as (1) thinner snow along the lowlands in the south of the domain and along the Arctic coast (the purple areas), (2) less snow in the bottoms of the drainage basins, (3) deeper snow in the higher elevations, (4) precipitation-shadow effects south of the Brooks Range, and (5) vegetation-related snow distributions (e.g. forests versus tundra). All of these important snow-distribution features are largely absent in the observations (figure 5(a)).

The data-model fusion example shown in figure 5(b) produced spatially- and temporally-continuous snow-property distributions that match our physical understanding of snow processes (e.g. snowfall/precipitation distributions, conservation of mass and energy, layer development within the snowpack, snow-canopy interception progression, blowing snow redistribution, snow-vegetation interactions, melt rates and timing, snow-covered-area evolution, etc) and the available field observations. Only one time (day; 1 April 2015) is shown, but the model simulated the snow-depth distribution every 3 h from 1 September 2014 through 31 August 2015. In addition, while this simulation example only assimilated snow-on-the-ground observations, remote sensing observations could have also been included in the data-model fusion.

This data-model fusion approach promises to be broadly applicable to a wide range of ecosystem studies in the ABR, but it is a particularly attractive option for studying wildlife-snow interactions because we expect it to deliver on all of our most pressing needs via one, consistent and effective approach. In fact, we believe there is no other option. There is no evidence to suggest we have the financial resources or ability to modify existing snow monitoring technologies (i.e. ground-based or remote-sensing systems) such that comprehensive and complete mosaics of wildlife-relevant snow variables are available for the purposes of wildlife science and management research. This forces us to turn to models. But it is also clear that these models can benefit from the observations we are making. The logical next step, and the solution to our data-deficiency problem, is to merge the data with the models and create something more than exists without this coupling. The data-model fusion approach can produce spatially detailed and temporally continuous datasets of multiple snowscape properties and variables relevant to wildlife over large areas. It can do this by merging (1) the detailed temporal coverage of snow and meteorological data from ground observations, (2) the vast spatial coverage offered by remotely sensed observations, and (3) the large suite of simulated snowscape variables from wildlife-relevant snow modeling tools.

Other ABR research efforts will also benefit from this data-model fusion approach. The spatial snow products that are critical to understanding wildlife behavior and fitness are also essential to understanding the influence of snow dynamics on people in the ABR. Humans are affected by the same suite of snowscape properties as wildlife. For example, dynamics in snowscape properties influence travel by residents of rural subsistence communities, and thus impact the ease and safety of access to provisioning ecosystem services such as fish, wild game, drinking water, hydro power, and firewood (Callaghan et al 2011, Brinkman et al 2016). As other examples, changes in snowscape conditions in wildlife habitat may alter availability of wild game to subsistence communities (Tyler 2010, Langlois et al 2017), or limit skiing opportunities and snowmobile operation, thus affecting winter tourism (Abegg et al 2007). At the same time, reported experiences and qualitative observations made by people working and living in ABRs can be formally included in our understanding of human-snow interactions. Currently, however, there is a dearth of information on the topic because the spatial and temporal resolution of data on both humans and snow are insufficient to estimate causal relationships. The data-model fusion approach we advocate herein is a viable mechanism to correct this data void.