Accuracy assessment of late winter snow depth mapping for tundra environments using Structure-from-Motion photogrammetry
Publication: Arctic Science
17 November 2020
Arctic tundra environments are characterized by a spatially heterogeneous end-of-winter snow depth resulting from wind transport and deposition. Traditional methods for measuring snow depth do not accurately capture such heterogeneity at catchment scales. In this study we address the use of high-resolution, spatially distributed, snow depth data for Arctic environments through the application of unmanned aerial systems (UASs). We apply Structure-from-Motion photogrammetry to images collected using a fixed-wing UAS to produce a 1 m resolution snow depth product across seven areas of interest (AOIs) within the Trail Valley Creek Research Watershed, Northwest Territories, Canada. We evaluated these snow depth products with in situ measurements of both the snow surface elevation (n = 8434) and snow depth (n = 7191). When all AOIs were averaged, the RMSE of the snow surface elevation models was 0.16 m (<0.01 m bias), similar to the snow depth product (UASSD) RMSE of 0.15 m (+0.04 m bias). The distribution of snow depth between in situ measurements and UASSD was similar along the transects where in situ snow depth was collected, although similarity varies by AOI. Finally, we provide a discussion of factors that may influence the accuracy of the snow depth products including vegetation, environmental conditions, and study design.
Les milieux de la toundra arctique sont caractérisés par une profondeur de neige hétérogène dans l’espace en fin d’hiver consécutive au transport par le vent et du dépôt. Les méthodes traditionnelles de mesure de profondeur de neige ne saisissent pas avec précision une telle hétérogénéité aux échelles des bassins hydrographiques. Dans cette étude, nous nous penchons sur l’utilisation de données à haute résolution, réparties dans l’espace, de la profondeur de neige pour les milieux arctiques par l’application de systèmes aériens sans pilote (UAS). Nous appliquons la photogrammétrie structure à partir du mouvement aux images recueillies à l’aide d’un UAS à voilure fixe pour réaliser un produit d’une résolution de 1 mètre sur sept zones d’intérêt (ZI) dans le « Trail Valley Creek Research Watershed », dans les Territoires du Nord-Ouest, Canada. Nous avons évalué ces produits avec des mesures in situ de l’élévation de la surface de neige (n = 8434) et de l’épaisseur de la neige (n = 7191). Lorsqu’une moyenne a été calculée à partir de toutes les ZI, l’écart moyen quadratique (EMQ) des modèles d’élévation de la surface de neige était de 0,16 m (biais de < 0,01 m), semblable à l’EMQ du produit de profondeur de neige (UASSD) de 0,15 m (biais de + 0,04 m). La répartition de l’épaisseur de neige entre les mesures in situ et l’UASSD était similaire le long des transects où l’épaisseur de neige in situ a été recueillie, bien que la similitude varie selon les ZI. Enfin, nous présentons une discussion sur les facteurs qui peuvent influer sur l’exactitude des produits d’épaisseur de neige, y compris la végétation, les conditions écologiques et la conception de l’étude. [Traduit par la Rédaction]
Arctic tundra environments are characterized by a spatially heterogenous end-of-winter snow distribution resulting from wind erosion, transport and deposition over the winter (Pomeroy et al. 1997). Large spatial variations in snow depth and density across the landscape are controlled primarily by topography and vegetation cover (Sturm and Benson 2004; Pohl and Marsh 2006; Derksen et al. 2009). Earlier studies measured snow distributions across Arctic tundra landscapes using in situ snow depth and density measurements (Stuefer et al. 2020) either within land cover types or along long transects. These measurements were then weighted by land cover area to estimate mean watershed snow water equivalent (Steppuhn and Dyck 1974). Although these methods are useful for testing non-distributed hydrologic models or evaluating snowfall undercatch (Pohl and Marsh 2006; Pan et al. 2016; DeBeer and Pomeroy 2017), they are not sufficient for capturing the spatial variability in snow depth, which is required when testing hyper-resolution hydrological models (Peters-Lidard et al. 2017). To address these requirements, there is a need to develop new methods to measure distributed snow depth at high spatial resolution.
Recent advances in unmanned aerial systems (UAS) and Structure-from-Motion (SfM) technology have made it possible to create metre- to centimetre-resolution snow depth products in low-vegetation prairie and alpine environments (Bühler et al. 2015, 2016, 2017; Vander Jagt et al. 2015; De Michele et al. 2016; Harder et al. 2016). However, most studies have been restricted to relatively small (<1 km2) unvegetated areas and utilize small validation data sets. For example, Vander Jagt et al. (2015) presented one of the first applications of UAS mapping of snow depths for an alpine environment using a quad-copter style UAS and achieved an estimated snow depth error of roughly 0.10 m. However, this study was conducted over a 0.007 km2 area, using only 20 in situ snow depth validation points. In a more recent study, Harder et al. (2016) used a real-time kinematic (RTK) global positioning system (GPS) fixed-wing UAS over sparsely vegetated prairie (0.65 km2) and alpine (0.32 km2) landscapes and found that the presence of vegetation reduces the accuracy of UAS-derived snow depths.
In this study, we use a fixed-wing UAS and SfM photogrammetry to map late winter snow depth at high resolution (1 m) across multiple large Arctic shrub-tundra areas, ranging from 0.75 to 2.35 km2. We address the shortcoming of previous studies by first testing the use of fixed-wing UAS SfM for mapping snow depths in both mid-winter and spring conditions. Second, we assess the accuracy of both the SfM snow surface elevation and the snow depth products using a large in situ data set (n = 15 625).
Snow depth mapping was performed at the Trail Valley Creek research watershed (68.74°N, 133.49°W) located within the Inuvialuit Settlement Region east of the Mackenzie Delta (Fig. 1). The Trail Valley Creek region is characterized by short, cool summers and long, cold winters, with an average annual air temperature of −8.2 °C measured at the Inuvik airport 50 km to the south (Environment and Climate Change Canada 2020). Nearly 40% of the annual precipitation in the region falls between late August and October, with rain dominating the August–September months and snow in October, and with over half the annual precipitation falling as snow (Pan et al. 2016; Mann 2018). The watershed is located along the northern fringe of the taiga–tundra ecotone and is underlain by ice-rich continuous permafrost (Rampton 1987; Lantz and Kokelj 2008). Vegetation consists of erect low-shrub tundra (Walker et al. 2005), consisting of mosses, lichen, grasses, low-lying shrubs (0.2–0.70 m), and areas of patchy tall shrubs (1–2 m) (Wilcox et al. 2019).
The end-of-winter snow cover varies spatially, ranging from 0.1 to 4 m (as shown in Fig. 2), with a large volume of snow stored in large hillslope and channel drifts. These topographic drifts occur where a break in slope results in windblown deposits of snow accumulating over the winter months (Fig. 2b), and cover 8% of the watershed, mostly located along eastern- and northeastern-facing aspects with slopes greater than 9° (Marsh and Pomeroy 1996). However, previous studies have not been able to accurately map snow drifting for all slope angles, aspects, and vegetation types. Field observations (Pomeroy et al. 1997) found that drifts occur in the same locations annually, although drift size varies from year to year.
Seven areas of interest (AOIs) within Trail Valley Creek were delineated to sample a representative range of topography and vegetation in collaboration with airborne and terrestrial radar campaigns led by the UK Meteorological Office and Environment and Climate Change Canada during the 2018 field campaign at Trail Valley Creek (Rutter et al. 2019). For each AOI, high-resolution imagery was collected using a fixed-wing UAS. AOI 2 to AOI 7 were only sampled during March 2018, whereas AOI 1 was only sampled in April 2018. The UAS was a senseFly EBEE Plus carrying an integrated onboard senseFly S.O.D.A. 20-megapixel camera (senseFly 2017). Flights were programmed using the proprietary eMotion 3 software (senseFly 2019; version 3.2) with a flight elevation of 100 m above take-off altitude in a series of transects perpendicular to the predominant wind direction at the time of take-off. Images were collected using 70% horizontal (between flight lines) and 80% vertical (along flight line) overlap resulting in ground sampling distances ranging from 0.028 to 0.05 m. Due to the large areal footprint of each AOI, multiple flights were required to fully cover the desired study area. Flights were conducted back-to-back at each AOI to reduce potential changes to the atmospheric lighting conditions and changing snow surface due to blowing snow.
The imagery was then used to produce a high resolution digital orthorectified mosaic (orthomosaic) and digital surface model (DSM) for each AOI using Pix4Dmapper Pro SfM photogrammetry software (version 4.2.26). Despite a high albedo snow cover and limited surface features (vegetation and snow-free surfaces), enough surface features were present for the SfM photogrammetry software to distinguish common pixel tie points across the imagery and produced mosaic and DSM outputs. Tall vegetation protruding from the snowpack was removed in the photogrammetry software by manually editing the point-cloud products. After editing, a new DSM was created representing the snow surface elevation and each DSM was rescaled to 0.10 m ground sampling distance. The total areal extent of all AOIs included in this study is 9.6 km2.
Ground control points (GCPs) were distributed across each AOI on the snow surface before each UAS flight and georeferenced using a Leica Global Navigation Satellite System RTK GPS system (Leica Geosystems 2018) with a 3-dimensional accuracy of ± <0.02 m. GCPs increased the accuracy of the DSMs, specifically for the z axis (elevation), which is the most integral component for estimating snow depths. Without incorporation of manual GCPs, accuracy along the horizontal and vertical axes are 3 m, with 5 m elevation accuracy resulting from the standard onboard GPS receiver equipped on the UAS. A summary of the processed AOIs is included in Table 1.
During the March 2018 campaign, the average near-surface air temperature during data acquisition was −13 °C as reported at the nearby Trail Valley Creek weather station (Table 1), with half-hourly temperatures ranging from −5 °C to −20 °C. April 2018 saw similarly low temperatures with an average air surface temperature of −16 °C during data acquisition (Table 1). However, through strategic management of our UAS and the Li-Po batteries we were able to continue flight operations well below the recommended operating temperatures. This was primarily accomplished through continued heating of the UAS body and batteries using battery-powered heat packs and standard hand-warmers stored in an insulated travel cooler. This enabled us to travel to our AOIs via snowmobile, deploy and survey GCPs, fly multiple flights to cover the AOI, and continue to another site without returning to the research station base.
SfM snow depth mapping
Snow depth was calculated as the difference between two elevation data sets: a snow-surface DSM produced using the SfM photogrammetry and a snow-free bare-ground elevation model (eq. 1). Rather than differentiating between a UAS snow-covered and snow-free surface model (e.g., Harder et al. 2016), we used a bare-ground LiDAR digital elevation model (DEM) produced in 2008 for the Trail Valley Creek watershed. This product was created using an aircraft-mounted Optech ALTM 3100 laser scanner, with a surface elevation error of 0.25 m (Hopkinson et al. 2008). Care was taken to ensure the LiDAR DEM and the SfM DSMs were accurately tied to the same datum reported by Hopkinson et al. (2008) using the GCP inputs surveyed across each AOI. Ground surface elevations were collected for 33 of the 50 GCPs across all AOIs using the RTK GPS system revealing an average difference in bare-ground surface elevations of 0.04 m (standard deviation 0.11 m), with the LiDAR underestimating elevation on average relative to the in situ observations. This slight negative bias may be linked to removal of surface vegetation during the multi-wave LiDAR processing (Hopkinson et al. 2008) resulting in a bare-ground DEM that underestimates surface height due to removal of upper vegetation layers. However, it is expected that this is accounted for in the error estimate of 0.25 m reported by the authors. We did not feel it necessary to systematically adjust the LiDAR elevation model for this bias due to the spatial variations in the measured bias and low quantity of validation observations.
The bare-ground LiDAR DEM has a spatial resolution of 1 m, and therefore required the UAS product to be upscaled to 1 m resolution from the original DSM pixel resolution of 0.10 m. Resampling was completed using bilinear interpolation to match the bare-ground LiDAR DEM. First, a Low-Pass filter was applied using ArcMap GIS software (v10.5) to reduce noise caused by small protruding shrubs above the snowpack. Snow depth for each 1 m raster cell is then calculated as:
Where: hs is snow depth for each pixel cell, x is the date of the UAS flight, DSMsnow is the snow surface elevation from the UAS and DEMground is the elevation of the bare ground with no vegetation from LiDAR. SfM maximum snow depth was limited to fit the observed maximum depth based on field observations across the AOIs, which ranged from 2 to 4 m. Pixels featuring a negative snow depth, resulting from a mismatch between elevation models, were adjusted to 0 m snow depth.
In situ observations of snow surface elevation and snow depth across multiple transects were collected to validate the SfM snow depth products for each AOI. Snow surface elevations were collected along transects at 2 m intervals using an RTK GPS across each AOI (3-dimensional accuracy ± 0.018 m). As the operator held the rover slightly above the snow surface as they measured from the moving snowmobile, we systematically corrected each real-time kinematic snow surface elevation (RTKSS) measurement by subtracting a measured 0.05 m bias as this was the average height of the rover above actual snow surface conditions during transit.
Snow depth along the same transects was measured using a GPS Magnaprobe (Sturm et al. 1999; SnowHydro 2013), which consists of a metal probing rod and sliding basket that relays the snow depth along with the corresponding GPS position to a Campbell Scientific datalogger. The Magnaprobe Global Navigation Satellite System receiver has a reported 5–10 m spatial GPS accuracy and a 0.01 m depth precision (SnowHydro 2013). Berezovskaya and Kane (2007) demonstrated the potential for over-probing into the upper unfrozen organic layer in Arctic regions, which can cause large overestimations in snow depth by 11% to 31%. To quantify this for our study site, over-probing into the upper layer of the unfrozen organic layer was assed at an open tundra site, revealing an average of 0.07 m between the top of the organic layer and the upper portion of the frozen ground (Environment and Climate Change Canada 2018). Technical and logistical limitations result in an uneven distribution of validation points across each AOI, where AOIs located closer to the research station feature a higher density of validation points due to ease of access and ongoing collaborative research in those regions with other researchers. Magnaprobe data was simultaneously collected by the authors of this study and collaborators who have made their data set publicly available (Environment and Climate Change Canada 2018); however, only points within the boundary of each AOI are assessed in this study. The resulting validation data set consists of 7191 GPS-referenced Magnaprobe snow depths and 8434 snow surface elevation points across all seven AOIs (Table 1).
SfM snow depth
Snow depth maps produced using SfM photogrammetry highlight the spatial heterogeneity of snow depths across the AOIs (Fig. 3). Deep snow drift features are easily identified and are commonly found on the leeward slope of steep hills, valleys, and within tall shrub patches where blowing snow is deposited and entrained. In contrast, regions with shallow snow depths, such as those typically found across open tundra environments, are displayed in blue.
Snow surface elevation validation
RTK snow-surface elevations (RTKSS) were compared directly to underlying UAS DSM snow-surface elevations (UASSS). Results demonstrate the UASSS was accurately able to represent the snow-surface elevation when assessed against RTKSS (Fig. 4); however, the similarity between RTKSS and UASSS varies both within and across the AOIs. When points from all AOIs are aggregated, the mean bias is <1 cm between RTKSS and UASSS. There was no positive or negative bias overall, but three of the seven AOIs had negative biases and skewed density distributions, meaning the UASSS underestimated the measured snow surface elevation. RMSE values ranged from 0.09 to 0.26 m across AOIs, with an overall RMSE of 0.16 m (n = 8434).
In situ snow depth validation
In a similar study using SfM to estimate snow depth, Nolan et al. (2015) suggested that weak correlations often found with point-to-point comparisons resulted from a co-registration error between the high spatial accuracy (± 0.02–0.05 m, resampled to 1 m) of the SfM-derived snow depth product and the low spatial accuracy of the Magnaprobe-measured snow depths (SD) (± 5−10 m). To address this issue, we only compare MagnaprobeSD to the mean UASSD within a 5 m buffer around each MagnaprobeSD point providing a footprint roughly equivalent to the spatial accuracy of the Magnaprobe GPS. This comparison (Fig. 5) results in an overall Pearson’s correlation coefficient of 0.73 and RMSE ranging from 0.097 m to 0.20 m with an overall average of 0.15 m (Table 2); however, the strength of these relationships is clearly not uniform across all AOIs, as demonstrated by a the wide range in Pearson’s correlation coefficients in Table 2.
Snow depth differences between co-located UASSD and MagnaprobeSD points were plotted by AOI (Fig. 6) to assess the distribution of variations between the two snow depth methods. Here, we observe a positive bias on behalf of UASSD in five of seven AOIs and with the exclusion of AOI1, all AOIs feature a positive mean bias under 0.10 m. When we assess the variation between UASSD and MagnaprobeSD differences we observe 73% of the values are within one standard deviation of 0 m, with a high percentage (92%) contained within± 0.25 m for all AOIs.
Snow depth distributions
UASSD was sampled within a 50 m radius of each Magnaprobe survey transect to compare snow depth distributions between the two methods. The resulting data set provides a better sample datum for comparison purposes as in situ surveys fail to capture the full spatial variability in snow depths across the AOIs. The distributions of MagnaprobeSD and UASSD appear similar for most AOIs (Fig. 7). Kolmogorov–Smirnov (KS) tests were used to assess similarity between the two distributions. These analyses suggest statistical similarity in snow depth distributions in five of the seven AOIs at the 99% confidence interval. For AOI4 and AOI9, low p values and high critical values (D) suggest a statistically significant difference between the two distributions. Overall, we can conclude the UAS SfM snow depth maps mostly capture the distribution of snow depths along the in situ snow survey transects. However, the UAS captures a wider range of snow depths and features a higher number of snow depths exceeding 1 m.
Our results demonstrate the use of UAS SfM snow depth mapping over “large” areas of the Arctic during late-winter seasons, clearly showing the utility of this method for both research and operational purposes. Here, we demonstrate similar error estimates of the snow depth products as those found in previous studies when comparing in situ to UAS snow depths (Bühler et al. 2017; Harder et al. 2016; Nolan et al. 2015). However, we have applied the method to larger areas, with a more comprehensive set of validation data, across an Arctic tundra landscape, and under harsh late winter weather conditions. We were able to successfully map snow depths across a wide range of snow cover conditions, ranging from shallow (<0.3 m) wind-scoured areas to large topographic drift features (up to 4 m depth) at a 1 m spatial resolution. At this scale, common remote sensing products including airborne Radar and satellite Earth Observing platforms often fail to capture such variation in snow depths at high resolutions (Bhardwaj et al. 2016; Deems et al. 2013; Dong 2018). The ability to measure snow depth with such high spatial resolutions is integral to capturing small-scale heterogeneity of snow-covered tundra snowpacks, specifically for validation and improvement of ultra-fine resolution spatially distributed hydrological models. Recent studies (Clark et al. 2017; Peters-Lidard et al. 2017; Sivapalan 2018; Marsh et al. 2020) have outlined the ongoing advancement of ultra-fine hydrological models; however, these studies illustrate the need for high-resolution field data sets to validate such models. Advancements in UAS applications for snow hydrology contribute to improved validation of snow redistribution models through field observations of snow distribution at scales directly comparable with ultra-fine models.
Potential sources of error
Vegetation, such as tall shrubs, rises above the snowpack in some regions, which can result in inflated (false) snow-surface elevations. We attempted to correct this by editing the point-cloud products and removing features such as trees and tall erect shrubs. However, it is unlikely that all protruding shrubs and individual trees were removed, especially in dense shrub patches found on hillslopes and in riparian zones across all AOIs. This may result in an overestimation of snow depth in these regions and could be the source of UASSD bias found in AOI3 and AOI4, for example. Coincidently, these two AOIs also feature larger patches of spruce trees that may affect the accuracy of snow depths. The extent to which vegetation may influence snow depth estimates is not known and is difficult to calculate for the purposes of this study without a priori knowledge of the vegetation cover, including vegetation height and species, individual shrub age, and patch density. From observations during the winter field campaign it is estimated that shrubs protrude through the snow by up to 0.20–0.30 m, but only in a limited number of areas, a phenomenon that is quite common and difficult to quantify (Marsh et al. 2010; Menard et al. 2014).
In situ snow depth sample design
Observations of snow surface elevation demonstrate SfM can accurately represent the snow surface consistently across the study areas. However, the effectiveness of the SfM snow depth products shows variable agreement when assessed against in situ snow depths (Fig. 5), revealing a bias towards shallow snow depths (0.20 m to 0.60 m) within AOIs 3, 4, 7, 8, and 9. Weak linear relationships and relatively larger RMSE values within these AOIS could be partially attributed to the sample design within these AOIs whereby MagnaprobeSD surveying took place exclusively in shallow open tundra areas, rather than distributed across a representative sample of the study region. Shallow snow depths, such as those used to for these AOIs, lie towards the lower limits of the SfM methodology, as reported in previous studies by Harder et al. (2016) and Nolan et al. (2015), among others, and are inherently prone to proportionally larger snow depth accuracy errors when compared against in situ observations. This potential source of error was compensated for in the in situ sample design for AOI1 that took place a month later. Coincidently, here we find the strongest linear relationships, relatively low RMSE, and statistically similar snow depth distributions between UASSD and Magnaprobe snow depths suggesting some of the error estimates and weak agreement across some AOIs is derived from poor in situ sample design.
Ground control points and tie points
Errors in both RTKSS and MagnaprobeSD are minimal within 100 m of a GCP point (Fig. 8a), with increasing error beyond 200 m, suggesting that the SfM accuracy may degrade for areas in excess of this distance from a GCP due to a decrease in georeferenced input during SfM processing. For future studies we suggest increasing GCP density so that no single point is beyond 200 m from a GCP. However, for large study areas this may not always be achievable due to technical or accessibility restrictions. One solution would be to incorporate the use of RTK-enabled UAS as it increases the overall UASSS 3-dimensional accuracy without the need for surveyed GCP input (3-dimensional accuracy of ∼0.05 m based on previous experience). However, we suggest it is still good practice to incorporate georeferenced GCPs across a study area even when RTK-enabled UAS are available as the accuracy of the DSM is still subject to elevation errors that may reduce accuracy of snow depth estimates.
Although we programmed the UAS to collect images using 70/80 horizontal/vertical overlap, poor image quality resulting from strong winds and poor lighting can reduce image overlap that can reduce the frequency of common tie points, increasing spatial error estimates (Strecha 2014). As the number of overlapping images and common tie points reduces towards the outer boundary of the snow depth map, we assessed the difference between UAS products and both RTKSS and MagnaprobeSD as a function of the distance from the outer boundary of each AOI (Fig. 8b). Accuracy decreased in both RTKSS and MagnaprobeSD towards the outer bounds of the flight area (0–50 m from outer boundary). Based on the results in Fig. 8b we recommend a buffer zone of at least 50 m around the proposed study area as errors are minimal beyond this threshold. It should be noted that the final SfM DSM often has a smaller area than the original AOI flight boundary due to a low number of common tie points around the boundary. As such, these regions of the AOI were not included in this study as they featured no UASSS values for the SfM photogrammetry software.
Environmental conditions and challenges
Availability of daylight during the polar winter presents a significant issue that cannot easily be overcome using optical imagery collected with UAS. Our measurements were made in March and April when lighting conditions were sufficient for multiple flights per day. However, other environmental factors such as low near surface air temperatures and strong winds presented challenges (Table 1). During flight acquisitions winds periodically reached an excess of 12 m/s, beyond the operating capability of our UAS. Based on the experiences with flight acquisitions conducted during this study we suggest flight operations be suspended when winds speeds exceed 8 m/s as the quality of the imagery degrades, and battery life is greatly diminished beyond this wind speed.
Average near-surface air temperatures during UAS flight acquisitions (Table 1) also provided many challenges, with half-hourly surface temperatures ranging from −5 °C to −20 °C during March and averaging −16 °C during April. This is below the manufacturers recommended minimum operating temperature of −15 °C (senseFly 2017) representing a challenge for mapping Arctic snow using UAS. However, as this study demonstrates some of these challenges can be overcome with strategic heating of the UAS and battery management. For future studies we suggest limiting data acquisition to climate conditions within the manufacturer’s recommended settings to ensure safe and efficient imagery collection.
UAS SfM photogrammetry has proven effective for mapping snow depth in alpine and prairie environments. However, no studies had applied this method in Arctic tundra environments, where there is more vegetation and greater spatial variability in snow depth. We assessed the ability of the UAS to measure snow-surface elevation using a large sample (n = 8434) of snow surface elevation measurements. Our results conclude that the UAS is able to accurately capture the snow-surface elevation with relatively small error margins (RMSE = 0.16 m with a mean bias <0.01 m). We also used Magnaprobe snow depths (n = 7191) to assess the ability of the UAS to quantify snow depth and found the UAS SfM snow depth maps to accurately represent both the point-to-point snow depth (RMSE = 0.15 m) and the overall snow depth distributions of comparable subsets, although the strength of the relationships varies by AOI. Assessment of snow depth accuracies agree with previous UAS SfM studies conducted in prairie and alpine regions (RMSE ranging between 0.07 m and 0.30 m). However, here we present one of the largest validation data sets to date across multiple large (0.75 km2 to 2.35 km2) tundra study sites.
High-resolution snow depth maps produced with UAS SfM photogrammetry offer a unique opportunity to capture small-scale changes in snow distributions across the landscape. Such data are unique as the spatial heterogeneity of Arctic tundra snow cover has proven difficult to capture using traditional in situ and satellite-based remote sensing techniques. It is hopeful that these data will contribute to improvement of current and future distributed snow models by providing comparable validation data sets.
The authors would like to thank Chris Derksen, Peter Toose, Arvids Silis, Nick Rutter and Richard Essery for their assistance in the field. We would also like to acknowledge those working from the Trail Valley Creek Research Station and the Aurora Research Institute for logistical support. The authors would like to thank Tyler de Jong and Philip Mann for assistance over the years in developing and testing the above methodology. This study was made possible by financial support from the Natural Sciences and Engineering Research Council of Canada (NSERC), Polar Continental Shelf Program (PCSP), ArcticNet, Canada Foundation for Innovation, the Canada Research Chair Program, Wilfrid Laurier University, and Polar Knowledge Canada. This project was conducted with approval issued by Aurora Research Institute– Aurora College (License No. 16237). EJW was supported by the W. Garfield Weston Foundation. The authors would like to acknowledge that this study occurred within the Inuvialuit Settlement Region located in the Western Canadian Arctic.
Accompanying data sets are stored and openly available at the following DOI: https://doi.org/10.5683/SP2/PWSKKG.
Berezovskaya, S., and Kane, D.L. 2007. Measuring snow water equivalent for hydrological applications: part 1, accuracy of observations. In 16th International Northern Research Basins Symposium and Workshop. Petrozavodsk, Russia, 27 August – 2 September 2007. pp. 29–35.
Bhardwaj A., Sam L., Bhardwaj A., and Martín-Torres F.J. 2016. LiDAR remote sensing of the cryosphere: Present applications and future prospects. Remote Sens. Environ. 177: 125–143.
Bühler Y., Marty M., Egli L., Veitinger J., Jonas T., Thee P., and Ginzler C. 2015. Snow depth mapping in high-alpine catchments using digital photogrammetry. Cryosphere, 9: 229–243.
Bühler Y., Adams M.S., Bosch R., and Stoffel A. 2016. Mapping snow depth in alpine terrain with unmanned aerial systems (UASs): Potential and limitations. Cryosphere 10: 1075–1088.
Bühler Y., Adams M.S., Stoffel A., and Boesch R. 2017. Photogrammetric reconstruction of homogenous snow surfaces in alpine terrain applying near-infrared UAS imagery. Int. J. Remote Sens. 38: 3135–3158.
Clark M.P., Bierkens M.F.P., Samaniego L., Woods R.A., Uijlenhoet R., Bennett K.E., et al. 2017. The evolution of process-based hydrologic models: Historical challenges and the collective quest for physical realism. Hydrol. Earth Syst. Sci. 21: 3427–3440.
De Michele C., Avanzi F., Passoni D., Barzaghi R., Pinto L., Dosso P., et al. 2016. Using a fixed-wing UAS to map snow depth distribution: An evaluation at peak accumulation. Cryosphere, 10: 511–522.
DeBeer C.M. and Pomeroy J.W. 2017. Influence of snowpack and melt energy heterogeneity on snow cover depletion and snowmelt runoff simulation in a cold mountain environment. J. Hydrol. 553: 199–213.
Deems J.S., Painter T.H., and Finnegan D.C. 2013. Lidar measurement of snow depth: A review. J. Glaciol. 59: 467–479.
Derksen C., Silis A., Sturm M., Holmgren J., Liston G.E., Huntington H., and Solie D. 2009. Northwest territories and nunavut snow characteristics from a subarctic traverse: implications for passive microwave remote sensing. J. Hydrometeorol. 10: 448–463.
Dong C. 2018. Remote sensing, hydrological modeling and in situ observations in snow cover research: A review. J. Hydrol. 561: 573–583.
Environment and Climate Change Canada. 2020. Canadian climate normals 1981-2010 station data: inuvik a [WWW Document]. Available from http://climate.weather.gc.ca/climate_normals/results_1981_2010_e.html?stnID=1669&autofwd=1 [accessed 20 February 2020].
Environment and Climate Change Canada. 2018. Snow depth data from trail valley creek.
Harder P., Schirmer M., Pomeroy J., and Helgason W. 2016. Accuracy of snow depth estimation in mountain and prairie environments by an unmanned aerial vehicle. Cryosphere 10: 2559–2571.
Hopkinson, C., Pietroniro, A., and Pomeroy, J.W. 2008. Hydroscan: airborne laser mapping of hydrological features and resources. Canadian Water Resources Association, Saskatoon, SK, Canada.
Lantz, T.C., and Kokelj, S. V., 2008. Increasing rates of retrogressive thaw slump activity in the Mackenzie Delta region, N.W.T., Canada. Geophys. Res. Lett. 35, L06502.
Mann, P. 2018. Spatial and temporal variability of the snow environment in the Western Canadian Arctic. Wilfrid Laurier University.
Marsh C.B., Pomeroy J.W., Spiteri R.J., and Wheater H.S. 2020. A finite volume blowing snow model for use with variable resolution meshes. Water Resour. Res. 56: 1–28.
Marsh P., Bartlett P., MacKay M., Pohl S., and Lantz T. 2010. Snowmelt energetics at a shrub tundra site in the western Canadian Arctic. Hydrol. Process. 24: 3603–3620.
Marsh P. and Pomeroy J.W. 1996. Meltwater fluxes at an arctic forest-tundra site. Hydrol. Process. 10: 1383–1400.
Menard C.B., Essery R., Pomeroy J., Marsh P., and Clark D.B. 2014. A shrub bending model to calculate the albedo of shrub-tundra. Hydrol. Process. 28: 341–351.
Nolan M., Larsen C., and Sturm M. 2015. Mapping snow-depth from manned-aircraft on landscape scales at centimeter resolution using Structure-from-Motion photogrammetry. Cryosph. Discuss. 9: 333–381.
Pan X., Yang D., Li Y., Barr A., Helgason W., Hayashi M., et al. 2016. Bias corrections of precipitation measurements across experimental sites in different ecoclimatic regions of western Canada. Cryosphere, 10: 2347–2360.
Peters-Lidard C.D., Clark M., Samaniego L., Verhoest N.E.C., Van Emmerik T., Uijlenhoet R., et al. 2017. Scaling, similarity, and the fourth paradigm for hydrology. Hydrol. Earth Syst. Sci. 21: 3701–3713.
Pohl S. and Marsh P. 2006. Modelling the spatial-temporary variability of spring snowmelt in an arctic catchment. Hydrol. Process. 20: 1773–1792.
Pomeroy J.W., Marsh P., and Gray A.D.M. 1997. Application of a Distributed Blowing Snow Model To the Arctic. Hydrol. Process. 11: 1451–1464.
Rampton, V.N. 1987. Surficial geology, Tuktoyaktuk Coastlands, District of Mackenzie, Northwest Territories, Geological Survey of Canada, “A” Series Map no. 1647A.
Rutter N.J., Sandells M., Derksen C., King J., Toose P., Wake L., et al. 2019. Effect of snow microstructure variability on Ku-band radar snow water equivalent retrievals. Cryosphere, 13: 3045–3059.
Sivapalan M. 2018. From engineering hydrology to Earth system science: Milestones in the transformation of hydrologic science. Hydrol. Earth Syst. Sci. 22: 1665–1693.
SnowHydro. 2013. GPS snow depth probe [WWW Document]. Available from http://www.snowhydro.com/products/column2.html.
Steppuhn, H., and Dyck, G.E. 1974. Estimating true basin snowcover. Interdiscip. Symp. Adv. Concepts Tech. Study Snow Ice Resour. 314–328.
Stuefer S.L., Kane D.L., and Dean K.M. 2020. Snow water equivalent measurements in remote arctic alaska watersheds. Water Resour. Res. 56: 3.
Sturm M. and Benson C. 2004. Scales of spatial heterogeneity for perennial and seasonal snow layers. Ann. Glaciol. 38: 253–260.
Sturm, M., Holmgren, J., and Liston, G.E. 1999. Self-recording snow depth probe. 1741.
Vander Jagt B., Lucieer A., Wallace L., Turner D., and Durand M. 2015. Snow depth retrieval with UAS using photogrammetric techniques. Geosciences, 5: 264–285.
Walker D.A., Reynolds M.K., Daniëls F.J.A., Einarsson E., Elvebakk A., Gould W.A., et al. 2005. The Circumpolar Arctic vegetation map. J. Veg. Sci. 16: 267–282.
Wilcox E.J., Keim D., de Jong T., Walker B., Sonnentag O., Sniderhan A.E., et al. 2019. Tundra shrub expansion may amplify permafrost thaw by advancing snowmelt timing. Arct. Sci. 5: 202–217.
Information & Authors
Volume 7 • Number 3 • September 2021
Pages: 588 - 604
Received: 9 March 2020
Accepted: 7 October 2020
Published online: 17 November 2020
This article is part of a collection on Unoccupied Vehicle Systems in Arctic Research and Monitoring jointly published by Arctic Science and Journal of Unmanned Vehicle Systems.
© 2020 The Author(s). This work is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author(s) and source are credited.
Metrics & Citations
BrandenWalker, Evan J.Wilcox, and PhilipMarsh. 2020. Accuracy assessment of late winter snow depth mapping for tundra environments using Structure-from-Motion photogrammetry. Arctic Science. 7(3): 588-604. https://doi.org/10.1139/as-2020-0006
If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download.
1. Monitoring 13 years of drastic catchment change and the hydroecological responses of a drained thermokarst lake
2. Winters are changing: snow effects on Arctic and alpine tundra ecosystems1
3. High‐resolution snow depth prediction using Random Forest algorithm with topographic parameters: A case study in the Greiner watershed, Nunavut
4. Characterizing tundra snow sub-pixel variability to improve brightness temperature estimation in satellite SWE retrievals
5. Unpiloted Aerial Vehicle Retrieval of Snow Depth Over Freshwater Lake Ice Using Structure From Motion
6. Snow water equivalent measurement in the Arctic based on cosmic ray neutron attenuation