Spatiotemporal patterns of rain-on-snow and basal ice in high Arctic Svalbard: detection of a climate-cryosphere regime shift

Arctic winters have become increasingly warmer and rainier. Where permafrost prevails, winter rain (or rain-on-snow) is known to occasionally cause extensive ice layers at the snow/ground interface, i.e. ‘basal ice’ or ‘ground ice’, with potentially large ecological and socio-economic implications. However, an overall lack of field data has so far restricted our predictive understanding of the environmental conditions shaping spatiotemporal variation in basal ice. Here, we use time-series of spatially replicated snowpack measurements from coastal (Ny-Ålesund area; 2000–2017) and central Spitsbergen (Nordenskiöld Land; 2010–2017), Svalbard, to analyze spatiotemporal patterns in basal ice and how they are linked with topography, weather, snowpack and climate change. As expected, both the spatial occurrence and thickness of basal ice increased strongly with the annual amount of winter rain. This effect was modified by accumulated snowfall; a deeper snowpack restricts ice formation following a minor rain event, but enhances ice formation following heavy rain due to an increased contribution of snowmelt. Accordingly, inter-annual variation in snow depth was negatively related to basal ice thickness. Annual fluctuations in basal ice thickness were strongly correlated in space (average correlation ρ = 0.40; 0–142 km distance between plots) due to strong spatial correlation in winter rain (ρ = 0.62; 14–410 km distance between meteorological stations). Models of basal ice based on meteorological time-series (1957–2017) suggested that ice-free winters (i.e. mean basal ice <0.1 cm) had virtually not occurred since 1998, whereas such winters previously (1957–1998) occurred every three–four years on average. This detected cryosphere regime shift was linked to a parallel climate regime shift with increased winter rain amounts. Svalbard is regarded a bellwether for Arctic winter climate change. Our empirical study may therefore provide an early warning of future changes in high-arctic snowpacks.


Introduction
In the Arctic, winter warm spells with near-surface air temperatures above 0°C are becoming more frequent due to global warming (Moore 2016, Graham et al 2017). This warming, in combination with enhanced surface evaporation due to the loss of sea-ice cover and poleward atmospheric moisture transport, contributes to an overall increase in precipitation over Arctic land areas (4.5% increase per degree of temperature rise), particularly in late autumn and winter (Serreze et al 2009, Zhang et al 2012, Bintanja and Selten 2014. Recent climate models indicate that rainfall will likely become the dominant form of winter precipitation Original content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. during the 21st century (Rennert et al 2009, Bintanja andAndry 2017). The rain-on-snow (ROS) phenomenon (Putkonen and Roe 2003) is of particular concern as it can lead to avalanches (Conway and Raymond 1993), increased ground surface temperature and near-surface permafrost thawing (Isaksen et al 2007, Westermann et al 2011, and ice-wedge cracking if water freezes after infiltrating the soil (Christiansen et al 2013). Furthermore, heavy ROS events can lead to thick ice layers within the snowpack or at the snow/ ground interface, i.e. 'basal ice' or 'ground ice', across the landscape (figure 1; e.g. Vikhamar-Schuler et al 2013).
Several studies have documented negative effects of ROS and basal ice on small and large herbivores (Kausrud et al 2008, Stien et al 2012, soil invertebrates (Coulson et al 2000), vegetation growth and reproduction (Bjerke 2011, Preece et al 2012, Milner et al 2016, Bjerke et al 2017, and even whole vertebrate communities (Hansen et al 2013). The most noticeable consequences are the occasional die-offs of Arctic ungulates due to mass starvation, as the impenetrable layer of ice makes vegetation inaccessible (e.g. Parker et al 1975, Forchhammer and Boertmann 1993, Miller and Gunn 2003, Kohler and Aanes 2004, Hansen et al 2011, Langlois et al 2017. More frequent ROS and basal ice formation in space and time may therefore also have serious socio-economic impacts for reindeer herders (Bartsch et al 2010, Forbes et al 2016, Riseth et al 2016, but also for the tourism industry and local communities in the Arctic (Hansen et al 2014).
Basal ice is formed when liquid water, deriving from snowmelt and/or rainfall (ROS), pools at the bottom of the snowpack and freezes as latent heat is transferred to the surrounding snowpack and frozen ground (Woo et al 1982, Putkonen and Roe 2003, Westermann et al 2011. In winter, small amounts of rainfall or melted snow resulting from solar radiation and warm air temperatures (Hock 1999) can be absorbed within the snowpack, where it freezes along lateral flow channels (Marsh and Woo 1984). However, as the snowpack becomes saturated, sufficient amounts of rainfall can percolate vertically through the snowpack and freeze in contact with the frozen ground (Woo et al 1982, Marsh and Woo 1984, Conway and Benedict 1994. During ROS events, a considerable proportion of total snowpack runoff can be caused by snowmelt (27% on average in the Swiss Alps, Würzer et al 2016). Heat fluxes from warm, moist, and windy conditions can be responsible for the main snowmelt energy input (e.g. Marks et al 1998). However, advective heat from rain is a potentially large additional source for melt energy (Würzer et al 2016). This is particularly the case in the Arctic, where shortwave radiation is not available for melting during the polar night, and near-surface air temperatures need to be several degrees above freezing point to sufficiently increase snowmelt rates (Hock 1999, Putkonen andRoe 2003). Accordingly, in high-arctic environments where permafrost prevails, the formation of basal ice layers is typically associated with heavy ROS and concurrent snowmelt during warm spells (Putkonen and Roe 2003, Kohler and Aanes 2004, Bartsch et al 2010, Vikhamar-Schuler et al 2013, Hansen et al 2014.
Because of a general lack of field data time-series from the high Arctic (but see Kohler and Aanes 2004, Hansen et al 2011for Svalbard, and Bulygina et al 2010, our empirical understanding of the spatial and temporal variation in basal ice occurrence remains poor. Some studies have used proxies or simulations based on meteorological or satellite data (e.g. Grenfell and Putkonen 2008, Bartsch et al 2010, Vikhamar-Schuler et al 2013, Langlois et al 2017, without validation from in situ field data. Furthermore, little is known on the spatial extent of basal ice events and how they are linked to large-scale meteorological patterns (but see Bartsch et al 2010, Forbes et al 2016.
In this study, we address this by taking advantage of time-series of spatially replicated snow and basal ice measurements from coastal (Ny-Ålesund area; 2000-2017) and central Spitsbergen (Nordenskiöld Land, near Longyearbyen;-2017, Svalbard, combined with data from local meteorological stations. First, we investigated combined effects of topography, temperature and precipitation patterns, and climate change on basal ice occurrence and thickness. To make our approach applicable across study areas and scientific disciplines, we built simple regression models based on commonly available data from meteorological stations and topographic information from digital elevation models. We then investigated how the spatial extent of basal ice events across our study area was linked to spatial correlation in winter rain. Finally, we modelled and predicted historical occurrence and thickness in basal ice using our regression models accounting for weather and topography, and investigated temporal changes linked with the recent warming in high Arctic Svalbard (Hansen et al 2014.

Study area
The two study areas are located in coastal and central Spitsbergen, Svalbard (74-81°N, 10-35°E; figure 2(a)). The coastal study area in North-West Spitsbergen, hereafter referred to as 'NW coast', consists of the mountainous peninsula Brøggerhalvøya, where the research settlement Ny-Ålesund is located, and the two peninsulas Sarsøyra and Kaffiøyra, which are characterized by flat coastal planes with steep mountains to the East ( figure 2(b)). The study area in Central Spitsbergen is located 17-33 km south of the meteorological station at Svalbard Airport, Longyearbyen, and covers the valleys Colesdalen, Semmeldalen and Reindalen (figure 2(c)). The climate in these parts of the Svalbard archipelago, and particularly the NW coast, is influenced by the North-Atlantic current and the coupled sea-ice-ocean atmosphere system (Benestad et al 2002). Annual total precipitation and mean temperature for 1970-2015 were on average 417 mm and −5.1°C for Ny-Ålesund, and 194 mm and −5.0°C for Svalbard Airport.
In our study, winter was defined as the period between 1 November (preceding year) and 31 March, i.e. just prior to the basal ice measurements. The main reason for the 1 November cut-off is that the onset of ground surface freezing, although depending on the landform, generally occurs some time during October (Eckerstorfer et al 2017). In addition, during the study period (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017), daily air temperatures were often <0°C in November (mean±SD=−6.5±5.3°C for Svalbard Airport and Ny-Ålesund combined), and less so in October (−3.2±4.4°C). Total winter precipitation and mean winter temperature were on average 204 mm and −11.2°C in Ny-Ålesund, and 88 mm and −11.8°C at Svalbard Airport. However, a strong winter warming has been identified during the last decades (Førland et al 2011), and warm spells with above-zero temperatures now occur more frequently Climate-cryosphere data Snow profiles and basal ice were sampled in April/ early May (earliest 29 March, latest 9 May). Snow pits were dug manually using a spade and, if basal ice was present, an axe or drill was used to reach the ground surface (figure 1(a)). Snow depth, basal ice thickness and the total thickness of ice layers within the snowpack were measured. Data were collected for a total of n=2 539 observations during the period -2017in Central Spitsbergen and 2000-2017(except 2001 ; see supplementary material 1, available online at stacks.iop.org/ERL/14/015002/mmedia). In Central Spitsbergen, sampling was conducted annually at the same sites (n=128), which were spatially structured at eight different locations following a hierarchical block design (see Loe et al 2016). Plots at each location covered ridges and sub-ridges at the smallest scale (5 m apart), and valley bottoms and flat hilltops at the largest scale (500 m apart). On the NW coast, snow-ice data were combined from three studies with varying sampling design (see supplementary material 1 for details). These included both randomly placed snow pits along transect lines (2000-2007, 2012-2015Kohler and Aanes 2004) and repeated measures at fixed sites following either a randomly placed grid system Daily average air temperature and amount of precipitation were obtained for seven meteorological stations across Svalbard (figure 2(a); table S1.2). All stations were used to analyze spatiotemporal patterns of winter rain, while only the meteorological stations in Ny-Ålesund and at Svalbard Airport were used for the basal ice analyses of the NW coast and Central Spitsbergen study areas, respectively. Winter (November-March) precipitation was classified as rain when daily average temperature1°C (Stien et al 2012, Hansen et al 2013. Note that a large rain event (R. Aanes pers. obs.) occurred on 24 March 2007 in Ny-Ålesund (67.3 mm), however with air temperatures ranging from −3.3°C to 3.8°C (mean=−0.3°C), which was included when calculating the variable Rain, i.e. total amount of winter rain (mm; figures 2(d), 3(c)). The 'peak rain event' for each winter was defined as the largest rainfall over a three-day period. Daily precipitation classified as snow (i.e. at temperatures <1°C) was then summarized from 1 November until this event to obtain the variable Snow_P, i.e. the accumulated amount of snowfall until the peak rain event (mm; figure 3(d)). Winter heat sum was calculated by accumulating daily temperatures >0°C for November-March (figure 3(e)).

Data analysis
First, we compared annual basal ice occurrence (presence/absence, where presence 0.5 cm basal ice) and thickness (cm on natural logarithmic (hereafter log) scale) between the NW coast and Central Spitsbergen for the period in which the time-series overlap (2010-2017). We also investigated how average snow depth measured in late winter (April/early May) was related to average observed basal ice thickness, accounting for accumulated snowfall (November-March), in the two study areas (see supplementary material 2 for details on the data analysis).
Second, we used mixed-effects regression models to investigate the effects of climate and topography on the occurrence and thickness of basal ice. One major advantage of mixed models is that they can account for temporal and spatial autocorrelation as well as unbalanced repeated observations (Cnaan et al 1997). A priori, we expected to find a positive effect of winter rain on ice occurrence and thickness (Putkonen and Roe, 2003), possibly modified by the amount of accumulated snowfall (Würzer et al 2016). We also tested for an effect of total accumulated heat sum as snowpack warming may lead to snow melting and refreezing at the bottom of the snowpack. Elevation (m a.s.l.) and Slope (degrees) at the sample plot level, derived from a Digital Elevation Model with a 20 m resolution (http://geodata.npolar.no), were included to account for topographic effects on basal ice. To evaluate the best fitting parameters, we performed model selection over candidate models using Akaike's information criterion (Burnham and Anderson 2002; see supplementary material 2 and 4).
Third, to quantify the spatial scales at which winter rain and basal ice occur, we analyzed spatial correlation in annual fluctuations of winter rain and ice thickness between pairs of meteorological stations or sampling sites, respectively. For this, we implemented a nonparametric covariance function on pairwise correlations (ρ) as a function of distance (Bjørnstad and Falck, 2001). To investigate the contribution of rain to the spatial correlation in basal ice, we first fitted loglinear models of basal ice thickness with rain as a predictor for every sampling site, and then analyzed the spatial correlation in annual fluctuations of the (d) Accumulated amount of snowfall from 1 November until the peak rain event (Snow_P). (e) Winter heat sum, i.e. accumulated temperature above 0°C (November-March). Basal ice data from the NW coast (2000-2008-2017) and Central Spitsbergen (2010-2017, and weather data from their local stations in Ny-Ålesund and at Svalbard Airport, are shown in blue and red, respectively. residuals from these models (see supplementary material 2 for details).
Finally, we investigated temporal changes in winter rain and basal ice. For this, we modelled and predicted past basal ice occurrence and thickness by using historical weather data (since 1958 and 1970 for Svalbard Airport and Ny-Ålesund, respectively) applied to our top-ranked regression models (supplementary material 4). Preliminary analyses suggested that changes over time were clearly not following linear trends. Thus, we tested for regime shifts (i.e. inter-decadal fluctuations in average levels; Overland et al 2006) in the time-series of winter rain and modelled basal ice (supplementary material 2).

Results
The average amount of winter rain (mm; log scale) was more than twice higher on the NW coast than in Central Spitsbergen (paired t-test: t=4.27, d.f.=47, P<0.001). Similarly, average basal ice thickness (cm; log scale) on the NW coast was on average twice the estimate for Central Spitsbergen (estimated ratio=2.0±0.7; figure 3(b)), while the spatial occurrence of basal ice was also more extensive during most winters ( figure 3(a)). Furthermore, late winter snow depth (measured in April/early May) was on average 53% deeper on the NW coast, and was more shallow in years with thick basal ice layers, independent of snowfall amount ( figure 4, table S3).
Annual fluctuations in basal ice occurrence and thickness largely followed the fluctuations in winter rain ( figure 3). Accordingly, the model selection for the regression analyses (supplementary material 4) indicated that the amount of winter rain was the strongest predictor of basal ice, with a positive effect on both occurrence and thickness (table 1, figure 5). As expected, there was a positive interaction between rain and accumulated snowfall (Snow_P) due to melting and refreezing processes in the snowpack (e.g. Woo et al 1982, Würzer et al 2016. Large amounts of snowfall (high Snow_P, i.e. presumably deep snow accumulation) prevented the formation of basal ice when there was only minor rainfall, but resulted in thicker basal ice when associated with major rain events. Accordingly, shallow late-winter snow depths recorded in years with particularly icy conditions (figure 4) were likely associated with more rain-induced snowmelt. At Table 1. Parameter estimates (β) with standard errors (SE) of standardized fixed effects covariates from the generalized linear mixed model (GLMM) and linear mixed model (LMM) of basal ice occurrence (on logit scale) and thickness, respectively. Rain, Snow_P (i.e. accumulated snowfall until the peak rain event) and Ice thickness were log-transformed in the analysis. Standard deviations (SD) and number of groups (n) are given for random effects on the intercept. Marginal and conditional R 2 indicate variance explained by the fixed effects and by both fixed and random effects, respectively (Nakagawa and Schielzeth 2013).

Ice occurrence (GLMM)
Ice thickness (LMM) the local scale, basal ice occurrence and thickness decreased with higher elevation and steeper slopes (table 1). Furthermore, the positive effect of rain on basal ice thickness decreased with increasing elevation, because precipitation is more likely to fall as snow at higher altitudes. Annual fluctuations in winter rain were strongly correlated over distances up to 410 km (i.e. the max distance between meteorological stations), with an average spatial or 'regional' correlation of ρ=0.62 [95% CI: 0.48, 0.76] ( figure 6(a)). However, the spatial correlation in winter rain gradually decreased with distance between the meteorological stations. Similarly, annual measurements of basal ice thickness were strongly correlated between sampling sites within and between the two study areas. The regional correlation in basal ice thickness was higher on the NW coast (ρ=0.61 [95% CI: 0.53, 0.68]) than in Central Spitsbergen (ρ=0.36 [95% CI: 0.31, 0.41]; figures 6(b)-(c)), but remained high across the two study areas (ρ=0.40 [95% CI: 0.36, 0.45]; figure S6(a)). After accounting for the effect of rain, the regional correlation in basal ice thickness (i.e. the correlation in model residuals) was close to zero (ρ=0.06 [95% CI: 0.02, 0.10]; figure S6(b)). Accordingly, and as expected, the strong spatial correlation in temporal variation of basal ice thickness across the landscape and study areas was mainly due to the strong spatial correlation in winter rainfall.
Historical weather data from Svalbard Airport and Ny-Ålesund (continuous time-series available since 1957 and 1969, respectively) indicated that winters with virtually no rain occurred approximately every third to fourth year until 1998 (figures 7(a)-(b)). Based Figure 5. Annual variation in basal ice is explained by the interaction between winter rain and snow. Estimated effects (regression slopes) of winter rain (log scale) are shown for basal ice (a) occurrence (probability) and (b) thickness (log scale), and at low (blue; mean -1SD) versus high (red; mean+1SD) amounts of accumulated snowfall until the peak rain event (Snow_P). Shaded areas indicate 95% CI from 1000 parametric bootstrap iterations. Observed means (one dot per year and study area) follow a colour gradient from low (blue) to high (red) amounts of Snow_P. on change point analyses, we detected a climate regime shift with an overall increase in the average amount of winter rain (log scale) around the turn of the century. Accordingly, when our best-ranked mixed-effects models were used to model basal ice occurrence and thickness based on historical weather data, we also detected a climate-related regime shift in the frequency of icy winters (figures 7(c)-(f)). Rain-and icefree winters have virtually not occurred since 1998. On the NW coast (Ny-Ålesund), the modelled occurrence of basal ice in the landscape (at an arbitrary elevation of 50 m a.s.l.) was on average 49% before 1987, 20% between 1987 and 1998, and 80% after 1998 (figure 7(c)). Modelled average basal ice thickness increased almost three-fold from 0.99 to 2.88 cm before and after 1998, respectively (figure 7(e)). Also, for Central Spitsbergen (Svalbard Airport), a recent change point in the time-series of modelled basal ice occurrence was detected in 1999, when average ice occurrence increased from 40% to 58% before and after, respectively ( figure 7(d)). Here, no change point was detected for ice thickness, but winters with very low ice thickness have been virtually absent since 1998 ( figure 7(f)).

Discussion
In the present study, we used spatiotemporal modelling of long-term field data from the high Arctic archipelago of Svalbard to demonstrate the occurrence of a major climate-cryosphere regime shift ( (a) 1987, 2000; (b) 1982, 2009; (c) 1987, 1998; (d) 1973, 1999; (e) 1998; (f) no change point observed). and snowpack simulations (e.g. Vikhamar-Schuler et al 2013). However, the very few long-term in situ studies in the high Arctic are from Svalbard (Kohler and Aanes 2004, Hansen et al 2011, Loe et al 2016. By combining these empirical time-series data, we have improved our predictive mechanistic understanding of the environmental conditions causing variation in basal ice in time and space. As expected, basal ice occurrence and thickness increased strongly with the amount of winter rain, particularly so when the accumulated snowfall until the largest rain event was high (figure 5, see also Putkonen and Roe 2003). Accordingly, for a given total snowfall amount, annual snow depths were negatively related to basal ice thickness (figure 4). Since ROS events generally occur over a substantial area (e.g. Rennert et al 2009) and are strongly correlated across Svalbard (figure 6(a)), observed basal ice thickness was also strongly correlated across sampling sites (figures 6(b)-(c), S6). This supports previous studies showing that icing events tend to occur across large spatial scales (Bartsch et al 2010, Hansen et al 2014, Forbes et al 2016. The positive interaction effect of rain and accumulated snowfall on basal ice likely relates to the complex thermodynamic process of latent heat exchange as rain percolates through the snowpack (Woo et al 1982, Marsh and Woo 1984, Westermann et al 2011. For low amounts of rain falling on top of a deep snowpack, water may fill in the pore space between the snow grains and freeze, or may flow laterally through subsurface layers Woo 1984, Conway andRaymond 1993), thus forming ice layers within the snowpack rather than on the ground surface (supplementary material 7). In contrast, substantial quantities of rainfall during a warm spell will likely percolate through the entire snowpack, as it becomes saturated (Conway and Raymond 1993, Putkonen and Roe 2003, Westermann et al 2011. Furthermore, latent heat exchange between rain and snow may increase the cumulative water runoff (Würzer et al 2016), thus increasing the formation of basal ice as melted snow and rainwater freezes on the frozen ground.
Winter rain may occur at near-surface temperatures below 0°C (i.e. 'freezing rain') due to vertical isothermal gradients in the atmosphere with melting layers at higher altitudes and sub-freezing layers towards the ground (Roberts and Stewart 2008). Particularly in coastal regions such as Ny-Ålesund, warmer moist air from the sea may be transported to higher atmospheric layers above land. However, most heavy rain events occur during extreme warm spells coupled with increased wind and air moisture (Hansen et al 2014), which boost turbulent fluxes and snowmelt. Therefore, our winter rain variable based on a 1°C threshold is likely to capture the combined effect of rain and increased temperature on basal ice. The western and southern coastal regions of Spitsbergen generally experience more rainfall (Van Pelt et al 2016), and thus heavier basal ice formation, compared to the inner fjord areas of Svalbard. Furthermore, on a more local scale, the positive effect of rainfall amount on basal ice thickness decreased at higher elevations (table 1), because winter precipitation is more likely to fall as snow due to lower ambient temperature at higher altitudes (see also Van Pelt et al 2016). Such an effect has also been proposed as a main explanation for why reindeer tend to climb up steep mountain slopes to forage when pastures at lower elevations are covered by ice (Hansen et al 2010).
Our models suggest that heavy rainfall that occurs early in winter is less likely to result in basal ice, as little snow has accumulated early in the season. Furthermore, the effect of winter rain on basal ice depends not only on snow cover, but also on other snowpack properties, such as initial liquid water content (Würzer et al 2016), linked with variation in air and ground temperatures as well as past icing events. Indeed, increased water run-off and drainage, rather than ice development, would be expected when ground temperatures are close to or above zero and/or air temperatures are not sufficiently low following warm spells. We should therefore expect more complex, non-linear relationships-especially under future winter climate scenarios (Bintanja and Selten 2014, Moore 2016, Bintanja and Andry 2017)-than those explored here.
While recent winters in Svalbard have become increasingly icy (figure 7), this has mainly been evident through a reduction in the frequency of ice-free winters, rather than increased average ice thickness. With continuously warmer and wetter winters, particularly for the early snow season (López-Moreno et al 2016, Bintanja and Andry 2017), and increased permafrost temperatures , rapidly warming sites like Svalbard may soon pass a tipping point where extreme winter rainfall actually results in less basal ice. Nevertheless, the increased frequency in ROS events and the coupled increase in basal ice, leading to a recent climate-cryosphere regime shift, are strong indicators of ongoing and near-future changes in high-arctic terrestrial environments, with potentially wide ecological and socio-economic implications.
with logistics, and numerous people for digging snow pits.