A synthesis of thermokarst lake water balance in high-latitude regions of North America from isotope tracers

Numerous studies utilizing remote sensing imagery and other methods have documented that thermokarst lakes are undergoing varied hydrological transitions in response to recent climate changes, from surface area expansion to drainage and evaporative desiccation. Here, we provide a synthesis of hydrological conditions for 376 lakes of mainly thermokarst origin across high-latitude North America. We assemble surface water isotope compositions measured during the past decade at five lake-rich landscapes including Arctic Coastal Plain (Alaska), Yukon Flats (Alaska), Old Crow Flats (Yukon), northwestern Hudson Bay Lowlands (Manitoba), and Nunavik (Quebec). These landscapes represent the broad range of thermokarst environments by spanning gradients inmeteorological, permafrost, and vegetation conditions. An isotope framework was established based on flux-weighted long-term averages of meteorological conditions for each lake to quantify water balance metrics. The isotope composition of source water and evaporation-to-inflow ratio for each lake were determined, and the results demonstrated a substantial array of regional and subregional diversity of lake hydrological conditions. Controls on lake water balance and how these vary among the five landscapes and with differing environmental drivers are assessed. Findings reveal that lakes in the Hudson Bay Lowlands are most vulnerable to evaporative desiccation, whereas those in Nunavik aremost resilient. However, we also identify the complexity in predicting hydrological responses of these thermokarst landscapes to future climate change.


Introduction
Thermokarst lakes and ponds (hereafter referred to collectively as lakes) are plentiful across permafrost terrain, occupying 15%-50% of the landscape in northwestern Canada, Siberia, and Alaska (e.g., Mackay 1988;Rampton 1988;Frohn et al. 2005;Grosse et al. 2005;Plug et al. 2008).Thermokarst lakes form as ice-rich permafrost thaws and surface water accumulates where subsidence occurs.These shallow waterbodies (generally <10 m deep and frequently <2 m) are a key component of northern hydrological and biogeochemical cycles, provide habitat and resources for wildlife and waterfowl populations, and support the traditional lifestyle of many indigenous communities.During the past few decades, increasing air temperatures and changes in precipitation patterns have been observed throughout much of the Arctic (e.g., ACIA 2004;IPCC 2013).Understanding the effects of climate change on thermokarst lake water balance is particularly important, as the greatest effects on aquatic ecosystems will occur indirectly via alteration of hydrological processes and their cascading influences on limnology, biogeochemistry, and aquatic ecology rather than from simply air temperature rise (Rouse et al. 1997;Prowse et al. 2006;Schindler and Smol 2006;Tranvik et al. 2009).Indeed, numerous studies have sought to document the hydrological status of thermokarst lakes.Many of these studies indicate that thermokarst lake hydrology is changing rapidly (e.g., Smith et al. 2005;Carroll et al. 2011), but along varying trajectories including surface area expansion, rapid drainage, and evaporative desiccation (e.g., Yoshikawa and Hinzman 2003;Riordan et al. 2006;Labrecque et al. 2009;Rowland et al. 2010;Bouchard et al. 2013).
Understanding the myriad of potential responses of thermokarst lake hydrology to ongoing climate change requires knowledge of their water balances (ΔS/ΔT, i.e., change in storage (S) over time (T)), which can be generally characterized as follows (Turner et al. 2010): Positive contributors to thermokarst lake water balance include snowmelt (P S ), rainfall (P R ), subsurface inflow (I GW ), and surface channelized inflow (I S ), whereas lake water loss may occur via evaporation (E), subsurface outflow (O GW ), and surface outflow (O S )the latter potentially occurring catastrophically.Relative roles of hydrological processes that control thermokarst lake water balances may be influenced by a variety of drivers, including meteorological and permafrost (continuous, discontinuous, and sporadic) conditions as primary drivers (e.g., Riordan et al. 2006;Plug et al. 2008;Labrecque et al. 2009).Changes in temperature can alter rates of evaporation (E), while changes in precipitation regimes can lead to direct fluctuations in snowmelt (P S ) and rainfall (P R ) input, surface channelized inflow (I S ), and surface outflow (O S ).Consequently, high rates of evaporation with low snowmelt or rainfall supply can cause lakes to desiccate, while low rates of evaporation and abundant supply from precipitation may result in attaining maximum basin capacity, which can lead to shoreline erosion and lake expansion or even rapid lateral lake drainage (e.g., Riordan et al. 2006;Hinkel et al. 2007;Plug et al. 2008;Marsh et al. 2009;Turner et al. 2010;Jones et al. 2011;MacDonald et al. 2012;Bouchard et al. 2013).A warming climate also causes increased permafrost degradation, which can influence thermokarst lake hydrological status (e.g., Yoshikawa and Hinzman 2003;Smith et al. 2005).For many thermokarst lakes, continuous permafrost impedes subsurface inflow (I GW ) and outflow (O GW ) from contributing significantly to lake water balance.However, as permafrost degrades, subsurface flow pathways can develop, which can lead to vertical lake drainage (e.g., Yoshikawa and Hinzman 2003;Jepsen et al. 2013).Additionally, landscape characteristics, such as catchment vegetation, strongly influence thermokarst lake water balance (Bouchard et al. 2013;Turner et al. 2014).For example, densely forested areas entrap snow, which results in enhanced snowmelt runoff to lakes during spring (P S ) compared to runoff generated in more sparsely vegetated areas.
Deciphering the relative influence of hydrological processes represented in eq. 1 is challenging, especially for lake-rich permafrost landscapes where there may be substantial spatial heterogeneity among lakes and their catchments.Due to logistical constraints of field work in remote locations, it is often impractical to perform direct conventional measurements of hydrological processes on a spatially extensive set of lakes that is required to capture the potential diversity of prevailing conditions.Alternatively, and especially for multiple lake studies across landscapes, measurement of water isotope composition (δ 2 H and δ 18 O) and application of isotope mass-balance models can be used to provide information of hydrological interest, as has recently been demonstrated for the continental United States (Brooks et al. 2014).For remote locations in particular, analysis of lake water isotope compositions is an excellent alternative to more instrument-intensive hydrological approaches.Surface water samples can easily and quickly be obtained during fieldwork, and their isotope compositions are sensitive to hydrological processes that influence lake water balances because systematic and well-understood isotopic fractionation of water occurs as it passes through the hydrological cycle (Edwards et al. 2004;Darling et al. 2006).Thus, the isotope composition of water provides quantitative information on lake water balance conditions, including the relative contributions of input waters (e.g., snowmelt, rain, and permafrost thaw waters as "δ I values") and the relative importance of evaporation (frequently expressed as an evaporation-to-inflow ratio (E/I)).Water isotope analysis has been applied in several northern and remote landscapes on thermokarst as well as other shallow lake systems, yielding novel insight into the diversity and importance of hydrological processes on lake water balances spanning multiple environmental gradients (e.g., Gibson and Edwards 2002;Brock et al. 2007;Yi et al. 2008;Turner et al. 2010Turner et al. , 2014;;Anderson et al. 2013;Tondu et al. 2013;Arp et al. 2015).
As an outcome of the Natural Sciences and Engineering Research Council of Canada Discovery Frontiers ADAPT (Arctic Development and Adaptation to Permafrost in Transition) project (Vincent et al. 2013), we provide a synthesis and snapshot of water balance conditions for 376 lakes in high-latitude North America that mainly formed by thermokarst processes.Specifically, we assemble surface water isotope compositions measured during summers of the past decade from mainly thermokarst lakes across five expansive lake-rich permafrost landscapes.From west to east, these include Arctic Coastal Plain (Alaska) (Arp et al. 2015), Yukon Flats (Alaska) (Anderson et al. 2013), Old Crow Flats (Yukon) (Turner et al. 2010(Turner et al. , 2014;;Tondu et al. 2013), western Hudson Bay Lowlands (Manitoba) (Bouchard et al. 2013), and Nunavik (Quebec) (Narancic et al. 2017).We use isotope-mass balance modeling to determine lake input-water isotope compositions and E/I ratios and explore their relations among landscapes and with environmental drivers.Results provide a unique opportunity to rank hydrological vulnerability of these lake-rich permafrost landscapes and to predict hydrological responses to various climate change scenarios.While most of these data have been previously published as part of individual landscape hydrological studies, to our knowledge, the present analysis is the first, broad spatial synthesis of lake water balance status across lake-rich permafrost landscapes of North America.

Study areas
The five study regions (Fig. 1) selected for this synthesis collectively span broad gradients in permafrost, catchment vegetation, and meteorological conditions and contain abundant thermokarst lakes that have been previously sampled and analyzed for water isotope composition.The Arctic Coastal Plain (ACP) north of the Brooks Range in Alaska including lands between Barrow and Prudhoe Bay contains abundant shallow lakes mainly of thermokarst origin, is underlain by continuous permafrost, and contains tundra vegetation (Arp and Jones 2009).The Yukon Flats (YF) spans ~118 000 km 2 and is set along the Yukon River floodplain and its terraces south of the Brooks Range in Alaska.This lowland interior landscape is located within the zone of discontinuous permafrost and contains over 40 000 lakes of thermokarst, fluvial, and eolian origin (Williams 1962;Arp and Jones 2009).Catchment vegetation includes grassy meadows and muskeg to spruce and birch forests (Anderson et al. 2013).Old Crow Flats (OCF) spans 5600 km 2 and is situated ~55 km north of the village of Old Crow in northern Yukon Territory.This low-relief landscape is located within an area of continuous permafrost and contains over 2700 shallow primarily thermokarst lakes (Lauriol et al. 2002;Turner et al. 2014).Vegetation in OCF is variable and captures a gradient from spruce forest to tall shrubs to tundra vegetation (Turner et al. 2014).The western Hudson Bay Lowlands (HBL) spans 475 000 km 2 and contains over 10 000 shallow mainly thermokarst lakes.The HBL is underlain by discontinuous permafrost in the southwest and continuous permafrost in the northeast.Vegetation ranges from boreal spruce forest in the southwest to arctic tundra in the northeast (Rouse 1991;Duguay and Lafleur 2003).Nunavik (NUN), located north of the 55°parallel along the eastern coast of Hudson Bay in northern Quebec, contains abundant thermokarst lakes.Permafrost ranges from sporadic in the south to discontinuous in the north (Allard and Séguin 1987;Brown et al. 2002).Vegetation is mainly spruce-lichen forest in the south and shrub tundra in the north.
A common gridded climate database was used to compile regional meteorological records for comparative purposes, to provide necessary parameters for water isotope mass balance modeling, and to gain insight of meteorological influence on lake water balances.The New et al. (2002) gridded climate database was selected due to the availability of lakespecific meteorological data and the ease of use of the database for a large data set, even though it predates our sampling intervals.Mean annual, summer, and winter temperatures and precipitation vary substantially among the five landscapes, based on mean monthly values for 1961-1990 (Fig. 2).Mean annual temperature ranges from −10.5 °C (ACP) to −5.3 °C (NUN) and annual precipitation ranges from 141 mm (ACP) to 580 mm (NUN) (Fig. 2a).ACP and OCF have lower mean annual temperature and precipitation than the other landscapes.YF has relatively low mean annual precipitation but high mean annual temperature, while HBL and NUN have relatively high mean annual temperature and precipitation.Similar patterns exist for mean winter temperature and winter precipitation, with mean winter temperature ranging from −19.2 °C (OCF) to −16.2 °C (NUN) and winter precipitation ranging from 53 mm (ACP) to 182 mm (NUN) (Fig. 2b).Mean summer temperature ranges from 5.6 °C (NUN) to 9.8 °C (YF) and summer precipitation ranges from 88 mm (ACP) to 399 mm (NUN) (Fig. 2c).Compared to mean annual and winter meteorological data, similar patterns for ACP, YF, and OCF are evident for summer temperature and summer precipitation.However, HBL has a more midrange mean summer temperature and NUN has the lowest mean summer temperature.

Isotope hydrology
We assembled water isotope compositions (δ 2 H and δ 18 O) for 376 lakes sampled during the past decade in the five study regions.Forty-four lakes were sampled during August 2015 in ACP, 149 lakes were sampled once during the summer between 2007 and 2011 in YF (Anderson et al. 2013), 53 lakes were sampled each summer from 2007 to 2009 and four additional lakes were sampled in 2007 and 2009 in OCF (Turner et al. 2010(Turner et al. , 2014)), 40 lakes were sampled in the summer of 2010 in HBL and 37 lakes were sampled in the summers of 2011-2012 (Bouchard et al. 2013), and 86 lakes were sampled from one to four times during summers 2011-2014 in NUN (Narancic et al. 2017).Samples were collected at 10-15 cm water depth in either 30 mL high density polyethylene bottles or 20 mL scintillation vials with plastic cone-shaped caps.Samples were transported back to the field base and then shipped to the Alaska Stable Isotope Facility at the University of Alaska Fairbanks (ACP), University of Arizona Environmental Isotope Laboratory (YF), or the University of Waterloo Environmental Isotope Laboratory (OCF, HBL, and NUN) for determination of hydrogen and oxygen isotope compositions using standard mass spectrometric techniques (Epstein and Mayeda 1953;Morrison et al. 2001), with the exception of NUN samples collected in 2014, which were analyzed using Off-Axis Integrated Cavity Output Spectroscopy.Isotope composition results are reported in δ notation, which represents deviations in per mil from Vienna Standard Mean Ocean Water (VSMOW) and are normalized to −428‰ and −55.5‰ for δ 2 H and δ 18 O, respectively, for Standard Light Antarctic Precipitation (Coplen 1996).We restricted our analysis to July and August sample collection time periods to reduce seasonal effects caused by the influence of snowmelt while also capturing the expected midsummer peak in evaporation.For lakes that were sampled more than once per summer (July and August), or over multiple summers, we used the average value in our analyses.

Isotope framework development
Raw water isotope compositions were initially assessed in conventional δ 18 O-δ 2 H space, superimposed upon an "isotope framework" consisting of the Global Meteoric Water Line (GMWL) and the Local Evaporation Line (LEL) predicted for each landscape (Fig. 3).The GMWL, described by δ 2 H = 8δ 18 O + 10 (Craig 1961), reflects the isotopic distribution of global precipitation.The position of amount-weighted precipitation along the GMWL is mainly dependent on the distillation history of atmospheric moisture contributing to precipitation and commonly leads to snow plotting along an isotopically depleted portion of the GMWL relative to rain (Fig. 3).Surface water isotope compositions, including lakes, typically plot along a LEL, which generally has a slope of 4-6 (Fig. 3).The LEL for a given landscape, as applied in this context, defines the expected isotopic evolution of a surface waterbody undergoing evaporation, fed by waters representing the average annual isotope composition of precipitation (δ P ) for that region.Displacement of water compositions along the LEL from δ P reflects evaporative loss, while deviation from the LEL is often indicative of mixing with source waters such as snowmelt or rainfall, which tend to plot along the GMWL (Fig. 3).Key reference points along the LEL include the terminal (i.e., closed-drainage) basin steady-state isotope composition (δ SSL ), which represents the special case of a waterbody at hydrologic and isotopic steady-state in which evaporation exactly equals inflow and the limiting non-steady-state isotope composition (δ*), which indicates the maximum potential transient isotopic enrichment of a waterbody as it approaches complete desiccation (Fig. 3).
For each landscape, the LEL was predicted using the linear resistance model of Craig and Gordon (1965) following similar approaches presented in Brock et al. (2007) and Wolfe et al. (2011).Hereafter, we refer to this as the "landscape-predicted LEL." Predicting the LEL, rather than the more commonly used empirical technique of applying linear regression through measured lake water isotope compositions, allows lake water isotope compositions to be interpreted independently based on their position along (degree of evaporation) and about (i.e., above/below; relative influence of different input waters such as snowmelt and rainfall) the LEL (e.g., see Tondu et al. 2013;Turner et al. 2014).
The following equations were used to develop the landscape-predicted LELs and are expressed in decimal notation.The equilibrium liquid-vapour fractionation factors (α*) Fig. 3. Schematic δ 18 O-δ 2 H diagram illustrating an approach for the interpretation of lake water isotope data within a region.Key features include the Global Meteoric Water Line (GMWL), the landscape-predicted Local Evaporation Line (LEL), average annual isotope composition of precipitation (δ P ), the terminal basin steadystate isotope composition (δ SSL ), the limiting non-steady-state isotope composition (δ*), lake water isotope composition (δ L ), input water isotope composition (δ I ), and the isotope composition of evaporated vapour from the lake (δ E ).
for oxygen and hydrogen are dependent on temperature and have been determined empirically by Horita and Wesolowski (1994), where for δ 2 H, where T represents the interface temperature in K. ε* is the temperature-dependent equilibrium separation between liquid and vapour water given by ð4Þ e Ã ¼ a Ã À 1 and kinetic separation (ε K ) is expressed by where constant enrichment values (C K ) for oxygen and hydrogen are 0.0142 and 0.0125, respectively, and h is relative humidity (Gonfiantini 1986).δ AS is the isotope composition of ambient open-water season atmospheric moisture, often assumed to be in isotopic equilibrium with evaporation-flux-weighted local open-water season precipitation (δ PS ) (Gibson et al. 2008) such that The limiting isotopic enrichment of a waterbody approaching desiccation (δ*) has been defined by Gonfiantini (1986) and can be determined from δ SSL represents the isotope composition of a terminal basin in which evaporation is exactly compensated by inflow, as defined by Gonfiantini (1986) where the isotope composition of inflow, δ I , is assumed to be equal to δ P .The landscapepredicted LEL was determined by linear regression through δ P and δ*.

Water-balance metrics
The water balance metrics, δ I and E/I ratios, were determined for each of the 376 lakes using the Yi et al. (2008) coupled-isotope tracer approach, which assumes conservation of mass and isotope composition during evaporation.According to mass conservation, the isotope composition of evaporated vapour from a lake (δ E ) will lie on the extension of the lake-specific LEL to the left of the GMWL (Fig. 3) and was determined from the formulation provided by Gonfiantini (1986), where δ L is the measured lake water isotope composition: Values for δ I were derived from calculating lake-specific evaporation lines and their intersection with the GMWL, which reasonably assumes that input waters are nonevaporated and plot on the GMWL and that all lake-specific evaporation lines converge at δ* (Yi et al. 2008) (Fig. 3).The relative contributions of rainfall and snowmelt were then assessed by evaluating the position of δ I compared to the landscape value of δ P along the GMWL.For example, δ I values that were more isotopically enriched than δ P were categorized as rainfall-dominated lakes and δ I values that were more isotopically depleted than δ P were categorized as snowmelt-dominated lakes.For some YF lakes, very low δ I values are interpreted as lakes fed primarily by permafrost thaw waters (see below and Anderson et al. 2013).E/I ratios, which provide a snapshot of water balance through the mass-balance relation of evaporation to inflow, were calculated from Gibson and Edwards (2002): An E/I ratio of 0.5 represents lakes where 50% of the inflow has evaporated, and we use this threshold to define evaporation-dominated lakes (after Tondu et al. 2013).As applied here, E/I ratios estimate net evaporative loss in midsummer and can indicate whether lake water volumes are increasing (E/I << 1) or decreasing (E/I > 1) where no drainage outlet exists.This approach assumes a well-mixed lake at isotopic steady-state; thus, values greater than 1 are inconsistent with the assumptions in the model but are used comparatively to identify lakes strongly influenced by evaporation.
Model input climate parameters, T and h, for calculation of the landscape-predicted LELs and lake water balance metrics were derived from the New et al. (2002) gridded climate database, which provided output for individual lake coordinates.This approach was used in the isotope mass-balance modeling of the individual lakes to account for spatial gradients in meteorological conditions within and among landscapes.Monthly T and h averages for the open-water season were flux-weighted according to potential evaporation using Thornthwaite (1948) for each landscape and for each of the 376 lakes.Values for δ P (to anchor the landscape-predicted LEL) and δ PS (used to determine δ As ; eq. 6 for both the landscape-predicted LEL and each individual lake to account for spatial variations) were extracted from "The online isotopes in precipitation calculator" (waterisotopes.org;Bowen 2016).This database uses global precipitation oxygen and hydrogen isotope data to calculate average monthly and annual δ P values for any given location and elevation (Bowen et al. 2005).Sampling year(s) meteorological conditions (temperature, relative humidity, and precipitation) for a representative location from each landscape were extracted from the NCEP North American Regional Reanalysis (NARR 2015) monthly composites and compared with the 1961-1990 landscape averages from the New et al. (2002) gridded database to assess the representativeness of meteorological conditions during the specific sampling years.
The influence of catchment vegetation on the water-balance metrics was assessed after land cover for each lake was broadly classified as tundra dominant or forest dominant.Tundra-dominant vegetation included catchments with high proportions of dwarf shrubs and areas of sparse vegetation, while forest-dominant vegetation included lake catchments with high proportions of deciduous and coniferous woodland or forest and tall shrub vegetation.Vegetation classes for ACP and YF were determined using the USGS National Land Cover Database of Alaska.For OCF, vegetation classes were simplified based on analysis of a Landsat 5 TM mosaic (Turner et al. 2014).Vegetation type for HBL and NUN was identified based on visual observations during field work.
Nonparametric Kruskal-Wallis statistical tests were conducted to assess whether E/I distributions differed among lakes in different permafrost zones (continuous, discontinuous, and sporadic) and between vegetation categories (forest dominant versus tundra dominant) and whether δ 18 O I values differed among lakes in the different vegetation categories.When Kruskal-Wallis tests involving the permafrost zones produced a significant result (P ≤ 0.05), pairwise comparisons were conducted using Dunnett's post hoc tests.All statistical tests were performed using the software SPSS version 20.E/I values for lakes that were evaporating under strongly non-steady-state conditions (E/I > 1) were set to 1.5 for boxplot analyses and the statistical tests.

Meteorological conditions during sampling years
Comparison of specific sampling year meteorological conditions (NARR) with the 1961-1990 average values (New et al. 2002) for each landscape reveals some similarities and differences (Fig. 4).Summer temperatures were higher (1.8-3.3 °C) during sampling years for YF, OCF, HBL, and NUN landscapes than the 1961-1990 average.At ACP, the summer temperatures were lower (0.9 °C) than the 1961-1990 average.Humidity values for the sampling years were similar to the 1961-1990 averages at all five landscapes.Precipitation shows the greatest difference between the values for the sampling years versus the 1961-1990 average.During the sampling years, ACP, YF, and OCF had consistently higher summer (54-121 mm) and winter precipitation (78-115 mm) than the 1961-1990 average, while HBL had higher summer precipitation (115 mm) during sampling years when compared to the 1961-1990 averages.In contrast, NUN had lower summer precipitation (32 mm) and higher winter precipitation (119 mm) during sampling years when compared to the 1961-1990 averages.

Isotope hydrology
Lake water isotope compositions (δ L ) Lake water isotope compositions (δ L ) from all of the assembled data range from −20.5‰ to −2.4‰ and from −168.7‰ to −53.0‰ for δ 18 O and δ 2 H, respectively (Appendix; Fig. 5).The wide range of δ L values reflects the diverse lake hydrological conditions at the time of sampling in these high-latitude regions.YF has the greatest range of δ L values, indicating substantial within-landscape variability, while NUN has the smallest range, signifying that lakes possess a narrower range of hydrological conditions in this landscape.For each of the five landscapes, δ L values form a linear trend that typically plot along a similar trajectory as the landscape-predicted LELs, supporting the contention that the frameworks are reasonable approximations of isotopic evaporative trajectories (Fig. 5).Indeed, the landscape-predicted LELs are in close agreement with the empirically defined LELs, except for ACP (Table 1).For ACP, δ L values plot along a trajectory with a somewhat steeper slope than the landscape-predicted LEL, likely due to high rainfall immediately prior to sampling (Fig. 4).The LELs and δ L values for OCF and YF, and HBL and NUN, are positioned in similar δ 18 O-δ 2 H space, likely reflecting similar latitudes and the associated well-known effect on isotope composition of precipitation (Rozanski et al. 1993).ACP, the most northerly landscape, does not follow this pattern, perhaps due to its closer proximity to the Arctic coast and associated reduced continental influence on precipitation isotope composition.δ L values from NUN are evenly distributed about the LEL, while δ L values from HBL and ACP typically plot above their respective LELs, suggesting a stronger influence of rainfall compared to snowmelt on lake water balance.Conversely, δ L values from YF and OCF generally plot below their respective LELs, reflecting a stronger influence of snowmelt compared to rainfall on water balances.Additionally, a small group of lakes (n = 15) from YF have δ L values that plot on a particularly low trajectory (i.e., parallel to, but offset below, the landscape-predicted LEL), which Anderson et al. (2013) suggested reflect more dominant input by isotopically depleted water from permafrost thaw in this region (elaborated on in the next section).δ L values from HBL are positioned farthest away from the GMWL on the LEL, with many lakes plotting beyond δ SSL and some approaching and surpassing the landscape-predicted δ*, indicating strong non-steady-state evaporative isotopic enrichment at the time of sampling.In contrast, δ L values from NUN are positioned closest to the GMWL on the LEL, indicating that lakes in this landscape are least influenced by evaporation.

Source water identification (δ I )
The isotope composition of lake-specific input water (δ I ) was calculated for each lake in the five landscapes to evaluate the relative roles of different source waters on lake hydrological conditions (Fig. 6).Lake-specific δ 18 O I values range from −38.6‰ to −7.2‰ and lakespecific δ 2 H I values range from −298.4‰ to −47.7‰.The large range in δ I values illustrates   For each landscape, δ I values were compared with the mean annual isotope composition of precipitation value (δ P ) to classify lakes as snowmelt (δ I ≤ δ P ) versus rainfall (δ I > δ P ) dominated (Fig. 6).YF and OCF have the highest proportions of snowmelt-dominated lakes, 89% and 72%, respectively, indicating the strong influence of snowmelt on lake water balances in these landscapes, even during midsummer sampling.Of note, there was a small group of lakes in YF with particularly low δ I values, likely due to input from snowmelt and permafrost thaw (Anderson et al. 2013).YF is underlain by discontinuous permafrost, and the observed values were within the range of values for permafrost thaw waters in this area (Meyer et al. 2010;Lachniet et al. 2012;Anderson et al. 2013).Slightly more than half of the lakes (52%) in NUN are snowmelt dominated, indicating a more even distribution of snowmelt versus rainfall source waters throughout the landscape.Some rainfalldominated lakes in NUN may also be fed by permafrost thaw waters (Narancic et al. 2017).Rainfall-dominated lakes are the overwhelming majority in HBL (80%) and ACP (91%), reflecting the strong influence of rainfall on lake water balances in these landscapes at the time of sampling.

Evaporation-to-inflow (E/I) estimates
Evaporation-to-inflow ratios (E/I) were calculated for each lake in the five different landscapes to evaluate the relative importance of vapour loss on lake hydrological conditions (Fig. 7).The 376 lakes span a wide spectrum of E/I values, from close to 0 to much greater than 1, illustrating a range of water balances from those dominated by input waters to those dominated by evaporation.Overall, 219 lakes (58%) have E/I < 0.5, while 157 lakes (42%) have E/I > 0.5 (i.e., >50% evaporative water loss), which we consider as evaporation dominated.Calculated E/I distributions vary among landscapes (Fig. 7).For NUN and ACP, the vast majority of lakes (95% and 98%, respectively) have E/I < 0.5.Lakes in OCF have a relatively even distribution with E/I < 0.5 in 46% of lakes.In contrast, a majority of lakes in YF and HBL have E/I > 0.5 (63% and 68%, respectively), and these two regions have the largest proportion of lakes with E/I > 1 (30% and 40%, respectively).In HBL and YF, E/I > 1 is consistent with field observations of lakes throughout the landscape having undergone desiccation by midsummer (Anderson et al. 2013;Bouchard et al. 2013).

Discussion
Thermokarst lakes have been undergoing hydrological transitions in response to recent climate change (e.g., Yoshikawa and Hinzman 2003;Smith et al. 2005;Riordan et al. 2006;Labrecque et al. 2009;Rowland et al. 2010;Carroll et al. 2011;Bouchard et al. 2013).Our analysis of water isotope compositions and calculations of δ I and E/I ratios for 376 lakes at five lake-rich permafrost landscapes (ACP, YF, OCF, HBL, and NUN) in arctic and subarctic North America indicate that the importance of input types (rainfall, snowmelt, and permafrost) and evaporation are highly variable.Results show that striking similarities and differences in thermokarst lake hydrology exist among landscapes.Large gradients in δ I occur within and among landscapes and identify that lakes in HBL and ACP are mainly rainfall dominated, whereas lakes in OCF and YF are mainly snowfall dominated.Lakes in NUN have roughly equal proportions of rainfall-and snowfall-dominated lakes.Snowfall-dominated lakes from YF also likely include lakes with substantial contributions from permafrost thaw water (and possibly also in NUN, although these are isotopically indistinguishable from rainfall-dominated lakes; Narancic et al. 2017).E/I values span from almost 0 to much greater than 1.Most lakes in ACP and NUN have E/I < 0.5, while the majority of lakes in YF and HBL are evaporation dominated despite higher-than-normal  precipitation during sampling years.These findings underscore the strong hydrological gradients that exist across thermokarst lakes from high-latitude regions.In the discussion below, we first acknowledge assumptions and uncertainties in the isotope modeling approach.Then, relations of δ I and E/I ratios with climate and catchment characteristics among the five landscapes are explored, which provide the basis for anticipating how thermokarst lake hydrology in these northern regions may change in the future.

Assumptions and uncertainties
The nature of this broad continental-scale meta-analysis necessarily assumes that lakes sampled are representative of their landscapes and required decisions to ensure a consistent modeling approach given availability of existing data.Water balance metrics derived in this study were calculated from a single lake water isotope measurement or an average of July and August lake water isotope measurements over multiple years and thus they represent a snapshot of conditions.Also, the specific sampling years varied among the five landscapes.Although comparing water isotope data from different years for the five landscapes may result in some inherent variability, it is unlikely that the interannual variability for a single lake would exert a strong influence on comparisons within and among the five landscapes given the large range of lake water isotope compositions and E/I and δ I values across the landscapes.We explored this for landscapes where multiple years of summer water isotope measurements were available (OCF, HBL, and NUN), and indeed, spatial variability far exceeded annual summer variability of individual lakes.For OCF, the range in δ 18 O and δ 2 H values for all lakes was 11.9‰ (minimum = −21.0‰,maximum = −9.1‰)and 64.6‰ (minimum = −172.3‰,maximum = −107.7‰),respectively.In contrast with the large spatial variability, the greatest range for an individual lake in OCF over the 3 year sampling period was 2.3‰ and 11.9‰ for δ 18 O and δ 2 H, respectively.For HBL, the range in δ 18 O and δ 2 H values for all lakes was 10.6‰ (minimum = −12.0‰,maximum = −1.4‰)and 51.5‰ (minimum = −100.3‰,maximum = −48.8‰),respectively, whereas the greatest range for an individual lake in HBL over the 3 year sampling period was much lower (4.7‰ and 20.5‰ for δ 18 O and δ 2 H, respectively).Similarly for NUN, the range in δ 18 O and δ 2 H values for all lakes was 6.6‰ (minimum = −14.4‰,maximum = −7.8‰)and 35.2‰ (minimum = −107.7,maximum = −72.5),respectively, while the greatest range for an individual lake over multiple years was much lower (2.8‰ and 19.8‰ for δ 18 O and δ 2 H, respectively).
The availability and quality of climate records also varied among the five landscapes, and we used a common gridded climate database to extract meteorological conditions.These data were used to calculate water balance metrics for each individual lake, which allowed for a consistent approach to modeling of all lakes.However, this also added some uncertainty to the model output given that the gridded data set estimates a 30-year average , which was used to represent meteorological conditions during the recent years of actual water sampling.Fortunately, the gridded 30-year averages for humidity were well aligned with sample year humidity, which is a parameter that the isotope-mass balance model is sensitive to.Yet, precipitation during the sampling years was generally higher than the 1961-1990 averages (Fig. 4).Relatively wet conditions may have led to an underestimation of some of the E/I values relative to expected long-term averages, particularly for ACP.Additionally, summer temperature was warmer during the sampling years than the 1961-1990 estimates, with the exception of ACP.Different data sources were used to demarcate catchment vegetation among landscapes (field observations, remote sensing, aerial photographs), which also result in some further uncertainty to comparisons we make below.
Our attempt to develop a consistent modeling approach that could be applied to all lakes and landscapes results in some differences in values presented in this paper compared to the previous landscape-specific studies.For example, estimates of δ P produced using waterisotopes.org(Bowen 2016) were lower than local precipitation isotope data utilized by Narancic et al. (2017), which placed some lakes in different classifications (snowmelt-versus rainfall-dominated categories).However, both approaches robustly identify that lakes in NUN experience a low degree of evaporation.Assumptions and limitations of data availability were unavoidable, but they are more likely to influence individual lake behaviour than the large-scale spatial patterns within and among landscapes (the primary aim of this paper), which clearly emerged.

Drivers of hydrological conditions
Meteorological conditions exert a strong influence on water balance of thermokarst lakes (e.g., Riordan et al. 2006;Plug et al. 2008;Labrecque et al. 2009).For temperature and precipitation, mean annual, mean winter, and mean summer values vary greatly among the five landscapes (Fig. 2).Previous water isotope studies of lakes in northern Canada and the continental United States (Gibson and Edwards 2002;Brooks et al. 2014) found that colder regions typically have lower E/I values compared to warmer regions.This is likely in response to more rapid evaporation at higher temperature and perhaps differences in the length of the open-water season.Variation in ice-out timing within a region due to lake morphometry and among years and regions due to spring temperatures can also strongly affect evaporation season duration (Arp et al. 2015).Based on differences in mean summer temperature of the five landscapes in this study, one might anticipate the lowest E/I values at ACP and NUN and the highest values at OCF and YF.Indeed, lakes in ACP and NUN have the lowest E/I values and YF has some of the highest E/I values, but lakes in OCF have more moderate E/I values (Fig. 7).However, HBL has a much higher percentage (40%) of lakes with E/I > 1 compared to OCF (4%) and YF (30%).
The amount of snowmelt and rainfall input to lakes (direct to the lake surface and via runoff) affects the water balance of thermokarst lakes through the degree of water replenishment that offsets evaporative losses (Schindler and Smol 2006;Bouchard et al. 2013).It may be anticipated that YF has the greatest proportion of lakes with E/I > 1, owing to higher temperatures and relatively low mean annual winter and summer precipitation available to offset evaporation.In contrast, NUN was expected to have the lowest E/I values because it has the lowest mean summer temperature and highest mean winter and summer precipitation.In general, the results are consistent with these expectations; 30% of lakes in YF have E/I values >1 and 95% of lakes in NUN have E/I values <0.5.However, HBL, with moderate temperature and precipitation, has the overall greatest proportion of lakes with E/I > 1 (40%).Thus, although HBL has the second highest mean annual summer and winter precipitation relative to the other landscapes, precipitation inputs do not offset midsummer evaporative losses for many lakes compared to the other landscapes, evidently even during years of apparent high summer precipitation.Bouchard et al. (2013) came to a similar conclusion that many lakes in HBL do not receive adequate precipitation, particularly snowmelt runoff, to offset midsummer evaporation leading to lake level decline.Snowmelt bypass, which occurs when snowmelt passes through a lake basin while the water mass is still frozen as ice, has been observed in some arctic lakes (e.g., Bergmann and Welch 1985) and may also serve to enhance E/I ratios in the absence of diluting effects of rainfall (Edwards and McAndrews 1989).
Source waters to lakes in both HBL and ACP were dominated by rainfall at the time of sampling (Fig. 6), but there is a large difference in amount of mean summer precipitation (Fig. 2).Similarly, lakes in OCF, YF, and NUN have snowmelt-dominated source waters, but again, these landscapes differ strongly in their mean winter precipitation.Thus, factors other than seasonal precipitation amounts must play a role in the relative importance of rainfall versus snowmelt inputs to thermokarst lakes in these landscapes.Interestingly, Fig. 2 shows that YF has relatively high temperature and low precipitation compared to the pattern observed for the other landscapes.YF is also the only landscape with lakes with input isotope compositions distinctly characteristic of water from permafrost thaw (Anderson et al. 2013).Higher temperatures since the early 1980s may be promoting more intense permafrost thaw in YF (Anderson et al. 2013).Overall, the data suggest that climate normals are not the best predictor of hydrological classification of thermokarst lakes when used alone.
Permafrost conditions, which are influenced by climate, affect surface area of thermokarst lakes throughout the Arctic and Subarctic.For example, studies have shown that lake surface area is decreasing in regions of discontinuous permafrost (Yoshikawa and Hinzman 2003;Smith et al. 2005) but expanding in areas of continuous permafrost (Smith et al. 2005).Decreasing lake surface area has largely been attributed to drainage, but increased evaporation in response to climate warming may also play a role (Riordan et al. 2006).The five landscapes in this study span permafrost classifications from sporadic to continuous as well as from water balances indicative of increasing or stable lake water volume (E/I < 0.5) to water balances indicative of decreasing lake water volume due to evaporation (E/I > 1).E/I values for lakes from the three permafrost categories are highly variable (Table 2; Fig. 8).Lakes in terrain with sporadic permafrost (NUN) have the lowest E/I values, whereas lakes in terrain with discontinuous permafrost (YF, HBL, and NUN) have the highest E/I values.However, lakes from regions classified as having continuous (ACP, OCF, and HBL) and discontinuous (YF, HBL, and NUN) permafrost do have a wide and comparable range of E/I values spanning from close to 0 to greater than 1.Thus, relations among permafrost zones, lake surface area, and lake water balance are not straightforward.
Lake surface area and depth, as imparted by permafrost or other factors, can influence lake water balance.For example, in ACP, Arp et al. (2015) identified that lakes tend to experience longer ice-free seasons if they are shallow enough to have bedfast ice.In YF, Anderson et al. (2013) proposed that lakes with high E/I values are more likely to be relatively shallow.Although specific lake depth measurements were not available for the entire data set in this study, lakes in HBL were by far the shallowest of the five landscapes and, analogous to observations of Anderson et al. (2013) for YF, had the highest E/I ratios.
Of the permafrost categories, lakes located in the discontinuous permafrost zone, where average temperatures are warmer, have the highest proportion classified as evaporation dominated (E/I > 0.5) and 25% had E/I > 1.This suggests that evaporation in response to climate warming is likely playing an important role in the observed decline of surface area of thermokarst lakes in discontinuous permafrost zones and that lake drainage (lateral or internal) is likely not the sole cause.The dominance of low E/I values (<0.5) in lakes located within the region of sporadic permafrost may be due to ground thaw, which allows increased lateral hydrological connectivity to offset effects of evaporation.Permafrost thaw and the subsequent increased lateral hydrologic connectivity have been shown to maintain positive lake water balances (low E/I values) in Churchill, Manitoba (Wolfe et al. 2011).Overall, E/I results suggest that large-scale predictions of changes in lake area based strictly on permafrost zonation throughout the Arctic and Subarctic likely would not account for the apparent spatial heterogeneity in thermokarst lake hydrological conditions.Additionally, our analysis suggests that lake drainage is not the only cause of lake level decline for thermokarst lakes in discontinuous permafrost zones and that increased evaporation associated with air temperature increase is likely playing an important role in observed water level changes.
Studies from northern regions have suggested that lakes in low-relief, tundra landscapes are more vulnerable to evaporative losses and desiccation than lakes in forested landscapes (e.g., Brock et al. 2009;Turner et al. 2010Turner et al. , 2014;;Bouchard et al. 2013).In forested landscapes, taller and denser vegetation entraps greater amounts of wind-redistributed snow than areas of sparse tundra vegetation (Pomeroy et al. 1997;Liston and Sturm 1998;McFadden et al. 2001;Sturm et al. 2001;Brock et al. 2009).In spring, snowmelt runoff to lakes helps to offset evaporative losses throughout the summer.Based on these observations, it could be reasoned that in this study, lakes located in forest-dominant catchments should have lower E/I and δ I values than lakes located in tundra-dominant catchments among the five landscapes.Results show that lakes from HBL display the clearest separation of E/I values between the two catchment vegetation classes with tundra-dominant catchments having higher E/I values followed by OCF, while lakes from YF have more similar ranges of observed E/I values for both vegetation classes (Table 2; Fig. 9).Additionally, more lakes in OCF, HBL, and YF have E/I > 1 in tundra-dominant landscapes compared to lakes situated in forest-dominant catchments.In fact, YF is the only landscape that has forest-dominant lake catchments with E/I > 1.In contrast, lakes from NUN do not follow this pattern (Fig. 9).Within this landscape, E/I values span similar albeit low ranges for lakes from both vegetation classes, but E/I ratios are significantly higher in lakes with forest-dominant catchments compared to lakes with tundra-dominant catchments (Table 2; Fig. 9).For ACP, all E/I values are relatively low despite that all lakes are situated in tundra-dominant catchments.Results also show that lakes with tundra-dominant catchments in YF, OCF, HBL, and NUN all had higher median δ 18 O I values compared to lakes in these landscapes situated in  forest-dominant catchments (Table 2; Fig. 9).Thus, tundra-dominant catchments appear to favour greater relative input of rainfall than snowmelt source waters to lakes.Overall, the data suggest that while vegetation appears to influence the composition of thermokarst lake input waters in YF, OCF, HBL, and NUN, the role of vegetation in vapour loss appears to be more important in HBL and OCF, and to a lesser degree in YF, than in NUN.
Interactions among meteorological conditions and catchment characteristics, such as vegetation and permafrost classifications, likely play key roles in promoting the similarities and diversity of hydrological conditions observed among the five landscapes.For example, variability in year-to-year meteorological conditions likely has the ability to mask the expected lake responses to other drivers such as vegetation and permafrost characteristics.ACP lakes all had E/I values that were relatively low despite all lakes being situated in catchments dominated by tundra vegetation.These values may have been lower than expected due to relatively high rainfall prior to sampling compared to long-term averages (Fig. 4).We speculate that in years of more typical precipitation in ACP, catchment vegetation may play a larger role in thermokarst lake hydrology and E/I values may be higher and perhaps more comparable to the tundra landscapes observed in HBL, OCF, and YF.Conversely, catchment vegetation may also mediate changes in meteorological conditions.For example, thermokarst lakes in HBL and ACP are mainly rainfall dominated but are also coastal landscapes with the majority of lakes located within open tundra.The coastal tundra settings may promote more wind redistribution of snowfall in these landscapes compared to the more inland snowmelt-dominated landscapes of YF and OCF, perhaps causing the lakes in these tundra landscapes to be more susceptible to hydrological changes in response to yearly fluctuations in rainfall.We further contend that because permafrost and overlying vegetation are influenced by climate conditions, precisely identifying discrete roles of permafrost and vegetation is difficult.For instance, within a landscape, water balance differences may be due to the climatic conditions that result in vegetation differences rather than caused solely by vegetation.Recognition of the complex interactions and relative importance among different drivers of thermokarst lake hydrology throughout high-latitude regions is required to anticipate future hydrological trajectories (Turner et al. 2014).

Future hydrological trajectories
During the next century, northern regions are expected to experience continued rise of air temperature, longer duration of the ice-free season, and changes in the amount and timing of precipitation (Kattsov et al. 2005;Prowse et al. 2006;AMAP 2011).Increased temperatures and longer ice-free seasons will promote greater vapour loss from lakes during summer (Schindler and Smol 2006;Arp et al. 2015), leading to increased E/I values.If increases in precipitation do not occur at a similar rate, this could cause widespread desiccation of thermokarst lakes (Bouchard et al. 2013), which has also been observed in shallow non-thermokarst lakes in Canada's High Arctic (Smol and Douglas 2007).Spring snow cover has declined over many areas of northern North America and this pattern is expected to continue, although with substantial spatial and temporal variability (AMAP 2011;Dersken and Brown 2012;Krasting et al. 2013), which may result in a reduction of runoff available for offsetting vapour loss.Thermokarst lakes in HBL have already begun to desiccate during the ice-free season, and analysis of a sediment core from one desiccated lake in HBL indicates that this recent drying trend is unprecedented in the context of the past 200 years (Bouchard et al. 2013).YF, OCF, and perhaps ACP may also evolve towards this scenario under conditions of continued climate warming.Based on the E/I results of this study, field observations, and the degree to which hydrological conditions in each landscape appear to be influenced by meteorological conditions as outlined above, we suggest that HBL is the most vulnerable of the five landscapes to widespread lake desiccation in the future followed by YF, OCF, and ACP, while NUN is likely the least vulnerable.Interestingly, the landscapes at the two ends of this lake hydrological spectrum lie on opposite sides of Hudson Bay, and this may be related to the more maritime conditions in NUN on the eastern shore (Fig. 2; Narancic et al. 2017).
Increases in shrub growth in response to longer ice-free seasons and warmer temperatures have been observed along tundra-taiga transition zones (Myers-Smith et al. 2011;Lantz et al. 2013).Increased shrub growth may result in an increase in the number of lakes having snowmelt-dominated input waters and, conversely, a decrease in the proportion of rainfall-dominated lakes.This increase may result in greater water replenishment for some lakes in HBL, OCF, YF, and possibly ACP, where tundra-dominated landscapes typically have higher E/I values.However, the ratio of catchment area to lake size of individual lakes will determine whether sufficient snowmelt runoff can be generated to offset evaporative losses.Furthermore, with more vegetation productivity, increases in terrestrial evapotranspiration may dampen this response.
Greater permafrost thaw throughout high-latitude regions of North America (Osterkamp and Romanovsky 1999;Burn and Kokelj 2009) may result in lake-level declines via increases in vertical lake drainage (e.g., Yoshikawa and Hinzman 2003), or it may result in increased lateral hydrological connectivity, which may offset water losses due to evaporation and vertical drainage, ultimately causing a net increase in lake surface area (Avis et al. 2011;Wolfe et al. 2011).However, previous studies showed that lakes in YF with hydrological connections to the drainage network tend to experience greater fluctuations in intra-and interannual water balances (Chen et al. 2012(Chen et al. , 2013)).A subset of thermokarst lakes in YF show evidence of source waters derived from permafrost thaw, suggesting that this landscape may be particularly sensitive to further changes in permafrost.E/I values of lakes in the sporadic permafrost zone of NUN may also be illustrating the effects of increased hydrological connectivity offsetting vapour loss.As the continuous permafrost warms in ACP, HBL, and OCF, these lakes may also become increasingly influenced by permafrost thaw waters.Overall, thermokarst lakes throughout permafrost regions of North America are unlikely to follow a uniform hydrological trajectory in response to amplified climate change.Rather, the hydrology of thermokarst lakes is likely to display dynamic and individualistic responses depending on their unique set of landscape and climate conditions and drivers.

Conclusions
We compiled water isotope data obtained during the past decade from 376 lakes of mainly thermokarst origin situated in arctic and subarctic permafrost landscapes across North America (Arctic Coastal Plain (Alaska), Yukon Flats (Alaska), Old Crow Flats (Yukon), northwestern Hudson Bay Lowlands (Manitoba), and Nunavik (Quebec)).Our results, as well as those derived from calculation of isotope-based water-balance metrics (including source water isotope compositions and evaporation-to-inflow ratios), demonstrate a substantial array of regional and subregional diversity of lake hydrological conditions characterized by varying influence of snowmelt, rainfall, permafrost thaw waters, and evaporation.Thermokarst lake hydrology is driven by complex interactions among prevailing temperature and precipitation, catchment vegetation, and permafrost status.Some regional patterns emerged, such as the strong role of open-water evaporation on thermokarst lakes of the Hudson Bay Lowlands and Yukon Flats in particular, yet these hydrological drivers are all "moving targets" with ongoing climate change.Thus, they are likely to have pronounced influence on future thermokarst lake hydrological trajectories at a wide range of spatial and temporal scales, challenging our ability to anticipate their consequences for water resources, aquatic ecosystems, and biogeochemical cycling.

Fig. 1 .
Fig. 1.Location of study regions and their relations with permafrost category.Permafrost spatial data are from Brown et al. (2002).

Fig. 2 .
Fig. 2. Average landscape values for (a) mean annual temperature and mean annual precipitation, (b) mean winter temperature and mean winter precipitation, and (c) mean summer temperature and mean summer precipitation extracted from the New et al. (2002) climate database.Winter and summer intervals were defined by mean monthly temperatures below and above 0 °C, respectively.

Fig. 4 .
Fig. 4. Comparison of average sampling year (solid bars, NARR 2015) and 1961-1990 average (grey bars, New et al. 2002) values for summer temperature, summer relative humidity, summer precipitation, and winter precipitation for the five landscapes.

Fig. 5 .
Fig.5.Water isotope compositions (δ L ) from 376 lakes superimposed onto the landscape-specific isotope frameworks.The data defining the landscape-predicted LELs are shown in Table1.
the high variability in the average proportion of source water type (i.e., rain and snowmelt) to all lakes within and among the landscapes.YF lakes possess the largest range of δ I values, indicating substantial within-landscape variability in the proportions of source water types, while lakes in NUN have the smallest range of δ I values, signifying less variability in proportions of source water type.Lakes with the lowest δ I values are found in YF, while lakes in HBL have the highest δ I values.

Fig. 6 .
Fig. 6.Isotope compositions of lake-specific input water (δ I ) for each of the five landscapes.Classification of snowmelt-dominated lakes (δ I ≤ δ P ), rainfall-dominated lakes (δ I > δ P ), and permafrost thaw-dominated lakes (δ I <≤ δ P ) are represented by the diagonal lines.Note the different axis scales.

Fig. 7 .
Fig. 7. Cumulative proportions (lines, right y-axis) and frequency (bars, left y-axis) distributions of E/I values for thermokarst lakes from the five landscapes.The vertical broken line represents E/I = 0.5.Water balance of lakes with E/I > 0.5 is considered evaporation dominated.Note the varying vertical scales.

Fig. 8 .
Fig. 8. Boxplots comparing evaporation-to-inflow ratios (E/I) for all 376 lakes among permafrost types.The broken line represents E/I = 0.5, the threshold for evaporation-dominated lakes.

Fig. 9 .
Fig.9.Boxplots comparing evaporation-to-inflow ratios (E/I, top) and oxygen isotope composition of lake-specific input water (δ 18 O I , bottom) for thermokarst lakes at each of the five landscapes between the two vegetation classes (forest versus tundra).The broken line in the upper panel represents E/I = 0.5, the threshold for evaporationdominated lakes.

Table 2 .
Results of Kruskal-Wallis tests, which compare evaporation-to-inflow ratios (E/I) or oxygen isotope composition of lake-specific input water (δ 18 O I ) values for the different permafrost zones (continuous, discontinuous, and sporadic) and vegetation categories (forest versus tundra dominant).