Introduction

Substantial emissions of carbon dioxide (CO2) into the atmosphere make inland waters critical components of atmospheric CO2 budgets (Cole and others 2007; Tranvik and others 2009; Raymond and others 2013). Northern latitude lakes, in the boreal and arctic region, play a particularly important role in atmospheric CO2 budgets, as a recent estimate of CO2 emissions from boreal and arctic inland waters (between latitudes 50°–90°N) suggests that 0.15 Pg C y−1 is evaded into the atmosphere, of which 0.11 Pg C y−1 is from arctic and boreal lakes and reservoirs (Aufdenkampe and others 2011). CO2 emissions into the atmosphere are strongly influenced by the partial pressure of carbon dioxide (pCO2) at the water–atmosphere interface. Many studies on pCO2 in arctic and boreal watersheds have been conducted on a catchment (Kling and others 1991; Kelly and others 2001; Laurion and others 2010) and regional scale (Humborg and others 2010; Weyhenmeyer and others 2012; Campeau and Del Giorgio 2014); however, most of these studies have a sampling bias towards the open water season. This sampling bias is particularly problematic in northern latitudes, as lakes that may be ice-covered for up to 7 months of the year (Prowse and others 2012) can accumulate a substantial amount of CO2 below ice, subsequently leading to high CO2 emissions into the atmosphere at ice-melt (Striegl and others 2001; Huotari and others 2009; Ducharme-Riel and others 2015). A recent study by Karlsson and others (2013) found that in twelve small lakes in subarctic Sweden the CO2 emitted at ice-melt accounted for 12–56% of the annual CO2 emitted from these lakes. However, on a regional scale, the contribution of CO2 emitted at ice-melt in terms of annual CO2 emissions has yet to be documented.

The growing interest in global CO2 emission estimates from inland waters emphasizes our need to consider the dynamics of lakes in a landscape context. Across the boreal and arctic region, studies have shown latitudinal variations in lake water CO2 during the open water season to be related to differences in catchment characteristics, lake morphometry, dissolved organic carbon (DOC), nutrient concentrations, and climate variables (Kelly and others 2001; Sobek and others 2003; Rantakari and Kortelainen 2005; Kortelainen and others 2006, 2013; Roehm and others 2009; Lapierre and del Giorgio 2012; Ducharme-Riel and others 2015; Finlay and others 2015). Differences in climate and catchment characteristics influence the loading of carbon and nutrients bound in organic matter (OM) to lakes, and in turn differences in lake size and shape affect stratification and oxygenation and therefore the utilization and transformation of OM, including microbial respiration of DOC into CO2. The interaction between DOC and nutrients in relation to DOC degradation and subsequent CO2 is still unclear on a regional scale (Roehm and others 2009) as nutrients have been found to increase productivity, decreasing CO2 via photosynthesis (for example, del Giorgio and others 1999; Hanson and others 2003), but also to stimulate the degradation of DOC, increasing pCO2 via respiration (for example, Huttunen and others 2003; Smith and Prairie 2004; Ask and others 2012). Further, climate drivers related to winter conditions, for example, ice cover duration and snow cover, are commonly neglected in landscape-scale studies. Thus, in order to understand regional scale variability of below ice pCO2, climate variables related to the winter period need to be included.

During the winter period, the physical structure of ice-covered lakes differs from the open water season as ice limits wind-induced lake mixing and gas exchange. In late winter, when ice has reached its maximum growth, mainly heat flux from sediments and penetration of solar radiation through the ice drives circulation and water column mixing (Kirillin and others 2012). Snow accumulation on ice-covered lakes further reduces light availability, minimizing water column mixing and primary production below ice (Belzile and others 2001). Over the ice cover period, minimized mixing can lead to stratification, that is, where surface and bottom waters become disconnected with warmer waters (4°C) found near the bottom of the lake. During winter, CO2 has been found to build up in hypolimnetic bottom waters, indicating that sediment respiration is an important source of CO2 to ice-covered lakes (Striegl and Michmerhuizen 1998; Kortelainen and others 2006). Across 15 temperate and boreal ice-covered lakes, Ducharme-Riel and others (2015) found that benthic-derived CO2 had a relatively greater role in shallow lakes, likely due to the larger sediment surface area-to-water volume ratio and smaller distance between bottom sediments and surface waters in shallow lakes (for example, Kelly and others 2001). Thus, the role of benthic-derived CO2 in below ice CO2 accumulation should further be investigated on a regional scale across different lake morphometry types.

Because climate-induced changes and associated feedbacks are accelerated in northern environments, particularly during winter (Callaghan and others 2010), understanding the regional drivers of pCO2 across ice-covered lakes is not only important for understanding present-day CO2 emissions from lakes but also for predicting the consequences of a changing climate and cryosphere to future CO2 emissions. Thus, the main aim of this study was to identify the drivers of below ice pCO2 on a large regional scale across lakes, and within a lake between surface and bottom waters. We hypothesized that ice cover length is significantly related to variations of below ice pCO2 across lakes. We further hypothesized that within a lake below ice pCO2 is significantly higher in bottom waters compared to surface waters because sediment respiration acts as an additional source of CO2 to bottom waters. Additionally, we hypothesized that pCO2 below ice is significantly higher in small and shallow lakes compared to large and deep lakes due to differences in the dilution of CO2 in the water column. To test our hypotheses, we compiled data on below ice water chemistry, lake morphometry, catchment characteristics, lake position (see methods), and climate for 506 ice-covered lakes across Sweden and Finland.

Methods

Study Region

Our study lakes were distributed along a north–south climate gradient between the latitudes 56°N and 69°N of boreal and subarctic/arctic region of Sweden and Finland, where permanent snow and ice cover duration ranges from 102 days in the south to 234 days in the north (Figure 1) and long-term average annual air temperatures range from +6.5 to −3.5°C. Lakes cover about 10% of the total area of Finland and 9% of the total area in Sweden (Raatikainen and Kuusisto 1990; Henriksen and others 1998). The topography of the study region is relatively flat with higher elevations (up to 2100 m) in northwest of Sweden. The bedrock is predominantly Precambrian igneous and metamorphic rock. Land-cover patterns are similar in Sweden and Finland with highest agriculture area found in the south, extensive forest in the interior and tundra or open land in the north. In Finland, peatlands cover one-third of the land area, half of which have been ditched, mostly for forestry (Finnish Statistical Yearbook Forestry 1997).

Figure 1
figure 1

Map of Sweden and Finland depicting the locations of below ice surface water pCO2 (pCO2surface) and the corresponding ice cover duration (D ice).

Database Description

The databases used in this study are available from the Swedish National Lake Inventory Programme (http://www.slu.se/vatten-miljo), and the published studies of Sobek and others (2003), Rantakari and Kortelainen (2005), and Kortelainen and others (2006), which together cover a broad geographical range spanning across Sweden and Finland and represent gradients in both trophic state and humic matter content: total phosphorus (TP) 11, 4–53 μg L−1; total nitrogen (TN) 460, 180–1400 μg L−1; and DOC 9, 3–21 mg L−1 (all values are reported as median and 5 and 95 percentiles). The median lake area (LA) was 0.7 km2 and more than 90% of the lakes were smaller than 100 km2 (Table 1). Although most lakes were small, large lakes existed in the dataset (max LA of 1,538 km2), as Rantakari and Kortelainen (2005) included all lakes in Finland larger than 100 km2.

Table 1 Distribution (Reported as Median, 5 and 95 Percentiles) of All Variables Used in the PLSall Analysis Including Below Ice Surface Water Chemistry, Lake Morphometry, Catchment Characteristics, Landscape Position, and Climate Parameters

pCO2 Data

From the different databases, all lakes with below ice pCO2 measured during the ice cover period were included. The data were split into 5 groups, abbreviated DataSweden, DataSwedendirect, DataFinland, DataFinlandTP, and DataFinlandlarge (Table 2), and described in the following.

Table 2 Description of the Five Data Groups Used in This Study

DataSweden (n = 224) represents lakes from the Swedish National Lake Inventory Programme database from which pCO2 was calculated based on alkalinity (Alk), pH, water temperature (T w), and altitude (Alt) according to Weyhenmeyer and others (2012). To reduce the influence of acidification, recent liming, alkaline lakes, or algal bloom conditions which might bias pCO2 (for example, Humborg and others 2010), we excluded observations with Alk < 0 or ≥ 1 mEq L−1 or pH > 8. From this database, we selected below ice data by assuming that any sample collected between January and March with a surface water temperature ≤4°C was sampled below ice. If data for a particular lake were not available during January and March, we selected data from April only if the temperature requirement of ≤4°C was met. We used surface water samples (in most cases at 0.5 m and in all cases <2 m) and bottom water samples (1 m above the deepest point of the lake). Whenever a lake was sampled more than once during the ice cover period (that is, within an ice cover period or across years), we calculated the maximum and median value for the ice cover period.

DataSwedendirect (n = 42) is an additional set of Swedish lakes from Sobek and others (2003) from which pCO2 was directly measured on an infrared gas analyzer. During the ice cover period, these lakes were sampled once between February and April 2001, surface water at 1 m below the ice and bottom water at 1–2 m above the sediment.

DataFinland (n = 175) represents below ice pCO2 data from Kortelainen and others (2006) which were collected once during the winters of 1998–1999, with lakes sampled at the end of the winter stratification (approximately between the end of March and April). pCO2 was calculated from total inorganic carbon (TIC), pH, and T w, using Henry’s law constants corrected for temperature and atmospheric pressure (Plummer and Busenberg 1982). Headspace TIC was measured with gas chromatography, where water samples were acidified to convert all TIC to CO2. Water samples were collected at the deepest point of the lake, surface waters were sampled at a depth of 1 m and bottom waters were sampled at 20 cm from sediment surface. Because many of the Finnish lakes smaller than 100 km2 are shallow, bottom water measurements made 20 cm from sediment surface were used to ensure that the difference between surface and bottom waters was represented.

DataFinlandTP (n = 28) is a subset of lakes sampled by Kortelainen and others (2006), which represent eutrophic lakes in the Nordic lake survey with the highest TP. Because a majority of Finnish lakes, as well as boreal lakes, are located in forested catchments with relatively minor human disturbance, eutrophic lakes are rather rare. Therefore, data from these lakes were included in analyzing the relationship between below ice pCO2 and other variables (Table 1; Figures 1, 2, 3, 4) but were excluded from comparisons between Sweden and Finland (Table 3) and from the estimation of CO2 emission because keeping them in the analysis would exert too strong an influence by quite a small population of lakes. The sampling techniques and pCO2 calculations follow DataFinland.

Figure 2
figure 2

Partial least squares loading plot of below ice pCO2surface observations for Sweden and Finland (PLSall; n = 506). The loading plot depicts the correlation structure between pCO2surface (Y-variable) and X-variables (for explanation of variable abbreviations, see Table 1). The greater the distance a variable is from the origin, the greater its overall influence (see Table 4 for VIP scores).

Figure 3
figure 3

Relationship between Swedish (black circles) and Finnish (open circles) below ice pCO2surface and A TP (log y = 6.7 + 0.5 log x), B TN (log y = 4.1 + 0.6 log x), C Z avg (log y = 9.0 + 0.75 log x), D Vol (log y = 8.2 + 0.14 log x), E DOC (log y = 7.2 + 0.4 log x), F D ice (log y = 8.2 + 0.05 log x), and G pCO2bottom (log y = 1.5 + 0.75 log x). All data were log-transformed.

Figure 4
figure 4

Distribution (median, 1st and 3rd quartile) of Swedish (black circles) and Finnish (open circles) below ice pCO2surface between A average depth classes, B lake area classes, and C data groups. One outlier (>20,000 μatm) was removed for clarity. Wilcoxon each pair test results are letter-coded, where groups not sharing a letter are significantly different.

Table 3 Below Ice Surface Water Chemistry and Lake Morphometry Across Lakes in Sweden (DataSweden and DataSwedendirect) and Finland (DataFinland) Reported as Median, 5 and 95 Percentiles

DataFinlandlarge (n = 37) represent the largest lakes in Finland and are from Rantakari and Kortelainen (2005). The pCO2 was calculated from TIC, pH, and T w, using Henry’s law constants corrected for temperature and atmospheric pressure (Plummer and Busenberg 1982). TIC in the water was measured with a carbon analyzer. Below ice data were collected during the winters of 1998–1999, with lakes sampled at the end of the winter stratification (approximately between the end of March and April). Samples were collected at the deepest point of the lake, surface waters were sampled at a depth of 1 m, and bottom waters were sampled 1 m above the sediment. Lakes were sampled at least twice during the ice cover period, and thus a maximum and median value was used for these lakes. Data from these lakes were excluded from comparisons between Sweden and Finland because DataSweden only included two lakes with an LA greater than 100 km2.

We acknowledge that the methodological differences between the five groups of data (Table 2) could cause deviation in pCO2 between groups, probably mainly between calculated versus measured pCO2 in acidic and organic-rich lakes (Abril and others 2014; Wallin and others 2014). We therefore examined a possible methodological bias by comparing pCO2 between DataSweden and DataSwedendirect. We also compared DataSweden and DataFinland to evaluate possible methodological issues between the Swedish and Finnish datasets (see “Results” section).

Altogether, surface water data from 506 ice-covered lakes were available for our analyses. Of the 506 lakes, some lakes did not have measurements from bottom waters and therefore a sub-database, of lakes with both surface and bottom water measurements, was created (n = 311 lakes). To differentiate between below ice samples collected from surface waters and bottom waters, we used abbreviations pCO2surface and pCO2bottom, respectively.

Additional Variables

In addition to below ice pCO2, we used data on pH, T w, Alk, conductivity (Cond), TN, TP, and total organic carbon (TOC). TOC in boreal lakes usually contains 97 ± 5% DOC (von Wachenfeldt and Tranvik 2008), thus TOC in this study can be seen as the equivalent to DOC. Water sampling and analyses were performed by the accredited water analysis laboratory at the Swedish University of Agricultural Sciences and by the accredited laboratories of the Finnish Regional Environment Centers, according to standard limnological methods. Sobek and others (2003) carried out manual water sampling and analysis according to certified methods.

Further, GIS-derived data on lake morphometry, catchment characteristics, landscape position, and climate variables were included. Lake morphometry and catchment characteristics were acquired from topographic maps combined with land-use data on satellite images using the Arc View georeferencing software (for example, Kortelainen and others (2006)), for Finnish lakes, and from the Swedish National Lake Inventory Programme, as regularly released by the Swedish Meteorological and Hydrological Institute (SMHI, http://www.smhi.se), for Swedish lakes). Lake morphometry comprised data on lake surface area (LA; km2), volume (Vol; Mm3), average depth (Z avg; m) (calculated as Vol/LA), lake shoreline length (SL; km), and shoreline development length (DL) (calculated as the SL/√2ΠLA, Wetzel 2001). For lakes whose Vol was not in the registries, the Vol was determined using a calibrated lake volume estimate for Swedish lakes according to Sobek and others (2011) (ln Vol = 1.39 + 1.12 * ln LA). For Finnish lakes, we recalibrated the Swedish estimate using data from 58 Finnish lakes (ln Vol = 1.11 + 1.09 * ln LA; R 2 = 0.91; n = 58; p < 0.0001).

Catchment characteristics included catchment area (CA; km2), drainage ratio (DR), % wetland/peatland in catchment (Peat), % agriculture in catchment (Agr), % urban in the catchment (Urb), % forest in catchment (For), and % water in the catchment including the lake itself and upstream water bodies (Wat). The drainage ratio (DR) was determined by dividing catchment area by lake area.

As an indicator of landscape position, in addition to altitude (Alt; m), X-coordinate (X-coord; °N), and Y-coordinate (Y-coord;°E), we defined lake hydrology (LH), following the protocol described in Martin and Soranno (2006), by assigning each lake to one of three categories; isolated, that is, have no connecting stream or lake (LH (Isolated)), headwater (LH(Head)), or flow through (LH(Flow)). Using ArcGIS (Version 10.1), each lake was assigned a category for the landscape position metric using the Swedish (VIVAN 2007, 298,215 lakes and 933,675 streams) and Finnish (53,511 lakes and 40,051 streams) network of rivers and lakes for flow-based modeling database. LH measures the overall surface hydrological position of a lake by incorporating connection both to lakes and streams. Altitude was calculated from a raster-based digital terrain model (DTM).

Climate variables, ice duration (D ice), and average annual air temperature (T avg) were assigned for each lake. Average annual air temperature for each lake was based on an averaged 1961–1990 temperature value (from SMHI for Sweden and Finnish meteorological institute (FMI, http://www.ilmatieteenlaitos.fi) for Finland). Although regional ice cover data are available for Sweden and Finland, the scale of the data is coarse and therefore we used a more robust measure of ice cover duration for each individual lake. The number of days a lake is covered by ice (D ice) was calculated using an air temperature function, which was calibrated and validated for Swedish lakes (Weyhenmeyer and others 2013):

$$ D_{\text{ice}} = \frac{365.25d}{\pi } \times \arccos \left( {\frac{{T_{\text{m}} }}{{T_{\text{a}} }}} \right), $$
(1)

where d is days, T m is the altitude-adjusted average air temperature, and T a is the altitude-adjusted average air temperature amplitude. T m and T a were estimated (Weyhenmeyer and others 2013; for abbreviations see above) as

$$ T_{\text{m}} = T_{\text{avg}} - \frac{{{\text{Alt}} \times 0.6}}{100} $$
(2)
$$ T_{\text{a}} = - 0.66 \times T_{\text{m}} + 14.32. $$
(3)

Although the primary aim of this study was to investigate regional scale patterns, we also addressed the temporal dimension by estimating the number of days a lake was ice-covered prior to sampling (S ice). This was done by subtracting the predicted ice-on date from the sampling date. Ice-on data for Swedish lakes were obtained from SMHI and for Finnish lakes from Finland’s Environmental Administration (http://www.ymparisto.fi/), both providing regional ice-on dates for small and medium/large lakes. For lakes that were sampled more than once during the ice cover period (DataSweden and DataFinlandlarge), S ice was calculated for the maximum pCO2.

Statistical Evaluation

In order to identify the drivers of below ice pCO2surface and below ice pCO2bottom, partial least square regression (PLS) was used. PLS, a method for relating how X correlates to Y by a linear multivariate model, offers a more robust technique compared to other multiple linear regression analyses as data can have missing values, they can co-correlate, and they do not need to be normally distributed (Eriksson and others 2006). In PLS, X-variables are classified according to their relevance in explaining Y, abbreviated as VIP values (Wold and others 1993). We considered VIP scores ≥1.0 as highly influential, between 0.8 and 1.0 moderately influential, and <0.8 less influential. The performance of the PLS model was expressed as R 2 Y, representing how much of the variance in Y is explained by X, and Q 2 Y, which is a measure of the predicative power of the PLS model. In the PLS models, data were log(10)-transformed if they were highly skewed (skewness >2.0 and min/max <0.1). An observation was excluded from the model if it fell outside the 99% confidence region of the model (that is, hotelling T 2) (Eriksson and others 2006). PLS modeling was carried out in the SIMCA-P 13.0 software (Umetrics AB, Umeå, Sweden).

We ran an initial PLS using data from all lakes with below ice pCO2surface observations (out of 506 lakes, 9 observations fell outside the 99% confidence range of the model and therefore were removed), termed PLSall. A total of 22 X-variables were included in the PLSall model with pCO2surface set as the Y-variable (Table 1); because Alk, pH, T w, and TIC were used to calculate pCO2, they were removed from the PLS analyses to avoid autocorrelation. We ran an additional PLS model, using the maximum pCO2surface for all lakes as the Y-variable and including an additional X-variable, S ice, accounting for the days a lake was ice-covered prior to sampling.

A subset of 311 lakes, having both pCO2surface and pCO2bottom data, was used to investigate if drivers of below ice pCO2surface were different from the drivers of pCO2bottom. Two separate PLS models were run for surface waters (PLSsurface; out of 311 lakes 5 observations fell outside the 99% confidence range and were removed from the model) and bottom waters (PLSbottom; out of 311 lakes 6 observations fell outside the 99% confidence range and were removed from the model) with pCO2surface and pCO2bottom set as the Y-variable, respectively.

Further statistical calculations were carried out in JMP, version 11.0.0. For determining the relationship between below ice pCO2surface and below ice lake chemistry, lake morphometry, and ice cover variables, Pearson’s correlation coefficients were used, where all the input data were log-transformed due to non-normal distribution of the data (Shapiro–Wilk test: p < 0.05 indicating data are non-normally distributed). To test if below ice pCO2bottom was significantly higher than pCO2surface, we applied a matched-pair t test with log-transformed data where below ice pCO2bottom and pCO2surface were paired for each lake (n = 311 lakes). To determine whether below ice pCO2surface differed between data groups (DataSweden DataSwedendirect, DataFinland, DataFinlandTP, and DataFinlandlarge), mean lake depth (<2.5, 2.5–3.5, 3.5–4.5, >4.5 m), and lake area classes (<0.1, 0.1–1, 1–10, >10 km2), we applied non-parametric Wilcoxon tests and Wilcoxon each pair test where a significant difference between a class is reached when p < 0.05.

Results

Below Ice pCO2 in Surface Waters

Of the below ice pCO2surface reported for the 506 lakes sampled, 504 were supersaturated in CO2. Highest below ice pCO2surface was found in small eutrophic Finnish lakes. In Finland (that is, DataFinland), below ice pCO2surface was on average about twice as high as in Sweden (Table 3). Also below ice nutrients in surface waters (median TP of 9 and 12 μg L−1 and TN of 404 and 510 μg L−1, for Sweden and Finland, respectively) were higher in Finland than in Sweden, while DOC was similar (median DOC of 9 mg L−1 for both countries). Further, Finnish lakes were generally smaller and shallower, while the Swedish lakes covered a larger altitude range (Table 3).

When we modeled variations in below ice pCO2surface across all 506 Finnish and Swedish lakes (PLSall), we received a good model predictability (Q 2 = 0.58) with two components able to explain 60% of the variation in pCO2surface (R 2 Y = 0.60). In the PLSall model, the first component (that is, horizontal axis) explained 53% of the variation in pCO2surface, representing lake morphometry and water chemistry variables (Figure 2). Lake morphometry (Z avg, Vol, LA, SL, DL) was negatively related to pCO2surface, whereas water chemical variables (TP, TN, Cond, DOC) were positively related to pCO2surface. The second component (that is, vertical axis) represented regional climate (i.e., D ice and T avg) and latitude and only explained 7% of the variation in below ice pCO2surface. When ice cover duration prior to sampling (S ice) was included as an additional X-variable in the PLS for the prediction of maximum pCO2surface, we found that the model remained similar, without an influence of S ice on the model (Q 2 = 0.59, R 2 Y = 0.61).

Overall, TP was the most influential variable, followed by lake morphometry (Z avg,Vol, LA, SL), TN, Cond, CA, LH(isolated), and Y-coord (Table 4). TP alone was able to explain 30% of the variation in pCO2surface (Figure 3A). Also TN (Figure 3B), Z avg (Figure 3C), and Vol (Figure 3D) had a high explanatory power. DOC only had a moderate influence on the PLSall model (Table 4), and by directly relating DOC to pCO2surface, we found a weak positive relationship (Figure 3E). Average depth was negatively related to TN (r 2 = 0.09, p < 0.001 n = 500), TP (r 2 = 0.08, p < 0.001 n = 500), and DOC (r 2 = 0.07, p < 0.001 n = 500). D ice was not an influential variable for the model performance (Table 4), and when relating D ice to pCO2surface we found no relationship (Figure 3F). However, D ice was significantly positively related to Cond (r 2 = 0.30, p < 0.001 n = 498), DOC (r 2 = 0.14, p < 0.001 n = 500), TN (r 2 = 0.13, p < 0.001 n = 443), and TP (r 2 = 0.01, p < 0.01 n = 498).

Table 4 Variable Importance in the Projection (VIP) Scores for Partial Least Squares (PLS) Models

Median below ice pCO2surface was significantly different across varying lake areas (Wilcoxon test: χ 2 = 85, p < 0.0001, n = 506) and average depths (Wilcoxon test: χ 2 = 180, p < 0.0001, n = 506). According to the Wilcoxon each pair test, median below ice pCO2surface was significantly different between each lake area and average depth class, that is, pCO2surface was higher in shallow lakes (Z avg < 2.5 m) compared to deep lakes (Z avg > 4.5 m) (Figure 4A) and in small lakes (LA < 0.1 km2) compared to large lakes (LA > 10 km2) (Figure 4B).

Below Ice pCO2 in Bottom Waters and Its Relation to pCO2 in Surface Waters

From a subset of 311 lakes having pCO2surface and pCO2bottom measurements, two separate PLS models were run for surface waters (PLSsurface) and bottom waters (PLSbottom) with pCO2surface and pCO2bottom set as the Y-variable, respectively. The PLSsurface model explained 64% (R 2 Y = 0.64) of the variation in pCO2surface and model predictability was good (Q 2 = 0.61). The model predictability of PLSbottom was similar (Q 2 = 0.64) and explained 67% (R 2 Y = 0.67) of the variation in pCO2bottom. The major difference between PLSbottom and PLSsurface was that pCO2bottom was slightly more influenced by water chemical variables, in particular TN, and pCO2surface was slightly more influenced by lake morphometric variables, in particular Z avg (Table 4).

When relating pCO2bottom to pCO2surface, we found that 62% of the variation in pCO2surface could be explained by pCO2bottom (Figure 3G). We found that the residuals of the regression (residuals log(pCO2surface)) were related to lake morphometry (Vol: r 2 = 0.11, p < 0.001, n = 311; LA: r 2 = 0.10, p < 0.001, n = 311) supporting the concept that pCO2surface is a function of pCO2bottom and the recipient volume of water. pCO2bottom differed significantly from pCO2surface (matched-pair t test result: t = 22.7, p < 0.05, number of pairs = 311) with median pCO2bottom more than twice as high as the median pCO2surface (7187 and 3206 μatm, respectively). Further, when relating maximum pCO2 to S ice, we found a weak positive relationship for pCO2surface (r 2 = 0.01, p = 0.037 n = 506) and a stronger positive relationship for pCO2bottom (r 2 = 0.13, p < 0.01, n = 311).

Comparison Between Data Groups

Comparing all five data groups, we found that pCO2surface was significantly different between groups (Wilcoxon test: χ 2 = 199, p < 0.0001, n = 506). When comparing DataSwedendirect, i.e., directly measured pCO2, with DataSweden, that is, calculated pCO2, we did not find a statistically significant difference (Wilcoxon each pair test: p > 0.05). In contrast, we found a significant difference between the Swedish dataset, DataSweden, and the Finnish dataset, DataFinland, (Wilcoxon each pair test: p < 0.0001) with higher below ice pCO2surface found for DataFinland (Figure 4C).

Discussion

Drivers of Below Ice pCO2 on a Spatial Scale

Using a multivariate approach comparing 506 ice-covered lakes across Sweden and Finland, we were able to identify key variables influencing below ice pCO2surface and pCO2bottom. Lake morphometry (Z avg, Vol, LA, SL) and lake water chemistry (TP, TN, Cond) were most important in explaining variations of below ice pCO2surface and pCO2bottom across lakes. Lake morphometric variables were situated opposite to water chemical variables along the first component axis in the PLS loading plot, suggesting that these variables are tightly negatively associated with each other and have more influence on pCO2surface and pCO2bottom than regional scale climate variables such as ice cover length and air temperature (that is, variables that lie along the secondary principal component axis; Figure 2). Negative relationships between water chemistry and Z avg reflect that small shallow lakes have proportionally higher chemical concentrations during winter compared to large deep lakes.

The identified drivers of below ice pCO2 are well known drivers for water chemical concentrations and for pCO2 during the open water season (for example, Kelly and others 2001; Sobek and others 2003). Lake morphometric variables reflect the degree of dilution, water mixing, catchment, and lake internal loading. Although these processes are minimized during the ice cover period, they are very important in determining initial pCO2 in the water column before ice cover disconnects the catchment and the atmosphere from the lake. Many processes determine initial pCO2 in the water before the ice cover period begins, one being the intensity and length of autumn water turnover. For example, a complete autumn turnover prior to ice-on vents CO2 from the lake, resulting in similar pCO2 throughout the water column (López Bellido and others 2009). If ice-on comes early, an incomplete autumn turnover will result in elevated pCO2 during winter (particularly in bottom waters). In addition, high precipitation prior to ice-on transports OM from the catchment to the lake, enhancing DOC and nutrient availability in the lake and hence conditions for degradation during ice cover (for example, Rantakari and Kortelainen 2005). Huotari and others (2009) support these ideas, as they found that longer autumn turnover resulted in lower CO2 below ice, and a wet autumn resulted in elevated below ice CO2 compared to a dry autumn.

Therefore, potentially our below ice pCO2 variations could simply reflect spatial pCO2 variations that already occur during the open water season. However, below ice pCO2surface with a median of 2168 µatm in Swedish lakes and a below ice pCO2bottom with a median of 2853 µatm in Swedish bottom waters (in Finland 4397 and 9943 µatm, respectively, Table 3) were substantially higher than the median pCO2 of less than 1500 µatm that was observed in the same regional area during the open water season (Weyhenmeyer and others 2012). Consequently, CO2 is most probably further produced within the lake during the ice cover period.

CO2 can be produced below ice cover by microbial mineralization of OM, mainly in the sediment where the availability of OM and nutrients regulates the microbial CO2 production (del Giorgio and others 1999; Kortelainen and others 2006). Because we observed a positive correlation between pCO2bottom and TP, TN, and DOC, we suggest that microbial mineralization in the water column and sediments frequently occurs below ice cover in our study lakes. Because additional nutrients and DOC can be released from the sediments into bottom waters under changed bottom water redox conditions (Mortimer 1941; Gonsior and others 2013; Yang and others 2014), microbial respiration in bottom waters might even be enhanced (Peter and others, submitted). Our study suggests that nutrients have a positive effect on pCO2bottom in ice-covered lakes; however, if this relationship is causal (enhanced DOC degradation) or correlative (sediment release of both nutrients and CO2), it needs to be investigated further.

Additional CO2 below ice compared to the open water season can also result from groundwater/weathering inputs (Striegl and Michmerhuizen 1998; Humborg and others 2010), here indicated by a positive correlation between pCO2bottom and Cond. Although groundwater can be an important source of CO2, particularly during early and late ice cover, it is likely that most of the groundwater entering the lake is from shallow soil water flow, opposed to deep groundwater flow. As this shallow soil water is generally below 4°C, it entrains the surficial layer of the lake and usually does not mix with bottom waters (Kirillin and Terzhevik 2011). Further, deep groundwater flow through sediments into lake bottom waters is less probable, confirmed by Pulkkanen and Salonen (2013) who found that the contribution of groundwater to deep lake waters during winter was minimal in five ice-covered Finnish lakes.

Our results did not support the hypothesis that below ice pCO2surface and pCO2bottom is significantly related to variations in ice cover duration across lakes. In general, regional scale climate (that is, T avg) effects on below ice pCO2surface or pCO2bottom were lacking. These results were unexpected because ice cover length and temperature are known to regulate landscape-scale processes (that is, growing season length). A similar case of a lacking direct temperature effect along a large spatial gradient on lake pCO2 was found during the open water season for a global lake database (Sobek and Tranvik 2005), and it was reasoned that the lack of relationship was due to the fact that the temperature sensitivity of a process depends on the substrate supply (Pomeroy and Wiebe 2001). On a regional scale, substrate supply can be quite variable, demonstrated by the wide range of TP, TN, and DOC concentrations found in this study (Table 1), whereas on a local or individual lake scale the variation in substrate supply is smaller. Therefore, on a local and also on a temporal scale we might expect ice cover length and climate to play a much more important role in influencing below ice pCO2. In this study, we may have missed some localized climate variability as snow cover and ice thickness data were not available for most of the lakes. Nonetheless, in a study of below ice pCO2 by Sobek and others (2003), localized climate variables (air temperature, precipitation, ice thickness, and snow cover) were less influential on below ice pCO2 compared to physiochemical and morphometric variables.

Our analyses had one temporal component included and this was the duration of ice cover prior to sampling (S ice). We found that S ice had a stronger relationship with pCO2bottom compared to pCO2surface. In a recent study, Denfeld and others (2015) found that in a small boreal lake below ice CO2 did not constantly increase throughout the winter period, but remained constant during the late winter period. This suggests that in some lakes there might only be a small change in surface water CO2 in late winter below ice, and therefore might explain why we did not find a strong relationship between S ice and pCO2surface in this study. However, this may not be true for all lakes (for example, Huotari and others 2009), particularly in bottom waters, as sediment respiration is likely to continue to contribute to CO2 accumulation throughout the whole winter (for example, Ducharme-Riel and others 2015). This suggests that on a temporal scale, ice cover days prior to sampling may be more important to pCO2bottom, whereas other factors, such as substrate supply and light below ice, may be more important in determining pCO2surface in late winter.

Geographical Differences Including Hot Spots of Below Ice pCO2

In agreement with our hypothesis, below ice pCO2 was generally highest in small shallow lakes and in bottom waters of lakes, suggesting that these waters are hot spots of gas accumulation during the ice cover season. Small shallow lakes and bottom waters have earlier been identified as biogeochemical hotspots during the open water season (Hanson and others 2007; Humborg and others 2010; Weyhenmeyer and others 2012), whereas this study further stresses their importance in carbon cycling during the ice cover period. Small shallow lakes dominate the Swedish (Sobek and others 2011) and Finnish landscape (Kuusisto and Hakala 2007) and are the most abundant lake type on earth (Downing and others 2006; Verpoorter and others 2014). Further, small shallow lakes are important systems to monitor in a changing cryosphere as they are particularly sensitive to changes in temperature and precipitation (Rautio and others 2011).

Across Sweden and Finland, we found that below ice pCO2 variation was more pronounced longitudinally (east–west) than latitudinally (north–south). Longitudinally, comparing below ice water chemistry and lake morphometry between Sweden and Finland (Table 3), we found that pCO2, TP, and TN were higher in Finland. Finnish lakes had twice as high median pCO2surface and three times as high median pCO2bottom compared to Swedish lakes. Part of this variation may be due to differences in lake selection, sampling approaches, and methodology (Table 2). For example, Finnish lakes could have higher pCO2 because they were on average sampled later into the ice cover period, and for bottom waters some Finnish lakes (DataFinland and DataFinlandTP) were sampled closer to the sediment (0.2 m above) than Swedish lakes (1–2 m above) (Table 2). If the selection of lakes, sampling approaches, and methods between the two countries were exactly the same, the observed differences in below ice water chemistry between Sweden and Finland might have been smaller, yet it is likely that Finnish lakes would still have higher below ice pCO2 because of different lake characteristics. Previous comparisons between the two countries during the open water period have found that Swedish lakes have higher pCO2 than Finnish lakes (Weyhenmeyer and others 2012). However, for Finnish lakes the average winter CO2 has been reported to be five times that of the average summer CO2 (Kortelainen and others 2006), while the difference between winter and summer CO2 may rather be around a factor of 2 for Swedish lakes (Sobek and others 2003; median below ice pCO2 of 2244 μatm in this study and median open water pCO2 of 1048 μatm reported in Weyhenmeyer and others (2012)). The fact that Finnish lakes were on average smaller and shallower than Swedish lakes (Table 3; Figure 4A) likely resulted in the observed higher below ice pCO2 and nutrients found in Finland. Concerning the differences in methods for determination of pCO2, we did not find a significant difference between directly measured and calculated pCO2 for Swedish lakes, indicating that any potential methodological bias to the data could not be detected statistically.

Latitudinally, across Sweden and Finland, the inputs of DOC, TP, and TN from the catchment are smaller in the north compared to the south (for example, Kortelainen and others 1997) and are related to ice cover length. In northern Scandinavia, low temperatures and a shorter growing season limit litter production in the catchment and seasonally frozen soils limit OM loading to adjacent surface waters (Laudon and others 2012). Because lake water chemistry was related to pCO2surface and pCO2bottom in our PLS models, a latitude influence on below ice pCO2 was expected. Nevertheless, in our PLS models, latitude (X-coord) did not have a large influence on below ice pCO2. A possible explanation is that clear water lakes (which can be found in northern Sweden) may have relatively higher pCO2 during the winter, as benthic algae produced during the open water season can act as an additional source to degradation (Karlsson and others 2008). Further, although land-cover types generally differ with latitude across Sweden and Finland (for example, highest agriculture area found in the south), the land-cover type in the catchment (Agr, For, Peat, and Urb) was not an important variable in explaining below ice pCO2. The OM input to lakes is likely a mixture of different sources based on the proportion and connectivity of the different land-cover types in the catchment, and therefore a single important land-cover type may not be individually important on below ice pCO2.

Implications of Below Ice pCO2 on CO2 Emissions at Ice-Melt

To provide a broad estimate of the importance of the ice-melt period across Swedish and Finnish lakes, we estimated the CO2 emission during the ice-melt period. Based on the relationship between below ice pCO2surface and lake area (Figure 4B), we estimate that the CO2 emission across Swedish and Finnish lakes during the ice-melt period corresponds to 1.2–1.5 Tg C y−1 (based on lake area size classes <0.1, 0.1–1, 1–10, >10 km2, lake area adjusted k 600 from Raymond and others (2013), lake area-specific below ice pCO2surface of this study (median pCO2surface and maximum pCO2surface for lower and upper range, respectively), and an ice-melt period lasting for 19 days according to Denfeld and others (2015) for a small boreal lake in Sweden). At ice-melt, CO2 accumulated in bottom waters can be emitted to the atmosphere by bottom water convective turnover. If we assume that CO2 accumulated in the bottom waters during ice cover is emitted at ice-melt, the CO2 emission during the ice-melt period would double to 2.5–2.9 Tg C y−1. It is likely that these estimates are conservative as pCO2 measurements were made in late winter prior to ice-melt and do not account for the dynamic ice-melt period when CO2 from the catchment can be sourced to the lake (Miettinen and others 2014; Denfeld and others 2015). Nevertheless, the amount of 1.2–2.9 Tg C y−1 emitted from Swedish and Finnish lakes at ice-melt is comparable to estimates of average Holocene C accumulation in boreal lake sediments, 2–3 Tg C y−1 (Kortelainen and others 2004), that is, annually, as much C was estimated to be emitted from Swedish and Finnish lakes at ice-melt as was estimated to annually accumulate across all boreal lakes during the Holocene. Increased human activity during recent decades has, however, resulted in increasing accumulation rates (Anderson and others 2013).

The CO2 emitted from Swedish and Finnish lakes at ice-melt contributed to 18–30% of the annual CO2 emission, of which the open water CO2 emission corresponded to 6.63 Tg C y−1 (based on lake area size classes <0.1, 0.1–1, 1–10, >10 km2, lake area adjusted k 600 from Raymond and others (2013), open water lake area pCO2surface from Humborg and others (2010) for Swedish lakes and from Kortelainen and others (2006) and Rantakari and Kortelainen (2005) for Finnish lakes, and an open water period lasting for 199 days, that is, the median open water period in this study). A contribution of 18–30% annual CO2 flux from Swedish and Finnish lakes at ice-melt is similar to previous estimates (22%) for Finnish lakes, based on different assumptions (Kortelainen and others 2006), and is a bit less than previously found for individual lakes (ranging from 3 to 80% in 15 temperate and boreal lakes and from 12 to 56% in 12 subarctic Swedish lakes, Ducharme-Riel and others (2015) and Karlsson and others (2013), respectively), indicating that CO2 emission during ice-melt in individual lakes may have an even stronger impact on annual CO2 emission from lakes than estimated in this study.

Conclusion

Our study demonstrates that lake water chemistry and lake morphometry were more important factors in determining below ice pCO2 on a regional scale than climate variables (air temperature and ice cover duration). CO2 accumulates in ice-covered lakes, most probably due to nutrient- and OM-driven microbial mineralization in the sediments, and CO2 inputs from the catchment prior to ice-melt. We conclude that on a regional scale, carbon cycling in ice-covered lakes and subsequent CO2 emission at ice-melt are important components of annual CO2 emission estimates. Thus, given the potential for significant ecosystem changes to ice-covered lakes, adequately integrating the ice cover period in global CO2 emission estimates involves monitoring changes not only to ice cover but also to changes in the trophic status of lakes.