Predicting organic carbon in lakes from climate drivers and catchment properties

Authors


Abstract

[1] We combine a synoptic data set on organic carbon (C) concentrations in ∼1000 Norwegian lakes with detailed information on catchment properties derived from digital maps of elevation, climate, vegetation density, and land use. The resulting regression model explains 83% of the variance in total organic C from eight predictor variables: mean air temperature, area specific runoff, terrain slope, vegetation density, atmospheric nitrogen deposition, water residence time, and catchment area fractions of bogs and arable land. Vegetation density, expressed as the normalized difference vegetation index (NDVI), was by far the strongest predictor, while area fraction of bogs in the catchment also gave a strong positive effect on organic C concentrations and export fluxes. The unbiased sampling of lakes relative to whole population of catchments in the region and the general robustness of the model, as witnessed by imputation and resampling tests, allowed confident prediction of organic C concentrations and export fluxes for all 20,000 catchments of the Norwegian mainland. The method used in this study can, with some modifications, be extended to larger geographical regions and holds promise for predicting organic C fluxes under changed climatic conditions.

1. Introduction

[2] The fate of terrestrially fixed carbon (C) is determined by a range of biological, physical and chemical parameters. The majority of the organic C, which is not accumulated or oxidized on site, is exported by runoff to rivers, lakes and marine areas. Catchment properties and processes, along with climatic drivers, are thus the key determinant of concentrations and fluxes of total organic carbon (TOC, of which typically 90% is in dissolved form: DOC) in lakes. There may however be a significant loss of TOC within the freshwater systems by retention (i.e., sedimentation and burial), modification (e.g., recalcitration), and photooxidation or microbial mineralization [Raymond and Bauer, 2001; von Wachenfeldt and Tranvik, 2008; Algesten et al., 2004; Jonsson et al., 2007]. These processes alter the amount, quality and distribution of terrigenous organic C and ultimately its fate in the global carbon cycle [Battin et al., 2008; Cole et al., 2007; Hope et al., 1994].

[3] Allochthonous TOC is also a key determinant of lake ecosystem properties [Prairie, 2008; Williamson et al., 1999]. TOC has a plethora of direct and indirect impacts on lake ecosystem functioning. For example, TOC is the predominant energy source for the functioning of the microbial loop, which shunts energy to higher trophic levels, independent of primary production [Pace et al., 2004], and thus crucial for the structure, dynamic and productivity of the lake food web [Hessen et al., 1990; Jansson et al., 2007; Karlsson et al., 2009]. The light-absorbing property of TOC reduces lake primary production and the depth of the euphotic zone, and also the vertical heat distribution which determines the mixing depth and the duration of summer stratification [Snucins and Gunn, 2000]. TOC effects on the light regime of lakes have also been linked to shifts in predator distributions leading to cascading effects on the ecosystem [Yan et al., 2008].

[4] Because TOC concentration tends to reduce the vertical extension of primary production and increase overall lake respiration, it is an important factor for the trophic status and net CO2 exchange of lakes. High concentrations of organic C (and thus high ratios of C:N and C:P) cause a predominantly heterotrophic activity and high respiratory activity [Hessen and Anderson, 2008]. Studies suggest that lakes shift from net autotrophic to net heterotrophic at a concentration around 5 mg TOC L−1 [Jansson et al., 2008; Sobek et al., 2006], although this threshold is clearly also dependent upon lake productivity in terms of bioavailable P [Hanson et al., 2004] and lake stoichiometry [Hessen, 2005].

[5] An improved understanding of the underlying factors determining the fluxes and concentration of TOC in lakes is thus imperative for assessing the consequences of climate driven changes in TOC fluxes on ecosystem properties, and also important for understanding this component of the global C cycle. There are several studies that have addressed the role of peatlands and wetlands as donors of organic C for watersheds and lakes [Kortelainen et al., 2006; Xenopoulos et al., 2003], yet also the role of coniferous forests [Lent et al., 1998] and landscape form [Dillon and Molot, 1997] have been highlighted as determinants of catchment export of TOC.

[6] The aim of this study was to use detailed drainage basin properties of a large number of “pristine” lakes representing wide range in physical, biological and climatic variables, to establish a robust model of inland water TOC concentrations. This model can then be combined with runoff data to get an estimate on the hydrology-driven flux of organic C. While recognizing that several previous studies have explicitly linked TOC (or DOC) concentrations in lakes to catchment properties, we believe that this study significantly add to current knowledge by not only including a very comprehensive list and wide range of relevant catchment properties and climatic parameters, and a very large number of lakes (∼1000), but also by including remote sensing data such as indexes of vegetation density. In our study, data on drainage basin properties was derived from publicly available digital maps for geographic information system (GIS); hence the methodology should be applicable to (and relevant for) drainage basins globally.

2. Methods

2.1. Introduction

[7] This study is based on the data from a national lake survey of Norwegian lakes conducted by the Norwegian Institute for Water Research (NIVA) [Henriksen et al., 1998]. The survey included 1016 lakes, which were sampled for a wide range of water quality parameters, among them TOC, which is the focus of this study.18 lakes with missing values for one or more variables were omitted and this study comprises 998 lakes. Each lake was represented by a single sample taken at the time of autumn overturn (Figure 1, left). Most lakes were sampled in October 1995 with the exceptions sampled in the adjacent weeks. TOC concentrations (mg C L−1) were quantified by infrared detection of CO2 after catalytic high-temperature combustion A national lake register maintained by the Norwegian Water Resources and Energy Directorate (NVE) was used for stratified random selection of lakes by regions and lake size classes. The actual catchments of the sampled lakes cover only a small fraction of the total area of Norway, making it possible that the representations of variables are biased. Since both the sample and population distributions of predictor variables are available we were able to make a graphical comparison how well the sampled catchments represent the whole population (Figure 2).

Figure 1.

(left) Map of the Nordic countries showing the locations of the sampled lakes. (right) Insert shows a blowup of the red rectangle, showing REGINE drainage areas' boundaries (lines) and sampling locations (triangles).

Figure 2.

Kernel density smoothed probability distributions of the catchment property variables described in Table 1. Solid lines represent sample distributions for the catchments in this study, while dashed lines represent population distributions of all 20,000 catchments in the REGINE database covering the whole of Norway.

[8] The drainage area for each lake was delineated in a single digital polygon layer. Drainage area boundaries were based on NVE's catchment database, “REGINE”, which is a hydrographic register of approximately 20000 catchments covering the entire mainland of Norway. REGINE defines the major drainage basins, and subdivisions on lower hierarchical levels, as polygons on a digital feature map (Figure 1, right). When the REGINE unit on the lowest hierarchical level did not correspond to the sampled lake, the drainage area was corrected manually using a digital topographic map and a digital line feature map of rivers and streams as guide in the delineation process. The resulting layer, with drainage area delineations for all 998 lakes in the survey, was superimposed on other GIS layers to extract catchment averages (raster data, such as temperature, precipitation) or catchment area fractions (polygon features, such as land use classes). The same procedure was used to compile statistics for all REGINE units. Spatial analyses were performed with Hawth's Analysis Tools 3.27 (http://www.spatialecology.com/htools), which is an extension for ESRI ArcMap 9.2.

[9] A summary of the model variables are listed in Table 1. Theoretical water residence time (REST) was calculated as the ratio between model estimates of lake volume (m3) and total yearly runoff (m3 yr−1) from the drainage area. A separate data set acquired from NVE comprising 490 Norwegian lakes were used to make a simple linear model to estimate lake volumes from lake surface areas (see Figure S1 of Text S1). Terrain slope (SLOPE) was defined as the catchment maximum rate of change between adjacent cells in a 1 × 1 km digital elevation model of Norway, as computed by the ArcGIS tool “Slope”. Atmospheric N deposition (NDEP) was averaged for each catchment polygon from a digital map of yearly, accumulated total atmospheric N deposition (including dry deposition) for 1995. The N deposition map was constructed by spatial interpolation (kriging with a spherical semivariogram model) on 1° × 1° gridded output data from the Unified EMEP MSC-W modeling system (http://www.emep.int/). Normalized Difference Vegetation Index (NDVI) was calculated from satellite imaging data, based on the assumption that vegetation absorbs visible light and reflects light in the near-infrared. The difference between visible and near-infrared light is normalized by the total amount of light in the two channels. NDVI raster data was acquired as monthly composites (1992–1993) obtained from the USGS Eurasia Land Cover Characteristics database (http://edc2.usgs.gov/glcc/). Annual mean temperatures (TEMP) were averaged from the 1 × 1 km WorldClim raster on mean annual temperature for the period 1960–1990.WorldClim data sets are based on interpolated climate data from a worldwide set of weather stations. The data layers are averaged by month and interpolated to a 30 arc-second resolution grid (∼1 km2) [Hijmans et al., 2005] (http://www.worldclim.org/). Land use statistic was extracted from digital 1:50000 maps from the Norwegian Mapping Authority (Statkart). Land use classes included fraction of bog (BOG), arable land (ARABLE), and forest (FOREST); the remainder represented largely nonforested moor or grassland. Runoff (RUNOFF) was averaged for each catchment polygon using 1960–1990 averages of area-specific runoff on a Norwegian 1 × 1 km NVE raster map.

Table 1. Abbreviations, Descriptions, Units, Transformations, and Sources for Catchment Property Variables That are Used in the Regression Model in Table 2
VariableDescriptionUnitTransformationSource
TEMPAnnual mean air temperature°CelsiusNo transformationWorldClim, Hijmans et al. [2005]
RUNOFFThe precipitation or snowmelt that runs off the land into streams or other surface water.mm yr−1Log10Norwegian Water Resources and Energy Directorate (NVE)
NDVIAn indicator of terrestrial vegetation density based on satellite imagingIndex [0:200]No transformationMonthly NDVI composites from the USGS Eurasia Land Cover Characteristics database
NDEPDeposition of reactive nitrogen (N) species from the atmosphere to the biospheremg N m−2 yr−1Log10EMEP MSC-W modeling system
SLOPEMaximum rate of change in z axis between a cell and the neighboring cells in a digital elevation map.DegreeLog10Digital Elevation Model: Norwegian Mapping Authority (Statkart)
BOGFraction of watershed covered with bogs: Nonforested area with peat vegetation.PercentageArcsine square rootDigital 1:50000 Land Use Map (Statkart)
ARABLEFraction of watershed suitable to be farmed or cultivatedPercentageArcsine square rootDigital 1:50000 Land Use Map (Statkart)
RESTLake water residence time, estimated from lake and catchment areas and specific runoffYearLog10Statkart, NVE

2.2. Statistical Analysis

[10] Land use fractions were arcsine square root–transformed to stabilize variances and to make distributions more symmetrical. Strongly skewed, positive variables like REST, SLOPE, RUNOFF and NDEP were log transformed, while NDVI and TEMP were not transformed. The dependent variable (TOC) was log transformed in all models. All statistical analyses were performed with the R statistical computing environment version 2.6.1 (http://www.R-project.org). An initial screening phase removed highly correlated variables (e.g., temperature and altitude). Regression models were then constructed by stepwise backward elimination using the Bayesian Information Criterion (BIC) for model selection [Johnson and Omland, 2004]. An alternative model including first-order interactions between the independent variables was also investigated (data not shown). Since the BIC selection resulted in a model with eight interaction terms, but only achieved a slight increase in explanatory power, a model without interactions was chosen for further investigation. Independent variables in the final model of TOC concentrations (hereafter referred to as the N1k model) were TEMP, RUNOFF, NDEP, NDVI, REST and SLOPE as well as the catchment area fractions BOG and ARABLE (Table 2).

Table 2. Regression Coefficient Estimates, Standard Errors, t Values, and Significance Probabilities of N1k Model, Explaining 83% of the Variance in log(TOC) With a Residual Standard Error 0.22 on 952 Degrees of Freedoma
 EstimateStandard Errort ValuePr(>∣t∣)
  • a

    Variable names and definitions are given in Table 1.

Intercept−2.0200.255−7.9097.19e-15
NDVI0.0260.00119.759<2e-16
REST−0.0890.013−6.9447.05e-12
RUNOFF−0.3690.042−8.745<2e-16
TEMP0.0190.0053.5580.000392
ARABLE−0.3610.094−3.8050.000151
SLOPE−0.1100.227−4.8681.32e-06
NDEP−0.2090.026−7.9774.30e-15
BOG0.7840.51215.139<2e-16

[11] The predictor variables are in different physical units and vary on widely different numerical scales, obfuscating a direct comparison of the model coefficients. Quartiles of the predictor values represent representative high and low values within the ranges of the predictor variables. The relative changes in predicted TOC within the interquartile ranges of the predictor variables thus represent the contrast between “typical” low and high values of the independent variables, which allows an effective comparison of the signs and magnitudes of the different predictor effects.

[12] The REGINE data set of all catchments on the Norwegian mainland includes the total lake surface area within each REGINE unit, but not the distribution of the total lake surface area among lakes. Thus, the lake volume model (Figure S1 of Text S1) could not be applied to REGINE units and, consequently, it was not possible to estimate residence times (see Text S1). An alternative model omitting REST as a predictor variable (hereafter referred to as the REGINE model) was therefore used to estimate TOC concentrations for all REGINE catchments. TOC export was calculated as the product of the estimated TOC concentration at the catchment's outlet and yearly runoff from the respective catchments.

3. Results

[13] The catchment properties of the surveyed lakes differed in many respects. Most of the surveyed lakes were deliberately selected by having pristine catchments in the sense that there were no direct major anthropogenic impact and none or only marginal agricultural activity. Cultivated land constituted on the average only 0.5% of total area, while bogs covered on the average 5%. Forested areas represented 28% of total area, the rest being nonforested, either as a result of high elevation or latitude and/or being located in coastal areas without forest cover. Open water (lake surface) was on the average 12% of the total watershed area. The statistical distribution of the catchment areas was positive skewed covering 6 orders of magnitude from 0.003 to 3676 km2 with a median of 2.8 km2. Lake areas ranged from 0.02–365 km2 with a median of 0.13 km2. Most of the dependent variables had low correlations, with the exception that TEMP was correlated with NDEP and highly correlated with NDVI (Figure 3). All the major predictor variables showed distinct geographical patterns, with for example NDVI being high in lowland areas of the south and low in alpine and northern areas. RUNOFF ranged from 100 to 2000 L m−2 yr−1 with a regional variability mirroring the precipitation pattern: wet in the west and comparatively dry in the east. TEMP ranged from −5 to +8°C and followed the expected pattern with decreasing temperature as latitude or altitude increases. The distribution of BOG had strong positive skewness with median and mean values of 1.6% and 4.7%, respectively. This was expected as the distribution must be nonnegative while 22% of the REGINE units did not have any bog at all. ARABLE had an even more pronounced skewness with less than 0.5% coverage of the total area and the land use class being absent in 55% of the REGINE units. NDEP ranged from 100 to 2000 mg N m−2 yr−1. Atmospheric nitrogen deposition in Norway is mainly dependent on nitrogen emission from central Europe, which is reflected in NDEP showing a clear latitudinal gradient with high values in the south.

Figure 3.

Scatterplot matrix of all the eight predictor variables described in Table 1, with LOWESS smoothed trend lines below the diagonal and product-moment correlation coefficients above the diagonal.

[14] The final model of TOC concentrations (the N1k model) included eight predictor variables: (NDVI, REST, RUNOFF, TEMP, ARABLE, SLOPE, NDEP and BOG), which explained 83% of the variation in TOC (Table 2). The interquartile effect size of NDVI was striking (Figure 4), and is also reflected by this variable alone accounting for 60% of the observed variation in TOC. BOG was also positively correlated with TOC, though the effect size was less than for NDVI. TEMP had a positive, but relatively small, effect on TOC. RUNOFF, NDEP, REST and SLOPE all had negative interquartile effect sizes. RUNOFF had the strongest negative effect, followed by NDEP and REST. ARABLE has a statistically significant positive effect (Table 2), albeit an interquartile effect size of zero. This is due to the highly positive skewness of this variable with > 75% of the drainage areas having no arable land at all: when the interquartile range is zero, the interquartile effect size becomes zero too.

Figure 4.

Effect size plot for the N1k regression model for predicting log(TOC) listed in Table 2 (solid bars). Effect sizes are computed as regression coefficient times the interquartile range of each predictor variable. Dashed bars are equivalent effect sizes for the REGINE model without water residence time (REST) as predictor.

[15] The high explanatory power of the N1k model, together with general robustness revealed by multiple imputation, resampling, bootstrapping, and cross validation tests (see Text S1) justifies an extrapolation of the model to drainage areas outside the sample pool. Since the sample distributions of predictor variables were congruent with their nationwide population distributions (Figure 1), it is possible to apply the model to the whole REGINE catchment data set without extrapolating the model beyond its original domain. As explained in Text S1, REST could not be calculated for the REGINE data set. To be able to estimate TOC for all REGINE units, we also fitted an alternative model (the REGINE model) with residence time omitted from the independent variables. The REGINE model still explained 82% of the variation in TOC and revealed no apparent systematic deviation from the N1k model. Furthermore, the remaining predictor variable coefficients from the REGINE model had identical signs and were within the confidence limits of the original N1k model (Figure 4).

[16] Figure 5 demonstrates the result of applying the REGINE model to all Norwegian catchments. The variation in predicted TOC followed a clear geographic pattern, with high values in the southeastern and central regions of Norway. High TOC concentrations were also predicted for the inland regions of northern Norway. TOC flux follows a regional pattern that is qualitatively different from TOC concentration, with highest rates in coastal zones in southern Norway while the inland regions displayed relatively low flux rates (Figure 5). The sum of TOC export estimates on for all distinct drainage areas was 0.8 Mt TOC yr−1 represents a rough estimate of the total yearly flux of TOC from the Norwegian mainland, which translates to an average area specific flux of 2.4 g TOC m−2 yr−1.

Figure 5.

Maps showing geographical distribution of the eight predictor variables described in Table 2, as well as predicted TOC concentration (TOC) and TOC flux (FLUX). Color mapping is from low (green) to high (red) in all maps.

4. Discussion

[17] Given that the ultimate source of allochthonous C in lakes is primary production by terrestrial vegetation, it seems reasonable that the vegetation index NDVI turned out to be the most important predictor variable in the model. The overall importance of this parameter suggests that recent terrestrial primary production is a major source of aquatic TOC, which is also supported by studies based on other methods [McKnight and Aiken, 1998] Schiff et al., 1998]. This has also implications for the C use efficiency in the aquatic recipients since forest TOC contains a far higher fraction of low molecular weight compounds accessible to aquatic bacteria compared with TOC derived from bogs [Giesler et al., 2007]. Catchment area fraction of bog (BOG) was also a significant predictor in this model, but this presumably more recalcitrant pool of organic C explained far less of the TOC variation, with an effect size less than 50% that that of NDVI. This somewhat contrasts other studies from northern, boreal catchments, where peatlands by far have been the best predictor of TOC export [Dillon and Molot, 2005; Gorham et al., 1998; Kortelainen et al., 2006; Schiff, 1998], and the fraction of forested wetlands seem in general to serve as the best predictor of TOC export across a wide range of biomes [Lent et al., 1998]. There was a positive correlation between NDVI and BOG in our study, but it is relatively weak (r = 0.45) and with strong indications of heteroscedasticity. Bogs accumulate organic carbon due to the balance between production and mineralization as well as a low proportion of fixed C returned by respiration, hence peatland accumulation and export of C are competing processes [Frost et al., 2006]. The comparatively lower importance of peat relative to forest in the Norwegian catchments may simply reflect a wider variability of forest cover due to coastal and high-elevation areas compared to other boreal areas. It may also reflect a more variable topography and younger ecosystems (<8000 years since last glaciation) and thus less peat accumulation. The fraction of peatlands was on average was no more than 5% in the Norwegian catchments, while close to 30% in the Finnish survey [Kortelainen et al., 2006], and it is also noteworthy that average TOC export from the Finnish catchments was 6200 kg m−2 yr−1, compared to 2 400 kg C km−2 yr−1 for the Norwegian catchment. This should not necessarily be interpreted as a higher area-specific TOC export from peatlands than from forests, since the average forest cover also is substantially higher in Finnish (and most other boreal) catchments.

[18] Runoff was the major negative contributor to the TOC concentration. Although runoff is the main vector for transporting terrestrial organic carbon to lakes, it also act as a dilution agent since it originates from precipitation which is essentially devoid of organic carbon. If terrestrial TOC production was constant everywhere, the only effect of runoff on TOC concentration would be an inverse proportionality due to dilution. When using log(runoff) as the only linear predictor for log(TOC), the slope is −0.97 which is not significantly different from what is theoretically expected (−1), but the explanatory power is low (R2 = 0.32). However, the role of runoff in the N1k model shows more facets of the variable since the model coefficient (−0.37) signifies that the effect of runoff is less potent than that of a pure dilution agent. Runoff serves not only as a vector for allochthonous carbon in lakes; it is also a denominator in the equation for residence time, as reflected by the negative correlation between the two variables. As increasing residence time causes a decrease in TOC, this implies that runoff has dual antagonistic effects on TOC. The direct effect decreases TOC by dilution, while the indirect effect increases TOC by reducing the water residence time. The two opposing effects of runoff do not cancel out since the volume of water entering a lake is a product of area specific runoff and catchment area. Moreover, variability in lake volume, the nominator in the residence time formula, further blurs the relation between catchment runoff and residence time among lakes.

[19] The average catchment slope (SLOPE) had a negative effect on lake TOC concentration. This could be attributed a decreasing fraction of bogs in catchments with steeper topography and higher runoff (Figure 3). Both correlations point to a possible confounding between the hydrological variables. Still, catchment hydrological properties determine the pathway, contact time and infiltration ratio of water discharge, which all have the potential to influence the concentration of organic carbon [Andersson and Nyberg, 2009].

[20] A number of in-lake physical, photochemical and biological processes directly or indirectly remove organic carbon from the water column. Residence time (REST) determines the period TOC is subjected to in-lake removal processes such as sedimentation and permanent burial [Dillon and Molot, 1997]. It also affects the exposure to solar radiation, which directly mineralizes dissolved organic carbon via photooxidation [Granéli et al., 1996] and stimulates flocculation [von Wachenfeldt et al., 2008]. In addition, photobleaching mitigates the recalcitrance of organic carbon for microbial metabolism [Scully et al., 2002]. In-lake mineralization as a result of microbial metabolism has also been shown to be a significant predictor variable when modeling temporal DOC dynamics in lakes on a subannual scale [Futter et al., 2007]. It thus seems reasonable that the effect of residence time on lake TOC concentrations is real and not just confounded with other variables such as runoff.

[21] Temperature (TEMP) was a significant predictor variable for lake TOC concentrations, but the positive effect was relatively small. Temperature affects decomposition rates of litter fall and the production of dissolved organic carbon [Marschner and Bredow, 2002]. Studies also suggest that leaching of dissolved organic carbon increases as the temperature increases [McDowell and Likens, 1988] and that the leaching rate is exponentially related to temperature [Christ and David, 1996; Neff and Asner, 2001]. Temperature per se, determines the effective growth days for vegetation and is the ultimate cause for altitudinal and latitudinal tree limits [Kullman and Oberg, 2009]. The positive and strong correlation with NDVI and TEMP (r = 0.85; see Figure 3) indicates a possible confounding. The bootstrap analysis in Figure S4 of Text S1 shows that NDVI is robustly present in all models from resampled subsets while TEMP dropped out from 23% of the cases. This can be taken as an indication that the estimated effect size of TEMP is less reliable than the one for NDVI.

[22] The fraction of arable land (ARABLE) was a significant predictor variable in the model and was included in 82% of the models in the bootstrap analysis (Figure S4 of Text S1). Nevertheless, there was no effect of ARABLE within the interquartile range of the variable since less than 25% of the catchments had any arable land at all. This underlines the relevance of scale when interpreting the effect of predictor variables. In this case, at the scale of single catchments, the fraction of arable land is an important factor for the amount of TOC in lakes. Considering the entire area of Norway, where less 0.5% of the total area is arable, the importance of arable land becomes negligible. Thus, even though arability is a statistically significant factor, it is not ecologically significant at the nationwide scale.

[23] The mechanism behind the effect of nitrogen deposition (NDEP) on TOC concentration is not obvious. NDEP was positively correlated to both NDVI and TEMP, although this does probably reflect a geographical correlation rather than a causal relationship. Since N deposition stimulates terrestrial primary production, one would expect a positive effect of N deposition on TOC concentration in lakes. This has also been seen in laboratory studies where long-term NO3 additions increased the DOC export [Pregitzer et al., 2004] yet our data suggested a significant negative, albeit not very strong, net effect (cf. Figure 4) for which we are unable to offer any mechanistic explanation. The role of nitrogen deposition on terrestrial ecosystem carbon storage and cycling is a subject still under debate [Dijkstra et al., 2004; Hobbie, 2008; Keeler et al., 2009; Smemo et al., 2007]. In the words of Keeler et al. [2009, p. 1]: “Long-term nitrogen addition experiments have found positive, negative and neutral effects of added N on rates of decomposition”. Whether the effect of nitrogen deposition on lake TOC concentration is confounded with other variables (e.g., sulphate deposition [Monteith et al., 2007]) or has an effect per se, requires further investigation.

[24] The graphical representation of the nationwide predictions (Figure 5) revealed a clear regional pattern in TOC concentrations of inland waters. High concentrations are found in lakes of the central and southeastern parts of Norway, while low values are found at higher altitudes and in northern regions. Given that NDVI was the most dominant predictor variable, it was no surprise that the two patterns resembled each other. But the high values of NDVI along the coast in southern and central parts of Norway are not reflected in correspondingly high TOC concentrations. The cause for this discrepancy is offered by the high TOC fluxes in these areas (Figure 5). Even though the NDVI map suggests that the areas should have high TOC concentrations, it is probably diluted by the high runoff. In general TOC flux follows a regional pattern that is qualitatively different from the regional TOC concentration pattern (Figure 5). The high TOC values in the northern inland regions do not translate into high flux values since the runoff values are low. Vice versa, regions in the west with low TOC concentrations still have substantial flux values due to the intense runoff. Thus, it is not possible to deduce the TOC flux from TOC concentrations alone.

[25] The flux of TOC from all the REGINE catchments sums up to 0.8 Mt C yr−1, which corresponds to an area specific flux of 2.4 g C m−2 yr−1. The estimates of TOC flux were based on a single sample from each lake taken at the time of autumn overturn. While this in theory could imply both a seasonal or annual bias, and clearly the single sample approach calls for caution regarding the absolute TOC flux from the lakes, the relative flux pattern and drivers behind should be unaffected by the sampling regime. Different renewal rates could also influence the results, yet there is no easy way to scale sampling frequency with renewal rate for such a huge data set, and we feel confident that the sampling design combined with the large number of lakes and catchments provide a fairly robust prediction. Long-term monitoring (1990–2004) of 160 Norwegian and Swedish lakes have not revealed any pronounced oscillations in TOC, and with exception for the southernmost lakes where a slight upward trend (0.06–0.13 mg C L−1 yr−1) have been observed, TOC has remained stable over this period [Skjelkvale et al., 2007]. Hence we feel confident that the 1995 data are representative, and this was also confirmed by the separate testing of 112 lakes and catchments sampled 9 years later and not included in the original data set. Still, the model predicted TOC in these lakes with 78% accuracy (for comparison, see Figure S4 of Text S1).

[26] The catchments in this study span a wide range of variability with respect to geographic locations, size, altitude, climate, and vegetation cover. Still, eight publicly available and accessible variables were able to explain more than 80% of the variation in observed TOC concentrations. Several studies have used drainage area characteristics to model the concentration of organic carbon in lakes, but this study is the first which includes NDVI as a predictor variable. [Sobek et al., 2007] used a large number of globally distributed lakes to analyze factors influencing TOC concentrations in lakes. NDVI was not included in this study, but one of the significant predictor variables were soil carbon density, which may be closely correlated with NDVI [Puzachenko et al., 2006; Zhou et al., 2007]. NDVI can be used as a proxy for rates of primary production and several algorithms have been developed for estimating primary production from NDVI. The ultimate source of mobile organic carbon in soils, and hence allochthoneous carbon in lakes, is terrestrial primary production. This study suggests that the relationship between NDVI (and derivatives thereof) could have a general potential for increasing of the understanding of carbon cycling at the land-water interface. All TOC model predictor variables except SLOPE can be expected to be affected by climate change. Runoff and temperature are primary climate parameters, while NDVI is largely controlled by climate related variables (i.e., precipitation and temperature).

[27] We believe that the basic approach and model used in this study with perhaps some regional adjustments for parameter choice should be applicable not only to the boreal zone, but also other biomes. Not the least the use of remote sensing for obtaining detailed catchment properties, including refinements of NDVI to account also for forest characteristics and not merely forest density, should provide more accurate predictions of TOC export. Thus, future research in this field should aim to incorporate climate scenarios to construe the consequences of global warming on TOC concentrations in lakes and thus TOC exports to the ocean.

Acknowledgments

[28] We acknowledge the work of the Norwegian Institute of Water Research (NIVA), in charge of the original survey that the TOC data originated from. This work was covered by grants from the Norwegian research Council: to D.O.H., project 165139 “Biogeochemistry in Northern Watersheds, a Reactor in Global Change”; and project 185109 “Forecasting Ecological Effects of Climate Change: Integrating Functional and Correlative Models (FECIMOD)”; and to T. A., project 196336 “Biodiversity, Community Saturation, and Ecosystem Function in Lakes.”