Peat plateaus are widespread at high northern latitudes and are important soil organic carbon reservoirs. A warming climate can cause either increased ground subsidence (thermokarst) resulting in lake formation or increased drainage as the permafrost thaws. A better understanding of spatiotemporal variations in these landforms in relation to climate change is important for predicting the future thawing permafrost carbon feedback. In this study, dynamics in thermokarst lake extent during the last 35–50 years has been quantified through time series analysis of aerial photographs and high-resolution satellite images (IKONOS/QuickBird) in three peat plateau complexes, spread out across the northern circumpolar region along a climatic and permafrost gradient. From the mid-1970s until the mid-2000s there has been an increase in mean annual air temperature, winter precipitation, and ground temperature in all three study areas. The two peat plateaus located in the continuous and discontinuous permafrost zones, respectively, where mean annual air temperatures are below −5°C and ground temperatures are −2°C or colder, have experienced small changes in thermokarst lake extent. In the peat plateau located in the sporadic permafrost zone where the mean annual air temperature is around −3°C, and the ground temperature is close to 0°C, lake drainage and infilling with fen vegetation has been extensive and many new thermokarst lakes have formed. In a future progressively warmer and wetter climate permafrost degradation can cause significant impacts on landscape composition and greenhouse gas exchange also in areas with extensive peat plateaus, which presently still experience stable permafrost conditions.
 Peatlands in the northern circumpolar permafrost zone are important reservoirs of soil organic carbon, containing ∼277 Pg carbon [Schuur et al., 2008; Tarnocai et al., 2009]. Most of the perennially frozen peat carbon is found in northern boreal and subarctic peatlands where the combination of relatively thick peat deposits and extensive permafrost occurs. In these regions peat plateaus are common landscape features. Peat plateaus are perennially frozen peatlands that have been uplifted above the water table of the surrounding wetlands by frost heave [Zoltai and Tarnocai, 1975]. They have relatively flat surfaces and can cover several square kilometers. Peat plateaus are commonly interspersed with fens and thermokarst lakes, formed as a result of ground subsidence following thawing of ice-rich permafrost. Accelerated permafrost thawing in peat plateaus as a result of higher temperatures or increased precipitation can result in two alternative or parallel scenarios: (1) increase in thermokarst, such as formation of new ponds or thermal erosion along already existing lake shorelines, and (2) lake drainage, as changes in hydrology occur when the underlying or surrounding ground thaws. These landscape changes have significant impacts on the carbon-climate feedback. Thermokarst formation and lateral expansion causes remobilization of previously frozen peat carbon. From expanding lakes greenhouse gases, particularly methane (CH4), can be emitted to the atmosphere [Walter et al., 2006]. Drained lakes, on the other hand, can act as carbon sinks when fen vegetation begins to grow in the old lakebeds and renewed peat accumulation takes place [Payette et al., 2004].
 In peatlands thermokarst features have different forms and developmental sequences compared to mineral ground [Dredge and Nixon, 1979]. Collapse scars with fen vegetation can form in boreal peat plateaus. Their consequences for the carbon sequestration and release have been discussed by Camill [1999, 2005] and Turetsky . Palsas are a landscape feature found mainly in the isolated and sporadic permafrost zone. They are small, circular to elongated mounds of peat with a permafrost core [Zoltai and Tarnocai, 1975]. In palsa peatlands, where thermokarst features in the form of collapse scars and ponds are smaller, a few studies have been performed. In Scandinavia, Sollid and Sørbel  and Zuidhoff and Kolstrup  reported extensive permafrost collapse and increase in ponds from ∼1960 until 1997. From a palsa peatland east of Hudson Bay, Canada, a substantial increase in pond and fen area has been described by Laberge and Payette  and Payette et al. . To our knowledge, there are no published studies of thermokarst lake dynamics in peat plateaus.
 More studies have been performed in mineral soils. An increase in thermokarst has been recorded from the discontinuous permafrost zone in Alaska [Osterkamp et al., 2000] and east central Canada [Fortier and Aubé-Maurice, 2008], as well as from the continuous permafrost zone in Siberia [Smith et al., 2005a; Walter et al., 2006]. Other studies, however, suggest that lakes located in the discontinuous permafrost zone in Alaska and Siberia are shrinking or disappearing as a result of drainage [Yoshikawa and Hinzman, 2003; Smith et al., 2005a; Riordan et al., 2006]. A majority of these studies have used medium-resolution satellite images such as Landsat Multispectral Scanner (MSS) or Landsat Thematic Mapper (TM) which have a high spatial coverage and thereby are good for mapping larger areas. However, these data lack the high spatial resolution that is needed to detect changes in lake extent caused by erosion along shorelines or thaw slumping [Grosse et al., 2008]. In order to detect small-scale changes, including new formation of small ponds, images with much higher spatial resolution such as IKONOS or QuickBird are needed. For historical data the only available source is often panchromatic aerial photographs, which in most cases have a good contrast for land-water separation and a spatial resolution which is sufficient for mapping at meter scale. Peat plateaus and thermokarst lakes are particularly suitable for this type of analysis, because their edges are characteristically steep and high making lake area less sensitive to yearly variations in precipitation.
 In this study, we have performed a remote sensing time series analysis of historical panchromatic aerial photographs and modern high-resolution satellite imagery of three peat plateau sites spread out across the northern circumpolar region. The objective has been to quantify dynamics in thermokarst lake extent during the last ∼35–50 years along a climatic and permafrost gradient, and to relate the results to climatic and permafrost temperature trends. By increasing our knowledge about thermokarst lake dynamics in organic terrain under past and present climatic conditions, better predictions can be made for future dynamics in perennially frozen peatlands and their consequences for the carbon-climate feedback.
2. Site Descriptions
 Selected field sites are located in the Hudson Bay Lowlands (central Canada), Rogovaya (northeastern European Russia) and Tavvavuoma (northern Sweden). Despite the vast distances between the three studied peat plateau/thermokarst lake complexes, they represent similar ecosystems and share a number of common features (Figures 1 and 2). All three sites are located in landscapes with a low relief underlain by marine or glaciolacustrine deposits. The peatlands originally developed as permafrost free fens and were later in their succession transformed into bogs, which were uplifted above the surrounding water table by frost heave as permafrost aggradation took place [Kuhry, 1998; Oksanen et al., 2001; Kuhry, 2008; Hempel, 2009]. The peat depth is ∼1–2 m at all three sites [Dredge and Nixon, 1979; Oksanen et al., 2001], and the present vegetation is mostly xerophytic consisting of low shrubs (e.g., Ledum spp., Empetrum nigrum and Vaccinium vitis-idaea), mosses and lichens. Trees are absent from the peat plateaus. The only exception is occurrence of sparse, dwarfed black spruce (Picea mariana) and larch (Larix laricina) at the site in Hudson Bay Lowlands [Kuhry, 1998]. Thermokarst lakes are common, and block erosion of peat along lake margins is observed at all sites. Fens are found in drainage gullies, as narrow strips along the edges of thermokarst lakes and on drained thermokarst lake surfaces. In the fens Eriophorum spp., Carex spp. and wet growing Sphagna (e.g., Sphagnum riparium and S. balticum) are found [Kuhry, 1998; Oksanen et al., 2001]. Field observations indicate that thermokarst lake surfaces are generally located 1–2 m below the adjacent peat plateau surfaces (Figures 2a–2c). In Rogovaya and Tavvavuoma lake depths have been measured to less than 1 m near the peat plateau edges. In general, thermokarst lakes in peat plateaus are shallow and flat bottomed and do not exceed 2–3 m in depth [Dredge, 1979; Sjöberg, 2009].
 At all three sites the climate is subarctic with cool summers and cold winters [Christopherson, 2006]. However, there are some climatic differences between the sites. These have impacts on the ground temperatures that could affect the landscape responses to future climate changes, as will be described next.
2.1. Hudson Bay Lowlands
 The Hudson Bay Lowlands study area (57°53′N, 94°10′W) is an extensive flat plain in northeastern Manitoba, Canada, located at approximately 85 m asl and covered by a vast (>10,000 km2) peat plateau/thermokarst lake complex (Figures 1 and 2a). The peat plateau is located in the southernmost part of the continuous permafrost zone [Zoltai, 1995]. It is underlain by silty tills and marine silt that was deposited in the Tyrrell Sea which covered the area after the retreat of the Laurentide Ice Sheet around 9500 calibrated years B.P. [Dyke and Prest, 1987]. The low permeability of the soil, together with climatic conditions suitable for peat growth has facilitated extensive peatland formation across the lowland [Dredge, 1979]. According to a study of peatland development and permafrost dynamics, peat formation at McClintock, just south of our study site, began around 6800 calibrated years B.P. [Kuhry, 2008]. After the land had emerged from the sea, as a result of isostatic rebound, it was first colonized by fen peat, later followed by bog peat [Kuhry, 1998; Dredge and Mott, 2003]. The permafrost history at the McClintock site has been quite dynamic. The first permafrost aggradation took place around 2250 calibrated years B.P., but was shortly after followed by collapse and nonpermafrost fen conditions [Kuhry, 2008]. A second phase of permafrost development occurred ∼500 calibrated years B.P. and was, again, followed by partial permafrost degradation. The present peat plateau stage formed very recently [Kuhry, 2008]. At the McClintock site the peat depth is 1.7 m [Kuhry, 2008]. The general peat thickness in the Hudson Bay Lowlands study area is 1.5–2.0 m [Dredge and Nixon, 1979].
 In Churchill (58°44′N, 94°04′W), located at 29 m asl ∼90 km north of the study area, the mean annual air temperature (MAAT) is −6.9°C (1971–2000). The total mean annual precipitation (1971–2000) is 432 mm and the mean winter precipitation, November–April, is 118 mm (Environment Canada, http://www.climate.weatheroffice.ec.gc.ca/; Royal Netherlands Meteorological Institute, http://climexp.knmi.nl; Tu Tiempo, http://www.tutiempo.net/en/Climate/). At our study site, temperatures can be expected to be 0.9°C warmer based on regional interpolation (see section 3.5). The permafrost temperature in the peat plateau area is at present (2007–2009) below −4°C [Smith et al., 2010].
 The >20 km2 peat plateau/thermokarst lake complex in Rogovaya (67°16′N, 62°08′E), northeastern European Russia (Figure 2b), is located in the discontinuous permafrost zone and is bounded to the west by the Rogovaya River and to the east by the Pyatomboiyu River. The peat plateau complex is located at approximately 80–85 m asl on a glaciolacustrine plain that formed in the proglacial Lake Komi in the Early to Middle Weichselian [Mangerud et al., 1999]. According to Oksanen et al. , who performed a study of peat plateau development in the complex based on macrofossil analyses of peat profiles, peat formation in the area began around 10,600 calibrated years B.P. The permafrost history has been dynamic within the peat plateau complex [Oksanen et al., 2001]. The first permafrost aggradation took place ∼3300 calibrated years B.P., but was shortly after followed by permafrost collapse. Permafrost formation also occurred around 2100 calibrated years B.P., followed by degradation. A third period of permafrost aggradation in the peat plateau complex took place ∼600–100 calibrated years B.P. [Oksanen et al., 2001]. The average peat depth at the Rogovaya study site is 1.2 m (0.6–1.9 m, n = 4) [Oksanen et al., 2001].
 In Khoseda (67°05′N, 59°23′E), located at 84 m asl ∼120 km west of the study site the MAAT (1971–2000) is −4.9°C. The total mean annual precipitation (1971–2000) is 464 mm and the mean winter precipitation, November–April, is 162 mm (data from the Center for Hydrometeorological and Environmental Monitoring in the Komi Republic). Based on regional interpolation, the temperatures at the Rogovaya study site are expected to be 0.5°C colder than at Khoseda (see section 3.5). Permafrost temperatures in peat plateaus of the Rogovaya area are reported at around −2°C (S. Marchenko, University of Alaska Fairbanks, personal communication, 2010) [see also Romanovsky et al., 2010].
 Tavvavuoma in northernmost Sweden is located in the sporadic permafrost zone and has one of the most extensive peat plateau/thermokarst lake complexes in northern Fennoscandia (Figures 2c and 2d), even though the sizes of the peat plateaus are much smaller than in Hudson Bay Lowlands and Rogovaya. The study site (68°28′N, 20°54′E) is located in a broad, flat valley at 550–560 m asl. The ∼2 km2 large peat plateau has numerous small thermokarst lakes and fens and is surrounded by low hills with mountain birch (Betula pubescens ssp. czerepanovii) to the west and a large lake in the east.
 Radiocarbon dating and plant macrofossil analysis of one peat profile suggest that a fen developed at the site around 10,100 calibrated years B.P. [Hempel, 2009]. The transition to bog peat and the permafrost development was probably recent, occurring around 250–100 calibrated years B.P. [Hempel, 2009], suggesting that the peat plateau is of young age. Recent permafrost aggradation has been described from other peatlands in northern Scandinavia. Zuidhoff and Kolstrup  suggest that palsas in Laivadalen, northern Sweden, formed during the late Little Ice Age, around 270–20 calibrated years B.P. Late Holocene permafrost aggradation (∼600 calibrated years B.P.) is also reported by Oksanen , who suggests an initial period of permafrost development around 2500 calibrated years B.P. in a palsa peatland in northern Finland. The average peat depth at the Tavvavuoma study site is 1.6 m (1.1–2.1 m, n = 4).
 In Karesuando (68°27′N, 22°30′E), located at 327 m asl ∼65 km east of the study area, the MAAT is −1.9°C (1971–2000). The total mean annual precipitation (1971–2000) is 451 mm and the mean winter precipitation, November–April, is 152 mm (Swedish Meteorological and Hydrological Institute, http://www.smhi.se/klimatdata/meteorologi/). At our study site, temperatures can be expected to be 1.2°C colder based on a consistent bias between observations at Tavvavuoma and the Karesuando records during recent years (see section 3.5). The present (2007–2009) permafrost temperature at the interface between the peat and the underlying mineral substrate at 2 m depth is −0.3°C [Christiansen et al., 2010].
3.1. Image Data and Selection of Study Sites
 Across the northern circumpolar region three study areas were selected along a climatic-permafrost gradient, based on the availability of historical remote sensing imagery and some permafrost monitoring data. The three sites were very similar in microtopography, peat depth, long-term peatland development history and present vegetation. At all sites thermokarst lakes were common and erosion was active along parts of the shorelines. For the time series analysis high spatial resolution panchromatic remote sensing images were obtained from the three sites (Table 1). The Hudson Bay Lowlands data set comprised historical panchromatic aerial photographs from 1954 and 1974, obtained from the National Air Photo Library of Natural Resources Canada, and a QuickBird multispectral and panchromatic scene from 2006. For Rogovaya the oldest available panchromatic aerial photographs were from 1974 and a QuickBird multispectral and panchromatic scene was acquired in 2007. A CORONA satellite image from 1961, U.S. Geological Survey, was considered for the analysis, but because of the low spatial resolution it was only used for detecting overall changes in lake extent prior to 1974. The Tavvavuoma data set contained historical panchromatic aerial photographs from 1963 and 1975 obtained from Lantmäteriet (the Swedish mapping, cadastral and land registration authority) and a multispectral and panchromatic IKONOS scene from 2003. In the analysis of the QuickBird and IKONOS images only the panchromatic band was used in order to have a comparable data set over time. The image processing and analysis were performed using ENVI 4.5 and ENVI 4.7 geospatial image processing software from ITT Visual Information Solutions and ArcGIS 9, geographical information system software from ESRI.
Table 1. Remote Sensing Images Used in the Time Series Analysis of Thermokarst Lake Dynamics
 For the extensive peat plateau/thermokarst lake complexes in Hudson Bay Lowlands and Rogovaya a 2 km × 2 km study area with a characteristic lake distribution was selected. Areas affected by human impact were avoided. In Hudson Bay Lowlands, areas close to the railway were excluded as the railroad embankment probably has an impact on the surface hydrology which in turn affects the permafrost dynamics [Kuhry, 2008]. In Rogovaya, an area with artificially drained lakes was avoided. Within these 4 km2 study areas, 1 km × 1 km study areas were laid out in the center. In Tavvavuoma, the peat plateau complex is smaller than 4 km2 and therefore only a 1 km × 1 km study area was selected (Figure 3). For all three sites the aerial photographs were rectified and georeferenced to the respective satellite image using first-degree polynomial fit and cubic convolution resampling. The pixel resolution in all analyzed images was set to 0.6 m, since a test with multiple operators manually digitizing lake shorelines showed that oversampling of the historical aerial photographs (from 1.0 m to 0.6 m pixels) did not increase the relative uncertainty [Sannel and Brown, 2010].
3.2. Manual Delineation of Lakes
 According to a methods test by Sannel and Brown , semiautomatic remote sensing classification and delineation techniques resulted in misclassifications of both land and water. Instead manual digitalization, in combination with binary encoding of transects perpendicular to the lake margin, proved to be more accurate for high spatial resolution mapping of thermokarst lakes. Therefore, this latter technique was used in our study looking at small-scale changes in thermokarst lake extent over time. The labor-intensive nature of this manual digitalization necessarily limits the area that can be covered by these analyses. Within the 4 km2 and 1 km2 study areas at the three sites the shorelines of all lakes and ponds >6 m in diameter (>28 m2) were digitized, first in the aerial photographs and then in the satellite scene. For lakes that cross study area borders only the part of the lake inside the study area was included in the analysis. Fen vegetation growing along lake edges can make the land-water separation difficult [Sannel and Brown, 2010]. To aid the manual delineation, spectral profiles were measured across the lake shoreline at ∼50–100 m intervals, with care taken to approximately balance the number of water and land pixels. Binary encoding of the profiles identified an objective point of separation between land and water based on the mean of the spectral profile. Since it was most difficult to distinguish the exact location of the shoreline where lake margins were adjacent to fens, in bays and at spits, transects with binary encoding were used more frequently (at ∼50 m intervals) in these areas. Where lakes were surrounded by peat plateau and the shoreline was more distinct the distance between profiles was increased to 100 m or more [Sannel and Brown, 2010].
3.3. Thermokarst Lake Data Set
 The resulting “maps” contain information about the distribution of land versus water in the study areas. Since the aim of this project was particularly to study dynamics in thermokarst lake distribution in perennially frozen peat plateau areas, a second data set was produced, consisting of only the lakes where >75% of the shoreline was surrounded by peat plateau. In this data set all lakes and water bodies in permafrost free, fen-like areas were removed. The removal of lakes surrounded by fen also reduced the intra and interannual variability in lake extent, which can be a problem in time series analysis of thermokarst lakes with flat shorelines [Plug et al., 2008]. Thermokarst lakes surrounded by peat plateau mostly have steep lake margins, and thereby less variability in lake extent. The distinction of peat plateau and fen was made through visual interpretation of the panchromatic QuickBird and IKONOS images using the multispectral image as additional decision support. This classification was then applied to the aerial photographs, as the lower radiometric and spatial resolution and lack of multispectral data in the aerial photographs did not allow separation of peat plateau and fen with the same confidence. Peat plateaus were identified in the imagery by their absence of fen vegetation (which has low reflectance) and characteristic texture. A preliminary analysis of the aerial photographs, with reference to the satellite images, showed that some fen dynamics are evident in these data. However, QuickBird/IKONOS scenes improved our ability to exclude lakes that are not predominantly surrounded by peat plateau in the historical aerial photographs. The relative uncertainty in the manual digitalization of shorelines for thermokarst lakes predominantly surrounded by peat plateau was ±0.6 m [Sannel and Brown, 2010].
3.4. Quantification of Thermokarst Lake Dynamics
 To characterize the different sites, thermokarst lake parameters including number of lakes, lake perimeter and extent were calculated for all three study sites. Furthermore, changes in thermokarst lake extent were calculated for the different time intervals between image acquisitions. Approximate 95% confidence intervals for the mean change were derived under the assumption that the true area for one specific lake is certainly between a minimum and a maximum lake size using ±1.2 m buffer zones, and that the measurement error within these limits is uniformly distributed (this is a conservative error estimate since it assumes complete dependency between the errors all around the lake). The confidence interval was then calculated using normal quantiles, since mean values of independent (uniformly distributed) random errors are asymptotically normal distributed, according to the central limit theorem in probability theory.
 The distribution of fens in the QuickBird and IKONOS scenes, retrieved from supervised parallelepiped classifications of the panchromatic bands, was used for calculating rates of infilling with fen vegetation [Sannel and Brown, 2010]. Lateral expansion of thermokarst lakes and new lake formation was also quantified. For the raster images showing infilling with fen vegetation and lateral expansion during the time intervals between image acquisitions, 100% confidence intervals were calculated. The raster images were converted into vector format, and maximum and minimum infilling/expansion was calculated using ±1.2 m buffer zones around the polygons. Since it is theoretically possible that a maximum infilling/expansion of +1.2 m has occurred not only in the masked out areas but also along the remaining parts of all thermokarst lake perimeters, this was considered when calculating maximum infilling/expansion. For lakes that clearly showed changes in extent over time, maximum rates of infilling and lake expansion were calculated.
3.5. Climate Trends
 Monthly data from the nearby meteorological stations Churchill, Khoseda and Karesuando has been used to calculate trends in MAAT at the study sites during the time intervals between image acquisitions. To calculate the bias in MAAT for the study sites, regional interpolations of data from surrounding meteorological stations have been performed. For Hudson Bay Lowlands, temperature data (1980–1999) from Churchill (58°44′N, 94°04′W, 29 m asl) (Environment Canada, http://www.climate.weatheroffice.ec.gc.ca/; Royal Netherlands Meteorological Institute, http://climexp.knmi.nl) ∼90 km north of the study area and Gillam (56°21′N, 94°42′W, 145 m asl) (Royal Netherlands Meteorological Institute, http://climexp.knmi.nl) ∼170 km south of the study area was used. According to the interpolation, the site in Hudson Bay Lowlands is expected to have a MAAT that is 0.9°C warmer than in Churchill. For Rogovaya, temperature data (1980–1999) from Khoseda (67°05′N, 59°23′E, 84 m asl) (data from the Center for Hydrometeorological and Environmental Monitoring in the Komi Republic) ∼120 km west of the study site and Vorkuta (67°48′N, 64°01′E, 172 m asl) (data from the Center for Hydrometeorological and Environmental Monitoring in the Komi Republic) ∼85 km east of the study area was used for the interpolation, according to which the site in Rogovaya is expected to have a MAAT that is 0.5°C colder than in Khoseda. For the study site in Tavvavuoma (∼555 m asl) the MAAT is expected to be colder than in Karesuando (68°27′N, 22°30′E, 327 m asl) ∼65 km east of the study area because of the altitudinal lapse rate. This was also measured as a consistent bias in MAAT (2006–2008) of −1.2°C in Tavvavuoma (A. B. K. Sannel, unpublished data, 2010) compared to Karesuando (Swedish Meteorological and Hydrological Institute, http://www.smhi.se/klimatdata/meteorologi/).
 For the precipitation it was difficult to make interpolations between stations as the mean annual and winter (November–April) precipitation data, in contrast to the MAAT data, was very variable between the different stations. Therefore, the closest meteorological stations with a long precipitation record (Churchill, Khoseda and Karesuando) were chosen to calculate trends in mean winter precipitation at the study sites during the time intervals between image acquisitions. In general, the actual snow depth on top of the peat plateau surfaces can be expected to be far less than suggested by the winter precipitation data from nearby meteorological stations. Since peat plateaus have elevated, relatively flat surfaces and little vegetation trapping the snow, much of the snowfall is redistributed by winds into adjacent depressions such as lakes and fens (Figure 2d).
3.6. Ground Temperature Trends
 Limited data was available for calculating trends in ground temperature at the study sites. In the Hudson Bay Lowlands study area, the permafrost temperature in the mineral soil below the peat was −4.5°C in 1978 [Dredge, 1979]. During the International Polar Year (2007–2009) the permafrost temperature in the Wapusk peat plateau (57°48′N, 93°24′W) just southeast of the study site was below −4°C [Smith et al., 2010], suggesting that the increase in ground temperature has been relatively modest during recent decades. This fits well with the general trend from midwestern Canada (55°N–65°N, 100°W–110°W) reported by Zhang et al. , which suggests a warming of ∼0.05°C/yr between 1954 and 1973 and ∼0.02°C/yr from 1974 to 1994. In our study, the trend from 1974 to 1994 presented by Zhang et al.  has been extrapolated until 2005 in accordance with the continuous increasing trend in ground temperature for the whole of Canada [Zhang et al., 2006]. In Rogovaya, mean annual ground temperatures (MAGT) have been recorded at two locations in the peat plateau from 1983 to 2007, with a gap in the measurements between 1996 and 2005 (S. Marchenko, University of Alaska Fairbanks, personal communication, 2010). Since this data set showed a consistent bias in MAGT (1983–1995) of −1.0°C when compared to data from a peat plateau located ∼55 km east of our study area [Oberman and Mazhitova, 2001], the latter data set, adjusted for the mean bias, has been used to fill the gap in the data set for Rogovaya between 1974 and 1983. The combined data sets suggest ground warming of ∼0.03°C/yr between 1974 and 2007. In the peat plateau in Tavvavuoma ground temperatures have been recorded since 2005 [Christiansen et al., 2010]. For extrapolation back in time, the linear increasing trend of 0.04°C/yr at 1 m depth between 1980 and 2002, reported from permafrost peat mires around Abisko ∼90 km west of Tavvavuoma [Johansson et al., 2008], has been used.
4. Results and Interpretation
4.1. Lake Characterization and Dynamics
 Thermokarst lakes surrounded by peat plateaus (>75%) cover ∼13–36% of the studied areas (Table 2). The total limnicity (lake area/study area) is higher when lakes surrounded by permafrost free fens along >25% of the shoreline are included in the data set. In Hudson Bay Lowlands the median lake size is bigger than in Rogovaya and Tavvavuoma (Figure 3 and Table 2). However, the limnicity is much greater in Rogovaya because of a few, very large lakes. In Tavvavuoma the mean lake size is smaller than at the other two sites, but the total lake perimeter is greatest. The long perimeter could be caused by the smaller lake size, or by a greater complexity in lake shorelines or lake shapes.
Table 2. Lake Characteristics at the Three Study Sites
 At all three sites both lateral expansion and infilling with fen vegetation along lake margins has been going on during the time intervals between the images (Figure 4). Infilling with floating vegetation mats has been observed in Rogovaya [Oksanen et al., 2001] and in Tavvavuoma, but infilling can also be caused by marginal strips of vegetation rooted in shallow bottoms. When comparing trends in lake extent during the period from the mid-1970s to the mid-2000s Rogovaya shows more or less stable conditions, with both limited infilling and lateral expansion (Table 3). In Hudson Bay Lowlands lateral expansion has been a little more extensive (∼1.4% per decade). Both in Rogovaya and Hudson Bay Lowlands lateral expansion has primarily occurred along shorelines of larger lakes. Within the 2 km × 2 km study areas no thermokarst lakes have drained in Hudson Bay Lowlands and Rogovaya since 1954 and 1974, respectively. Despite the relatively poor quality of the CORONA satellite image from 1961, it is also evident that lake drainage has not occurred in Rogovaya between 1961 and 1974. The most significant change in lake extent has occurred in Tavvavuoma, where infilling with fen vegetation has been extensive (∼7.6% per decade). Here, the infilling has not only taken place along shorelines but in a number of lakes drainage has facilitated more or less complete revegetation of former lakebeds (Figure 4). The highest lateral rate of infilling with fen vegetation (39 m/decade) was recorded from a lake in Tavvavuoma that drained between 1963 and 1975 (Table 3). The temporal resolution of the images does not allow for a more precise dating of the drainage event, but if the drainage was catastrophic the maximum infilling rates presented in Table 3 are probably underestimations as fen vegetation most likely occupied the drained lakebeds in only a few years time.
Table 3. Changes in Thermokarst Lake Extent, Infilling With Fen Vegetation and Lateral Erosion at the Three Study Sites During the Time Intervals Between Image Acquisitionsa
Change in Thermokarst Lake Area (95% Confidence Intervals) (m2)
Change in Thermokarst Lake Area (%/Decade)
Infilling With Fen Vegetation (100% Confidence Intervals) (m2)
For all study sites the changes in lake area were statistically significant (p < 0.001), except for in the 2 km × 2 km area in Rogovaya (p = 0.63). Here nc, not calculated (as fen vegetation was difficult to identify in the aerial photographs).
Excluding newly formed lakes.
For a lake that drained between 1975 and 2003.
For a new pond that formed between 1975 and 2003.
Excluding one lake (28,662 m2) where lateral expansion over fen vegetation occurred between 1963 and 1975.
 One lake in Tavvavuoma that drained between 1975 and 2003 expanded laterally over fen vegetation from 1963 to 1975, with a maximum expansion rate of 25 m/decade. The expansion could potentially be related to interannual variability in lake extent caused by variations in precipitation, which plays an important role for lake extent in flat terrain [Plug et al., 2008]. During the 12 month period prior to image acquisition in 1975, the precipitation in Karesuando was higher (542 mm) than the mean annual precipitation value of 451 mm (1971–2000) (Swedish Meteorological and Hydrological Institute, http://www.smhi.se/klimatdata/meteorologi/). However, it is not likely that the fen vegetation would be flooded by this relatively moderate increase in precipitation. An alternative explanation could be that the expansion was caused by frost heave in the surrounding peatland area resulting in new damming. In 1975 the lake area had expanded so that the lake margins were surrounded by peat plateau along >75% of the shoreline. The lake was therefore included in the thermokarst lake data set when calculating changes in thermokarst lake extent between 1975 and 2003, but excluded when calculating the extent of lateral erosion and changes in thermokarst lake extent from 1963 to 1975 and 1963–2003 (Table 3).
 In Tavvavuoma (1 km2) many new small lakes have formed, both between 1963 and 1975 and 1975–2003 (Table 4). In the 1 km2 study area in Hudson Bay Lowlands and Rogovaya no new lakes have formed from the mid-1970s to the mid-2000s. However, in the 4 km2 study areas limited new lake formation has occurred. In all three study areas the new lakes that formed were small (mean lake size <140 m2).
Table 4. Characteristics of New Thermokarst Lakes That Have Formed During the Time Intervals Between Image Acquisitions at the Three Study Sites
Number of New Thermokarst Lakes
Total Lake Area (m2)
Mean Lake Area (m2)
Total Perimeter (m)
Mean Perimeter (m)
Hudson Bay Lowlands 2 km × 2 km
Hudson Bay Lowlands 1 km × 1 km
Rogovaya 2 km × 2 km
Rogovaya 1 km × 1 km
Tavvavuoma 1 km × 1 km
4.2. Climate Trends
 During the time period 1971–2000 the MAAT was higher in Tavvavuoma (−3.1°C) than in Rogovaya (−5.4°C) and Hudson Bay Lowlands (−6.0°C). The total annual winter precipitation (November–April) was lowest in Hudson Bay Lowlands (118 mm) and relatively similar for Tavvavuoma (152 mm) and Rogovaya (162 mm). All three study sites show similar trends with increasing MAAT and winter precipitation (November–April) during the time intervals between image acquisitions in the mid-1970s and mid-2000s (Table 5 and Figure 5). The MAAT has increased with ∼1°C in Hudson Bay Lowlands and Rogovaya and by ∼2°C in Tavvavuoma. Hudson Bay Lowlands and Tavvavuoma have experienced a ∼60–70% increase in winter precipitation during the time interval. Even if excluding the unexpectedly high value recorded in Hudson Bay Lowlands (Churchill) in 2003, the increase is 28% (the high 2007 value is outside the period of analysis). In Rogovaya the increase in winter precipitation is ∼6% from the mid-1970s until the mid-2000s.
Table 5. Mean Annual Air Temperature, Mean Annual Winter Precipitation (November–April), and Approximate Permafrost Temperature, With Linear Trends, at the Three Study Sites During the Time Intervals Between Image Acquisitions
Mean Annual Air Temperature (°C)
Trend in Mean Annual Air Temperature (°C/Decade)
Mean Winter Precipitation November–April (mm)
Trend in Mean Winter Precipitation November–April (%)
Data from Hudson Bay Lowlands (1978) [Dredge, 1979] and general ground temperature warming in midwestern Canada (55°N–65°N, 100°W–110°W) of 0.047°C/yr (1954–1973) and 0.015°C/yr after 1974 [Zhang et al., 2005].
Data from Khoseda (data from the Center for Hydrometeorological and Environmental Monitoring in the Komi Republic) adjusted to represent Rogovaya.
Data from Khoseda (data from the Center for Hydrometeorological and Environmental Monitoring in the Komi Republic).
 In Tavvavuoma the trends with increasing MAAT and winter precipitation already occurred during the time interval 1963–1975, but were interrupted by some cooling in the mid-1970s (Figure 5). In Hudson Bay Lowlands the MAAT was stable (around −6.5°C) between 1954 and 1974, and during the same time interval the winter precipitation (November–December) decreased by ∼55%, promoting stable permafrost conditions.
4.3. Ground Temperature Trends
 In Hudson Bay Lowlands the permafrost temperature below the peat was −4.5°C in 1978 [Dredge, 1979]. The trends with increasing ground temperature presented by Zhang et al. [2005, 2006] suggest that the permafrost temperature in Hudson Bay Lowlands peat plateaus was around −5.5°C in 1954 and −4.1°C in 2005. This fits well with recent data showing that the present permafrost temperature in the area is below −4°C [Smith et al., 2010]. The mean annual ground temperature in the peat plateau in Rogovaya was −3.7°C at the depth of zero annual amplitude in 1974. By 2007 the temperature had increased to −2.1°C [Romanovsky et al., 2010; S. Marchenko, University of Alaska Fairbanks, personal communication, 2010]. MAGT measurements in Tavvavuoma show that at present (2007–2009) the permafrost in the peat plateau is near thawing. At 1 m depth the MAGT is −0.3°C [Christiansen et al., 2010]. Ground temperature trends can be expected to be similar in Tavvavuoma and Abisko because of the short distance between the two sites. When using the same linear trend in MAGT for Tavvavuoma that has been reported from the peat plateau area in Abisko [Johansson et al., 2008], the temperature at 1 m depth should have been −1.6°C in 1975 and −2.1°C in 1963, supposing that the trend can be extrapolated prior to 1980 (Table 5). However, it is also possible that the permafrost temperature in Tavvavuoma has been close to 0°C for some time. In a peat plateau located in sporadic permafrost in west central Canada (northern Alberta) where the ground temperature was around −0.2°C, Smith et al. [2005b] observed no increasing trend in permafrost temperature from 1989 to 2002, despite a warming trend in air temperature. The absence of a warming trend when permafrost temperatures are close to 0°C is probably caused by absorption of latent heat required for phase transition [e.g., Riseborough, 1990; Smith et al., 2010]. Based on available data, the offset between MAAT and ground temperature is ∼2°C at all our study sites.
4.4. Linking Lake Dynamics to Climate and Ground Temperature Trends
 At our three study areas, the mean winter (and summer) precipitation is similar for all sites, suggesting that the observed ground temperature differences are more dependent on MAAT values. According to Johansson et al.  there is a statistically significant correlation between MAAT and ground temperature. The trends in ground temperature for all three study regions show a similar pattern with increasing temperatures from the mid-1970s to the mid-2000s. Since there are gaps and uncertainties in the ground temperature data sets from our three study sites we have instead looked for correlations between MAAT and lake dynamics parameters.
 In the 4 km2 study areas in Hudson Bay Lowlands and Rogovaya, where the MAAT has been colder than −5°C (Table 5), new thermokarst lake formation has been very limited from the mid-1970s until the mid-2000s (Figure 6a). Between 1954 and 1974, when the MAAT was below −6°C, no new lakes formed at the Hudson Bay Lowlands site. In Tavvavuoma, where the MAAT was around −3°C (Table 5), formation of new thermokarst lakes has been relatively extensive both from 1963 to 1975 and 1975–2003 (Figure 6a).
 The relationship between MAAT and infilling with fen vegetation at the three study sites, from the mid-1970s until the mid-2000s, looks very similar. In Hudson Bay Lowlands and Rogovaya where the MAAT was colder than −5°C the infilling has been limited, whereas extensive infilling as a result of lake drainage has occurred in Tavvavuoma where the MAAT was around −3°C (Figure 6b). Field observations of the drained lakes in Tavvavuoma show that the drainage has occurred horizontally, as a result of lateral thawing along lake margins. After 2003 (2003–2008) extensive lake drainage has not taken place within the 1 km2 study area, according to field observations.
5.1. Changes in Lake Extent Over Time
 The time series analysis of remote sensing images from Hudson Bay Lowlands and Rogovaya shows relatively small changes in thermokarst lake extent from 1974 to 2006/2007. This result is coherent with a previous study of thermokarst lake dynamics in mineral soil in the continuous permafrost zone in Alaska by Hinkel et al. , who concluded that the landscape was stable and that drainage of lakes was very limited from the mid-1970s to around 2000. Smith et al. [2005a] suggested that climatic warming initially causes thermokarst expansion, which is later followed by lake drainage as permafrost degradation continues. This could explain why several studies have described lake expansion in the continuous permafrost zone from the mid-1970s to ∼2000 [Smith et al., 2005a; Walter et al., 2006], whereas lakes have been reported to shrink or disappear from the discontinuous permafrost zone [Yoshikawa and Hinzman, 2003; Smith et al., 2005a; Riordan et al., 2006]. Because of the different thermal properties of organic compared to mineral soil [e.g., Rinke et al., 2008], permafrost peatlands could be expected to respond slower to changes in the climate.
 Horizontal expansion of thermokarst lakes is caused by wave cut erosion and/or thaw slumping [Pelletier, 2005]. Thaw slumping can be triggered by rapid thawing or unusually high summer temperatures [Pelletier, 2005]. In our study the most extensive lateral expansion along lake margins from the mid-1970s to the mid-2000s occurred in large lakes (>20,000 m2) in Hudson Bay Lowlands and Rogovaya. Larger lakes are more likely to experience erosion as they have a longer fetch and thereby higher potential wave energy. In our study covering the last few decades, the maximum recorded lateral expansion rate was 7.3 m/decade in Hudson Bay Lowlands. This is lower than the expansion rate of ∼50 m/decade suggested by Hinkel et al.  for lakes in mineral soils in northern Alaska during the Middle to Late Holocene, but higher than average modeled expansion rates of 1.0–2.6 m/decade under present climatic conditions for lakes in mineral soils in northern Alaska and Yukon [Plug and West, 2009]. Also lake coalescence was most frequently occurring in Hudson Bay Lowlands and Rogovaya, possibly because coalescence is favored by a high lake density [Hinkel et al., 2007].
 Formation of new thermokarst lakes was most frequent in Tavvavuoma, where the MAAT and MAGT were highest. However, in all three study areas the newly formed lakes were small, with mean lake sizes of ∼60–140 m2. Possibly, the time intervals between image acquisitions have not been long enough for larger lakes to develop. These small-scale landscape changes (new lake formation, lateral erosion) would not be possible to detect by means of more commonly used satellite sensors with a much lower spatial resolution.
 Another type of small-scale landscape dynamics which requires high-resolution data and analysis is infilling with fen vegetation along thermokarst lake shorelines. This phenomenon has been recorded at all three study sites and is a part of the natural succession and dynamics in peat plateau/thermokarst lake complexes. However, the accelerated infilling rates which have been recorded in Tavvavuoma are coherent with the observed trend of shrinking ponds elsewhere in the sporadic and discontinuous permafrost zones [e.g., Yoshikawa and Hinzman, 2003; Smith et al., 2005a]. The reduction in lake extent may be caused by increased evapotranspiration during warmer and longer growing seasons, or by increased drainage as the permafrost thaws [Riordan et al., 2006]. In the western Canadian Arctic, rapid or catastrophic drainage of thermokarst lakes in mineral soils occur frequently [Marsh et al., 2008]. The drainage can either occur laterally in surface channels or by subsurface tunnel formation [Mackay, 1981, 1988; Marsh and Neumann, 2001]. In Tavvavuoma the extensive reduction in lake extent from 1963 to 2003 does not seem to be caused by thawing of the underlying permafrost. Field observations show that the lake drainage has occurred laterally.
5.2. Linking Changes in Thermokarst Lake Extent to Climate Trends
 It has been debated whether temperature or precipitation is the most important factor influencing permafrost degradation. Jorgenson et al.  argue that increased summer temperatures in combination with increased summer precipitation have caused the observed increase in permafrost degradation in Alaska between 1982 and 2001. Brewer et al.  suggested that lake drainage in Alaska was primarily due to increased precipitation and that increased temperature was not so important. In Scandinavia Johansson et al.  found that increases in ground temperature from 1980 to 2002 were correlated to increases in air temperature and summer precipitation. Conversely, Payette et al. , Agafonov et al.  and Hinkel and Hurd  claim that increased snowfall is the main climatic driver for thermokarst formation. In our study, the climatic trend at all three sites have been very similar during the time intervals between image acquisitions in the mid-1970s and the mid-2000s with both increasing MAAT and winter precipitation (November–April). During the same time intervals the summer precipitation (June–September) shows an increasing trend for Hudson Bay Lowlands, whereas no increase is recorded for Rogovaya and Tavvavuoma. This suggests that summer precipitation is probably not the most important factor causing the permafrost degradation in Tavvavuoma. In Hudson Bay Lowlands and Rogovaya no lake drainage and only limited new lake formation and lateral expansion along lake margins have occurred, indicating that the increases in temperature and snowfall have not been so extensive as to cause rapid and extensive permafrost degradation. However, permafrost thawing in peatlands is a slow process and there can be a significant time lag in the response to changes in climate [Halsey et al., 1995].
 From the mid-1970s to the mid-2000s the peat plateau complex in Tavvavuoma has experienced the most extensive new lake formation and lake drainage. Local microtopography, peat depth and vegetation cover are similar compared to Hudson Bay Lowlands and Rogovaya. The most pronounced difference between the three study sites (1971–2000) is the >2°C higher MAAT in Tavvavuoma compared to in Rogovaya and Hudson Bay Lowlands, which in turn affects the permafrost temperature (at present close to 0°C in Tavvavuoma compared to approximately −2°C in Rogovaya and −4°C in Hudson Bay Lowlands). The lower permafrost temperature in Hudson Bay Lowlands could be a result of a combination of low MAAT and low winter precipitation (1971–2000) in November–April. However, the difference in permafrost temperature between Tavvavuoma and Rogovaya can only be explained by the differences in air temperature as the winter precipitation is higher in Rogovaya (162 mm/yr) than in Tavvavuoma (152 mm/yr). Even though local variability in winter precipitation can be expected at the study sites, the snow depth at the uplifted and flat peat plateau surfaces is in all cases expected to be relatively shallow as a result of snow drift. Considering the large differences in thermokarst formation and drainage between Tavvavuoma on the one hand and Rogovaya and Hudson Bay Lowlands on the other hand there seems to be a threshold value in MAAT, and corresponding permafrost temperature, above which permafrost degradation significantly starts to affect lake extent and landscape patterns. According to our study, the critical threshold in MAAT is between −3°C and −5°C. This fits well with the threshold value of −3.5°C suggested by Halsey et al.  for permafrost degradation and collapse scar formation in forested peat plateaus in the discontinuous permafrost zone in boreal continental western Canada. In palsas, which are smaller permafrost landscape features located mainly in the sporadic and isolated permafrost zones, permafrost degradation has been extensive during the last ∼50 years [Laberge and Payette, 1995; Sollid and Sørbel, 1998; Zuidhoff and Kolstrup, 2000; Payette et al., 2004], further corroborating our observations. If the trend with increasing MAAT and winter precipitation continues in the future, there is a potential that extensive ground subsidence, lake expansion and lake drainage will occur also in peat plateau areas which are at present still experiencing stable permafrost conditions.
5.3. Consequences for the Carbon-Climate Feedback
 During the last deglaciation CH4 ebullition from newly formed thermokarst lakes was a major contributor to the high-latitude increase in atmospheric CH4 concentration [Walter et al., 2007]. Rapid and extensive development of northern peatlands during the Early Holocene caused a peak in the atmospheric CH4 concentration and a decrease in the carbon dioxide (CO2) concentration, suggesting that peatlands can have important impacts on the global carbon cycle [Yu et al., 2003; Smith et al., 2004; MacDonald et al., 2006]. Since permafrost peatlands store large quantities of soil organic carbon there is a substantial potential for increased greenhouse gas emissions under warmer climatic conditions [Schuur et al., 2008; Tarnocai et al., 2009].
 Surface hydrology is a key factor for greenhouse gas exchange in peatlands [Charman, 2002; Christensen et al., 2004; Johansson et al., 2006]. Thawing of permafrost as a consequence of increased warming can result in alternative or parallel scenarios for the carbon balance [e.g., Turetsky, 2004; Tarnocai, 2006]: (1) increased thermokarst creating a wetter landscape with more lakes or collapse scar fens, from which higher CH4 emissions can be expected as a result of increased anaerobic decay of newly available thawed out peat deposits; (2) drier surface conditions on the peat plateaus, as a result of increased evapotranspiration in combination with gradually increased active layer depth, causing increased CO2 emissions as a result of increased aerobic decay, but also increased fire frequency; and (3) renewed peat accumulation in drained lakes, causing an increase in the carbon uptake; however, it is likely that renewed peat accumulation can only compensate for the initial rapid loss of carbon from decomposition and leaching of thawed out peat deposits on a long-term basis.
 In Rogovaya only minor changes in lake extent have occurred from the mid-1970s to the mid-2000s, even though both lateral expansion and infilling with fen vegetation has occurred along lake margins. The carbon loss due to thaw slumping and collapse of the peat plateau can be expected to be more extensive than the carbon sequestration by photosynthesis in the expanding fen vegetation. CH4 emissions are expected to be particularly high close to newly eroded shorelines [Walter et al., 2006]. In Hudson Bay Lowlands the lateral expansion has been more widespread than the infilling. From 1954 to 2006 the thermokarst lake area has increased by ∼5%, exposing only moderately decayed organic carbon that had previously been locked in the permafrost to anaerobic decay. In Tavvavuoma new thermokarst lakes have formed, also exposing old soil organic carbon to anaerobic decay resulting in increased emissions of CH4 to the atmosphere. The extensive lake drainage that has occurred from 1963 to 2003 has possibly mitigated the climatic forcing, as increased uptake of CO2 can be expected from the revegetated drained lakebeds. On the other hand the recently drained lakebeds are still wet fens, from which high CH4 emissions can be expected [Christensen et al., 2004; Johansson et al., 2006]. The interactions between climate, permafrost, and surface hydrology are very intricate and it is therefore difficult to predict future responses of peatland ecosystems and their carbon balance to changing conditions.
 Peat plateau/thermokarst lake complexes are a common landscape feature in the sporadic to continuous permafrost zones, indicating that thermokarst is a common process across permafrost regions. The results of this study suggest that climatic warming and subsequent ground temperature increases can induce a rapid destabilization of these landforms when approaching critical thresholds. These are now being reached in more southern isolated and sporadic permafrost zones, but progressively warmer conditions over the next century could also affect the discontinuous and southern continuous permafrost zones where perennially frozen peat deposits are particularly abundant.
 The peat plateau in Tavvavuoma, located in the sporadic permafrost zone, has experienced both extensive lake drainage and new thermokarst formation in recent decades. During the same time interval there has been an increase in MAAT and snowfall, and the permafrost temperature has increased and is at present close to 0°C. Hudson Bay Lowlands and Rogovaya, located in the continuous and discontinuous permafrost zones, respectively, have also had increasing trends in MAAT, winter precipitation and permafrost temperatures. However, at both these sites, where the present permafrost temperature is below −2°C, changes in lake extent and landscape dynamics are less pronounced. In these peat plateau/thermokarst lake complexes limited expansion of fen vegetation along lake margins is almost as extensive as the moderate rates of erosion and thaw slumping, making the net changes in lake extent over time small.
 The results suggest that our peatlands located in the continuous and discontinuous permafrost zones where the MAAT is below −5°C have not been experiencing extensive variations in lake extent during the last 35–50 years, even though lateral expansion caused by erosion and infilling with fen vegetation has occurred along lake margins. In the peatland located in the sporadic permafrost zone, where the MAAT is around −3°C, extensive lake drainage and infilling with fen vegetation has taken place and at the same time many new thermokarst lakes have formed.
 In a future warmer and wetter climate permafrost degradation can cause significant impacts on landscape patterns and carbon exchange also from the extensive peat plateaus which at present still experience stable permafrost conditions. More studies of thermokarst lake dynamics and related landscape changes under changing climatic conditions are needed to better understand the complex interactions between climate, permafrost, ecosystems and surface hydrology. The thawing permafrost carbon feedback needs to be integrated in model projections of future climate and environmental change. Present-day climate models have a relatively coarse spatial resolution. To incorporate key periglacial processes occurring at the landscape level and transient landscape changes related to permafrost thawing are important challenges for the future.
 Financial support was provided by the Swedish Society for Anthropology and Geography, Helge Ax:son Johnson Foundation, Ahlmann Foundation, Foundation Ymer-80, and the Knut and Alice Wallenberg Foundation for acquisition of satellite images, aerial photographs, and remote sensing software. Jan-Olov Persson at the Statistical Research Group, Stockholm University, kindly provided assistance with the statistics.