Landsat scenes spanning 1978–2001 were used to classify thermokarst lake area and determine changes in lake coverage on the Tuktoyaktuk Peninsula in northwestern Canada. Changes in total lake area between scenes were substantial, spanning a 14% increase and 11% decrease, mostly owing to changes in the area of lakes >1.3 km2 in area. Increases in total lake area occurred primarily between 1978–1992, and decreases between 1992–2001. Differences in total lake area between scenes from different years depend strongly on cumulative precipitation in the 12 month period preceding scene acquisition (r2 = 0.82), and not on summer or mean annual air temperature (MAAT). Our results indicate that precipitation is the primary factor influencing the areal extent of lakes as detected by remote sensing. On the Tuktoyaktuk Peninsula, persistent lake area changes (if any) such as might occur by regional warming or changes in evaporation/precipitation balance over decades, are well-masked by short-term climatological changes.
 Landsat-7 imagery of the Alaskan Arctic Coastal Plain has been used to measure geometric properties of lakes and drained basins [Hinkel et al., 2005] and their areal extent, which is 20–40% [Frohn et al., 2005]. Cote and Burn  measured the current size, shape and orientation of 578 lakes on the Tuktoyaktuk Peninsula. Duguay et al.  used 1986–1999 Landsat data for the Old Crow Flats (Yukon), finding increases and decreases in individual lakes, but an insignificant change in overall lake extent. Smith et al.  used low resolution satellite data to measure changes between 1973 and 1998 for >10,000 large lakes across 515,000 km2 of Siberia. Total lake area declined by 6% but, in continuous permafrost, a 12% increase in total lake area and a 4% increase in lakes larger than 40 ha was observed. Rising Arctic temperatures might cause these changes, with lakes in continuous permafrost increasing in area owing to thermokarst acceleration, whereas lakes in discontinuous permafrost may be shrinking owing to thawing of underlying impermeable permafrost [Smith et al., 2005]. Riordan et al.  conducted change studies of lakes and ponds at selected sites across Alaska, finding a general decrease from 1950–2002, but only negligible change between 1954, 1978 and 1999 for the thaw lake region in continuous permafrost on the Arctic Coastal Plain.
 Here, we examine changes in lake cover from 1978–2001 for the Tuktoyaktuk Peninsula, NW Canada, a region of continuous permafrost and particularly rapid recent warming. Mean annual air temperature has increased by 1.7°C over the past century, the greatest recorded increase in Canada [Environment Canada, 2007]. We attempt to distinguish long term (decadal) change, if any, from changes in lake area and number caused by intra-annual and inter-annual variability in precipitation and temperature.
2. Tuktoyaktuk Peninsula
 Numerous field studies have examined permafrost and thermokarst features of the western Canadian Arctic region [Mackay, 1956, 1981, 1988, 1997; Rampton, 1988; Burn, 2002; Murton and Bateman, 2007]. The Tuktoyaktuk Peninsula is approximately 250 × 50 km. Relief is typically <60 m A.M.S.L and much of the landscape is poorly drained. Permafrost is 200–600 m deep [Rampton, 1988]. Mean annual air temperature is currently ∼−10°C (over the past 40 y at the Tuktoyaktuk meteorological station: Environment Canada ). Annual precipitation averages 140–160 mm, ∼50% of which is snow. Rainfall is highest in August and late July, normally as sustained low-intensity storms. Unconsolidated Pleistocene sediment is thick over most of the Peninsula and bedrock outcrops rare [Rampton, 1988]. The northeast part of the peninsula is an area of unconsolidated Pleistocene outwash sands and silts, which in part may be low in excess ice content [Murton and Bateman, 2007]. The lower portion of the peninsula was glaciated during the last ice advance [Rampton, 1988]. Lakes, many of thermokarst origin, cover 20–40% of the peninsula and adjoining mainland [Mackay and Burn, 2002]. As with other regions considered to have widespread thaw lakes, in which lakes may actually have formed chiefly by other processes such as inundation of depressions [Jorgenson and Shur, 2007], an unknown number of lakes in the study area may have initiated by processes other than thawing.
 Six Landsat scenes of the Tuktoyaktuk Peninsula from 1978–2001 were selected from data archives landsat.org and EROS based on availability, lack of obscuring clouds and season of acquisition (June 28th to September 19th) (Table 1). Surface water objects were classified by spectral signatures using all bands (auxiliary materials). Objects with water signatures that were not (or not likely) thermokarst lakes were eliminated, including rivers, fragments of rivers and other structures such as large oxbow lakes. These features formed <0.1% of objects and <0.001% of areal extent of waterbodies. Wetlands, inundated terrain near the coast, and coastal lakes which displayed connectivity to the ocean in one or more scenes, were excluded from analyses. The water objects in the remaining regions of the scenes were then separated into four area classes of small to increasingly large lakes: Class 1, 3250–12996 m2 (1–4 pixels); Class 2, 12997–263169 m2 (5–36 pixels); Class 3, 263170–1299600 m2 (37–80 pixels); Class 4, >1299600 m2 (>80 pixels). Because Class 1 objects are 1–4 pixels and subject to image noise and detection algorithm, we excluded these from analysis. The remaining objects classified as lakes by our procedure are believed to be a conservative but unambiguous representation of lakes. Classified images were overlaid to measure lake areas common to both images, area found only in the older image (lake shrinkage), and area found only in the newer image (lake growth).
Table 1. Data Sources and Areal Extent of Land and Thermokarst Lakes in Subregions of the Tuktoyaktuk Peninsula, NW Canada
Lower Peninsula (greater, 7120 km2) Recent change over max. area
 Because of image availability and differences in path coverage between MSS and TM/ETM, our results cover five different but not exclusive geographical regions. TM/ETM scenes spanned the entire Peninsula but MSS imagery has a gap in the central Peninsula where available Landsat-2 paths did not join (Figure 1). Hence, the greater Lower and greater Upper Peninsulas are the regions spanned by TM/ETM in 1990–2001, whereas the lesser Lower and lesser Upper Peninsulas have full temporal coverage (1978–2001) but the smaller areas covered by MSS (Table 1). We used the results of classifications to conduct several different sensitivity tests using these areas. In all our comparisons, the number of Class 4 lakes is >100 and in all but two comparisons >200 (and the number of smaller lakes, classes 2 and 3, even greater). Our results are therefore representative of the average behaviour of a large number of lakes.
4. Results and Discussion
4.1. Lake Area Changes Through Time by Region, 1978–2001
 Lakes occupy 22–34% of the total land area on the Peninsula, with lake cover increasing toward the Upper Peninsula (Table 1). In the Lower Peninsula, lake area expanded by +14% from 1979 to 1991 (1991 was a +10% anomaly from the mean, Figure 2a), followed by shrinkage of −11% from 1991 to 2001 (2001 had a −8% anomaly from the mean, Figure 2a). Consequently, there was only a small overall increase from 1979–2001. Large lakes (Class 4) saw the greatest changes in absolute lake area and so drove the overall change. For the greater Lower Peninsula (the larger area covered by TM/ETM data), changes in total lake area from 1991–2001 were consistent with those measured from the MSS data range (Tables S5 and S6) and also were driven by a reduction in Class 4 lake area. Numbers of Class 4 and Class 3 lakes both decreased, owing to lakes switching to lower classes as they shrink, causing an increase in Class 2 lake count. Similar to the Lower Tuktoyaktuk Peninsula, lake area in the Upper Peninsula increased substantially from 1978 to 1992 (+14%), and decreased from 1992 to 2000. Unlike the lower Tuktoyaktuk Peninsula data, from 1992–2000 lake area increased in smaller lakes (Classes 2–3) but there was an overall reduction in lake area due to shrinkage of Class 4 lake area. This was caused by Class 4 lakes that switched to Class 3 owing to shrinkage. Results for the greater Upper Peninsula (TM/ETM data range) are consistent with those from the MSS data range.
 Distinguishing shallow lake water from inundated sedge marshes and areas of wet exposed sediment at lake margins is complicated by their similar spectral signatures. Shallow water is ∼5% of total lake surface area in our classification (its greatest representation is in the NE peninsula where some lakes in sandsheet deposits have particularly shallow shelves). One implication is that the uncertainty of lake area measurements is ±≤5%, because measurements from deep water are robust -i.e. spectrally distinct from terrestrial surfaces. Also, because the magnitude of changes in total lake area through time exceeds shallow water area, large changes between years (±10%) cannot only be attributed to seasonal wetting and drying of wetlands that may have been misclassified as lakes.
4.2. Interannual Variability
 An estimate of inter-annual variability in lake extent was found by comparing image pairs from 1991 and 1992, and 2000 and 2001, for a 3910 km2 region of the Central Peninsula (Table 1). From June 28 of 1991 to July 18th of 1992, a change of only +0.3% occurred in total lake area, and none of the lake classes showed significant change (Table S11). A change of −4% from September 9, 2000 to September 19, 2001 was chiefly cause by a 6% reduction in area of the largest lakes (Table S12). The scenes are separated by only 10 days in terms of season, so the change between them is unlikely to be related to season of acquisition.
4.3. Correlation to Precipitation and Temperature
 The areal extent and number of lakes might depend on annual and seasonal climate factors that affect the water level in lakes. We compared changes in lake extent and lake number between image pairs to climate conditions for the year of acquisition, including MAAT, mean thaw season temperature (mean temperature from June–August), and cumulative precipitation in the 12 months preceding image acquisition [Environment Canada, 2007] (Figure S2). In 1970–2005, thaw season temperature varied between ≈7–12.6°C, MAAT from −4.8 to −11.3°C, and cumulative 12 month precipitation from 110 cm to 260 cm (Figure S2). Anomalies in the total extent and number of lakes in a given year (expressed in relation to respective means over the entire period of record) were compared to anomalies for climate parameters (Figures 2b–2d). Only cumulative precipitation in the 12 month period prior to Landsat scene acquisition has a correlation with total lake area change across image pairings (Table S13 and Figure 2d). A linear regression of the change in lake area against the change in 12-month precipitation has a coefficient of determination (r2) of 0.823, indicating a strong relationship (Figure 2e).
4.4. Comparisons With Other Thaw Lake Regions
 Our results suggest that no persistent change in lake extent has occurred on the Tuktoyaktuk Peninsula in recent decades, and that thermokarst lake level changes caused by inter-annual fluctuations in precipitation cause large changes (>10%) in the areal extent of thaw lakes on the Tuktoyaktuk Peninsula. Similarly, Riordan et al.  found no significant change in the area of water bodies in continuous permafrost on the Alaskan Arctic coastal plain between 1954, 1978 and 1999. Conversely, Smith et al.  reported a significant increase, 12%, in total lake area in a portion of Siberia underlain by continuous permafrost, when comparing images from one year in the early 1970s to one year in the late 1990s. The measured interannual variability in lake extent (presumably owing to precipitation/evaporation balance) for Siberia was 3%, comparable to the successive-year variability of 4% we measured for the Tuktoyaktuk Peninsula. However, this 4% is likely a minimum estimate, with the robustness of the relationship between precipitation and total lake extent (which varies by >10%) indicating the true magnitude of seasonal variability in lake area owing to precipitation changes probably also is >10% for the Tuktoyaktuk Peninsula. On the one hand, the Siberian estimate may similarly be an underestimate of typical seasonal variability, depending on the relative wetness and number of years used in sensitivity tests. On the other hand, differences in the sensitivity of lake extent to recent precipitation intensity might vary between regions, depending on the magnitude of interannual variability in precipitation, the sizes of the drainage areas that contribute to lakes, on basin geometry (particularly if lakes are closed basins), and on the bathymetry of thaw lakes-low-angle margins causing higher sensitivity of the areal extent of lakes to water inputs. Discerning the effects of these factors on remote sensing measurements, and on future behaviour of thaw lakes, requires additional information on the morphology of thaw lake basins and the landscapes in which they occur.
4.5. Processes Effecting Lake Surface Area
 The absence of a persistent change in thaw lake areal extent on the Tuktoyaktuk Peninsula, despite regional warming in the western Canadian Arctic, might be attributed to several causes. First, decreases in the water level in lakes might offset more rapid thermokarst expansion of basins. Remote sensing studies such as ours [e.g. Smith et al., 2005; Riordan et al., 2006], which measure surface water area, do not discern between these. Second, few lakes might be of thermokarst origin, or thermokarst processes are insensitive to temperature increases of the magnitude experienced in recent decades, so basin area change is absent or too small to distinguish given the magnitude of interannual variability in water level caused by precipitation. Third, the time-scale over which thaw lakes respond to a climate perturbation (such as recent warming) might be longer than that spanned by our remote sensing measurements; i.e. the dynamics of lake basins lag climate warming by decades or longer. The coupled thermal and geomorphic processes that operate on thaw lake margins might give rise to a complicated time-response of lakes to climate change. For example, thaw slump material released by deep thawing in a warm year may temporarily accumulate along lake margins, serving as a thermal blanket that buffers permafrost from the effects of high water levels and/or warmer temperatures in following years, until is redeposited into the deeper basin. Investigating the implications of these and other geomorphic/thermal feedback processes on lake expansion requires morphologic measurements and numerical models [West and Plug, 2008; L. Plug and J. West, Thaw lake expansion in a two-dimensional model of coupled model of heat transfer, thaw subsidence and mass movement, submitted to Journal of Geophysical Research, 2007] beyond the scope of our remote sensing. Such geomorphic/thermal feedbacks underline that the response of thaw lakes in continuous permafrost to warming probably is complicated and may vary between regions.
5. Concluding Remarks
 Our results indicate: 1) Short-term (annual) changes in the areal extent of thermokarst lakes on the Tuktoyaktuk Peninsula are up to 4% in one year, and maximum variability of 14% between all years sampled. There is no sign of a consistent increase in lake area over the period measured, such as might be caused by a warming climate. Instead, the changes in lake extent between years are best explained by anomalies in cumulative precipitation in the 12 month period preceding measurement. The lakes of the Tuktoyaktuk Peninsula therefore may be responding differently to high-latitude warming than those in other regions (in contrast to the expansion of lakes in continuous permafrost in Siberia [e.g., Smith et al., 2005]). 2) Because of the apparent sensitivity of lake extent to precipitation-driven water levels, measuring long-term trends in thermokarst lake behavior would benefit from identification of basin margins, demarcated by recent thawing and mass movement, rather than annually-varying extent of water coverage. 3) Because thaw lake dynamics are complicated, both because of coupled thermal and geomorphic processes within single lakes and interactions between lakes, the persistent responses of individual thaw lake regions to climate changes may be highly individual and difficult to decipher from decadal-scale remote sensing measurements.
 We thank T. Jorgenson, L. C. Smith, and an anonymous reviewer for useful comments and discussions. Research has been supported by an undergraduate research fellowship from the Natural Science and Engineering Research Council (NSERC) of Canada to BMS, and by NSERC awards to LJP. Equipment has been kindly provided to LJP by the Canadian Foundation for Innovation and IBM Canada.