5.1. Data Analysis
The combined MLC and OBIA classification into four classes (land, water, and temporal transitions between both) based on the 16-year (1999–2014) Landsat trend data yielded an excellent separation of the input data based on different metrics, such as RF’s internal accuracy estimator (OOB) or five-fold cross-validated classification accuracy and Cohen’s kappa. Area estimates of pure pixels of the defined classes can therefore be considered as highly accurate within the spatial resolution of the data.
However, as thermokarst lakes are characterized by a dynamic, asymmetric behavior with a large number of slowly growing lakes and a low number of quickly draining lakes, and Landsat’s spatial resolution of 30 m, sub-pixel analysis becomes crucial to properly account for the magnitude and direction of changes. Therefore, we included a sub-pixel analysis of changes along the lake margins based on the probability values of the MLC.
The processing step from pixel based classification to lake or lake-change objects worked generally well in most cases. Lakes embedded in tundra, boreal or transitional environments were clearly detected and separated without notable differences between the different eco-zones. However, more dynamic zones without a clear distinction of water and land, e.g., wetlands, are a potential error source for lake change calculations, due to their constantly changing surface conditions, local-scale permafrost landscape features conditions, and its transitional nature between water and non-water, which is a common issue identified in previous studies e.g., [25
]. These particular settings usually occurred in coastal or delta regions and river valleys or flat drained lake basins. Such regions might be more susceptible to larger errors than for most other regions. Particularly, the highly dynamic AKS study site was affected by this effect, where many small basins constantly changed their surface water conditions.
Few misclassifications were identified in regions of frequent wildfires, such as boreal taiga or forest tundra regions as wildfires were occasionally classified as change in either direction, due to its strong spectral change over time. The use of global forest cover change information [45
] and DEM data [89
] helped to detect and mask these regions by automatically discarding false positive classified lake objects. However, as large regions in the tundra-taiga transitional zone are not accounted for in the global forest cover change map, the removal of false positive lake objects due to fire may lack the reliability of boreal regions.
5.2. Comparison of Sites and Prior Studies
Each of the four analyzed regions shows a slightly different behavior in the dynamics of its lakes. The more northerly coastal regions NSL and KOL, have a small decline in lake area with 0.69% and 0.51%, respectively. Lake dynamics are locally varying, and spatial differences, based on several factors, such as geology (types of sediments, ice-content) and geographical setting (e.g., proximity to the coast or rivers) have an influence on the lake dynamics. The western Alaskan site (AKS) is subject to a decrease in lake area by 2.8% and is characterized by frequent dynamics, predominantly as lake drainage. As in the first two regions, clear small-scale local spatial patterns can be distinguished. The Central Yakutian study site (CYA) is characterized by extreme lake expansion of 48.48% between 1999 and 2014, which strongly deviates from the other sites.
The predominantly stable conditions on the Alaska North Slope (NSL) compare well with Jones et al. [59
], who found no significant long-term trends of lake area change between 1985 and 2007. Our characterization of mostly stable conditions over most of the North Slope study site confirms Hinkel et al. [31
] who found two lake drainage events per year from the mid 1970s to 2001/2002 on a similar sized and largely overlapping part of the North Slope. Considering the same criteria, we observed 2.44 drainage events per year. Arp et al. [16
] detected net lake expansion of 3.4% using high-resolution data and 4.1% using Landsat data, between 1979 and 2002 on a small sample of 13 lakes in a more dynamic subset on the YOCP, north of Teshekpuk Lake. For twelve of the same lakes together, we found a zero net change, but strong changes in both directions for individual lakes. The missing lake L195
drained catastrophically in summer 2015 [15
] and lost a significant portion of its 80 ha area. Due to the close proximity to the sea, the lake was automatically discarded in our processing, as it was detected as connected to the sea.
In the Kolyma Lowland region, our calculated net lake area decrease of 0.5% reveals a strong discrepancy to Walter et al. [33
], who detected a lake area increase of 14.7% along the Kolyma river for a longer time period from 1974 to 2000. The strong difference can be explained by the extent and setting of study areas, as we also found a strong lake area increase within the Kolyma floodplain. Their use of sensors with a different spatial resolution (Landsat MSS to ETM+), may have also had an influence on the results. Considering the discrepancy to the regional statistics, this particular subset is not representative for the entire region, which stresses the spatial variety of lake change processes, particularly within this region. Veremeeva and Gubin [92
] found a trend of lake area decrease in the range of 0.9 to 10.7% for a small region in the western Yedoma-Alas Complex from 1973 to 2001. Within this specific region, our results also indicate a strong loss of lake area, largely due to the partial drainage of the around 50 km² large Bolshoy Oler lake. The sharp boundary of lake growth trends and lake shrinkage trends coincided very well with boundary of geomorphology with old low-lying thermokarst basins in the Yedoma-Alas Complex region.
In a comparable Arctic coastal lowland setting, to NSL and KOL, with thick continuous permafrost on the NW Canadian Tuktoyaktuk Peninsula, Olthof et al. [27
] found a slight lake area increase of 0.64% over eight years with strong inter-annual fluctuations of up to 4%. Within the same region Plug et al. [32
] found substantial lake area fluctuations for large lakes of +14% from 1978 to 1992 and −11% from 1991 to 2001, however with an anomalously low lake area in 2001.
In the western Alaskan study site (AKS), located in the transition zone from continuous to discontinuous permafrost, we observed a net lake area loss of 2.8%. It compares well to Roach et al. [29
] who measured a lake area loss of 0.81% per year. Their study site largely covers the area affected by widespread lake drainage and trends of lake change nicely resemble the spatial pattern we found in our study, e.g., lake growth in the Selawik river delta and widespread loss in the northern Selawik river valley. In a more continental site east of the Kobuk Dunes
, Necsoiu et al. [42
] found an overall decreasing lake area trend from 1978 to 2005 within 22 lakes or ponds. We could detect 14 of these lakes and found a lake area loss of 13.4%, largely fueled by the partial drainage of one lake. Similar trends were detected on the nearby northern Seward Peninsula where Jones et al. [14
] found 11% lake area loss where the drainage events of few large lakes were the main drivers for a net lake area loss. Each of the coastal sites showed trends of lake area increase in near-shore areas. To our knowledge this effect has not been described or discussed in other lake change studies. It may be caused by several factors, such as local climatic effects; surface geology, e.g. sediments more prone to erosion; or very flat terrain, which is affected to sea water inundation.
The abundance of disappearing lakes in discontinuous permafrost regions in Central Alaska [28
] has been linked to increased connectivity to groundwater [38
]. Permafrost degradation or disappearance along the continuous-discontinuous permafrost interface in the AKS site would be a good explanation for the strong lake drainage trend within this region. Other causes in areas with relative high continentality, such as the central Alaskan Yukon Flats or northwest Canadian Old Crow Flats, include increased evapotranspiration [12
However, for the highly continental Central Yakutia site, we calculated a 48.48% increase in lake area for the 1999–2014 period, which is a significant outlier in terms of thermokarst lake dynamics. The same wetting pattern has been described by Boike et al. [34
], who used Landsat snapshots and found an increase of around 85% from 2002 to 2009 within the strongest wetting part of the CYA study site. The strongest increase in lake area occurred in 2007 after above-average precipitation in the prior year [44
]. Furthermore, this particular region has been subject to very strong rates of lake expansion over the last decades due to several factors, including anthropogenic activity and the change of climatic conditions [43
]. The recharge of lakes here may be connected to a wetter and warmer climate over the recent decades [34
] and shifting agriculture practices, where meadows and grasslands in alas basins are increasingly managed to produce richer pastures [71
]. In addition to the climatic conditions and anthropogenic influence, the local geological conditions seemingly had a strong influence on the lake area changes, where the terraces with ice-rich sediments showed a much more pronounced lake area expansion in comparison to the remaining area.
The comparison of different studies as well as the analysis of local trends highlights the variability of lake dynamics within the northern permafrost region.
The wide variety of spatial scales poses a large challenge for the comparison of different studies and regions, as results may vary strongly even for nearby locations. Furthermore, seasonal or short-term lake area fluctuations, which can exceed long-term trends [27
], may mask long-term trends in studies based on the widely applied practice of using snapshots [14
]. The trend analysis of single date lake masks [27
] can help to suppress short-term fluctuations and produce more reliable and comparable results. In our study we applied the trend analysis at an earlier stage and translated spectral trends to semantic information with MLC, which allowed us to accurately distinguish between zones of stable water or land and changing transition zones around lake margins. With the inclusion of classification probability values, we exploited sub-pixel information to detect permafrost region specific thermokarst lake growth.
Using the trend calculation helps generalize the input data regardless of its location and enables the comparison and upscaling across multiple spatial scales, starting from individual lakes up to very large regional scales. The successful application of the method to different study sites across the permafrost zone proved the transferability and scalability of the highly automated processing method and highlights its strong potential for applying it to the entire permafrost domain to fully characterize lake changes and associated permafrost dynamics.