Thermokarst lake expansion and drainage in permafrost terrain
Thawing ground ice creates thermokarst lakes that grow, migrate and drain on timescales from decades to hours. Multispectral time-series analysis over the Landsat archive quantifies area change, new formation and abrupt drainage events across Arctic and sub-Arctic terrain.
Sensors
- Landsat 4-9 (USGS/NASA): 30 m multispectral resolution, 16-day revisit per satellite. The archive runs from 1982, giving over four decades of lake-area change. Bands 5 and 7 (SWIR) sharpen water-land boundaries; the MNDWI index reliably delineates open water down to roughly one to two Landsat pixels in clear conditions.
- Sentinel-2 MSI (ESA Copernicus): 10 m visible and 20 m SWIR bands, 5-day revisit at high latitudes with two satellites. Extends the archive from 2015 and resolves small lakes (below one hectare) that Landsat misses. Band 11 (1610 nm SWIR) is the primary water-edge discriminator in low-turbidity conditions.
- Planet SuperDove: 3-5 m resolution, near-daily revisit over tasked areas. Useful for tracking rapid drainage events and monitoring individual lake margins at sub-pixel Landsat scale. No public archive; commercial tasking required.
- ALOS PALSAR / PALSAR-2 (JAXA): L-band SAR at 10-25 m resolution penetrates thin vegetation and provides soil-moisture context around lake margins. Useful for detecting waterlogged active-layer soils that precede new lake formation, and for mapping lake ice in winter when optical sensors are limited by polar darkness.
What a thermokarst lake actually records
Permafrost contains ground ice: wedge ice filling frost-crack polygons, pore ice cementing sediments, and massive ice lenses that can be metres thick. When the active layer deepens or surface disturbance removes insulating vegetation, that ice melts. The ground subsides. Water pools in the depression. A thermokarst lake is born.
The lake's area is not decorative information. It is a proxy for cumulative ground-ice loss beneath and around it. A lake that has grown by 20 percent over a decade records a specific volume of ice melt and organic carbon release from the thawed sediment column. Abrupt drainage, by contrast, records a different process: the lake has breached its bank, often into a river system or an adjacent drained lake basin, releasing stored water and resetting the local carbon cycle. Both signals are legible from orbit.
Forty years of evidence in thirty-metre pixels
The Landsat archive is the primary instrument for decadal thermokarst analysis. Landsat 5 TM data from the mid-1980s can be compared directly with Landsat 8 or 9 OLI imagery from the present day using consistent radiometric calibration and the same 30 m ground sampling distance. The Modified Normalised Difference Water Index (MNDWI), computed from green and SWIR bands, separates open water from surrounding tundra with reasonable reliability in clear-sky conditions. Automated change-detection algorithms applied to annual or seasonal composites can identify new lakes, measure area change in existing lakes, and flag sudden disappearances that indicate drainage.
The honest limit here is resolution. A lake smaller than roughly 0.09 hectares (three by three Landsat pixels) is effectively invisible. Lakes between that threshold and about one hectare carry significant mapping uncertainty because mixed pixels along the shoreline dominate the water-area estimate. Sentinel-2 at 10 m pushes the detection floor down to lakes of roughly 0.01 hectares, though the archive only begins in 2015. For the pre-2015 period, researchers have used Landsat 1-5 data going back to 1972, though earlier sensors have lower radiometric quality and less consistent calibration.
Cloud cover is a persistent problem at high latitudes. Growing-season cloud fractions over Siberia and Alaska regularly exceed 60 percent in any given overpass. The standard mitigation is seasonal compositing: selecting the clearest pixels across all acquisitions within a defined window, typically June through September, to construct a cloud-minimised surface. This works well for slow, decadal change but can obscure rapid drainage events that occur within a single season.
Abrupt drainage: the signal that matters most for carbon budgets
Gradual lake expansion is ecologically significant. Abrupt drainage is potentially catastrophic for local carbon accounting. When a lake drains, it exposes a lake-bed sediment that is rich in previously frozen organic matter. That material can oxidise rapidly, releasing CO2 and methane. Studies of the Yukon Flats and the western Siberian lowlands have documented drainage events that convert a lake to a drained basin within days to weeks.
Detecting these events from satellite requires dense time-series coverage. A 16-day Landsat revisit can miss a drainage that completes in under a week. Sentinel-2's 5-day revisit at high latitudes is more useful here, and Planet SuperDove's near-daily coverage over tasked areas is the most capable option for event capture. The spectral signature of a freshly drained basin is distinctive: exposed pale sediment with high reflectance in visible bands, low MNDWI values, and sometimes residual waterlogging detectable in L-band SAR backscatter. Combining optical and SAR observations reduces the ambiguity between a drained lake and a lake obscured by cloud or ice.
Where spectral methods break down
Turbid water confounds MNDWI-based mapping. Suspended sediment raises reflectance in the green and SWIR bands, compressing the index value toward zero and making shallow turbid lakes resemble wet soil. Glacially fed lakes in Alaska and parts of northern Canada are particularly problematic. Spectral unmixing, using multiple endmembers for clear water, turbid water, wet sediment and dry tundra, partially recovers accuracy but adds complexity and requires local calibration data.
Dense shrub vegetation overhanging lake margins is a second source of error. At 30 m resolution, a narrow fringe of alder or willow can mask several metres of actual shoreline. This matters most for small lakes where the vegetated margin is a significant fraction of the total area. Sentinel-2 at 10 m reduces but does not eliminate the problem. High-resolution commercial imagery is the only optical solution; alternatively, L-band SAR can penetrate low shrub canopy to some degree and provide an independent shoreline estimate.
Winter acquisitions are largely useless for water-area mapping because lake ice raises visible and NIR reflectance to values indistinguishable from snow-covered tundra. The analysis window is therefore constrained to the ice-free season, typically a 90 to 120 day window at latitudes above 65 degrees north.
Building a usable monitoring product
A practical thermokarst lake monitoring system combines three layers. First, a baseline lake inventory derived from the full Landsat archive, giving area and centroid for every detectable water body in the region of interest at five-year intervals going back to the 1980s. Second, an annual update using seasonal composites from Sentinel-2, resolving smaller lakes and providing a consistent recent record. Third, an event-detection layer that flags anomalous area changes between consecutive acquisitions, triggering higher-resolution tasking where drainage or rapid expansion is suspected.
The deliverable for a government client is typically a GIS polygon layer with area, perimeter, change rate and confidence class for each lake, alongside a summary report quantifying net lake-area change, the number of drainage events detected, and the area of newly formed lakes. For carbon or hydrological modelling, the polygon outputs can feed directly into published empirical relationships between lake area and thaw-bulb volume.
Satellize applies this workflow on open Sentinel and Landsat archives, with commercial SuperDove tasking added on client licence for event monitoring. The methodology is closely analogous to the spectral time-series approach used in the Tonga crop-estimation programme, adapted from agricultural field boundaries to Arctic water bodies.
Typical figures
| Primary spatial resolution | 30 m (Landsat 4-9); 10 m visible / 20 m SWIR (Sentinel-2); 3-5 m (Planet SuperDove) |
| Revisit frequency | 16 days per Landsat satellite; 5 days (Sentinel-2A+B combined at high latitudes); near-daily (Planet SuperDove, tasked) |
| Minimum detectable lake area | ~0.09 ha (Landsat, clear water); ~0.01 ha (Sentinel-2, clear water); smaller with SuperDove |
| Archive depth | Landsat: 1982 to present (Landsat 4+); limited Landsat 1-3 data to 1972. Sentinel-2: 2015 to present |
| Spectral bands used | Green, NIR, SWIR-1 (1610 nm), SWIR-2 (2200 nm) for MNDWI and water-edge mapping; L-band SAR (1.2 GHz) for soil moisture context |
| Analysis window (ice-free season) | Approximately June to September at latitudes above 65°N; shorter at higher latitudes |
| Cloud-cover limitation | Growing-season cloud fractions commonly 40-70% over Arctic lowlands; seasonal compositing required |
| Delivery formats | GeoPackage / Shapefile polygon layers, GeoTIFF change rasters, PDF summary report, CSV area-time series per lake |
Analytics Satellize can run
| Decadal lake-area change inventory | MNDWI thresholding on Landsat seasonal composites; automated polygon extraction and area computation at five-year intervals | GIS polygon layer with area, perimeter, change rate and confidence class per lake; PDF summary report |
| New lake formation detection | Binary change detection between baseline and current-year Sentinel-2 composites; minimum mapping unit set to 0.01 ha | Flagged polygon layer of newly formed water bodies with formation-year estimate |
| Abrupt drainage event alerts | Anomaly detection on inter-acquisition MNDWI difference images; threshold exceedance triggers commercial tasking request | Near-real-time alert with lake ID, estimated drainage date window and pre/post imagery chip |
| Lake-margin turbidity classification | Spectral unmixing using clear-water, turbid-water and wet-sediment endmembers; flags lakes where MNDWI-based area estimates carry high uncertainty | Confidence-class attribute on each polygon; turbid lakes flagged for manual review or SAR cross-check |
| Drained lake basin mapping | Post-drainage spectral signature detection (high visible reflectance, low MNDWI, L-band SAR backscatter change) to distinguish drained basins from cloud-obscured lakes | Drained-basin polygon layer with estimated drainage date and exposed-sediment area |
| Annual area time-series per lake | Per-lake area extraction from each cloud-acceptable Sentinel-2 acquisition within the ice-free season; smoothed trend fitting | CSV time-series file per lake; optional GeoJSON feed for integration with client hydrological models |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.