These authors contributed equally to this work.
We introduce an algorithm (Watta) which automatically calculates supraglacial lake bathymetry and detects potential ice layers along tracks of the ICESat-2 (Ice, Cloud, and Land Elevation Satellite) laser altimeter. Watta uses photon heights estimated by the ICESat-2 ATL03 product and extracts supraglacial lake surface, bottom, and depth corrected for refraction and (sub-)surface ice cover in addition to producing surface heights at the native resolution of the ATL03 photon cloud. These measurements are used to constrain empirical estimates of lake depth from satellite imagery, which were thus far dependent on sparse sets of in situ measurements for calibration. Imagery sources include Landsat 8 Operational Land Imager (OLI), Sentinel-2, and high-resolution Planet Labs PlanetScope and SkySat data, used here for the first time to calculate supraglacial lake depths. The Watta algorithm was developed and tested using a set of 46 lakes near Sermeq Kujalleq (Jakobshavn) glacier in western Greenland, and we use multiple imagery sources (available for 45 of these lakes) to assess the use of the red vs. green band to extrapolate depths along a profile to full lake volumes. We use Watta-derived estimates in conjunction with high-resolution imagery from both satellite-based sources (tasked over the season) and nearly simultaneous Operation IceBridge CAMBOT (Continuous Airborne Mapping By Optical Translator) imagery (on a single airborne flight) for a focused study of the drainage of a single lake over the 2019 melt season. Our results suggest that the use of multiple imagery sources (both publicly available and commercial), in combination with altimetry-based depths, can move towards capturing the evolution of supraglacial hydrology at improved spatial and temporal scales.
Ice loss from Greenland and Antarctica is the greatest current contributor to rising sea levels, and paleodata and modeling efforts indicate that enhanced mass loss of these ice sheets may become irreversible if certain major tipping points are passed (IPCC, 2019). Recent observations have shown that ice loss is accelerating faster than projected (Slater et al., 2020), with a sixfold increase since the 1970s and 1980s. In Antarctica, this was largely driven by increased ocean melting of outlet glaciers (Rignot et al., 2019), while on the Greenland Ice Sheet mass loss is further promoted by increased surface melt and runoff (Mouginot et al., 2019).
Owing to the nonlinear relationship between increasing summer air temperatures and surface melt (Trusel et al., 2018), meltwater production has increased rapidly on the Greenland Ice Sheet (van den Broeke et al., 2016). In the summer of 2019, advection of warm, wet mid-latitude air led to a summer mass loss unprecedented in the past 50 years, with widespread surface melt occurring up to the highest regions of the ice sheet (Tedesco and Fettweis, 2020; Sasgen et al., 2020). Concurrent with the increase in melt extent and duration, supraglacial lakes – which form when meltwater runoff collects in local topographic lows – are now a common feature on large parts of the ice sheets and have become more extensive and have advanced inland toward higher elevations in the past decades (Gledhill and Williamson, 2018; Leeson et al., 2015; Howat et al., 2013).
These meltwater lakes and streams are a key component of the hydrological
system of both ice sheets. Hydrological systems over Antarctic ice shelves
have been identified as a major factor for potential ice shelf collapse
(Bell et al., 2018). While mass loss in Antarctica over the next 100 years
is generally thought to be dominated by the basal melt under ice shelves
(Schlegel et al., 2018), emerging research has focused on the potential
importance of surface hydrology over Antarctica (Arthur et al., 2020).
Supraglacial lakes have been observed around the margin of the Antarctic Ice
Sheet up to high elevations (Stokes et al., 2019) and are likely to become
more prevalent on firn-depleted ice shelves in future warming scenarios,
which could potentially trigger their collapse and consequently lead to
accelerated sea level rise (Lai et al., 2020). Meltwater pathways can include surface flow into lakes and then streams, leading, in Greenland, to direct loss to the bed from lake drainage or the sudden termination of a stream into a moulin, or near-surface flow in which ice slabs can limit vertical motion (MacFerrin et al., 2019). The complex links between supraglacial hydrological systems and englacial or subglacial pathways can potentially be deduced by capitalizing on increasingly higher-resolution imagery and classification techniques of supraglacial feature types (Yang et al., 2017). Past remote-sensing work has derived lake volumes from high-resolution (
The ICESat-2 (Ice, Cloud, and Land Elevation Satellite) laser altimeter data, available since 2018, has now introduced the potential to replace the in situ measurements used in empirical (supraglacial lake depth) bathymetric methods with satellite laser bathymetric depths at a high vertical resolution, consequently extracting lake volumes from imagery (Parrish et al., 2019; Albright and Glennie, 2020; Thomas et al., 2021). Here, we present a new algorithm, titled “Watta”, using the ICESat-2 laser altimeter to derive properties of supraglacial lakes. Watta was first presented in Fricker et al. (2020), demonstrating both the potential for ICESat-2-based bathymetry estimates and the greater accuracy of empirically based lake depths in comparison to physically based estimates; the latter tended to underestimate lake depth by over 2 m. However, we note that physically based methods have the advantage of being dependent on imagery alone. In addition to bathymetry derived from the difference between the air–water and water–ice interface, this algorithm assigns a probability for surface type characteristics to photon returns along track. These types include lakes, refrozen lakes, and lakes with ice layers on top and under the surface. Watta also returns surface heights at the native resolution of the ATL03 photon cloud (0.7 m), allowing the algorithm to capture small-scale changes in surface relief when multiple passes are differenced. Additionally, we exploit a range of imagery data to validate the surface types and to derive spectrally driven depth estimates calibrated to ICESat-2-based depths, thereby providing an estimate for meltwater volume over the full image. We compare empirically based volume estimates derived from a single ICESat-2-based depth estimate but from multiple imagery sources with different spatial resolutions (and without an atmospheric correction) to better understand the importance of spatial resolution and radiometric calibration for the relative accuracy of depth volume estimates.
Study region over Sermeq Kujalleq and Sarqardliup Sermia. Top left panel: Lake Ayşe on 23 May using Planet SkySat Visual imagery. Bottom left panel: study region over western Greenland. Right panel: main region with repeat-tasking locations for Planet SkySat shown in grey boxes over annual velocity estimates from MEaSUREs (NSIDC). ICESat-2 reference ground tracks shown in white. Operation IceBridge flight on 15 May 2019 shown in green. Five lakes indicated in red discussed throughout text include the following: C: Lake Cecily; A: Lake Ayşe; J: Lake Julian; N: Lake Niels; Z: Lake Zadie.
The method is tested and refined using representative sections along the flowline of Sermeq Kujalleq (Jakobshavn Isbræ), one of the fastest-moving glaciers in Greenland, as well as the slower-flowing Sarqardliup Sermia. The repeat tasking of Planet SkySat imagery was designed to coincide with ICESat-2 tracks (Fig. 1), capturing lake depths at various stages of lake development during the unusually intense melt season of summer of 2019 (Tedesco and Fettweis, 2020). One of the major motivations for this tasking effort was its coincidence with several NASA Operation IceBridge (OIB) flights at the beginning and end of the summer. Data from multiple instruments on board OIB could potentially provide additional insight in future work, and within this study, we use OIB CAMBOT (Continuous Airborne Mapping by Optical Translator) imagery as a part of a focused multi-instrument study of the evolution of a supraglacial lake. The availability of simultaneous laser altimetry and high-resolution imagery over the season provided a rich test dataset with which to extract altimetry-based estimates of supraglacial lakes at various points in the season. Here, we present initial results exploiting this dataset, as well as introducing the Watta ICESat-2 surface feature detection algorithm.
Our Watta method relies on individual photon heights as measured by ICESat-2's (Ice, Cloud, and Land Elevation Satellite) ATLAS (Advanced Topographic Laser Altimeter System) instrument, distributed in the ICESat-2 ATL03 product, L2A, Global Geolocated Photon Data (Neumann et al., 2019). The polar orbiting ICESat-2 satellite was launched in September 2018 to continue the mission begun by ICESat (2003–2009) and bridged with the airborne Operation IceBridge mission, namely to provide ice sheet mass balance estimates at an unprecedented level of accuracy. ATLAS is a photon-counting 532 nm laser altimeter on board ICESat-2 split into six beams which are divided into three pairs (separated by 3.3 km), in which beams within each pair are separated by 90 m. Within a single track, the beam pair is designated by a number, e.g., “3” in “gt3r”. Each beam pair consists of a strong and weak beam, with the strong beam returning 0.6–3.9 signal photons per laser pulse vs. 0.6–1.0 signal photons per laser pulse for the weak beam (Neumann et al., 2019). The beam is designated with “r” or “l” depending on the orientation of the satellite, as in “r” in “gt3r”.
While the strong beam produces a stronger signal, we have developed the Watta algorithm to work effectively with both strong and weak beams. ATL03 produces a photon cloud in which each photon is geolocated to within a 6.5 m accuracy and a footprint size of
For imagery sources, in addition to freely available Landsat Operational Land Imager (OLI; 30 m) and Sentinel-2 (10 m) imagery, we incorporate very high-resolution imagery from Planet Labs, including Dove-R (3 m) and SkySat (
As a part of this project, SkySat imagery was tasked for repeat cycles of
The final set of lakes used for the development of the Watta method included 50 lakes captured by ICESat-2 (46 over Sermeq Kujalleq and Sarqardliup Sermia and 4 additional lakes in the southwest, not shown here), 14 of which coincided with very high-resolution imagery (SkySat) within a 3 d window. The date, time, and imagery IDs and ICESat-2 details for all data sources are presented in Table S1 in the Supplement.
We derive supraglacial lake volume from a given imagery source in four steps. We first calculate lake depths along an ICESat-2 beam using the Watta algorithm applied to the ICESat-2 ATL03 photon cloud. Secondly, we coregister Watta-based surface and lake bottom heights with the imagery source (itself coregistered to a common Landsat base) and delineate lake boundaries in the process (methods detailed in Sec. 3.2). Finally, we develop an empirical relationship between ICESat-2-based depths and coincident imagery which can be applied to calculate lake depths over the full image. The empirical relationship is based on the exponential decay of reflectance at water depth, as detailed by Box and Ski (2007). In the original work, in situ depth estimates (
Diagram of the Watta methods described in main text.
Watta is an algorithm which takes ICESat-2 ATL03 photon data as input and automatically detects supraglacial surface features with an associated likelihood. In its current state, the algorithm detects lakes and their associated surface, lake bottom, and corrected depth estimates, as well as subsurface ice when present (Fig. 2). We also exploit the algorithm for the detection of frozen streams in this study. The codebase for Watta is divided into a module which calculates surface and bottom returns (“Surface Detection”) at the native 0.7 m resolution of ICESat-2 and a second “Interpretive” module that resolves the calculated bottom and surface to specific supraglacial features, in this case lakes. The Surface Detection module determines, for a collection of photons surrounding any individual photon (75 collected before and 75 collected after; selected in step a), heights with the three strongest peak probabilities within a kernel density (step b). This provides estimates for (1) a height for the surface or top of a lake or refrozen pond, (2) a lake bottom, and (3) a third height value, which can potentially be subsurface ice (Fig. 2, step a). We note that photons are selected without regard to ATL03 confidence level. Although the bin width (and therefore vertical resolution) used to calculate heights is 0.1 m, we perform a second kernel density estimate calculation using a 0.3 m bin width to confirm robustness of the initial bottom estimate, i.e., that a coarser calculation produces a bottom height near that of the finer-resolution (0.1 m bin width) calculation. In post-processing step c, outliers are identified in comparison to surface and bottom heights within a larger horizontal window, whereby the number of standard deviations used to detect an outlier and the number of photons used to calculate a mean (window) increase over several steps. Where outliers are found, the kernel density estimate (steps a and b) are recalculated with a larger number of photons (in multiples of 75) to account for any erroneous calculations generated by insufficient photon density. Where values continue to be outliers, they are removed from the estimates to be interpolated instead. The final output of the Surface Detection module includes a calculated surface and potential lake bottom return at the native resolution of the ATL03 photon cloud.
Diagram of imagery processing steps accepting Watta outputs as input (top left panel) and producing lake depth estimates (bottom left panel). The Watta lake depth profile shown is for Lake Ayşe using RGT 841 on 23 May and Planet SkySat imagery on 22 May.
The Interpretive module uses output from the Surface Detection module to automatically determine locations of surface features (e.g., lakes or frozen streams), as well as characteristics of a lake, e.g., the presence of refrozen ice at the surface. First, remaining outliers are removed by calculating a local background and surface photon density for each ATL03 photon, as determined from a 5000 photon-count window surrounding the estimate location (step d). We then remove those estimated heights for top (surface) and bottom (potential lake bottom) where the photon density more closely resembles the background photon density than a surface density estimate. In step e, we detect breaks in the slope of the top (surface) to divide the satellite pass into segments which are potential lakes. For example, a semi-parabolic depression in topography, with a high absolute value of the slope, is broken by a lake surface where the slope approaches zero (e.g., as for the lake shown in Fig. 2, step a). For each of these potential lakes, we then perform several steps both to refine lake surface and bottom and to assign the lake a class based on its properties. To produce a lake bottom value with greater accuracy, we first perform a recalculation of the kernel density estimate (step f, equivalent to the Surface Detection step b), except here we limit the kernel density estimate to photons below the calculated surface and use 30 photons rather than 75 photons to better capture the lake bathymetry, which in general is more irregular than the lake surface. Subsurface ice layers near the edges of the lake are then reclassified as lake surfaces, thus sealing the bottom of the lake to the top at the lake edges (step g). In step h, we first perform a final smoothing, passing the resulting bottom photons through an iterative robust quadratic local regression (rloess) filter to remove outliers in the bottom estimates and then assign physical meaning to each photon (e.g., lake surface, bottom, surface ice, subsurface ice). The presence of surface ice is determined based on the variability in thickness of the lake surface (i.e., a bimodal distribution indicates surface ice at some locations). We identify subsurface ice by the presence of weak returns above the lake bottom but below the surface (see Fig. 2, step b). The surface and bottom photons are then used to derive lake depth, to which we apply a simple correction for refraction, described in Parrish et al. (2019), to produce a real corrected lake depth. Finally, in step i, we assign a final classification of a lake type using properties of the local surface slope and the strength of the bottom return (Table 1). For example, a segment with a surface slope smaller than 0.03 %, as well as a distinct bottom (a photon density far exceeding the density of a background return), is given a lake classification of “highly likely”, whereas segments passing the same slope threshold but not showing a strong bottom return are identified as “likely ice-covered” lakes. On the other hand, segments with a slope exceeding 0.3 % and no significant peak below the surface in the histogram are allocated to the “highly unlikely” lake class.
Definitions for Watta lake classes.
Two potential sources of ambiguity with the subsurface ice classification are (a) the possibility for specular returns and (b) apparent multiple
surface returns which result from instrument echo. Specular returns over
flat water (implying high energy return) return a strong surface, as well as
multiple layers below the surface spaced according to the ATLAS dead time
(0.45 m below the surface and a potential tertiary return below that). Echoes produced by electronic noise in the instrument, which also frequently
occur on very smooth water surfaces, can similarly produce a strong return at
the surface with double echoes at
A subsequent set of steps uses the lake depths extracted from the Watta algorithm to produce lake volumes from concurrent imagery, e.g., SkySat, PlanetScope, Sentinel-2, and Landsat OLI, requiring geolocation, as well as the
semi-automated identification of lake edges. Coregistration between ICESat-2
photon locations and imagery (Fig. 3, step j) is managed by registering ICESat-2 elevations with the GIMP-2 DEM as an intermediary step (GIMP-2 is
also used for georeferencing of Landsat) and by transforming the point cloud
using the iterative closest point algorithm, which minimizes the square error between the two datasets (Besl and McKay, 1992). The point cloud ICESat-2 is chosen to include a 0.2
To register imagery sources to one another, we standardize all imagery to the nearest Landsat image using the arosics library in Python (Scheffler et al., 2017), which detects and corrects misregistrations of an input image (based on a reference image) at the sub-pixel scale. However, the coregistration of all other imagery sources to Landsat OLI first requires the delineation of lake boundaries in order to exclude regions with moving surface water, which evolves rapidly and can be mistaken for fixed topography (which is more useful for geolocation). Here we calculated a normalized difference water index (NDWI) for each image, using a standard NDWI (with the green and NIR bands) to deliberately include regions with ice layers (as these are also detected by Watta) rather than the modified NDWI
The test dataset provides a diversity of lake types, with the largest surface area calculated at 5.6 km
Lake Ayşe was selected for closer examination because, despite the dense
photon cloud, its relief in surface and bottom, combined with the presence of ice cover, pose a challenge for our detection algorithm. Additionally, multiple imagery sources were available within a 5 d window at this location. Lake Zadie is chosen because it represents an ideal case for the algorithm, while Lake Cecily is chosen because two beams passed over the
same lake with SkySat imagery available 1 d afterwards. A basic assumption we make in this study is that the lake bottom remains relatively
consistent over several days, although past research on two lakes in western
Greenland has estimated lake bottom ablation rates at 6.5 cm d
The most rigorous weighting system used for the algorithm, using only lake classes 1, 2, 4, and 7 (see Table 1), succeeded in automatically detecting 49 out of the 50 lakes identified in the available imagery, with two likely false positives (i.e., not confirmed by NDWI values exceeding 0.2 in the imagery). A less rigorous weighting system, including lake class 3, detected the 50th lake but resulted in a large number of false positives in areas of steep and rough topography where abrupt changes in the photon elevations are misinterpreted as bottom reflections by the algorithm.
In the absence of simultaneous in situ data, we evaluate the performance of the algorithm based on visual inspection (comparing ATL03 heights with Watta-calculated depths as shown in Fig. S3). Additionally, where an empirical relationship with imagery is successful (a high correlation coefficient value between imagery-derived depths and Watta-derived depths), we take this consistency for partial evidence that ICESat-2 and imagery sources have detected bathymetry correctly, although we note that this metric is only applicable to the specific lake, not all the lakes in the region. The most successful bottom retrieval occurred where ice cover was minimal, the density of photons was high, and where the bottom slope was relatively uniform (e.g., Lake Zadie). The presence of ice near the surface (between the surface and 1 m below the surface) frequently obscured lake bottom detection (e.g., reference ground track – RGT – 1222, Lake 3 in Fig. S3), although in some cases only partially; however, the presence of subsurface ice did not always preclude the presence of a strong bottom return (e.g., Lake 7, RGT 1169, Fig. S3). The algorithm therefore indicates the presence of surface and near-surface ice but does not automate the removal of the calculated bottom return due to ambiguity. We can confirm the presence of an ice layer both by visual inspection of the imagery and by comparing standardized NDWI values calculated from imagery coincident with the ICESat-2 track (Fig. S1a). We note that for at least one case (RGT 1108, Lake 6, Fig. S3), the designation of “lake” was ambiguous as this could be treated as either a shallow lake containing a large amount of subsurface ice or as a slush layer (a number of which were identified elsewhere).
Total uncertainty for the empirically based depth estimates from imagery is
comprised of uncertainty in ICESat-2 geolocation, uncertainty from the Watta algorithm itself (which operates at a vertical resolution of 0.1 m), from the resolution of the imagery, from the uncertainty in alpha coefficients
calculated from the empirical method, and finally from physical changes in the lake occurring between the time that imagery is captured and the ICESat-2 pass. The empirical calculation is less likely to be affected by physical changes in the lake when the lake surfaces calculated by imagery vs. altimetry differ by less than a meter; here we estimate precision with a
simple
Comparing
Past work has considered either the red or green band for developing depth
estimates (Moussavi et al., 2020; Williamson et al., 2018), though in situ validation was limited at the time (Pope et al., 2016). Figure 4 compares
To demonstrate the robustness of the Watta algorithm, the impact of band choice, and the sensitivity to absolute lake depth, we show depths calculated from two beams passing over Lake Cecily on 13 June, followed by retrieval of SkySat imagery on 14 June and Sentinel-2 on 16 June (Fig. 5a and b and RGT 1169 Lake 5, 6 in the Supplement profiles). Over this spot, covered by the 3l beam on this pass, the green band shows higher
Watta-calculated and imagery-derived depths: Lake Cecily based on Sentinel-2
Within Fig. 6, we show the depth evolution of Lake Ayşe over 5 d, both along the ICESat-2 ATL03 and Watta-calculated profile and the lake volume estimates then constructed from imagery using the empirical equation. Increasing lake volume is demonstrated both by the expansion of the surface area of the lake through time (right column) and the rise in the lake level (cyan line, left column). Planet SkySat estimates at a 1 m resolution (Fig. 6c) show the greatest level of detail of crevassing at the bottom of the lake although PlanetScope estimates at a 3 m resolution (Fig. 6d and e) are comparable. We note that PlanetScope data showed variations in the fidelity to Watta-based estimates between the green band vs. the red band depending on the instrument. The bottom relief is maintained at depth with the PlanetScope resolution (3 m), and future work could generate more reliable depth estimates by calibrating empirically based depths to Sentinel estimates, which could provide very high-resolution depth estimates while also leveraging the high temporal frequency of PlanetScope data collection.
Lake Ayşe, filling over 5 d between 20 and 25 May with
ICESat-2 pass on 23 May. Left column panels: profiles with Watta-calculated and imagery-derived depths from the green, indicated by “G”, and red, indicated by “R”, bands (with corresponding
Our improved ability to track lake depth and volumes using the combination of ICESat-2 and multi-sensor imagery can potentially provide new insights into patterns of lake drainage. Lake Julian (Fig. 7a) was selected for closer examination because on 15 May, both the airborne Operation IceBridge mission and ICESat-2 passed over this region, providing a unique stack of both airborne and satellite data. While we show only very high-resolution Operation IceBridge CAMBOT imagery here, other instruments on board OIB could potentially provide valuable insight into the state of both surface hydrology and firn characteristics in future work. Additionally, there are cloud-free ICESat-2 RGT 727 passes over this lake both on 15 May and on 14 August 2019, providing a profile of the lake both when it was filled and after drainage. A second lake, Lake Niels (Fig. 7a), is examined briefly primarily to provide context. Although no altimetry estimates are available over Lake Niels, imagery sources reveal a very different evolution and drainage pattern despite its being located only 3500 m from Lake Julian and consequently subject to many of the same atmospheric drivers. Within the larger region shown in Fig. 7a (see also Fig. S4), the percentage of the ice sheet surface covered in liquid water, as measured by the percentage of the region where NDWI ice values exceed 0.2, remains constant at around 3 % from May through June, with meltwater being spread more uniformly (less visibly collected in large hydrological features) over the ice sheet early in the season and shifting to larger lakes later in the season (Fig. S4). We note that this measure of melt extent does not translate directly to consistent meltwater volume; meltwater underneath the snow cover on Lake Niels was not estimated earlier in the season, while the deeper lakes which are present later in the season will contain larger water volumes.
Although elevation decreases overall toward the northwest, both lakes coincide with large-scale surface depressions calculated from the GIMP-2 DEM (Fig. 7), and this region experiences comparatively low ice velocities. Lake Niels is located in a deep surface depression, whereas the corresponding surface depression for Lake Julian is relatively shallow. Because both imagery and ICESat-2 are coregistered to the GIMP-2 DEM, we presume that all inaccuracies will be consistent (i.e., even if geolocation is incorrect in absolute terms, imagery and ICESat-2 should overlap).
The volume of Lake Julian begins to increase substantially on 9 May and reaches a maximum volume between 25 and 29 May (Fig. 8j–l). After 1 June, the lake begins to lose volume until only remnants are present on 10 June, which disappear almost entirely by 19 June (Fig. 8m–o). The surrounding region (i.e., a kilometer to the north and west) contain smaller bodies of water connected by streams. We note that while larger streams can be captured by Sentinel-2 imagery (Fig. 8j), many of the smaller streams present later in the season can only be reliably detected with imagery at a 1 m resolution or below (e.g., Figs. 8r and S5). The progression shown in Fig. 8 captures the development of an efficient drainage system over the season.
Imagery over Lake Julian shown from 20 April through 24 September 2019. All images use Planet SkySat Visual data unless otherwise indicated by letter on bottom left, with S indicating Sentinel and L indicating Landsat.
Three potential drainage mechanisms can be observed over Lake Julian. Firstly, we note a small stream ending in a spray of snow (alternatively, an ice bridge) which is potentially indicative of a moulin (Fig. 9b). However, an overflight of Operation IceBridge on 15 May (Fig. 9a) does not definitively show a moulin at the end of the incision, allowing for the possibility that the actual drainage occurred under the overlapping ice bridge. We also identify the point labeled B in Fig. 9a and b (also in Fig. 10d and e) as another potential drainage point. Presuming that drainage occurs at either location, lake volume could still increase slowly (as it does between 9 and 25 May) if the inflow rate exceeds the outflow rate of the lake. This dynamic is captured in a previous in situ study of lake drainage, which indicated that drainage through a moulin decelerated as the hydraulic head between the lake and the moulin declined (Tedesco et al., 2013).
Mechanisms of lake drainage over Lake Julian.
ICESat-2 RGT 727 over Lake Julian.
However, following late May, SkySat imagery captures the development of a second stream directly south of the initial potential moulin (Fig. 9b). This small stream, which flows downstream (Fig. 7b), deepens throughout the season (Fig. 9c) with a very deep incision shown distinctly in imagery on 24 September (Fig. 8t). The development suggests that the relatively slow initial drainage from the potential moulin or point B accelerated due to increased drainage from a second small stream, draining Lake Julian almost entirely between 1 and 10 June (Fig. 8m–n). While smaller pre-existent streams to the right (Fig. 9b) may also have facilitated drainage, we assume that the newly incised stream to the left is the most likely cause for the rapid drainage due to the timing of its appearance. We note that smaller bodies of water are still apparent on the surface after the drainage of Lake Julian, some connected by very small stream networks which appear to be frozen-over by 24 September (Fig. 8t).
Lake Julian reached a volume of 268 120 m
A comparison between the two passes of ICESat-2 RGT 727 (the second on 14 August) indicate uneven lowering in this region and potential slight ice motion, e.g., a small southward shift in the lowest point of the depth profile (point B) calculated on 15 May vs. on 14 August (Fig. 10c). We estimate large-scale surface lowering around
In addition to locations where Watta calculates a lake surface (Fig. 10d and e, label “B”), Watta also identifies regions where ice cover is probable. These are shown in cyan in Fig. 10d at locations A, C, and D and in imagery in Fig. 10e. Whereas lake surfaces are calculated at the horizontal resolution of the ICESat-2 ATL03 photon cloud, the ice surface class is assigned at a coarser resolution. This is because the Interpretive module assigns the “ice surface” class based on the presence of a flat surface under an overlying layer with more varied topography. While currently the algorithm potentially overestimates the extent of these regions, a first automatic pass can be used to identify larger regions where ice surfaces exist, after which manual inspection can then identify specific ice layers. Shown in Fig. 10d are lake B and three additional points where we identify ice layers using both Watta and manual inspection (A, C, D). Point B is also captured in Operation IceBridge CAMBOT imagery on 15 May (Fig. 9a). This is potentially a drainage point which retains meltwater as late as 14 August. This location is also covered by a floating ice layer on
14 May, suggesting that an ice layer had formed at the same place and settled at this point following drainage in the previous season. Point C corresponds to a deeply incised stream (which is not captured in the Watta profile calculated on 15 May), while point D corresponds to a smaller stream; both of these points were covered by Lake Julian during the last half of May. We note that the designation of point A is more ambiguous as it is collocated with an incision which was previously a stream but is weakly resolved in both Watta-based estimates and in the SkySat imagery collected on 20 August. The Watta designation of “ice surface” here is likely, but not unambiguous, as this method will capture both real ice surface and false dual returns as detailed in Sect. 3.1. The main attributes typical of the false dual returns are a strong top surface over surface water in a flat region, followed by weaker returns at predictable intervals (
Lake Niels shown between 20 April and 4 September 2019. Panel
In comparison to Lake Julian, Lake Niels begins and ends the melt season as a frozen lake. By 9 May, the ice surface begins to melt, and the lake surface area expands substantially. However, by 13 May (Fig. 11d), imagery captures an insulating layer of snow, after which the lake expands until 26 June. By 20 July, slow lake drainage is evident via a stream which is identified in Fig. 11a and is first observed to contain a substantial quantity of liquid water on 19 June (Fig. 11h). We observe that the stream is clearly incised both on 20 April (Fig. 11b) and on 24 September (Fig. 11m), when the lake is likely frozen over, based on imagery. In comparison to Lake Julian, which was located in a relatively shallow depression, Lake Niels is located in a deep depression (Fig. 7), and neither drains early in the season nor connects to an efficient drainage system. Although these lakes are subject to similar atmospheric drivers, the differences in drainage patterns highlight how local topography and the corresponding depth of lakes can influence how meltwater is either retained on the ice sheet or drained downstream or into englacial or subglacial pathways. We note that surface topography, in turn, is strongly influenced by basal topography, thus linking total runoff to surface expression of bed topography (Ignéczi et al., 2018).
This study represents initial work developing the Watta algorithm for lake depth estimates, as well as subsurface ice detection, using a unique stacked dataset over western Greenland during the intense 2019 melt season. We demonstrate the potential of ICESat-2 for automated lake detection and depth estimation, as well as how empirically derived depths derived from a combination of imagery sources can complement each source's strengths and weaknesses. For example, while Landsat is only available at a low resolution, it provides a rich historical record, as well as high geolocational accuracy (at 5 m), which is leveraged here to better geolocate imagery from Planet Labs. Similarly, while PlanetScope data contain several known issues with radiometry and geolocation, imagery is available at a high spatial and temporal resolution. As demonstrated in our test cases, a time series constructed from multiple sources can provide valuable information about the evolution of ice cover and drainage mechanisms in addition to volume estimates. Given the accelerating sophistication of altimetry-based observations, ongoing efforts to improve geolocation, radiometric quality, or temporal frequency of high-resolution imagery are crucial. Additionally, the availability of simultaneous imagery and altimetry would enhance the capabilities of other satellite imagery sources to fill out the time series by providing a calibration standard.
While this initial study focused on lakes in grounded ice in Greenland, Watta can potentially be applied to Antarctic melt lakes as well. Additionally, lake depths calculated empirically can potentially be used to calibrate physically based methods towards developing ice-sheet-wide time series for the evolution of surface hydrology.
Our algorithm successfully detects a wide variety of lake types automatically and can be applied to the growing set of ICESat-2 and imagery data over large sections of Antarctica and Greenland. Identification of narrow stream features on sloping surfaces, however, still needs visual verification due to a large number of false positives. This will be addressed in future work, together with adding features to the interpretive layer, including slush layers as well as cracks, using the Planet SkySat imagery dataset for testing purposes. In addition to these improvements, Watta, which is currently written in MATLAB, is available in GitHub with documentation but will eventually be moved to an open-source language.
Code for Watta can be found here:
Imagery, IDs for which can be found in the Supplement, can be accessed at the following urls:
for Planet imagery: for Sentinel-2 and Landsat 7 and 8 imagery: for ICESat-2 ATL03:
The supplement related to this article is available online at:
Both authors contributed equally to the design of the study, writing and running of the code, analysis and interpretation of the results, and writing and proofreading of this article.
Bert Wouters is editor of
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
We would like to acknowledge Lauren Andrews for her help in identifying drainage patterns, Jeremy Harbeck for his assistance with Operation IceBridge CAMBOT imagery, and the NASA Small Satellite Data Buy Pilot Program, as well as Planet Labs for the use of Planet Imagery.
Bert Wouters acknowledges funding by NWO grants 016.Vidi.171.063 and OCENW.GROOT.2019.091. Rajashree Tri Datta was funded by the ICESat-2 Project Science Office.
This paper was edited by Louise Sandberg Sørensen and reviewed by Jennifer Arthur and one anonymous referee.