Remote sensing is commonly used to monitor supraglacial lakes on
the Greenland Ice Sheet (GrIS); however, most satellite records must trade off
higher spatial resolution for higher temporal resolution (e.g. MODIS) or vice
versa (e.g. Landsat). Here, we overcome this issue by developing and applying
a dual-sensor method that can monitor changes to lake areas and volumes at
high spatial resolution (10–30 m) with a frequent revisit time (
In the summer, supraglacial lakes (hereafter “lakes”) form within the ablation zone of the Greenland Ice Sheet (GrIS), influencing the GrIS's accelerating mass loss (van den Broeke et al., 2016) in two main ways. First, because the lakes have low albedo, they can directly affect the surface mass balance through enhancing ablation relative to the surrounding bare ice (Lüthje et al., 2006; Tedesco et al., 2012). Second, many lakes affect the dynamic component of the GrIS's mass balance when they drain either “slowly” or “rapidly” in the middle to late melt season (e.g. Palmer et al., 2011; Joughin et al., 2013; Chu, 2014; Nienow et al., 2017). Slowly draining lakes typically overtop and incise supraglacial streams in days to weeks (Hoffman et al., 2011; Tedesco et al., 2013), while rapidly draining lakes drain by hydrofracture in hours to days (Das et al., 2008; Doyle et al., 2013; Tedesco et al., 2013; Stevens et al., 2015).
Rapid lake drainage plays an important role in the GrIS's negative mass balance because the large volumes of lake water can reach the subglacial drainage system, perturbing it from a steady state, lowering subglacial effective pressure, and enhancing basal sliding over hours to days (Shepherd et al., 2009; Schoof, 2010; Bartholomew et al., 2011a, b, 2012; Hoffman et al., 2011; Banwell et al., 2013, 2016; Tedesco et al., 2013; Andrews et al., 2014), particularly if the GrIS is underlain by sediment (Bougamont et al., 2014; Kulessa et al., 2017; Doyle et al., 2018; Hofstede et al., 2018). Rapid-lake-drainage events also have two longer-term effects. First, they open moulins, either directly within lake basins (Das et al., 2008; Tedesco et al., 2013) or in the far field if perturbations in stress exceed the tensile strength of ice (Hoffman et al., 2018), sometimes leading to a cascading lake-drainage process (Christoffersen et al., 2018). These moulins deliver the bulk of surface meltwater to the ice-sheet bed (Koziol et al., 2017), explaining the observations of increased ice velocities over monthly to seasonal timescales within some sectors of the GrIS (Zwally et al., 2002; Joughin et al., 2008, 2013, 2016; Bartholomew et al., 2010; Colgan et al., 2011; Hoffman et al., 2011; Palmer et al., 2011; Banwell et al., 2013, 2016; Cowton et al., 2013; Sole et al., 2013; Tedstone et al., 2014; Koziol and Arnold, 2018). Second, the fractures generated during drainage allow surface meltwater to reach the subfreezing ice underneath, potentially increasing the ice-deformation rate over longer timescales (Phillips et al., 2010, 2013; Lüthi et al., 2015), although the magnitude of this effect is unclear (Poinar et al., 2017). Alternatively, the water might promote enhanced subglacial conduit formation due to increased viscous heat dissipation (Mankoff and Tulaczyk, 2017). Although rapidly and slowly draining lakes are distinct, they can influence each other synoptically if, for example, the water within a stream overflowing from a slowly draining lake reaches the ice-sheet bed, thus causing basal uplift or sliding, and thereby increasing the propensity for rapid lake drainage nearby (Tedesco et al., 2013; Stevens et al., 2015).
While lake drainage is known to affect ice dynamics over short (hourly to weekly) timescales, greater uncertainty surrounds its longer-term (seasonal to decadal) dynamic impacts (Nienow et al., 2017). This is because the subglacial drainage system in land-terminating regions may evolve to higher hydraulic efficiency, or water may leak into poorly connected regions of the bed, producing subsequent ice velocity slowdowns either in the late summer, winter, or longer term (van de Wal et al., 2008, 2015; Bartholomew et al., 2010; Hoffman et al., 2011, 2016; Sundal et al., 2011; Sole et al., 2013; Tedstone et al., 2015; de Fleurian et al., 2016; Stevens et al., 2016). Despite this observed slowdown for some of the GrIS's ice-marginal regions, greater uncertainty surrounds the impact of lake drainage on ice dynamics within interior regions of the ice sheet, since fieldwork and modelling suggest that increased summer velocities may not be offset by later ice velocity decreases (Doyle et al., 2014; de Fleurian et al., 2016), and it is unclear whether hydrofracture can occur within these regions, due to the thicker ice and limited crevassing (Dow et al., 2014; Poinar et al., 2015). These unknowns inland add to the uncertainty in predicting future mass loss from the GrIS. There is a need, therefore, to study the seasonal filling and drainage of lakes on the GrIS, and to understand its spatial distribution and inter-annual variation, in order to inform the boundary conditions for GrIS hydrology and ice-dynamic models (Banwell et al., 2012, 2016; Leeson et al., 2012; Arnold et al., 2014; Koziol et al., 2017).
Remote sensing has helped to fulfil this goal (Hock et al., 2017; Nienow et
al., 2017), although it usually involves trading off either higher spatial
resolution for lower temporal resolution, or vice versa. For example, the
Landsat and ASTER satellites have been used to monitor lake evolution (Sneed
and Hamilton, 2007; McMillan et al., 2007; Georgiou et al., 2009; Arnold et
al., 2014; Banwell et al., 2014; Legleiter et al., 2014; Moussavi et al.,
2016; Pope et al., 2016; Chen et al., 2017; Miles et al., 2017; Gledhill and
Williamson, 2018; Macdonald et al., 2018). While this work involves analysing
lakes at spatial resolutions of 30 or 15 m, the best temporal
resolution that can be achieved using these satellites is
Because of the problems associated with the frequency or spatial resolution
of these satellite records, it has been suggested that greater insights into
GrIS hydrology might be gained if the images from multiple satellites could
be used simultaneously (Pope et al., 2016). Miles et al. (2017) were the
first to present such a record of lake observations in West Greenland,
combining imagery from the Sentinel-1 Synthetic Aperture Radar (SAR)
(hereafter “Sentinel-1”) and Landsat 8 Operational Land Imager (OLI)
(hereafter “Landsat 8”) satellites, and developing a method for tracking
lakes at high spatial (30 m) and temporal resolution (
The Sentinel-2 Multispectral Instrument (MSI) comprises the Sentinel-2A (launched in 2016) and Sentinel-2B (launched in 2017) satellites, which have 290 km swath widths, a combined 5-day revisit time at the Equator (with an even shorter revisit time at the poles), and 10 m spatial resolution in the optical bands; Sentinel-2 also has a 12-bit radiometric resolution, the same as Landsat 8, which improves on earlier satellite records with their 8-bit (or lower) dynamic range. Within glaciology, Sentinel-2 data have been used to, for example, map valley-glacier extents (Kääb et al., 2016; Paul et al., 2016), monitor changes to ice-dammed lakes (Kjeldsen et al., 2017), and cross-compare ice-albedo products (Naegeli et al., 2017); this research indicates that Sentinel-2 can be reliably combined with Landsat 8 since they produce similar results. Thus, Sentinel-2 imagery offers great potential for determining the changing volumes of lakes on the GrIS, for resolving smaller lakes, and for calculating volumes with higher accuracy than is possible with MODIS (Williamson et al., 2017).
In this study, our objective is to present an automatic method for monitoring
the evolution and drainage of lakes on the GrIS using a combination of
Sentinel-2 and Landsat 8 imagery, which will allow the mosaicking of a high-spatial-resolution (10–30 m) record, with a frequent revisit time
(approaching that of MODIS), something only possible by using the two sets of
imagery simultaneously. The objective is addressed using four aims, which are
as follows.
We aim to trial new methods for calculating lake areas, depths, and volumes from
Sentinel-2 imagery and assess their accuracy against Landsat 8 for 2 days
of overlapping imagery in 2016. We aim to apply the best methods for Sentinel-2 from (1), alongside existing methods
for calculating lake areas, depths, and volumes for Landsat 8, to all of the
available 2016 melt season (May–October) imagery for a large study site
( We aim to identify lakes that drain rapidly (in We aim to quantify the run-off volumes routed into the GrIS both during the
lake-drainage events themselves and afterwards via moulins opened by
hydrofracture, for the small and large lakes.
Here, we describe the study region (Sect. 2.1), the collection and pre-processing of the Landsat 8 and Sentinel-2 imagery (Sect. 2.2), the technique for delineating lake area (Sect. 2.3), the methods used to calculate lake depth and volume (Sect. 2.4), the approaches for automatically tracking lakes and identifying rapid lake drainage (Sect. 2.5), and the methods used to determine the run-off volumes that are routed into the GrIS's internal hydrological system following the opening of moulins by hydrofracture (Sect. 2.6).
Our analysis focuses on a
A total of 17 Landsat 8 images from May to October 2016 (Table S2) were
downloaded from the USGS Earth Explorer interface
(
The
A total of 39 Sentinel-2A level-1C images from May to October 2016 (Table S1)
were downloaded from the Amazon S3 Sentinel-2 database
(
Summary of the methods applied to the Sentinel-2 and Landsat 8 input data to calculate lake areas using the NDWI and depths using the physically based (“PB” in this figure) method. In this figure, the interpolation techniques used are indicated by “NN” for nearest neighbour or “B” for bilinear. The methods applied to Landsat 8 are shown only once the images had been reprojected and batch cropped to the same extent as the Sentinel-2 images (Sect. 2.2.1). The lake area outputs are compared between the two datasets as described in Sect. 2.3, and the physically based lake depth outputs are compared as outlined in Sect. 2.4. When the empirical lake depth method for calculating Sentinel-2 lake depths was also evaluated, the final Landsat 8 depths at 10 m resolution were directly compared against the original Sentinel-2 input band data (at native 10 m resolution) within the lake outlines defined with the NDWI.
Figure 2 summarises the overall method used to calculate lake areas and
depths for the Landsat 8 and Sentinel-2 imagery, including the cloud-masking
procedure described above, and the resampling required because the data were
distributed at different spatial resolutions. We derived lake areas for the
two sets of imagery using the normalised difference water index (NDWI)
approach, which has been widely used previously for medium- to
high-resolution imagery of the GrIS (e.g. Moussavi et al., 2016; Miles et
al., 2017). There were two stages involved here. First, we applied various
NDWI thresholds to the Sentinel-2 and Landsat 8 images and compared the
delineated lake boundaries against the lake perimeters in the background RGB
images. We then qualitatively selected the NDWI threshold for each type of
imagery based on the threshold that produced the closest match between the
two. Based on this qualitative analysis, we chose NDWI thresholds of 0.25 for
both types of imagery (Fig. 2). By varying the thresholds from these values
to 0.251 and 0.249, the total lake area calculated across the
whole image only changed by
For each Landsat 8 image, we calculated the lake depths and volumes using the
physically based method of Pope (2016) and Pope et al. (2016), based on the original method for ASTER imagery from Sneed
and Hamilton (2007). This approach is
based on the premise that there is a measurable change in the reflectance of
a pixel within a lake according to its depth since deeper water causes
higher attenuation of the optical wavelengths within the water column. Lake
depth (
Since no existing work has derived lake depths using Sentinel-2, we needed to formulate a new method. For this purpose, we used the Landsat 8 lake depths as our validation dataset. We conducted the validation on the two dates (1 and 31 July 2016) with contemporaneous Landsat 8 and Sentinel-2 images (as described in Sect. 2.3). We chose to test both physically based and empirically based techniques to derive Sentinel-2 lake depths, noting at the outset that physical techniques are generally thought to be preferable over empirical ones since they do not require site- or time-specific tuning.
For the physically based technique, we tested whether the same method as
applied to Landsat 8 (Eq. 1) could be used on the Sentinel-2 TOA reflectance
data. However, since Sentinel-2 does not collect panchromatic band
measurements, we could only use individual Sentinel-2 bands to calculate lake
depths. We applied this physically based technique to the red- and green-band
data within the lake outlines defined with the NDWI (Sect. 2.3). We derived
the value for
Our empirically based approach involved deriving various lake depth–reflectance regression relationships to determine which explained the most variance in the data. We used the Landsat 8 lake depth data (dependent variable) and the Sentinel-2 TOA reflectance data for the three optical bands (independent variables) for each pixel within the lake outlines predicted in both sets of imagery to determine which band and relationship produced the best match between the two datasets. To compare these values, we first resampled (using nearest-neighbour interpolation) the Landsat 8 depth data from 30 to 10 m to match the resolution of the Sentinel-2 TOA reflectance data (Fig. 2).
To evaluate the performance of the empirical versus physical techniques, we calculated goodness-of-fit indicators for the Sentinel-2 and Landsat 8 measurements derived from the empirically based technique (applied to all optical bands) and physically based method (applied to the red and green bands) for the two validation dates (1 and 31 July 2016) when contemporaneous Landsat 8 and Sentinel-2 images were available.
As for Landsat 8, Sentinel-2 lake volumes were calculated as the sum of the individual lake depths, multiplied by the pixel areas, within the lake boundaries.
Once validated, the new techniques to calculate lake areas, depths, and
volumes from Sentinel-2, as well as the existing methods for Landsat 8
(Sect. 2.4), were applied to the satellite imagery within the Fully Automated
Supraglacial lake Tracking at Enhanced Resolution (FASTER) algorithm to
produce cloud- and ice-marginal-free 10 m resolution lake area and depth
arrays for each day of the 2016 melt season for which either a Landsat 8 or
Sentinel-2 image was available (Fig. 2). For the days (1 and 31 July) when
both Landsat 8 and Sentinel-2 imagery was available (as used for the
comparisons above), in the FASTER algorithm, we used only the
higher-resolution Sentinel-2 images. The FASTER algorithm is an adapted
version of the Fully Automated Supraglacial lake Tracking (FAST) algorithm
(Williamson et al., 2017), which was developed for MODIS imagery. The FASTER
algorithm involves creating an array mask to show the maximum extent of lakes
within the region in the 2016 melt season, by superimposing the lake areas
from each image. Within this maximum lake-extent mask, changes to lake areas
and volumes were tracked between each consecutive image pair, with any lakes
that were obscured (even partially) by cloud marked as no data. We only
tracked lakes that grew to
From the time series, a lake was classified as draining rapidly if two
criteria were met: (i) it lost
To determine how much extra information could be obtained from the finer-spatial-resolution satellite record, we compared the number of rapidly
draining lakes identified that grew to
Goodness-of-fit indicators for the empirical and physical techniques
tested in this paper for deriving Sentinel-2 lake depths, with validation
against the Landsat 8 lake depth measurements, on 1 and 31 July 2016.
Comparison of lake depths calculated using the physically based
method for Sentinel-2 (with the red band) and for Landsat 8 (with the average
depths from the red and panchromatic bands). Degrees of freedom (“df” in
this figure)
The empirical power-law regression (solid black curve, equation
Using the dual Sentinel-2 and Landsat 8 record, the locations and timings of moulin openings by “large” and “small” rapidly draining lakes were identified. Then, at these moulin locations, the run-off volumes that subsequently entered the ice sheet were determined using statistically downscaled daily 1 km resolution RACMO2.3p2 run-off data (Noël et al., 2018). Here, run-off was defined as melt plus rainfall minus any refreezing in snow (Noël et al., 2018). These data were reprojected from Polar Stereographic (EPSG: 3413) to WGS 84 UTM zone 22N (EPSG: 32622) for consistency with the other data and resampled to 100 m resolution using bilinear resampling. Then, the ice-surface catchment for each rapidly draining lake was delineated using MATLAB's “watershed” function, applied to the GIMP ice-surface elevation data (Howat et al., 2014). The elevation data were first coarsened using bilinear resampling to 100 m resolution from 30 m native resolution. For each of the days after rapid lake drainage had finished, it was assumed that all of the run-off within a lake's catchment reached the moulin in that catchment instantaneously (i.e. no flow-delay algorithm was applied) and entered the GrIS. This method therefore assumes that once a moulin has opened at a lake-drainage site, it remains open for the remainder of the melt season. This allowed first-order comparisons between cumulative run-off routed into the GrIS via the moulins opened by small and large lake-drainage events.
Table 1 shows the results of the lake depth calculations using the physically
and empirically based techniques applied to imagery from 1 and 31 July 2016
when contemporaneous Landsat 8 and Sentinel-2 images were available. The
physically based method applied to the red and green bands (Figs. 3 and S2,
respectively) performed slightly worse (for the red band:
Although the physically based method performed slightly worse than the empirical techniques, the physical method is preferable because it can be applied across wide areas of the GrIS and in different years without site- or time-specific tuning; it is likely that a different empirical relationship would have better represented the data for a different area of the GrIS or in a different year. We therefore carried forward the physically based method applied to the red band into the lake-tracking approach. We selected the red band instead of the green band because of the large difference between the depths calculated with the two satellites at higher values when using the green band (Fig. S2). We defined the error on all of the subsequently calculated lake depth (and therefore lake volume) measurements for Sentinel-2 using the RMSE of 0.555 m and treated the Landsat 8 measurements as ground-truth data, meaning they did not have errors associated with them.
Having verified the reliability of the lake area and depth techniques for
both Sentinel-2 and Landsat 8, the automatic calculation methods were
included in the FASTER algorithm to derive seasonal changes to lake areas and
depths, and therefore volumes. The FASTER algorithm was applied to the
Landsat 8 and Sentinel-2 image batches individually, as well as to both sets
when combined into a dual-satellite record. Using the dual-satellite image collection
produced an improvement to the temporal resolution of the dataset over the
melt season (1 May to 30 September) from averages of 9.0 days (for Landsat 8)
and 3.9 days (for Sentinel-2) to 2.8 days (for the dual-satellite record). The months
of June and July had the most imagery available (both with 14 images) within
the dual-satellite analysis. For the Landsat 8 individual analysis, the
algorithm tracked changes to 453 lakes that grew to
The largest lake size varied between the analyses: 4.0 km
Sample time series of lake volume to show seasonal changes for
Evolution to total lake area and volume across the whole study
region during the 2016 melt season. “Portion of region visible” measures
the percentage of all of the pixels within the entire region that are visible
in the satellite image, i.e. which are not obscured either by cloud (or cloud
shadows) or are not missing data values. Figure 7 presents total lake area
and volume after normalising for the proportion of region visible. Blue error
bars for lake area were calculated by multiplying the lake area RMSE of
0.007 km
Using the full Sentinel-2–Landsat 8 dataset, the FASTER algorithm produced
time series that documented changes to individual lake volumes over the
season, samples of which are shown in Fig. 5. Total areal and volumetric
changes across the whole region were calculated by summing the values for all
lakes in the region. However, we found that cloud cover (which was masked
from the images) often affected the observational record, and there were time
periods, such as early July and the end of August, with a lot of missing data
(Fig. 6). Figure 7 was therefore produced to normalise total lake areas and
volumes against the proportion of the region visible, and this shows the
estimated pattern of lake evolution on the GrIS: there was virtually no water
in lakes before June, steady increases in total lake area and volume until
the middle of July, and then a gradual decrease in total lake area and volume
through the remainder of the season, with most lakes emptying by early
September (Figs. 6 and 7). Dates with seemingly low total lake areas and
volumes were usually explained by the low portion of the whole region visible
in those images (Figs. 6 and 7). Finally, as in previous studies (e.g. Box
and Ski, 2007; Georgiou et al., 2009; Williamson et al., 2017), we found a
close correspondence between lake areas and volumes: comparing lake area and
volume values from all dates produced an
Estimates of evolution to total lake
Properties of rapid-lake-drainage events identified using the
satellite datasets individually and as part of a dual-satellite dataset.
Large lakes are defined as
Dates of rapid drainage events for small (circles) and large (triangles) lakes in 2016. The panel coverage and background are the same as that shown in Fig. 1. The extreme colour bar values include those dates outside of the range shown (i.e. before 19 June and after 5 September).
Table 2 shows the results of the identification of rapidly draining lakes using the three different datasets and indicates that the dual-satellite record was better for identifying rapidly draining lakes than the individual records. This was for two main reasons. First, the dual-satellite record identified 118 (or 91) more rapidly draining lakes than the Landsat 8 (or Sentinel-2) record in isolation (Table 2). When either record was used alone, Sentinel-2 (or Landsat 8) performed better (or worse), identifying 50.5 % (or 35.9 %) of the total number of rapidly draining lakes identified by the dual-satellite record. Second, with the dual-satellite dataset, drainage dates were identified with higher precision (i.e. half of the number of days between the date of drainage initiation and cessation; Sect. 2.5.2) than with the Sentinel-2 analysis (Table 2). However, the precision appears higher for the Landsat 8 analysis than either the dual-satellite or Sentinel-2 analysis, and this is because nearly all Landsat 8 lake-drainage events occurred on two occasions when the pair of images was only separated by a day, on 8–9 July (small lakes) and 13–14 July (large lakes) (Table 2).
The dual-satellite record also identified the rapid drainage of many small
lakes (
Finally, we tested how adjusting the thresholds used to define rapidly draining lakes would impact rapid-lake-drainage identification. Changing the critical volume loss required for a lake to be identified as having drained from 80 % to 70 % and 90 % resulted in the identification of only six more and four fewer rapid-lake-drainage events, respectively. Similarly, changing the critical-refilling threshold from 20 % to 10 % and 30 % resulted in identifying only eight fewer and five more rapidly draining lakes. However, adjusting the timing over which this loss was required had a larger impact, with adjustments from 4 to 3 and 5 days producing 37 fewer and 65 more rapid-lake-drainage events, respectively.
Each rapid-lake-drainage event from the dual-satellite record delivered a
mean water volume of
Lake water volumes measured using the physically based technique on
the days prior to rapid drainage, categorised into small (
Using the data from the dual-satellite record, and considering just the water
volumes delivered into the GrIS during lake-drainage events (and not
subsequently via the moulins opened), the drainage of small (
Frequency distribution of water volumes prior to rapid drainage for small and large lakes to show the lower and more tightly clustered water volumes contained within small lakes compared with large lakes. Natural logs of water volumes were taken for presentation purposes.
Cumulative run-off volume, from RACMO2.3p2 data (Noël et al.,
2018), entering the GrIS over the remainder of the melt season via the
moulins opened by rapid lake drainage for small (
The first and second aims of this study involved trialling and then applying a new method for calculating lake depths from Sentinel-2 imagery. We found an RMSE of 0.555 m for lake depths calculated with the physically based method applied to Sentinel-2's red band when compared with lake depths calculated for Landsat 8 using existing methods (Pope et al., 2016). When we applied the physically based method to Sentinel-2's green band and compared the depths with Landsat 8 measurements, we found a slightly lower RMSE, but the Sentinel-2 depths were unrealistically high compared with Landsat 8 values, and so this method was excluded (Table 1; Fig. S2). We selected the physical method over the empirical one because the empirical method cannot be applied without the site- or time-specific adjustments suggested in previous research (e.g. Sneed and Hamilton, 2007; Pope et al., 2016; Williamson et al., 2017) and might therefore perform more poorly in other years and/or for other regions of the GrIS. Given that the performance of the two methods was very similar, it therefore seemed most sensible to use the more robust physically based technique. In addition, the RMSE value (0.555 m) obtained here using the physically based method applied to the red band is only slightly higher than the error on lake depth calculations using the physical method for similar-resolution Landsat 8 data (0.28 m for the red band and 0.63 m for the panchromatic band; Pope et al., 2016). However, the RMSE on Sentinel-2 lake depths is less than half of both that produced using the physically based method applied to coarser-resolution (250 m) MODIS red-band data (1.27 m; Williamson et al., 2017) and that produced using an empirical depth–reflectance relationship for MODIS (1.47 m; Fitzpatrick et al., 2014). Therefore, using the Sentinel-2–Landsat 8 record over the MODIS one produces a much more reliable measure of lake water depths on the GrIS because of the improved spatial resolution. The dual-satellite record is even further strengthened by its high temporal resolution, which approaches that of MODIS (Sect. 4.2).
Despite the low overall error for the physically based lake depth
calculations from Sentinel-2, we observed different performances on the two
validation dates (1 and 31 July; Sect. 3.1): the depths calculated for
Sentinel-2 and Landsat 8 showed closer agreement on 1 July than on 31 July
(Fig. 3). This is likely because clouds obscured a large portion of the image
from 1 July. Although the lakes used for comparison were cloud free and
pixels within 200 m of a cloud-marked area were filtered, there were likely
adjacency effects (at distances
Alternatively, the presence of clouds on the 1 July image might be indicative of a difference in the atmospheric composition on that day, which could have affected the lake depth calculations with Sentinel-2, but not to the same degree with Landsat 8. This might be because of the difference in bandwidths between the satellites, or because Landsat 8's panchromatic band (used for calculating lake depths) is less sensitive to the presence of clouds on an image, therefore producing more reliable lake depth measurements. The effect of clouds on the atmosphere could have been better accounted for if the Sentinel-2 TOA reflectance data had been first converted to bottom-of-atmosphere (i.e. surface-reflectance) measurements. However, while surface-reflectance data are available for Landsat 8's optical bands, they are not for its panchromatic band, meaning that the lake depth calculation method used here could not have been applied to generate reliable ground-truth data. We therefore intentionally chose not to perform this correction on the Sentinel-2 TOA data because we wished to directly compare the measurements from the two satellites.
Finally, in this study, the Sentinel-2 lake depths were validated using Landsat 8 measurements, which were regarded as ground-truth data, in line with a previous study involving validation of depths calculated with MODIS (Williamson et al., 2017). This approach was justified since previous work (Pope et al., 2016) indicated a close agreement between Landsat 8 lake depths and DEM measurements. However, it is important to note that the Landsat 8 data have errors associated with them, including a possible under-measurement of the deepest water due to saturation of the red band within the water column (Moussavi et al., 2016; Pope et al., 2016). Future work involving Sentinel-2 lake depth calculations could therefore alternatively validate Sentinel-2 lake depth estimates using different ground-truth validation data, such as higher-resolution (e.g. WorldView-2) satellite imagery, high-resolution DEM measurements of lake basins, or field lake depth measurements.
The second aim of this research was to apply the new methods for calculating
lake areas, depths, and volumes from Sentinel-2 imagery alongside those for
Landsat 8 within the FASTER algorithm to produce time series for the
evolution of all lakes. Applying this algorithm to the dual-satellite record
allowed us to track the evolution of 690 lakes. The mean lake size
(0.137 km
The third and fourth aims of the work were to identify the lakes tracked by
the FASTER algorithm that drained rapidly, and to investigate the quantity of
run-off reaching the GrIS's internal hydrological system both during the
drainage events themselves and subsequently via the moulins opened by rapid
lake drainage since recent work (Banwell et al., 2016; Koziol et al., 2017)
has shown that the moulins opened by rapid-lake-drainage events allow much
greater run-off volumes to reach the subglacial system than the volumes
released during the actual drainage events themselves. Most research to date
has used MODIS imagery to identify rapidly draining lakes because the high
temporal resolution is required to separate rapidly draining lakes from those
draining slowly. Although this MODIS-based research has been helpful for
quantifying the characteristics of relatively large lakes (
Rapid drainage of both large and small lakes can be identified using the
FASTER algorithm with the dual-satellite record. Although the water volumes
associated with the drainage of small lakes into the GrIS amount to just
5.1 % of the total water volume associated with the drainage of all lakes
across the region, rapid drainage of small lakes is important because, like
large lakes, they open moulins that can direct surface run-off into the
GrIS's internal hydrological system over the remainder of the season. This
assumes that the moulins remain open for the rest of the melt season, and we
note that this may vary across the study region according to ice thickness or
stress state. However, acknowledging this assumption, with the dual-satellite
record, we identified 105 small rapid-lake-drainage events, thus providing
105 more input locations for surface run-off to reach the ice sheet's
internal hydrological system than would be identified by MODIS. The moulins
opened by small lake-drainage events are particularly important because in
total they deliver over half (61.5 %) of the total run-off delivered via
all moulins into the GrIS's internal hydrological system. This is because the
small rapidly draining lakes are more numerous (105 compared with 79) and
tend to be at lower elevations than the larger lakes (small lake mean
elevation
Over the 2016 melt season, 27 % of all lakes detected in the region
drained rapidly, compared with 21 % that drained rapidly in 2014 across
the slightly smaller Paakitsoq region contained within the region of this
study (Williamson et al., 2017). However, that earlier study used MODIS
imagery, so it omitted the rapid drainage of small lakes, which could explain
the lower percentage if it is assumed that these small lakes are more likely
to drain rapidly than the large ones, relative to the total numbers of lakes
in each category. Therefore, considering just the rapid drainage of large
lakes (i.e. which could be identified by MODIS) we found that 18 % of
large lakes drained rapidly, which is similar to the 21 % value in
Williamson et al. (2017). The 27 % value in this study compares well with
that of 22 % from Miles et al. (2017), who also tracked changes to small
lakes using a similar tracking threshold to that used here, albeit for a
different combination of satellite platforms (Landsat 8 and Sentinel-1), and
for a larger region of West Greenland in the 2015 melt season. The precision
of rapid-lake-drainage dates in this study (
We have presented the results of the first approach to combine two
medium-resolution optical satellite datasets (Sentinel-2 and Landsat 8) to
generate the highest spatial- and temporal-resolution record of lake area and
volume evolution on the GrIS to date. To achieve this, we have exploited the
increasing availability of medium- to high-resolution satellite imagery and
then combined these newly available data with recent techniques for
automatically tracking changes to lake areas and volumes and for identifying
rapid lake drainage. The resultant FASTER algorithm allows lake areas and
volumes to be calculated with high accuracy from Sentinel-2. For lake area,
the RMSE is 0.007 km
We have additionally taken advantage of new, and increasingly reliable, downscaled regional climate-model (RACMO2.3p2) output data (Noël et al., 2018) to provide insights into the run-off volumes entering the GrIS's englacial or subglacial hydrological systems after moulin opening was identified using the FASTER algorithm. Our results show that the water volumes released into the GrIS by small lakes during the lake-drainage events themselves are small (only 5.1 %) relative to the volumes released by all lake-drainage events, suggesting small lakes are less important in this sense. However, of the total water volume that subsequently reaches the GrIS's internal hydrological system via all moulins opened by lake drainage (from both large and small lakes), moulins opened by small lakes deliver 61.5% of the total run-off volume. This suggests that small lakes are important to include in future remote-sensing and modelling studies.
Resulting from the above, the FASTER algorithm holds great potential for generating novel insights into lake behaviour on the GrIS from remote sensing, including for small lakes that change quickly. Future work should focus on applying the FASTER algorithm to wider areas of the GrIS and comparing the results with increasingly available and reliable high-temporal-resolution ice velocity data (e.g. Joughin et al., 2018) to investigate the influence of lake drainage on the observed patterns of intra- and inter-annual velocity variations across the GrIS. Moreover, the high-spatial-resolution record could be used to identify the potential controls on the initiation of rapid lake drainage, something that could not be achieved with MODIS data, perhaps due to the data's coarse spatial resolution (Williamson et al., 2018). Finally, the water volumes delivered into the GrIS during the rapid-lake-drainage events identified with this record, the moulins that are assumed to open during such events, and the subsequent run-off that enters the GrIS via these moulins, could be used as forcing or testing data for subglacial hydrology models (e.g. Hewitt, 2013; Banwell et al., 2016) and linked hydrology–ice dynamics models (e.g. Koziol and Arnold, 2018). Ultimately, applications of the FASTER algorithm such as these could enable the GrIS's supraglacial and subglacial hydrology to be modelled more accurately in order to provide better constraints on future run-off, ice discharge, and sea-level rise from the GrIS.
All satellite imagery is open access (see Sect. 2.2), and regional climate-model output data are available as described in Noël et al. (2018). The full MATLAB source code for the FASTER algorithm used to process and analyse the imagery is freely available for download (Williamson, 2018b).
The supplement related to this article is available online at:
AGW conceived the study, designed and executed the method presented in the research, conducted the analysis, and drafted the original paper, all under the supervision of the other authors. All authors discussed the results and contributed towards editing the paper. AGW revised the paper following reviewer and editorial comments.
The authors declare that they have no conflict of interest.
Andrew G. Williamson was funded by a UK Natural Environment Research Council PhD studentship (NE/L002507/1) awarded through the Cambridge Earth System Science Doctoral Training Partnership and a Cambridge Philosophical Society research studentship. Alison F. Banwell was funded by a Leverhulme/Newton Trust Early Career Fellowship (ECF-2014-412). The Scott Polar Research Institute's B. B. Roberts Fund and the Cambridge Philosophical Society provided funding for Andrew G. Williamson to present this research at the European Geosciences Union General Assembly 2018. We are grateful to Allen Pope for discussing the results of the Sentinel-2 lake depth calculations with us and to Brice Noël for speedily providing the RACMO2.3p2 data. Katie Miles and Corinne Benedek are thanked for generally contributing to the idea for the study, and we thank Gareth Rees and Pete Nienow for providing thoughtful feedback on this work. Finally, detailed reviewer and editorial comments helped to significantly improve the quality of the paper. Edited by: Bert Wouters Reviewed by: Allen Pope, Samuel Doyle, and Kristin Poinar