2.1 Platform

This paper describes data collected during the 2016–2017 field season using an AS-350 helicopter to fly a VHF radar designed by the University of Texas Institute for Geophysics (UTIG; Fig. 2). Additional science instrumentation included a Renishaw laser altimeter and a Canon dSLR camera. For precise positioning of the data a Trimble Net-R9 dual-frequency carrier-phase GNSS and a Novatel SPAN IGM-1A inertial navigation system were used. After initial test flights, the instruments did not require an operator on board; the pilot had a power switch that was used to disable VHF transmission when necessary. All flights were based out of South Korea’s Jang Bogo Station and were supported by KOPRI.

Figure 2Helicopter in survey configuration. Radar antennas were mounted in the lateral booms, which measure 11.52 m from tip to tip; the front boom was empty. The laser altimeter and dSLR camera were mounted to the right of the pilot’s seat using a pre-existing window.

Download

2.2 Laser altimeter

A Renishaw ILM-1200-HR 905 nm laser altimeter was mounted to the right of the pilot’s seat, collocated with the camera and utilizing an existing downward-facing window in the helicopter. It provided raw range measurements at 1000 Hz with 1 cm precision, and its effective maximum range in Antarctic conditions was ∼900 m. The raw serial stream was recorded by the same acquisition system as the radar data, which provides synchronous timestamps with reference to GPS time.

2.3 Ice-penetrating radar

The new instrument described here is a direct descendant of a lineage of coherent radars that started with an experimental field season in 2001. The original system, termed the High-Capability Airborne Radar Sounder (HiCARS; Peters et al., 2005), was a hybrid of a JPL-designed coherent radar (Moussessian et al., 2000) and the Technical University of Denmark (TUD) 60 MHz airborne ice-penetrating radar system (Skou and Søndergaard, 1976). It was first mounted on a Twin Otter airplane in 2001 to perform surveys of the Siple Coast (Peters et al., 2005), South Pole, and the B15a iceberg (Peters et al., 2007b). This was followed by the 2005 Airborne Geophysical Survey of the Amundsen Sea Embayment, Antarctica (AGASEA), survey of Thwaites, also using Twin Otters (Holt et al., 2006). Since 2008, the International Collaborative Exploration of the Cryosphere by Airborne Profiling (ICECAP) project has been fielding similar radars using a DC-3T airplane, and UTIG entirely redesigned the electronics with a focus on using commercial, off-the-shelf components to create the HiCARS2 radar in 2010 (Blankenship et al., 2017a, b). In 2014, independent recording from each antenna was added to create the Multifrequency Airborne Radar Sounder with Full-phase Assessment (MARFA; Castelletti et al., 2017), in which digitizer improvements also enabled the replacement of local-oscillator-based down conversion with bandpass sampling. The system described in this paper uses the same electronics as MARFA but with custom antennas for installation on an AS-350 helicopter.

Figure 3Focused radargram DVG/IBH0c/GL0328c. The location of the transect is shown in Fig. 7, and its overlap with the D2 lake outline is highlighted in blue. Observed surface elevation change since 2003 is shown in red. The radargram has been depth-corrected based on laser surface elevations and truncated 20 m above the surface.

Download

The radar transmits 1 µs wide chirps, linearly sweeping the frequency from 52.7 to 67.5 MHz with a 6250 Hz pulse repetition frequency and 8 kW peak pulse power. Using separate 14-bit digitizers with low gain for the surface and higher gain for the bed, the system independently records the received signal from each antenna at 50 Msamples s⁻¹, with a total trace length of 3200 samples. The record is stacked 32 times in hardware, then written to disk with 16-bit precision at 196 Hz. This resulted in one raw trace every 18 cm along-track at an average helicopter ground speed of 70 knots, or 42 cm at the DC-3T ground speed of 160 knots.

The ability to compare data between HiCARS, HiCARS2, MARFA, and the present system has been considered to be of paramount importance in developing and fielding the new system. The required processing techniques are functionally identical, with differences confined to peak power output and gains as well as the platform-dependent antennas.

The airplane antennas have a heritage dating back to the TUD radar from the 1970s. They are center-fed flat-plate dipole antennas installed inside an airfoil and mounted 1∕4 wavelength under the wings. The helicopter’s antennas were designed to fit inside existing flight-certified booms originally designed for magnetometer surveys. These geometric constraints led to an end-fed design with an end plate installed in each lateral boom; the forward boom was empty. The lack of an airplane wing providing a ground plane means that the upward lobe is not reflected, yielding 6 dB lower total system gain. Additionally, the smaller separation between the antennas yields a wider central lobe, leading to increased surface scattering that can be mitigated by flying closer to the ice surface. There is no evidence in the data for time-varying interference due to the helicopter blades rotating at ∼400 rpm.

3.1 Positioning

Processing of GPS observations was performed using Novatel’s Waypoint GrafNav software, which reports ∼15 cm σ for precision. All data in this paper are reported with reference to the WGS84 ellipsoid.

3.2 Laser altimeter

All analyses presented in this paper used data that geolocated the median of 100 raw range measurements, which spans ∼3.6 m along-track.

3.3 Ice-penetrating radar

For this work, we used the 1D-focused processing for radargrams described in Peters et al. (2007a) for geometry and basal reflectivity, complemented with 2D focusing to derive specularity content (Schroeder et al., 2015). Figure 3 shows an example radargram that crosses D2. Focusing is performed by convolving a kernel with pulse-compressed radar data, where the kernel is generated based on the expected appearance (delay and phase) of a point scatterer at that location, which is a function of airplane height, ice thickness, and surface slope. Different aperture lengths are used for focusing, which correspond to the 1D and 2D nomenclature in Peters et al. (2007a). One-dimensional focusing uses a short enough aperture that range changes are less than a pulse width; for a longer aperture, a 2D kernel (in this case accommodating 1 µs of range change) is required to match the phase history, further improving resolution, collection of scattered energy, and detection of sloping interfaces at some cost to the signal-to-noise ratio.

3.3.1 Topography

For estimating the bed elevation and ice thickness, the radar reflection off the basal interface was identified by a manual process that labels the first-returned continuous reflector (Blankenship et al., 2001). The right-side radar antenna had a stall strip that raised its noise floor, so labeling was performed on the data from the left antenna only (further processing used the combined product). Due to the existence of side lobes in the transmit–receive beam pattern, the first return criteria may underestimate ice thickness in rough terrain, overestimate the width of mountains and ridges, and possibly fail to detect valley floors and lakes. Given the surface and bed horizons, ice thickness is calculated using 168.42 m µs⁻¹ as the speed of light in ice without a correction for the firn density gradient. The bed elevation product results from subtracting this ice thickness from the laser-determined surface elevation.

All intersecting lines where bed picks were recovered within a 100 m radius are used to characterize the uncertainty in bed elevation estimates. There is no attempt to reconcile differing bed estimates from intersecting transects in the labeling process: picking is purely based on the first-return criteria. This means that the computed crossover differences are valid for projecting to regions without crossovers but with equivalently rough topography. These crossover differences are shown in Fig. 4. Of the 450 locations where survey lines intersected and a bed was recovered, 76 % had differences less than 50 m, 88 % had differences less than 100 m, and 95 % were under 200 m. There was no clear spatial pattern to the distribution of errors, and inspection of the six intersections with a difference greater than 400 m revealed that the apparent errors were consistent with the observed along-track variation in bed elevation at length scales equivalent to the across-track beam width.

Generation of the digital elevation model (DEM) started with the full-density bed elevation points from every trace at which a bed return was detected. These profiles are preprocessed for gridding by sampling to a 500 m cell size using the Generic Mapping Tools’ (Wessel and Smith, 1998, 1991) blockmean function. Natural neighbor interpolation was performed on this decimated dataset using matplotlib’s (Hunter, 2007) griddata function, which is based on Delaunay triangulation. This interpolation retained artifacts along the flight lines, so a 1 km standard deviation Gaussian filter was applied as a final step. The DEM presented here reports the ice–water interface beneath floating ice; it makes no attempt to mask the grounding line or correct for the water column to determine bathymetry.

Figure 4Bed elevation DEM generated from KOPRI data, with 100 m contours and errors in bed elevation between intersecting lines. Lake outlines from Smith et al. (2009) are black.

Download

Profile-based ice thicknesses can be problematic to grid due to their anisotropic sampling density. This dataset’s line spacing does not support using a higher-resolution DEM; therefore, when compared to the raw data, there are sometimes significant gridding errors. They are particularly noticeable in higher-relief areas, where bed features are flattened and broadened. Unlike the crossover errors, these gridding errors follow a roughly normal distribution, with a standard deviation of 95 m. Of the approximately half million bed elevation measurements, 53 % differ from the gridded product by less than 50 m error, 79 % by less than 100 m, and 95 % by less than 200 m.

3.3.3 Reflection coefficients

The strong dielectric contrast between water and ice means that this reflection should be significantly brighter than one produced by ice and rock. This observation has been used in an attempt to identify subglacial water as early as Robin et al. (1969) and has been used frequently since (e.g., Oswald and Robin, 1973; Siegert et al., 1996; Carter et al., 2007).

The radar equation describes the amplitude of the returned signal at the antenna (P_r) in terms of system and environmental parameters (Peters et al., 2005) assuming a specularly reflecting interface:

(3) P r = P t λ 1 4 π 2 G t G r T 12 2 L ice 2 2 h + z / n 2 2 R 23 ,

where R₂₃ is the ice–bed reflection coefficient that we are interested in. T₁₂ is the air–ice transmission coefficient. Transmitted power (P_t), antenna gain due to cross section (λ14π), and the receiver and processing gains (G_t, G_r) combine to determine the system gain. The geometric spreading loss

(4) L s = 1 2 h + z / n 2 2

is a function of aircraft height above the ice surface (h) and ice thickness (z), both of which can be recovered directly from the interpreted radar data, along with the dielectric constant for glacial ice (n2=ϵ=1.78).

Finally, L_ice, the energy lost as an electromagnetic wave travels through a dielectric medium, is a function of its permittivity. For ice, this depends primarily on temperature and chemistry (Matsuoka, 2011). Across Antarctica, one-way depth-averaged dielectric ice loss (N_a) varies from 3 to 30 dB km⁻¹ (Matsuoka et al., 2012). This wide variation in physically feasible values is the dominant source of uncertainty when calculating reflection coefficients.

Some studies attempt to determine L_ice independently from the radar data, either deriving it from first principles based on modeled temperature profiles and salt content (Matsuoka et al., 2012) or extrapolating from measured properties at ice cores (MacGregor et al., 2007). Other studies estimate ice loss from the radar data: Peters et al. (2005) assumed that the brightest echoes correspond to water at the bed and that depth-averaged L_ice is constant across the survey area; Jacobel et al. (2009) assumed that the distribution of reflection coefficients is independent of ice thickness and obtained L_ice from the slope of ice thickness vs. geometry-corrected returned power. Recent work has refined these approaches to infer spatially varying patterns of dielectric ice loss across a survey at a resolution determined by the topographic variation (Schroeder et al., 2016a).

This paper does not attempt to use absolute reflection coefficients to verify the existence of water at the bed. Instead, they are used to compare relative bed properties under similar ice thicknesses. Absolute values would require calibration of the system gain (typically obtained both at the lab bench and by collecting data over open seawater) and validation to previous systems.

Given these goals, we modified the simplest approach of determining depth-averaged dielectric ice loss from the slope of geometry-corrected echo strengths vs. ice thicknesses. This approach relies on the assumption that basal reflection coefficients are independent of ice thickness, which is overly simplifying since basal temperature – and therefore the presence of water at the bed – is correlated with ice thickness. We see evidence of a slope change associated with the likely presence of water, so we restricted our linear regression to data in thinner ice. Since thinner ice is on average cooler than thicker ice, restricting the range of thicknesses used in the fit will result in an estimate that is a lower bound on dielectric ice loss.

Figure 5 shows reflection coefficient data derived from peak power measurements from all KOPRI radar bed picks in the study region. In order to address the uneven distribution of samples across ice thicknesses, it also shows the median reflection coefficient for 50 m bins of ice thickness, which is the input used in calculating the linear regression.

Figure 5Geometry-corrected reflection coefficients vs. ice thickness for all bed picks in the survey. Red dots show the median reflection coefficient for 50 m wide ice thickness bins that were used in calculating the slope. The solid red line shows the slope for −15.1 dB km⁻¹ (one-way) based on calculations for ice thicknesses less than 1300 m. The red rectangle shows the region where we assume that increased presence of water at the bed causes a broadening of the reflection coefficient distribution, and the gray triangle indicates the region where the signal-to-noise ratio could explain the absence of data. The dotted line has been shifted up 20 dB from the shallow-ice fit, representing the higher reflection coefficients expected if there is water at the bed. The blue line shows the cumulative probability of a thawed bed for a given ice thickness.

Download

At first glance, this does not appear to be a linear distribution; there is a higher slope for the thinner ice than for the thicker ice. It would be surprising if this were due to the distribution of dielectric losses since thinner ice is typically colder on average and thus has a lower dielectric ice loss. Instead, this distribution can be explained as a combination of the radar system’s noise floor and changing basal properties with depth.

The observed geometry-corrected reflection coefficients have a minimum value of around −110 dB, which serves to cut off the linear distribution. There are areas where the bed can only be identified as a disturbance in the background noise. While the manually interpreted bed picks include these regions, their computed reflection coefficients are not valid. This threshold is not a hard limit because the noise distribution varies from trace to trace. Additionally, the lower bound would be expected to have a slight positive slope due to the effects of correcting for spreading loss, which is apparent in the data.

Liquid water at the bed would be expected to increase the range of observed reflection coefficients, with a maximum value up to 20 dB above those observed on a dry bed. This can explain the observed widening and/or shift of the distribution at depths between 1300 and 1800 m since the existence of basal water typically requires the insulation provided by thicker ice. Due to the noise floor, a linear fit in this region will underestimate L_ice. However, the slope of the upper bound of the scatterplot at depths over 1700 m matches the average slope at depths under 1200 m, which supports a roughly constant L_ice.

Ice thickness required to reach the basal melting point can be estimated using the Robin model (Robin, 1955; Cuffey and Patterson, 2010), which is a 1D model that accounts for ice thickness, accumulation rate, surface temperature, geothermal flux, and basal heat generation. There are many degrees of freedom, and in an attempt to simplify and constrain the possible range of solutions, a Monte Carlo approach was adopted. The accumulation rate in ice equivalent was assumed to have a normal distribution with 1 standard deviation of 0.06±0.02 m a⁻¹ (van Wessem et al., 2014b). Geothermal flux was assumed to have a normal distribution of 0.06±0.01 W m⁻2 (An et al., 2015). Surface temperature was assumed to have a normal distribution of -35±5 ^∘C (van Wessem et al., 2014a). The standard deviation of the geothermal flux was expected to capture the effects of frictional heating of 0 to 0.01 W m⁻², which is appropriate for up to 50 m yr⁻¹ of basal sliding (Rignot et al., 2017) with 10 kPa of shear stress. A total of 20 000 solutions were generated, and the resulting cumulative probabilities of a thawed bed are shown in Fig. 5. This analysis indicates that the transition of the bed from predominately frozen to predominately thawed occurs with a sufficient degree of likelihood across ice thicknesses consistent with the observed change in basal reflectivity slope.

In combination, these effects can explain the shape of the distribution of observed reflection coefficients vs. ice thicknesses shown in Fig. 5. Performing a linear regression with an upper limit on ice thickness of 1300 m, where the basal water is hypothesized to start contributing, yields a one-way L_ice=15.1 dB km⁻¹ (σ=0.7), which should be a lower bound within the region.

4.1 Lake stage

Figure 6 shows the spatial distribution of time-normalized elevation changes as observed using crossings between GLAS lines and the new transects. Figure 7 shows elevation changes computed along individual profiles that followed GLAS ground tracks.

Figure 6Difference between the 2017 KOPRI surface elevations and the 2004–2009 ICESat elevations, normalized by time between observations. Blue is rising; red is falling. Dashed lines are nominal ICESat ground tracks. The background image is the MODIS mosaic.

Download

Figure 7(a) Individual profiles for the GLAS reflights showing surface elevations with respect to modeled surface in 2005. Vertical red lines show the intersections with the X transects, and blue shading shows where the profiles intersect the Smith et al. (2009) lakes. (b) Context map showing locations of selected tracks. The background is the ice surface velocity (Rignot et al., 2011b).

Download

Both the crossovers (Fig. 6) and reflight data (Fig. 7) are consistent in showing that the downstream part of the D1 outline (as defined by GL0194 and GL0158) has continued to rise, while the upstream portion is inconclusive.

While the region inside the original D2 outline appears to still be lowering with a total displacement of ∼5 m, it borders an area along line GL0328 with up to 15 m of elevation gain since the ICESat era. There is a similar area that is also lowering on the south side of the large positive anomaly, and all three extrema are observed in both profile and intersection data. Note that the two points to the west of D2 where the surface appears to be rising are from a single unrepeated GLAS track, so there is no time series associated with them, and we do not consider them to be a reliable signal.

Unfortunately, this survey alone has been unable to address the dHdt behavior in detail since the end of the ICESat era. However, there is no evidence that any of the Smith et al. (2009) lakes have switched from draining to filling or vice versa, and they have been established to be at a highstand relative to previous ICESat observations. Additionally, there is no evidence of an ice-dynamic-associated dHdt signal in the David Glacier region when compared to ICESat data. That is, patterns of surface elevation lowering are not associated with surface velocities or their gradients.

4.2 Hydraulic potential gradients

Figure 8DEM showing hydraulic potential with 10 m contours for the region around lakes D1 and D2. The uncertainty is estimated to be at least 10 m of hydraulic head, or one small contour line.

Download

Figure 8 shows gridded hydraulic potential over the survey. The most immediate observation in the new hydraulic potential map is that there is a ridge running through D1. This is consistent with surface observations of crevassing extending into D1, and it confirms that lake outlines based on interpolating between repeat-track surface elevation changes do not necessarily correspond to a large connected collection of water, which by definition would have to be at a constant potential. However, there is a broad low over the lower part of D1, which is consistent with the reinterpreted surface elevation record. Additionally, there is no clear potential minimum associated with the original D2 outline. Instead, this survey shows a small minimum to the south.

4.3 Reflection coefficient

Figure 9 shows the reflection coefficients that have been corrected for geometry and ice loss. While the highest reflection coefficients are in the main trunk of the glacier and found in areas of greater than ∼1700 m ice thickness, their distribution within those bounds is not obviously correlated with surface velocities or ice thicknesses. Instead, around lake D1, P_r tends to be higher in regions with lower gradients of hydraulic potential, consistent with water pooling. The region around lake D2 is more complicated, with bright beds corresponding to gradients of low hydraulic potential but not necessarily matching up with the observed surface deflections.

It is possible that a more sophisticated method of calculating the contribution of dielectric ice loss would lead to clearer results. Model-based approaches were not pursued: the continent-wide model from Matsuoka et al. (2012) has insufficient resolution, and integrating a dynamic model with the new topography is beyond the scope of this paper.

We also note that the span of reflection coefficients is still larger than would be expected for typical materials and cannot be explained by contributions of dielectric ice loss alone. The analysis presented here uses the radar equation for specular interfaces; a pure scattering interface would have a geometric spreading loss of 1∕r⁴ (Peters et al., 2005). Additionally, there could be englacial or surface terms not correlated with ice thickness that we are not accounting for. There is significant surface crevassing along the shear margins and over parts of D2, so correcting for surface scattering losses (Schroeder et al., 2016b) will likely yield an improvement in reflection coefficients. This topic warrants future investigation.

4.4 Specularity

Figure 10 shows the results of calculating specularity content over the lake region, plotted on top of the hydraulic potential contours. As with the reflection coefficient analysis, the regions of higher along-flow specularity are aligned with the regions of lower gradient. This is consistent with water collecting where it is flat and then being transported more efficiently under higher gradients. However, the clear anisotropy in the specularity signal is not characteristic of a typical radar lake, which would be expected to have an isotropically mirror-like surface (Young et al., 2016). Instead, we see higher specularity along-flow and lower across-flow.

Figure 9Relative reflection coefficients for the lake region, corrected for geometry and one-way dielectric losses of 15.1 dB km⁻¹. The background consists of 10 m contours for hydraulic potential, with hypothesized lake locations outlined in black (Smith et al., 2009).

Download

Figure 10Specularity along- (a) and across-flow (b) of the main trunk. The background is the same as Fig. 9.

Download

5.1 Reinterpretation of lake locations

The Smith et al. (2009) classification of D1 was based on three lines, and D2 was based on a single line. Their paper does not specify which GLAS tracks were used, but based on the number of repeats for each track, D1's outline was presumably derived from GL0194, GL0158, and GL0039 but not GL0075, and D2 was based on GL0292 and not GL0328. Lines GL0075 and GL0328 would have been left out of the Smith et al. (2009) analysis due to having insufficient repeats for the determination of cross-track slope.

Figure 11Observed time-normalized surface elevation changes are plotted on a DEM showing hydraulic potential with 10 m contours for the region around lakes D1 and D2. Transect segments where this study observed correspondence between a hydraulic potential minimum and a time history of surface elevation change are highlighted in yellow.

Download

Based on the new ice thickness and hydraulic potential grids discussed in Sect. 3.3.2, an extension of the ICESat surface elevation record, and an understanding of the sparse data spacing involved in the original outlines (Smith et al., 2009), this paper concludes that the original D1 outline is transected by an across-flow ridge with only the downstream portion associated with a significant surface elevation change signal. Additionally, the potential lakes do not extend as far south as the original outline.

Figure 11 shows time-normalized surface elevation change and hydraulic potential. It can be seen that D2, which was thought to be draining, is instead on the edge of a previously unexamined region with significant surface uplift that is not consistent with the effects of ice dynamics. The area of maximum uplift is consistent with a local minimum in hydraulic potential to the south of the previous outline and is bracketed by areas of significant but smaller subsidence. We interpret these as all being part of the same feature, corresponding to water accumulation. The surface expression of basal changes is not straightforward to determine: Sergienko et al. (2007) modeled this for a draining lake and found an evolving, nonmonopole pattern.

The analysis in our paper is able to include GLAS data from lines with fewer repeats and pointed farther off the nominal tracks than Smith et al. (2009) could because the crossover analysis agrees with the profile-based comparisons, and none of the elevation change signals were correlated with cross-track distance or aircraft roll. So, instead of observing the larger elevation change, Smith et al. (2009) identified an outline for D2 based on a line that only crossed the bordering subsiding region. This caused them to infer an offset boundary from what we observe as well as classify it as draining instead of filling.

Our results do not show any evidence of the D1 or D2 features switching between filling and draining, but it cannot be ruled out by a single snapshot 7 years after the end of the GLAS era. Siegfried and Fricker (2018) used CryoSat-2 in an effort to extend the surface altimetry record for a large subset of the lakes originally described by Smith et al. (2009), including David Glacier. They concluded that “small-magnitude height-anomalies on these lakes were in-phase with small height-anomalies in the region outside the lake boundaries’’ while pointing out that CryoSat-2 data are challenging to interpret in this region due to surface roughness. Their analysis does not agree with our results, where we see clear evidence of concentrated surface elevation change. We attribute the difference to laser altimetry's higher precision, making it the preferred tool for this region.

This reclassification of the potential boundaries of D1 and D2 provides an illustration of the pitfalls inherent in attempting to study active lakes based purely on the Smith et al. (2009) polygons or to use them to assume basal boundary conditions of temperature or basal shear stress. As described in the original paper, lake outlines are interpolations based on data that are increasingly sparse farther north. This is particularly relevant for planning and interpreting surveys consisting of a single radar transect over an active lake; a traverse planned directly across the middle of D1 could easily have resulted in a transect crossing the region with gradients of high hydraulic potential and no evidence of collected basal water in any form, missing the smaller region that has gradients of low hydraulic potential and anisotropic specularity.

5.2 Radar signature of active lakes

Consistent with most other radar investigations of suspected active lakes (Welch et al., 2009; Wright et al., 2014; Langley et al., 2011), the D1 and D2 surface features are not associated with the relatively bright and isotropically specular signature of a classic radar lake. There are three possible explanations for this mismatch: surveys are looking for water in the wrong place, at the wrong time, or using the wrong features in the radargram. Previous investigations consisting of a single line could be explained by the complicated transfer function from basal changes to surface expression or by the uncertainty in the lake outlines. Other surveys that did not include surface elevation measurements could be explained by hypothesizing that the active lake was at a lowstand. However, the laser altimetry presented here shows that both D1 and D2 are at an even higher stand than during the ICESat era, so inconclusive results cannot be explained as being due to drained lakes, and the spatial extent and density of the survey make it unlikely to have entirely missed the lakes. Thus, we conclude that active lakes cannot be expected to share the distinguishing physical features of radar lakes.

In interpreting reflection coefficients, there are a number of possibly complicating factors. In an active lake system, it is likely that there are significant portions of the bed at the pressure melting point, which would lead to lower contrast between the ice–water interface and the ice–bed interface. Depending on the depth of the lake, the roughness of the water–rock interface, and the salinity of its water, it is also possible that the radar return from the water–rock interface could interfere with that from the ice–water interface, lowering the observed reflection coefficient (MacGregor et al., 2011). Christianson et al. (2016) investigated anomalously low reflection coefficients in a region just offshore of the Whillans Ice Stream grounding zone and concluded that they were due at least in part to sediments entrained in the ice that had not yet melted out. Similarly, we could consider active lakes to be at one end of a continuum, where stable radar lakes are at the other end, and they are primarily differentiated by their water residence times. The more rapidly evolving features may not have existed for long enough to melt a smooth roof, so we could be observing the preserved imprint of the bed at lowstands or basal roughness advected from upstream of the lake. Supporting this view, some active lakes (Adventure Subglacial Trench) do appear in classic lake inventories, but they are typically the ones farther upstream, with longer cycle times.

Specularity is an appealing complement to reflection coefficient analysis for characterizing the basal interface. A classic lake would be expected to have a smooth, flat ice–water interface and appears as an isotropically specular surface, which requires decimeter-scale smoothness over hundreds of meters. This concept has been used in earlier work to characterize the distribution of subglacial water: Young et al. (2016) looked at the anisotropy of individual lines by comparing the specularity of the first return to the amount of scattering recorded afterward, while Schroeder et al. (2013) leveraged a gridded survey of Thwaites Glacier. Schroeder et al. (2013) reported a pattern of anisotropic specularity in Thwaites Glacier and concluded that it indicates “canals’’ of subglacial water pooling, aligned with the ice flow. Since canals require a sedimentary subglacial interface, this is consistent with both the hypothesis of Carter et al. (2017) that active subglacial lakes drain through canals and with the observations of Smith et al. (2017) of active lakes in the region of Thwaites Glacier where Schroeder et al. (2013) identified the water system transition. In the David Lakes region, we see an overall pattern of anisotropic specularity, including over the regions of surface elevation change, that is similar to that seen in Thwaites. Further work is needed to understand the overall hydrologic systems driving these active lakes, and the anisotropic specularity in this region provides an interesting constraint on possible organizations of water.