Glacier and debris characteristics

We carried out a combination of survey profiles and static field tests on the ground during consecutive spring melt seasons on the debris-covered ablation-area tongues of Lirung, Langtang and Ngozumpa glaciers in Nepal (Fig. 1), chosen to be particularly challenging targets for radar sounding due to a combination of thick and at least partially temperate ice, and a surface debris cover. Our tests coincided with a diurnal freeze-thaw cycle with a wet ice surface during the day (overlain by a wet snow layer on Lirung and Langtang glaciers), and we observed surface ponding of water at each site. Given the survey locations in the ablation areas of these glaciers, which have extensive ponding and englacial drainage networks (e.g., Gulley and Benn, Reference Gulley and Benn2007), we assume the ice to be temperate for much of its depth (Miles and others, Reference Miles2018). Our sites have a near-continuous debris cover, and we sampled the debris thickness using pit measurements and relatively high-frequency ground-penetrating radar (GPR) profiles (McCarthy and others, Reference McCarthy, Pritchard, Willis and King2017; Nicholson and others, Reference Nicholson, McCarthy, Pritchard and Willis2018).

Fig. 1. Glacier ice thickness profiling results from glaciers (a) Lirung (left) and Langtang (right), and (b) Ngozumpa. Location of intersecting profiles discussed later are shown by white arrows. Image source: ESRI (GeoEye, DigitalGlobe).

Debris-thickness pit measurements and surface observations at our sites in Nepal showed clasts ranging from silt to large boulders several metres in diameter, but predominantly gravels with air-filled voids near the surface grading to gravel with wet sand- and silt-filled voids in the centimetres to decimetres above the ice interface (McCarthy and others, Reference McCarthy, Pritchard, Willis and King2017; Nicholson and others, Reference Nicholson, McCarthy, Pritchard and Willis2018). The pit and GPR measurements showed debris thickness to be locally variable, ranging from 0.11 (range 0.01–0.34 m) to 2.3 ± 0.39 m and typically ~1 m on Lirung Glacier, and from 0.18 ± 0.14 to 7.3 ± 0.83 m, typically ~2 m on Ngozumpa. These debris characteristics are similar to and at least as thick as those found elsewhere, (e.g., Foster and others, Reference Foster, Brock, Cutler and Diotri2012), and are at least an order of magnitude smaller than the radar wavelengths tested here. These findings reassure us that our sites are reasonably representative of challenging debris-covered ablation areas, and that the debris layer is thin enough and sufficiently well drained through much of its depth that it is unlikely to modify significantly the dielectric or scattering properties of the glacier/air interface as sensed by our low-frequency radar system. Such debris layers are therefore unlikely to be a major impediment to detecting the glacier bed.

Radar profiles

Dipole radars are effective transmitters and benefit from a linear antenna design whose simplicity makes them relatively easy to incorporate into a robust system for dragging over glacier surfaces (e.g., Evans, Reference Evans1963). Design of a dipole radar for airborne deployment requires a compromise in antenna length that gives a centre frequency low enough to achieve bed detection in most expected conditions while still being short enough to allow portability of the system in the field. Our first goal was therefore to test several frequencies for bed detection in representative Himalayan glacial settings.

We surveyed for ice thickness on foot along a series of long- and cross-profiles (parallel and perpendicular to glacier flow, respectively) on our three test glaciers, primarily using the British Antarctic Survey DELORES dipole radar system. This is a mono-pulse radar capable of operating between 1 and 20 MHz, with a Kentech pulse transmitter that produces ±2000 V pulses at a repetition rate of 1 kHz (King and others, Reference King, Hindmarsh, Corr and Bingham2008; Hindmarsh and others, Reference Hindmarsh2011; King, Reference King2011). It uses resistively loaded antennas that can be physically swapped to give a different centre frequency depending on antenna length. We used 40 and 20 m long dipoles for repeated profiles, giving centre frequencies of 3.5 and 7 MHz, respectively (Table 1), with received data converted from analogue signals with a 12-bit Picoscope digitiser.

Table 1. Radar characteristics

Nominal vertical resolution is the quarter-wavelength typically used to estimate the minimum resolvable separation of two similar reflectors (e.g., Reynolds, Reference Reynolds1997). Optimal vertical resolution refers to the resolvable range to a single discrete, prominent reflector in an otherwise homogeneous medium (e.g., a smooth glacier bed below clean ice in a clutter-free environment) (King, Reference King2020). Practical vertical precision refers to the distance observed in these data over which the bed couplet rises from the first break to the first peak (approximately one-quarter of the couplet length). This falls between the optimal and nominal resolutions and is a conservative estimate of the vertical distance over which the bed first break can be picked with confidence in these data. The Fresnel zone radius is calculated at 280 m because this is approximately the ice thickness of Langtang Glacier, and is indicative or the radar footprint size at the bed.

We surveyed nine profiles over the lower tongues of Lirung and Langtang glaciers in March 2015, and eight profiles on Ngozumpa Glacier in April 2016, with sampling at 5 m intervals by stop-go surveys (Fig. 1). A stop-go approach rather than continuous surveys allowed us to control receiver/transmitter separation over rough ground. For most profiles, the transmitting and receiving antennas were arrayed line-astern with 30 to 50 m tip-to-tip antenna separation (a configuration familiar from Antarctic over-snow surveys), and each sample was taken as the stacked average of 4000 radar pulses. For two profiles on Ngozumpa Glacier, we experimented with a different system (Lambrecht and others, Reference Lambrecht, Mayer, Aizen, Floricioiu and Surazakov2014) with antennas aligned in parallel at a 3 m lateral separation. Given the slower pulse-repetition frequency and digitiser speed of this system, each sample was taken as the stacked average of eight shots.

The separation of dipole antennas and their relative positions affects the system geometry. Relative to the antenna-parallel configuration, the line-astern separation primarily used here increased the path length through ice by typically 6 m (typically ~1% of the two-way travel distance) and increased the reflection angle at the bed (the separation between incoming and reflected signals) by typically 16°. These are, however, modest changes compared to those introduced by the variability in glacier thickness and antenna orientation along our undulating survey profiles, given the roughness of the surface. The more significant difference between these configurations is practical. Line-astern antennas are well suited to towed surveys over relatively smooth glacier surfaces (on foot or by vehicle), while over rough surfaces it is considerably easier for transmitter and receiver operators on foot to communicate and maintain a constant antenna separation using the parallel configuration. The disadvantage in parallel configuration is that the receiver is exposed to high peak power pulses from the adjacent transmitter, generally precluding the use of amplifiers in the receiver. The parallel configuration is considerably shorter than line-astern, however, and therefore more likely to be suitable for an airborne system, hence we tested the practical application of both approaches in these tests.

Sample location was determined using post-processed differential GPS data. We processed radar data in ReflexW software using combinations of time-dependent divergence compensation, de-wowing (over a 300 ns moving window) and a band-pass frequency filter to aid manual detection of the glacier bed, and we corrected the resulting ice thicknesses for surface topography and system geometry. The band-pass filter lower cut-off, lower plateau, upper plateau and upper cut-off were 1, 2, 10 and 20 MHz for the 3.5 MHz profiles and 2, 4, 15 and 30 MHz for the 7 MHz profiles. We used a radar velocity in ice of 0.168 m ns⁻¹ to convert two-way-travel times to ice thicknesses. Velocity depends somewhat on depth-averaged water content, however, which is unknown. Velocities of 0.165 and 0.172 m ns⁻¹ have, for example, been ascribed to temperate and cold ice, respectively, above and below the equilibrium line of an alpine glacier (Macheret and others, Reference Macheret, Moskalevsky and Vasilenko1993).

Radar profiles at 3.5 and 7 MHz produced radargrams with prominent linear or quasi-linear signals at various return times, and among these returns we identified the bed manually from its consistently linear nature throughout, and the agreement in depth and relative signal strength between long- and cross-profiles (e.g., Figs 2 and 3). Long-profiles from Langtang Glacier (Fig. 2) demonstrate the detectability of the bed and the potential for higher horizontal resolution in the bed return at the higher frequency (comparing Figs 2b and d), as expected from the smaller Fresnel zone at higher frequencies (Table 1). They also highlight the ambiguity that arises in picking the bed return due to the presence of other nonbed signals.

Fig. 2. Long-profiles with de-wow, divergence compensation and band-pass filtering from the lower Langtang Glacier (Label A in Fig. 1) at 7 MHz (a) and with bed pick (b), and 3.5 MHz (c) and with bed pick (d).

Fig. 3. (a) Long-profile (Label C in Fig. 1), and (b) intersecting cross-profile (Label D in Fig. 1) from Ngozumpa Glacier, at 3.5 MHz. The red dashed line indicates a prominent ‘false bed’, the yellow/orange line, the true bed. The blue dashed lines highlight the trend of the dipping reflectors from the valley walls to the east and west of the glacier.

We class these nonbed signals as clutter of unknown surface or englacial origin and note that this is often of similar strength to the bed signal or stronger, and present at similar return times. This leads to the potential for systematic bed-misidentification particularly in radar long-profiles of valley glaciers, where the range to the lateral moraines or valley walls (potential sources of additional reflections at similar return times) may be relatively constant. A prominent, linear, bed-like reflection in a long-profile from Ngozumpa Glacier (Fig. 3a), for example, is absent in a crossing profile (Fig. 3b), which shows linear, diagonally dipping returns from both valley sides, indicating that these are the source of the bed-like reflection. The true, shallower bed reflection is apparent in both profiles and is weaker, which we attribute to the greater signal attenuation through temperate ice than air.

Multiple blank traces and several with misaligned start times are apparent (e.g., in the first 600 m of Fig. 2c). This resulted from mis-triggering of the radar receiver from the transmitted airwave, a problem that was common when surveying on the undulating glacier surface with the transmit dipole trailing the receive dipole (by ~50 m tip-to-tip). Occasional loss of receiver-transmitter line-of-sight intervisibility caused unpredictable distortion of the airwave in some locations, which is difficult to accommodate when setting airwave-triggering parameters during ground surveys. For thickness mapping we corrected mistimed traces in post-processing. Our survey profiles reveal ice thicknesses ranging from ~40 to 440 m, with the deepest ice present in the central trunk of Ngozumpa, Nepal's largest glacier (Fig. 1).

Static tests

Having established the ice thickness along our survey profiles, we also carried out static tests with and without receiver amplification at three frequencies (3.5, 7 and 14 MHz) in relatively flat areas of the glacier surface, to quantify the relationship between frequency and bed SNR and signal-to-clutter ratios (SCR). We defined the bed signal as the mean absolute voltage of the bed-return wave couplet, clutter as the mean absolute voltage of nonbed signals in the range from 1500 ns before to 600 ns after the bed return (after de-wow processing with a 300 ns window), and noise as the mean absolute voltage of the far-field recorded signal (covering a period of 2000 ns starting at 2000 ns after the end of the bed return), after de-wow and ReflexW's time-dependent divergence-compensation processing (Fig. 4). We chose the signal, noise and clutter time windows through manual interpretation of the traces. The window that defines clutter brackets the known bed depth, beginning after the prominent declining trend of signal amplitude in the early part of the record and ending as the signal drops close to the apparent noise floor at the higher frequencies tested (Fig. 4). The noise window begins after the onset of the noise floor at all three frequencies. These static tests over a bed of depth known from the profiling surveys allowed us to compare the bed SNR and SCR at three frequencies (Table 1, locations of intersections in Fig. 1).

Fig. 4. Radar traces collected at a point on Langtang Glacier (intersection of profiles A and B in Figs 1 and 6), showing (a) a 7 MHz, amplified 1000-stack raw trace, (b) the same trace after processing for de-wow and (c) de-wow plus divergence-compensation, with the windows used to quantify the bed-signal SNR and SCR. (d) shows in detail the de-wowed and divergence-compensated traces at 14 MHz, (e) 7 MHz, and (f) 3.5 MHz. The depth scale is corrected for the radar geometry and ice thickness is approximately 286 m.

We found that the detectability of the bed signal relative to both noise and clutter is primarily dependent on frequency, with a weaker dependency on stacking (Fig. 5). Clutter is the dominant constraint on bed detectability, with SCR not exceeding 3 while the SNR exceeded 70 in amplified data and 9 in unamplified data. Stacking helps to increase the SNR by averaging randomly varying noise but has little effect in boosting the SCR because the clutter signals, like the bed signal, are consistent through time. Note that because noise suppression through stacking is ultimately limited by the precision (bit-depth) of the digitiser, a higher SNR could be achieved by changing from the 12-bit Picoscope digitiser to a 14-bit system, which we subsequently did (see ‘Antenna configuration’).

Fig. 5. Bed SNR (solid lines, left axis) and bed SCR (dashed lines, right axis) varying with stacking for the three frequencies tested for (a) amplified and (b) unamplified data. Error bars show ±1 Std dev. of the ratios measured over 11 repeated stacks for each configuration.

The higher SNR of the bed at lower frequencies is expected given that attenuation increases in proportion to the square of the frequency in ice and fresh water (e.g., Page and Ramseier, Reference Page and Ramseier1975; Davis and Annan, Reference Davis and Annan1989), hence at lower frequencies less signal is lost, while in contrast, system noise is frequency-independent. There is also some weaker frequency-dependence apparent in the SCR (Fig. 5, dashed lines, right axis), but it is less clear why the clutter and bed signal-strength should respond differently with frequency.

This could arise if the bed signal is preferentially weakened at higher frequencies through attenuation as described above, while in contrast, the clutter comes primarily from the glacier surface, travels largely unattenuated through air and therefore changes little with frequency. It could alternatively arise if the dominant clutter reflectors are effectively point scatterers of similar size to the shorter wavelengths tested (on the scale of metres), and substantially smaller than the longest wavelengths (close to 100 m) (Table 1). For a flat-plate scatterer of a given physical area and dielectric properties, the radar cross section (i.e., backscatter strength) is inversely proportional to the wavelength squared (Bole and others, Reference Bole, Wall, Norris, Bole, Wall and Norris2014), hence clutter would be weaker at lower frequencies for such relatively small scatterers. Likely contributors to clutter that are within the ranges observed are the hummocky, wet and debris-covered glacier surface, as well as debris- and water-filled crevasses, wet ice cliffs and water-filled englacial channels and ponds. In addition, the lateral moraines should contribute airwave reflections with relatively little attenuation around the time of the returning bed signal, given the faster signal propagation rate and lower attenuation in air (Fig. 6). Further static tests in a variety of settings are needed to understand better the relationship between frequency and clutter.

Fig. 6. (a) Langtang Glacier and (b) Ngozumpa Glacier radar static-test sites (locations at the AB and CD intersections in Fig. 1). White circles show horizontal ranges that are equivalent to the bed depth for radar propagation speeds through ice (inner circle) and through air (outer circle). Surface features corresponding approximately with the outer circle could provide clutter with a similar time delay as the bed signal. Image source: ESRI (GeoEye, DigitalGlobe).

Antenna configuration

Based on the findings of the ground tests, we developed and tested an antenna configuration to maximise length while seeking a compromise with the practical constraints of an airborne system. At ~100 m long, the typical line-astern dipole configuration for ground surveys at the lower end of the frequency range considered is not feasible for helicopter deployment. We therefore considered switching between transmit and receive on a single dipole, but the high instantaneous pulse power, wide bandwidth and very low duty cycle of our transmitter (Kentech Instruments Ltd. GPR Pulser J10) made this impractical. Our ground tests (section ‘Radar profiles’) showed, however, that a separate and overlapping matching pair of resistively-loaded receiver and transmitter dipoles in parallel could also be used, provided that the receiver digitiser is protected from pulse overload when the antennas are brought close together. Because our antennas are resistively loaded (with resistance increasing towards the tips), we were able to limit the peak received power to below the operational limits of the digitiser by staggering the parallel receiver and transmitter dipoles by 4.8 m (Fig. 7).

Fig. 7. (a) Full-scale airframe prototype. (b) Schematic of radar components, shown configured for 8 MHz (top) and 16 MHz (bottom). Rx and Tx refer to the receiver and transmitter units and antennas, T to the tail, B to battery and G to GPS.

Deploying the antennas in parallel rather than line-astern reduces the system length by ~70%, though the close proximity of transmitter and receiver preclude the incorporation of amplifiers in the receiver system. We were able to compensate for this by introducing a high-precision, 14-bit Teledyne SP Devices digitiser (rather than the 12-bit Picoscope used in the ground tests) to record two receiver data channels, one with a large dynamic range of ±5 V covering a wide range of returning signals and the other covering ±1 V to give enhanced sensitivity to weak signals. The high precision also makes stacking more effective at lowering the noise floor, and we used this high-sensitivity channel for bed detection when the SNR was low.

Airframe structure

Informed by the bed-detection field tests regarding antenna length, and constrained by the need for a light, robust, modular, and electrically nonconductive frame design that can be operated safely by helicopter in high-mountain environments, we used modelling and prototyping to develop an elongated frame structure for suspension under an un-modified helicopter as a sling-load. We commissioned aerodynamic modelling (Graham and McRobie,Reference Graham and McRobie2017) to assess, for a frame-structure of up to 40 m long towed at airspeeds from 10 to 250 km h⁻¹, resilience to buckling and tendency to the oscillatory motions of fluttering, weather-cocking and penduluming that could cause structural fatigue and failure (Fig. 8).

Fig. 8. (a) Buckling, and the oscillatory modes (b) weather-cocking, (c) flutter and (d) penduluming.

Buckling forces are controlled by the angle between the guy-lines and the frame and the wind forces experienced by the structure, which depends on frame drag, frame attitude in flight and flying speed. Buckling calculations therefore informed decisions on frame length, rigidity, strength and weight (as a control on attitude), and on the height above the frame of the guy-line apex (yoke). The modelling identified potential weather-cocking and penduluming oscillatory motions but predicted both to be stable at all speeds tested, i.e., to decay with time rather than amplify. Their frequencies were predicted to differ (weather-cocking in most cases lower than penduluming) and although no mode-interactions were found to be unstable, such interactions are difficult to predict so we aimed for a design that kept these frequencies apart. Specifically, we aimed to maintain a relatively low weather-cocking frequency by providing sufficient resistance to yaw and by limiting the maximum flying speed, and a relatively high penduluming frequency by minimising the total suspended length, while maintaining a yoke height sufficient to prevent buckling. We also took into consideration test results showing that, due to drag effects, both oscillations decayed more rapidly as the number of guy-lines increased, and that flutter is reduced by increasing frame rigidity (Appendix).

For prototyping, we designed a modular system based on a set of box-frames constructed of polyester resin fibreglass tubes and glass-filled nylon couplings, with each frame 1.87 m × 0.35 m × 0.35 m and weighing 7 kg. Twenty-two frames strapped in an 11 × 2 configuration gives a 20.5 m-long airframe that retains some flexibility and an acceptably light weight for field deployment, and the length can readily be extended by using additional single fibreglass poles at each end. We suspended the whole frame from a set of high-breaking-strain, low-stretch guy-lines that meet 10 m above at a central yoke, and latch to a standard helicopter load sling for survey flights. To resist yaw, we doubled the height of the tail section and wrapped it in waterproof fabric (Fig. 7). Symmetrically in this frame, we mounted the radar transmitter, receiver and battery in waterproof plastic cases, with antenna wires attached to the frame structure. The total system weight is ~230 kg with a 100 Ah lead-acid battery. With a parallel transmit-and-receive dipole configuration, our 20.5 m-long system can provide transmitted radar frequencies as low as 7 MHz.

We suspended our prototype from cranes to test for rigidity and structural integrity, and conducted airborne helicopter flight tests for aerodynamic performance and structural integrity in straight flight at various speeds, and in turns, climbs and descents.

Structural design

In 5 hours of helicopter flight tests in the UK and Svalbard between February and April 2018, our frame design performed well aerodynamically at intended survey speeds of ~30 knots (55 km h⁻¹) and up to the maximum-tested 50 knots (90 km h⁻¹), maintaining an attitude aligned with the flight direction and reliably returning to this attitude after short-radius (100 m) turns through 180°. The frame showed no tendency to oscillate or rotate in flight and remained under good control, allowing surveys to be conducted at helicopter flying heights of 75–90 m (250–300 feet) with a ground clearance for the radar of 50–65 m (165–210 feet). Post-flight inspections showed no evidence of structural fatigue or damage.

Flight surveys

To test our full frame and radar system we flew three survey flights in Svalbard. The primary goal of these flights was to test our frame for aerodynamic issues experienced by the helicopter crew during prolonged flight operations, and to test for ease of shipping, construction and operational use, and for structural performance and robustness of the frame in cold field conditions. The secondary goal was to run our radar in-flight to test for electrical interference issues with the helicopter (both in terms of flight operations and radar data quality), and for robustness of our electrical system in survey use. The tertiary goal was to test radar-data acquisition and storage performance by surveying over our target glaciers using a simple radar configuration of known capability that we have used in previous polar surveys.

For simplicity, we configured the radar on our 20.5 m frame with separated (nonoverlapping) 8 m-long, line-astern transmit and receive antennas (giving a 16 MHz centre frequency), with the receiver unamplified and triggered from the airwave (Fig. 7). We flew this over the small, polythermal, land-terminating and largely un-crevassed Midtre Lovenbreen, Svalbard, where equivalent 16 MHz ground-survey data were previously collected by one of the authors (E. King) and so were available for validation if needed (not presented here). We also flew two flights over thicker, heavily crevassed tidewater Kronebreen, where ground survey is not possible (Fig. 9).

Fig. 9. Helicopter-surveyed longitudinal profiles over (a) the lower 13 km of Kronebreen, and (b) the lower 3 km of Midtre Lovenbreen, Svalbard. We calculated depths using a wave speed of 0.168 m ns⁻¹ in ice. For the heavily crevassed Kronebreen, post-processing with a horizontal averaging filter and a manual gain curve suppressed surface-crevasse clutter by 19 dB. Map data: ESRI.

These flight surveys demonstrated that our system design is aerodynamically and operationally robust and practical to deploy in field conditions, and did not suffer from interference with the helicopter. Even at a relatively high frequency (and relatively low stacking of 280), it detected the beds of the small, polythermal, land-terminating Midtre Lovenbreen and the large, heavily crevassed polythermal tidewater glacier Kronebreen throughout the survey profiles, through maximum encountered thicknesses of 175 and 450 m, respectively (Fig. 9).

In contrast to glacier conditions encountered in Nepal, the surfaces of these Svalbard glaciers were debris-free and frozen at the time of survey, however. Furthermore, physical decoupling of our airborne radar from the ground will introduce radar reflection losses from the glacier surface in addition to the attenuation losses present in our ground tests. Reflection is governed by the contrast in dielectric permittivity (ɛ) and therefore wave-speed between air (ɛ = 1) and the glacier. Over melt ponds (ɛ = 86 (Frolov and Macheret, Reference Frolov and Macheret1999)) we expect reflection losses of ~80% (Wightman and others, Reference Wightman, Jalinoos, Sirles and Hanna2003) but elsewhere we expect ɛ ≈ 3, dominated by the properties of ice and modified relatively little by the similar properties of the debris or wet snow cover (ɛ ≈ 3 to 5 (Frolov and Macheret, Reference Frolov and Macheret1999; Grima and others, Reference Grima, Blankenship, Young and Schroeder2014)), that are anyway thin relative to the wavelength. At incidence normal to the surface, this implies reflection losses of ~28% (Wightman and others, Reference Wightman, Jalinoos, Sirles and Hanna2003) (or up to 38% for ε = 5 (Grima and others, Reference Grima, Blankenship, Young and Schroeder2014)). Additional spreading losses are also introduced with an airborne system due to the 10–20% increase in range to the bed.

These factors decreasing the likely bed signal strength are partly offset by the smaller reflection angle (~0.004°) and fixed, horizontal antenna alignment of our airborne system relative to the line-astern ground surveys. The likely magnitude of these changes in signal strength relative to the SNR (unamplified) of 5–6 with achievable stacking rates (through ~300 m of ice at 7 MHz) that we observed in Nepal give us confidence that our airborne system is capable of detecting the beds of Himalayan glaciers.