We mapped a 250 m×250 m area of ice surface at each site using the methodology described previously in . Briefly, we integrated a MicaSense RedEdge multi-spectral camera onto a SteadiDrone Mavrik-M quadcopter (referred to hereafter as UAS). The camera was remotely triggered through the autopilot, which was programmed along with the flight coordinates in the open-source software Mission Planner. Images were acquired at approximately 2 cm ground resolution, with 60 % overlap and 40 % side lap. Mapping required two successive flights with a UAS battery change between them. Each flight lasted ∼10 min, was made at 30 m above the ice surface, and took place under clear-sky illumination conditions unless otherwise noted (Appendix ).

At S6 we made UAS flights over several successive days, requiring us to remove the effect of ∼0.5–1 md-1 of ice motion from the final orthomosaics. We therefore placed 15 ground control points (GCPs) and measured their X and Y locations on 21 July using a differential global navigation satellite system (GNSS) receiver, post-corrected by reference to the Kellyville International GNSS Service (IGS) GNSS station using IGS final orbits. We used the GCPs to constrain the horizontal geo-referencing of every orthomosaic to the same static georectification solution.

We applied radiometric calibration and geometric distortion correction following MicaSense procedures . We then converted from radiance to reflectance using time-dependent regression between measurements of the MicaSense Calibrated Reflectance Panel (and, at UPE, a Spectralon^® panel) acquired before and after each flight. The individual reflectance-corrected images were mosaicked using AgiSoft PhotoScan following the , yielding multi-spectral orthomosaics with 5 cm ground resolution. Finally, the orthomosaics were radiometrically adjusted to match directional reflectance measurements made by ground spectroscopy so that our surface classifier (Sect. ), which was trained using the directional reflectance measurements, could be applied to the orthomosaics.

The orthomosaics were used in three ways: (i) converted to albedo using a narrowband-to-broadband approximation , (ii) classified into surface types (see Sect. ), and (iii) digital elevation models derived photogrammetrically in Agisoft PhotoScan at 5 cm ground resolution.

Narrowband-to-broadband approximations for albedo calculations were employed because empirical bi-directional reflectance distribution functions (BRDFs) are not available for the surface types that we mapped. While these surfaces are highly anisotropic, scattering light preferentially in the forward direction , there are no datasets we know of that can accurately correct reflectance values gathered at the nadir. We therefore omit a BRDF correction, as existing BRDF datasets cannot be confidently applied to our sample surfaces.

We used the photogrammetric digital elevation models (DEMs) to derive (i) study area slope angle and (ii) local topographic variability. To calculate the slope angle we applied a Gaussian filter with a window of 0.25 m to remove very high-frequency topographic features; then we calculated the average slope across each study area after , as implemented in the RichDEM library . To examine local topographic variability (“roughness”), we applied a Gaussian filter with a window of 4.95 m and then subtracted it from the DEM to yield a detrended surface.

We took samples of the ice surface at each site to quantify the presence of glacier algal cells. At S6, samples were made immediately after collection of paired ground spectra (Sect. ) to enable direct upscaling by UAS imagery analysis. At UPE, widespread snow cover prevented us from utilizing the paired approach carried out at S6. Instead, on 26 July (2 d after the UAS flight), we cast a random 75-point sampling grid over our UAS flight area. We used a trowel to scrape the snow away to reveal the bare-ice surface beneath for sampling.

Samples were made by cutting a 30 cm×30 cm×2 cm volume out using a metal ice saw and trowel and transferring into a sterile Whirl-Pak bag, which was immediately placed in the dark to melt over a ∼24 h period at ambient air temperature. Following melting, samples were homogenized, sub-sampled into Falcon tubes and fixed with 2 % final-concentration glutaraldehyde. Samples were then returned to laboratories at the University of Sheffield and University of Bristol for counting by microscopic haemocytometers. Full details of the enumeration protocols used are in samples from 2017 and samples from 2018.

At S6, we measured daily surface lowering across a total of four plastic ablation stakes drilled into the ice. The poles formed a square ∼100 m ×100 m centred within the wider 250 m×250 m UAS survey area. At each pole we took two measurements: one west and one east. Measurements were made between 16:00 and 18:00 local time, except for 17 and 18 July at 12:00 local time and 19 July at 19:45 local time. Here we present daily mean surface lowering calculated from all the poles.

Clear-sky Sentinel-2 (hereafter S-2) data were available at the S6 site for 20 and 21 July. No clear-sky acquisitions were available, coincident with our field season at the UPE site. We downloaded S-2 L1C data from Sentinel Hub (Sinergise, Slovenia). We used all bands available at 10 and 20 m resolution by resampling those bands delivered at 10 m resolution to 20 m using the S-2 toolbox of the European Space Agency (ESA) SNAP platform. We processed the L1C data to L2A surface reflectance using the ESA Sen2Cor processor. The data were then (i) converted to broadband albedo using a narrowband-to-broadband approximation and (ii) classified into surface types (see Sect. ).

To classify images by surface type we used a supervised classification approach following , trained on ground spectra collected at S6 with a FieldSpec Pro 3 (Analytical Spectral Devices, Boulder, USA) during the 2016 and 2017 field seasons at S6. Briefly, we used 171 directional reflectance measurements. The measurements were labelled by visual examination as snow (SN), water (WA), clean ice (CI), light algae (LA), heavy algae (HA) and dispersed cryoconite (CC). After ground spectra were acquired we took destructive ground samples following procedures in Sect. . Clean-ice samples contained 625±381 cellsmL-1, light-algae samples 4.73×103±2.57×103 cellsmL-1 and heavy-algae samples 2.9×104±2.01×104 cellsmL-1, confirming the accuracy of our visual assessments of each surface type. We split the dataset randomly into training (70 %) and test (30 %) sets. These data were used to train a random-forest classifier, which had the highest performance of all classifiers tested . We trained the algorithm to predict surface type from (i) our UAS-acquired data, utilizing all five bands of data, and (ii) S-2 data, utilizing all nine bands at 20 m resolution. The confusion matrices (Appendix ) for the classifiers in this study were similar to those in . Against the test set, UAS classifier accuracy and recall were both 97 %, and S-2 classifier accuracy and recall were both 88 %.

We used the albedo retrievals contained within the MODIS/Terra Snow Cover Daily L3 Global 500 m Grid V006 “MOD10A1” data product . The two pixels which overlapped with our S6 UAS area were examined in their original sinusoidal projection. Precise overpass times were extracted from the granule pointer information contained within each product file (Appendix ). There were no cloud-free MOD10A1 data available at UPE during our field season.

To provide a local environmental context we used a point surface energy balance model to estimate net shortwave and longwave radiation fluxes, the turbulent sensible and latent heat fluxes, and the surface melt rate at a point on a melting ice or snow surface. The model was forced at an hourly time step by continuous measurements of shortwave radiation, vapour pressure, air temperature and wind speed made by IMAU S6 AWS and PROMICE UPE_U AWS . We used the albedo measured at each AWS, which at UPE_U was only for solar zenith angles below 70∘ and at S6 was only when downwelling shortwave radiation was >250 Wm-2; night-time values were therefore forward-filled from the last valid albedo observation. The surface roughness length was held constant at 1 mm, according to similar values for ablating ice surfaces . Daily melt fluxes were estimated from all time points when the air temperature was ≥0 ∘C. As the AWSs were located a few kilometres away, the computed melt rates should be interpreted as indicative of the meteorologically forced melting regime rather than as absolute melt rates experienced across the study areas.

Biological sampling and UAS observations showed that glacier algae were ubiquitous at both S6 and UPE. At S6, low albedo (Fig. a) was caused by extensive algal blooming (Fig. b) enabled by melting over several preceding weeks see. This finding is supported by radiative transfer modelling, which shows that mineral dusts local to S6 are weakly absorbing and strongly scattering, meaning that they locally increase albedo, whereas glacier algae have an albedo-reducing effect . At UPE, the albedo was higher (Fig. d) due to persistent snow cover obscuring the darker bare-ice surface (Fig. e). However, our ground sampling revealed up to 80 % LA+HA coverage of the survey area (Appendix ) on the bare-ice surface that was hidden from our aerial remote sensing by a layer of fresh snow. Ultrasonic ranging observations from the UPE_U AWS show that the winter snowpack had melted by 29 June 2018, revealing the bare ice beneath . Between bare-ice exposure and our arrival at the field site the surface had remained snow-free, and our energy balance modelling estimates that 35 cm w.e. of melt had occurred. These conditions promote algal growth , explaining the presence of algae beneath the recently deposited snow. These observations of spatially expansive populations of algae at both sites demonstrate that biological albedo reduction is important across the ablation zone of the western GrIS, including areas outside of the dark zone.

The two sites were distinct in their local topography and hydrology. S6 had an average slope of 5∘. Topographic features within the area principally consisted of (1) a ∼0.3 m wide ice-incised supraglacial stream and (2) a few isolated small (<2 m2) ice rises up to ∼0.2 m high. After detrending (Sect. ) 99 % of the area had topographic variability of <±0.05 m and 54 % of the area was within ± 0.01 m. The ice surface to ∼300 m upslope of the area was flatter and had several small moulins, reducing the area contributing to local flow. The shallow and ephemeral arterial hydrological pathways present across the study area during July were likely the result of a constant slope and negligible meltwater routed from upslope, reducing frictional stream incision . However, during June, winter snowpack retreat caused significant ephemeral water drainage pathways to develop, causing algal-cell concentration and redistribution (e.g. Appendix Fig. ) until upslope crevasses and moulins opened to route meltwater away englacially. This was likely important in distributing concentrated algal blooms growing in local niches over a wider area given that glacier algae lack a flagellated life stage and so are not independently motile .

Observations at UPE where there was substantial local surface roughness (Fig. a, b) showed that lower albedos were associated with local depressions (Fig. c; R2=0.71 with all data; R2=0.95 when albedo bins with <0.5 % area coverage removed). The higher the biomass loading (from CI through LA to HA), the lower the local elevation of the associated surface was (Fig. c). There are at least two possible reasons for the concentration of heavy algae in local depressions. One is entrainment and transport of algal cells in topographically higher areas by meltwater; once the competence of the meltwater flow drops in local depressions, then the impurities will be deposited. Another is that local depressions favour near- or at-surface availability of meltwater through ponding, especially if a weathering crust is well-developed at topographic highs. Surface meltwater reduces albedo , which results in favourable growth conditions for glacier algae , further reducing albedo and amplifying surface ablation.

There are strong indications that the local topography and near-surface hydrology at UPE resulted in a different surface state to S6. The shallower slope (1∘) than at S6 is likely to favour the evolution of perched meltwater ponds (Fig. e), as the lower gravitational potential is less conducive to runoff. Meanwhile, meltwater generated further upglacier flows through the area in streams incised to ∼0.6 m below the mean surface elevation (e.g. the stream running from north-west to south-east through the study area; Fig. a). Arterial meltwater pathways are thus likely to persist inter-annually, as little melting had occurred in the 2018 melt season prior to our measurements (Sect. ). This stream-dominated hydrological regime likely reduces the movement of microbial cells suspended in meltwater through the weathering crust compared to S6. A stream-dominated regime therefore also favours complex spatial and temporal patterns of albedo where most ice is weathered, persistently bright and strongly scattering due to minimal sub-surface meltwater, punctuated by low-albedo melt ponds and concentration of LAIs and water in topographic lows.

Our S6 field campaign captured a period of storm conditions in the middle of an otherwise shortwave-dominant energy regime, with concomitant impacts upon the ice surface state, including the weathering crust. Energy balance modelling indicates that shortwave energy fluxes dominated the 10 d preceding the start of our UAS observations on 15 July (mean daily maximum net shortwave flux: 273 Wm-2). Oblique photography from 15 July confirms that a weathering crust was ubiquitous throughout the UAS area (Appendix Fig. a). During 16–17 July, no melting was modelled (Fig. a) or measured (Fig. b), and ∼5 cm snow fell on 17 July. During 18–19 July, shortwave energy fluxes remained low while latent and sensible heat fluxes increased (Fig. a), associated with strong winds (up to 17.5 ms-1 at S6) and rainfall. We observed substantially more surface lowering (Fig. b) than estimated by our model (Fig. a). By the evening of 19 July much of the UAS survey area surface had been transformed into flatter, denser ice, overlain in places by ponded meltwater (Appendix Fig. b). The remainder of our field campaign saw a return to shortwave-dominant energy balance conditions. Sub-surface melting presumably dominated ablation, as no significant surface lowering occurred (Fig. c) despite relatively large modelled melt fluxes (Fig. b), resulting in weathering crust (re-)development (Appendix Fig. c).

We propose that changes in weathering crust state can be diagnosed through repeat measurements of reflectance in the near-infrared (NIR) part of the spectrum made by our UAS, centred on 840 nm. Absorption by LAIs such as glacier algae is concentrated in the visible part of the solar spectrum, while at 840 nm the albedo-reducing effect of glacier algae and other impurities are negligible , and so by deduction variations in the NIR are primarily due to changes in near-surface ice properties. Whilst there may be some residual reflectance reduction attributable to black carbon , by deduction, the dominant signal retrieved at 840 nm by our UAS is indicative of the weathering crust state, inclusive of ice grain sizes, ice density, porosity, and interstitial and ponded surface meltwater. Reflectance at 840 nm thus essentially provides an indication of the presence or absence of air–ice interfaces available for high-angle light scattering, whereby more interfaces result in higher albedo .

UAS observations from S6 show a widespread increase in 840 nm reflectance between 20 July and 21 July (Fig. a). True-colour UAV composites illustrate a transition from wet, polished and impermeable ice surfaces (Fig. c, f) to drained, whiter ice with meltwater draining through the porous near-surface (Fig. d, g), also shown by oblique surface photos (Appendix Fig. b, c). This change was coincident with the surface energy balance returning to a shortwave-dominant regime following 4 d of dramatically reduced net shortwave radiation (Fig. a) and 1–2 d of rainfall. Following regrowth of the weathering crust and drainage of ponded meltwater there were no further systematic changes in NIR reflectance (Fig. b) or true-colour composites (Fig. d, e). These findings are consistent with previous studies showing that weathering crust development versus decay is controlled primarily by the relative dominance of radiative or turbulent fluxes in the surface energy budget . Further, the reduction of albedo by rainfall through weathering crust stripping means that the melt-generating potential of cyclonic moisture intrusions which have been shown to account for ∼40 % of total precipitation over Greenland is likely to be higher if this rainfall–albedo feedback is accounted for in regional climate models.

Repeat UAS acquisitions at S6 showed that the proportional coverage of different surface classes varied significantly from one day to the next (Fig. ). Over the study period, LA coverage varied by 19 % and HA coverage by 11 %. The reduction in snow and CI between 15 and 20 July was caused by rainfall and high winds on 18 and 19 July, which resulted in high sensible heat fluxes (Fig. a) and rapid surface melting on 19 July despite low net shortwave radiation (Fig. b). Rainfall caused widespread reduction of the thickness of the porous near-surface weathering crust layer and transient cryoconite hole melt out, dispersing cryoconite granules and darkening the surface further . On 20 July only 5 % of the surface was CI, compared to ∼10 % on 15 July, with the majority of the area (87 %) classified as LA or HA. However, the data used to train our classifier have few examples of CI, which is dark due to very thin or absent weathering crusts, so it is likely that some CI surfaces may have been misclassified as LA. Subsequently, CI coverage increased on 21 July and was associated with a 9 % increase in albedo (Fig. d) and regrowth of the weathering crust (Fig. ).

From 21 to 23 July there was relatively little change in proportional surface cover. However, from 23 to 24 July there was a substantial increase in HA, together with the appearance of water and cryoconite and a 10 % albedo reduction (Fig. d). Furthermore, variable illumination conditions during the 24 July flight over the western half of the study area caused overestimation of reflectance, likely favouring classification as CI, and so we probably did not capture the full magnitude of surface darkening.

The apparent increase in HA coverage on 24 July was probably not driven entirely by algal growth. Population doubling times are estimated to be 5 d , longer than the 1 d here. Indeed, LA coverage declined on 24 July while CI remained constant, whereas we would expect both LA and HA to increase in the case of widespread population growth. Instead, cells in LA areas may have been mobilized by the abundant surface meltwater and then deposited downslope in higher concentrations: air temperatures stayed above 0 ∘C overnight from 23 to 24 July (Fig. b), associated with higher sensible heat fluxes (Fig. a), causing the most daily modelled melting of the observation period (Fig. b). Furthermore, the sensible heat flux increased the proportion of surface melting relative to sub-surface melting by shortwave penetration, likely thinning the weathering crust and further increasing the amount of liquid meltwater available on the surface, reducing albedo, and increasing the likelihood of misclassification as HA. However, we note that our classification approach relies on coarse surface categories. Any LA ice patch loaded with algae towards the upper bounds of 103 cells only needs a relatively small amount of growth to become loaded with 104 cells found in HA samples (Sect. ), and so in some pixels the algal population need not have doubled in order to switch from LA to HA. We therefore cannot rule out the role of algae in causing daily surface type changes.

We also found that albedo was a weak predictor of surface class, with considerable overlap in the albedo of the various classes (Fig. c, f). Broadband albedo alone is therefore not a reliable predictor of ice surface type and cannot be used to infer the presence of glacier algae or other LAIs.

These findings illustrate that there are two principal reasons why surface classes might change through time: (1) algal growth (and removal, for instance by flushing by meltwater) and (2) physical changes which result in (mis-)classification. We cannot uniquely distinguish between changes caused by algae versus those caused by the weathering crust. First, algal growth is associated with enhanced melting, which reduces the thickness of the weathering crust and liberates liquid water and nutrients, stimulating further growth . Second, changes in weathering crust optics occur beneath the algae, so any diagnostic algal feature present in our UAS images may change as the surface microtopography constituting the cell habitat changes. Third, there is uncertainty in spectral biomarkers unique to glacier algae. Theoretically, a simple band ratio, spectral feature identification or spectral mixing technique could be used to detect glacier algae, as has been achieved for snow algae . However, absorption by Mesotaenium berggrenii and Ancylonema nordenskiöldii , the species found on the GrIS, is dominated by phenolic compounds that absorb strongly across the visible wavelengths and obscure potentially diagnostic spectral features associated with other algal pigments . A subtle absorption feature related to Chlorophyll a is sometimes detectable using high-spectral-resolution measurements but is not visible in our multi-spectral imagery.

Our measurements at S6 were undertaken coincident with clear-sky observations by S-2 and MODIS MOD10A1. There was generally close agreement between UAS and satellite-derived albedo measured at S6 (Fig. d). We attribute discrepancies to unavoidable differences between the radiometric calibration and narrowband–broadband conversion techniques and the different degrees of spatial integration. Nevertheless, the direction and magnitude of albedo change between the UAS and S-2 showed good agreement, whilst in general the UAS and MOD10A1 agreed on the direction of albedo changes (Fig. d). In the following section we use our UAS data to understand variability in surface type and albedo measured by S-2 and MOD10A1.