Terrestrial LiDAR is an established, effective technique for monitoring surface roughness and a wide variety of snow processes (e.g. ). A fixed, automated, high-resolution 3D Terrestrial Laser Scanner (TLS; FARO^® Focus3D S120) scanned a delimited area of about 300 m2 in the immediate proximity of the Weissfluhjoch research station (2536 ma.s.l., Davos, Switzerland), at hourly resolution and over three snow seasons. The TLS occasionally required restarting or relocation, resulting in mostly minor data gaps; the largest occurred in 2023, when the TLS only became operational in April, with no acquisitions made beforehand. Overall, less than 2 % of the scans were discarded due to acquisition instabilities caused by wind gusts or heavy snowfall. Because the scanned area is large and contains obstacles such as instruments and fences, and because point density decreases with distance from the scanner as incidence angles become increasingly grazing, we reduced the point clouds to a region of interest (ROI) of approximately 10m×5m. During ablation, snow generally disappears earlier in the south-eastern part of the domain; therefore, to track suncup evolution over a longer period, we defined a smaller 5m×5m sub-domain within the north-western area of the ROI. We then project the ROI onto the horizontal plane. To remove points not belonging to the snow surface (be it snowflakes from snowfall or blowing snow, or random noise) which would disturb the interpolation, we applied the well-established cloth simulation filtering algorithm of , with a cloth resolution of 30 cm, a classification threshold of 10 cm, and maximum rigidness, as recommended for flat terrains. On average, we filtered out 1 % of points per scan. We then interpolated the filtered point clouds to a regular 5 mm grid using bilinear interpolation to produce digital surface models (DSMs). Finally, we removed the slope trend using Gaussian filtering (e.g. ), revealing microtopographic reliefs and concavities. The result is a time series of microtopographic DSMs, suitable for estimating surface roughness indices using raster-based approaches .

We computed snow surface roughness descriptors across two complementary scales. Vertical descriptors characterize the amplitude of surface relief; we parameterize them using the aerodynamic roughness length (z0). z0 is the theoretical height above a surface at which the mean wind-speed logarithmic profile extrapolates to zero. We estimated z0 from our Gaussian-detrended LiDAR DSMs using the geometric method of , adapting it to raster grids. First, we smoothed the interpolated DSMs from their native resolution of 5 mm to 1 cm, to reduce computational load and remove grain-scale roughness. Then, drawing conceptual inspiration from , we identified roughness elements as individual obstacles. While identify individual roughness elements using an innovative watershed segmentation approach, we chose a simplified method to reduce computational costs, given the need to process nearly 10 000 DSMs. We used the 75th percentile of the elevation distribution to classify cells as obstacles, and merged neighboring obstacle cells into single roughness elements. A percentile threshold, rather than a fixed height, provides a measure independent of the surface amplitude, which varies throughout the season. This threshold identifies individual suncups as obstacles and avoids aggressively merging multiple suncups. We also set a minimum threshold of 20 neighboring cells (i.e. 20 cm2) to identify an obstacle element. Physically, the resulting obstacles correspond to the suncup walls, which produce most of the drag. We assume the hollows between them are mostly sheltered from the wind and therefore do not enter the calculation. Then, for each obstacle i, we computed the maximum height hi above the mean detrended surface, the horizontal footprint area Si, and the vertical area si,wd facing each cardinal wind direction wd∈(N,E,S,W), where a cell counts as wind-facing if it is higher than its immediate upwind neighbor. For each wind direction, the aerodynamic roughness length of the whole ROI is: 1z0,wd=Cd⋅h‾⋅∑isi,wd∑iSi, with h‾=1n∑ihi

Where Cd=0.5 is Lettau's drag coefficient and n is the total number of obstacles. Unlike , we summed the areas first, giving larger elements more relative weight. Each scan thus yielded four directional z0 values, which we averaged to obtain the time series shown in Fig. a–c. The median spread of the four directional z0 values across all scans is less than 3 %, meaning there is almost no directional z0 anisotropy at our study plot.

Horizontal descriptors characterize the lateral scale of surface relief. To represent this scale, we use the correlation length (CL), which expresses the horizontal distance over which surface height fluctuations remain statistically correlated . We computed CL separately across (CLx) and along (CLy) the TLS beam direction. The TLS is facing north; therefore, the y axis is along the dominant wind direction. We define the anisotropy between CLx and CLy: 2aCL=max(CLx,CLy)min(CLx,CLy)

The ratio aCL describes surface directionality: anisotropic, preferentially oriented reliefs exhibit large values, whereas aCL≃1 (i.e. CLx≃CLy) indicates isotropy. The vertical and horizontal descriptors provide complementary information about the snow surface state. A ripple field exhibits large horizontal and moderate-to-low vertical descriptors: its reliefs are spatially extended, widely spaced, and strongly directional, giving large aCL. Suncups, conversely, show smaller horizontal and larger vertical descriptors, since reliefs of similar amplitude arrange periodically over short distances with generally no preferential direction: the surface tends towards isotropy, giving aCL→1.

Moreover, we computed raster-based descriptors of suncup formation and growth through image analysis of the DSMs, such as the number of features with a negative height, the suncup area, the suncup maximum depth, and the 2D equivalent of the roughness coverage fraction η used in . Unlike in , the non-concave portions of our natural surfaces are not flat, so we include ridges in the calculation.

Snow metamorphism decreases albedo by altering snow density, liquid water content, and SSA. We simulated the evolution of these snowpack properties with the 1D multi-layer, physics-based SNOWPACK model . SNOWPACK describes the driving processes impacting the snow energy balance, such as metamorphism and microstructure , phase changes, liquid water transport and preferential flow , and turbulent kinetic exchanges at the surface , among others. We applied the snow surface temperature as a Dirichlet upper boundary condition while the snow surface was sub-freezing, and used calculated energy fluxes as Neumann boundary conditions during melt. We forced SNOWPACK simulations with quality-controlled measurements from advanced meteorological sensors at the research station and its immediate surroundings . The SNOWPACK output includes energy fluxes and multilayer profiles of snow temperature, density, liquid water content, and optical grain size, among other properties. We analyzed surface energy fluxes at hourly resolution and multi-layer snow profiles at 3 h intervals (00:00, 03:00, …, 21:00 UTC) to improve computational efficiency.

The ability of SNOWPACK to reproduce the SSA and density of individual layers was thoroughly investigated at Weissfluhjoch . Over one winter season, the authors report overall underestimations of SSA by SNOWPACK, with the magnitude of the deviations depending on the measurement technique used for comparison. Regarding density, found that SNOWPACK overestimates the densification rate of lower snowpack layers and, conversely, underestimates the density of surface layers evolving from fresh snow to rounded grains. The comparison between our SNOWPACK simulations and the detailed surface measurements from confirms the above observations (Fig. k–l and n–o).

We used the Two-streAm Radiative TransfEr in Snow (TARTES; ) to simulate the spectral albedo (350–2200 nm) of a flat snowpack, with density and optical grain size prescribed by SNOWPACK. Because incident light is either absorbed or scattered away by snow grains within 10–20 cm from the snow surface , we prescribed numerical layers 1 cm thick for the top 10 cm, and a single layer of weighted-average properties below. For each layer, we specified the impurity type and concentration.

As impurity content measurements were unavailable at our site, we designed five scenarios (Table ) building upon the existing monitoring campaigns in the Alps, assuming that LAPs absorption is only attributable to black carbon and dust , which systematically co-exist in the Central Swiss Alps , although at strongly varying ratios. Specifically, dust dominates radiative forcing in dust-event years, while black carbon dominates otherwise . Therefore, we divided our scenarios into two families: mixed scenarios, in which black carbon and dust co-exist at comparable levels, and dominated scenarios, in which one species clearly outweighs the other.

Our medium black carbon scenario matches the observed maxima in the two-year campaign of and is consistent with several previous assessments . Our high black carbon scenario reflects extreme melt events in which surface concentration can multiply up to a factor of 5 via melt amplification . Our dust scenarios are based on measurements and simulations from , which considers Algeria as the main source area of Saharan dust (PM2.5) reaching Europe. The two scenarios in which either black carbon or dust dominates are based on the same literature mentioned above. TARTES computes the mass absorption efficiency of impurities, assuming the optical properties of black carbon and dust .

The current version of TARTES does not implement the optical properties of red snow algae. Therefore, we incorporated them as a custom light-absorbing impurity using the empirical in vivo properties measured by . For the algal bloom scenario, we estimated a mass mixing ratio from the minimum cell concentration detectable by Sentinel-2 (∼20 000 cellsmL-1; ), representing bloom intensities with a measurable impact on the surface reflectance; a cell dry mass of 10-12 kgcell-1 derived from the mean snow algal biovolume and dry density reported in , and the mean modelled snow surface density during the melt season (300 kgm-3); yielding a mass mixing ratio of 66 µg g-1.

We ran TARTES simulations under diffuse (overcast-sky) and direct (clear-sky) conditions and compared the results with measurements from Kipp and Zonen CM21 broadband pyranometer. We discarded measurements with grazing solar zenith angles (below 25° and above 65°) to avoid errors arising from the pyranometer's cosine response . In the absence of a dedicated diffuse-radiation sensor, we obtained the diffuse fraction by decomposing the measured global radiation using the formulations of , , and implemented in MeteoIO , the meteorological pre-processing library of SNOWPACK. Following , we then computed the ratio rλ between incoming diffuse and global horizontal irradiance and classified timesteps with rλ≤0.20 as direct-dominated and rλ≥0.90 as diffuse-dominated. The deliberately wide gap between the two thresholds excludes mixed illumination, so that each group represents a quasi-pure illumination condition.

We obtain modeled broadband albedo values for both illumination conditions by weighting the TARTES spectral albedo with modeled incident solar spectra computed with libRadtran 2.0.6 : direct-beam irradiance under clear sky for the direct class, and global downwelling irradiance under overcast sky for the diffuse class. Spectra were generated for the neighboring Totalp site (46.837° N, 9.815° E; 2470 ma.s.l.) on 7 March 2024 at 12:00 UTC (solar zenith angle 52.3°) using the Kurucz solar source , a midlatitude-summer atmosphere , the rural aerosol model of scaled via the Ångström relation (α=1.491, β=0.0482). Overcast conditions were represented by a fully overcast water cloud (liquid water content 0.2 gm-3, effective droplet radius 10 µm) between 3000–4500 ma.s.l. Both spectra were resampled to 1 nm between 350–2500 nm and normalised to unit integral before weighting.

To determine the effect of suncups on the apparent broadband albedo, we compare measured time series with TARTES simulations in diffuse and direct illumination conditions for all years (Fig. a–f). Because TARTES neglects the presence of liquid water, relying on pure-ice refractive indices, we evaluate the quality of the simulations in dry and wet conditions separately (Fig. g–i). Wet conditions are initiated when the daily average of the modeled LWC exceeded 1 % in the top 30 cm (2022), or based on direct field measurements through 2023 and 2024 . Because suncups form on a wet snow surface , any misrepresentation of wet-snow metamorphism by SNOWPACK or TARTES must be ruled out for the earlier flat-surface stage. Despite SNOWPACK sometimes underestimating surface SSA and systematically underestimating density (Fig. k, l, n, and o), TARTES simulations forced by either SNOWPACK or measured surface properties show mostly negligible differences in broadband albedo (Fig. ). This indicates that simulated albedo is insensitive to these input biases in the wet phase, before suncup development. We therefore attribute the deviation from measured broadband albedo primarily to TARTES not accounting for the optical effects of liquid water and wet-snow metamorphism on a flat surface. To do so, we define the index βwetsmooth as the mean offset between TARTES simulations and broadband albedo measurements, computed separately for each year and illumination condition, over the wet-smooth phase. Therefore, for each year y, the wet-smooth window Wysmooth=t:twet≤t<trough defines the evaluation set Sysmooth containing pairing broadband albedo values αTARTES(t) and αmeasured(t). The offset is the mean difference over the evaluation set: βwetsmooth(y)=1Sysmooth3×∑t∈Sysmooth(αTARTES(t)-αmeasured(t))

We evaluate this separately for direct and diffuse illumination conditions. Consequently, and similarly, we define βwetrough as the additional offset that appears once the suncups develop: βwetrough(y)=1Syrough∑t∈SyroughαTARTES(t)-αmeasured(t)︸mean raw offset over Wyrough4-βwetsmooth(y)

By subtracting βwetsmooth from albedo residuals in the rough phase, we assume that the systematic errors deriving from wet-snow representation in TARTES are accounted for to the first order, leaving a residual signal that we can attribute solely to the co-evolution of suncup roughness and impurity concentration.

To (partially) account for the effect of suncup roughness on the domain-averaged broadband albedo, we applied the effective albedo parametrization presented in , originally formulated for subgrid topographic shading and terrain reflections in large-scale models, but fully applicable to smaller scales to the first order, provided the surface statistics are described by the same geometric parameters. These parameters are the mean-squared slope μ and the sky-view factor ψ. According to these parameters, this approach modifies the flat-surface albedo computed by TARTES. To avoid propagating the wet-snow bias, this parametrization modifies the flat-surface albedo already corrected for βwetsmooth (Eq. ). Specifically, the direct irradiance is attenuated by a shading factor that accounts for the fraction of the surface in shadow as a function of the solar zenith angle and μ, including partial shadowing at low sun elevations when suncup walls shade adjacent terrain. The diffuse irradiance is reduced in proportion to ψ, whereas the obstructed fraction (1-ψ) adds up the terrain-reflected radiation. We derived both μ and ψ from the LiDAR surface scans.

As the shading factor depends on the solar zenith angle and decreases significantly under diffuse illumination (overcast sky), direct-dominated illumination conditions are those in which shading between suncup walls is the dominant roughness-induced albedo-reduction mechanism . Under an overcast sky, the remaining roughness-induced albedo reductions are instead governed by the limited sky view factor and single-terrain reflections, both of which the parameterization includes. The parametrization in does not account for multiple reflections under either condition. This practical compromise is motivated by the need for a quasi-analytical method that can be applied systematically across three seasons of hourly LiDAR roughness acquisitions, for which full 3D ray-tracing would be computationally prohibitive. Neglecting multiple reflections omits the photon-trapping contribution to albedo reduction. quantified this effect in the field at roughly 5 % of the total roughness-driven albedo decay on average, but found it exceeds the 20 % for late-ablation suncups: a regime our observations fully cover. Its omission therefore likely underestimates the modeled albedo reduction that increases with suncup growth.

Our time series highlight two distinct roughness regimes. The transition between regimes is marked by the onset of monotonically increasing vertical roughness index, occurring on 11 May 2022, 22 May 2023, and 3 June 2024 for the three seasons, respectively (Fig. a–c). We refer to the period prior to these thresholds as the smooth (or textured) phase, and to the period after as the rough phase.

During the smooth phase, the horizontal roughness index (aCL) varies widely, while the vertical roughness index (z0) remains comparatively stable. In detail, aCL peaks up to 4, whereas z0 changes by no more than 4 mm above the 1 mm smooth baseline across all years (Fig. a–f). Here, isolated peaks in aCL are decoupled from any vertical roughness signal and are often correlated with wind gusts (Fig. p–r). This suggests that wind-driven (dry) snow redistribution processes, such as surface erosion or deposition, significantly alter the snow surface roughness. , reported similar findings, albeit in Antarctic snowfields, where sastrugi can form. During the rough phase, this pattern reverses. Vertical roughness increases steeply, reaching values one order of magnitude above the smooth baseline. Across all years, z0 changes by up to 15 mm. Conversely, the horizontal roughness signal becomes highly isotropic, with aCL shrinking towards 1. The absence of a wind signature in aCL during this phase suggests that suncup development, rather than wind redistribution, governs the evolution of surface roughness: the surface is mostly a melt-refreeze crust and therefore relatively insensitive to wind redistribution processes. In the following, we focus on the rough phase, which coincides with the snow ablation season (Fig. s–u) and where suncup growth is the dominant surface degradation mechanism at our site.

We define conditions for the onset of suncup growth by combining roughness time series with meteorological forcings and snowpack state variables. The disappearance of cold content provides the first signal for roughness-prone conditions (Fig. g–i). Once the cold content reaches zero for the first time, the snowpack is isothermal and can then melt and refreeze, initiating the first melt-related surface reliefs such as melt-refreeze crusts and ice layers. After several melt-refreeze cycles, the wetting front propagates until the bottom of the snowpack, with increasing average liquid water content (LWC, Fig. g–i). The average daily snow surface temperature (TSS) progressively increases, together with the number of isothermal surface hours per day (Fig. j–l). Across all years, the transition from a smooth to a rough regime consistently occurs when TSS remains at 0 °C for approximately 20 hd-1, with an average snowpack LWC of around 2 %. This suggests that the persistence of melt-prone conditions, rather than their episodic occurrence, is the key threshold for suncup growth. Moreover, in all years, the suncup growth onset closely coincides with wet-bulb temperatures consistently above zero (with average values of around 3 °C over the entire period – Fig. m–o) and reduced wind speeds relative to the smooth phase (Fig. s–u). The warming reflects the seasonal shift in thermal conditions. With above-freezing air temperatures and high wind speeds, more heat likely reaches the suncup ridges than the hollows, reducing deepening. Together, these indicators mark favorable conditions for the onset of suncup growth: (1) a snowpack that has reached an isothermal state, (2) near-continuous daily surface melt, (3) close to constantly positive wet-bulb temperatures, and (4) reduced wind speeds. As noted above, the smooth phase corresponds predominantly to the accumulation period, whereas the rough phase coincides with snowpack ablation and higher melt rates (Fig. s–u).

Suncup growth was consistently observed across all ablation seasons on our study plot (Fig. a–c), with 170–200 suncups forming each season. Webcam and LiDAR acquisitions (Fig. a–f) show that the 2D equivalent of the roughness coverage fraction η used in approached 100 % under natural conditions: the portions of the sub-domain not occupied by suncup hollows are structured into ridges, leaving little to no flat area. Consequently, η alone cannot describe suncup development under natural conditions, motivating the use of suncup area and maximum depth as additional descriptors.

We define the suncup growth window as the interval between the onset of the rough phase and reaching maximum depth. Its initial phase is marked by a steady increase in the number of suncups; subsequently, individual features begin to merge, and further growth proceeds through merging and deepening. Spring snowfall events periodically interrupt this progression, temporarily suspending growth by producing abrupt decreases in depth and increases in equivalent diameter. Table  summarizes the suncup growth descriptors, the growth-window duration, and the corresponding meteorological and radiative conditions for each season.

Suncup growth was most pronounced in 2022, which recorded the highest average depth of 7 cm (at least 2 cm greater than in the other seasons) and the narrowest mean suncup areas (Fig. g, j, and m). Deepening was also fastest that year (0.4 cmd-1), compressed into a 17-d growth window, 5 and 15 d shorter than in 2023 and 2024, respectively. Despite being the shortest window, it combined extremely favorable radiative and thermal conditions. The net shortwave gains were the highest (+7.1 MJm-2d-1; 7 % and 20 % above 2023 and 2024 – Fig. p–r), and the wet-bulb temperature the warmest (+4.0 °C – Fig. s), and therefore the overall most melt-permissive. Winds were the weakest (1.7 ms-1), showing decent constancy (C‾w=0.5, ) and relatively low directional variability (σθw=±69°) around a mean direction from the NW (Fig. v). This combination of favorable radiative and thermal conditions produced deep, small, circular, and evenly distributed suncups (Fig. a and d).

In 2023, suncups reached comparable maximum depths (11 cm) but were on average 2 cm shallower and 50 % wider than in 2022, developing over a longer 23-d window (Fig. h, k, and n). Radiative conditions remained favorable (+6.6 MJm-2d-1 net shortwave), yet this season carried the largest net longwave loss (-1.7 MJm-2d-1 – Fig. q) with the overall coolest wet-bulb temperature (+2.8 °C – Fig. t) which, whilst still favoring melt, did so least of the three seasons. Despite average wind speeds and directions being comparable to 2022, in 2023 wind showed overall greatest constancy (C‾w=0.7, ) and lowest directional variability (σθw=±52°) around the NW direction (Fig. w). This produced notably larger, elongated features oriented along the prevailing NW wind direction (Fig. b and e), consistent with the large size variability of that season, where the standard deviation of suncup area (σA=1982cm2) approached twice the mean (μA=1119cm2).

Radiative conditions in 2024 were the least favorable, with the overall lowest net shortwave gains (+5.7 MJm-2d-1), and the highest net longwave gain from frequent cloud cover (+0.4 MJm-2d-1 – Fig. r). However, thermal conditions were still favorable: the wet-bulb temperature was the second-highest and exceeded 2023 by nearly 1 °C (Fig. u). Winds were the strongest (2.3 ms-1) but also the most variable (C‾w=0.05) around a bimodal NW/SE regime (Fig. x). Suncups were 25 % shallower than in 2022 (Fig. m and o). The highest directional variability (σθw=±118°) precluded the development of a preferred morphological axis despite the greatest overall cumulative wind forcing, producing similarly wide but less elongated suncups than in 2023 (Fig. b, c, e, f, k, and l).

Together, these results indicate that suncup deepening rate and depth were most closely associated with the radiative energy balance and with melt-permissive (warm, humid) near-surface conditions. In contrast, orientation and elongation were most closely associated with wind speed and directional constancy. Table  summarizes all quantitative indices.

Under smooth, dry conditions, TARTES overestimates and slightly underestimates measured albedo under diffuse and direct illumination conditions, respectively, with biases βdrysmooth+0.04 and -0.01 (Fig. a–c and d–f), which may reflect inaccuracies in the input snowpack and model limitations. As the snow wets and SSA decreases, TARTES overestimates albedo under all illumination conditions before suncup growth, with average biases across years of βwetsmooth=+0.06 (diffuse) and +0.04 (direct). Despite SNOWPACK systematically underestimating the density of the uppermost layer and, occasionally, its SSA (Fig. j–o), this doesn't seem to introduce important biases in SNOWPACK-forced simulations with respect to measurement-forced ones (Fig. ). Therefore, we attribute βwetsmooth to TARTES's misrepresentation of wet-snow metamorphism on a flat surface.

At the onset of suncup growth, we further partition the contributions to albedo reduction and account for impurities and algae on wet, heavily metamorphosed snow. We quantify the suncup signature, βwetrough, by subtracting βwetsmooth from the residual between measured and TARTES broadband albedo (Eq. , Fig. p–q). Assuming pristine snow, suncups generate albedo decreases of 0.04–0.07 and 0.08–0.1 for direct and diffuse conditions, respectively. The consistently larger diffuse signature (Fig. q), evident across all scenarios, reflects the visible-weighting of the overcast spectrum (Fig. r; see below). When assuming mixed, low-to-medium impurity scenarios (or the black-carbon-dominated scenario, which reduces albedo comparably), suncup signatures decrease to 0.01–0.06 and 0.05–0.09 under direct and diffuse conditions, respectively. This suggests that suncup roughness and impurities jointly contribute to the albedo deficit that TARTES cannot reproduce when assuming a flat surface. When assuming mixed-high scenarios (or the dust-dominated scenario, which reduces albedo comparably) and algal bloom scenarios, βwetrough tends to, and often falls below, zero under direct illumination conditions (Fig. p). These impurity scenarios cancel the suncup signature, suggesting that the combined effect of suncups and impurities is comparable to that of the hypothesized impurity content on a flat surface.

βwetrough is, on average, twice as large under diffuse as under direct illumination (Fig. p and q). This asymmetry cannot be attributed to suncup roughness alone. At the visible-NIR boundary, the two insolation weights intersect and cross (Fig. r): the diffuse spectrum over-weights in the visible, the direct over-weights in the NIR. However, suncup mechanisms do not favor diffuse illumination: the effective-angle effect activates only under direct illumination, and the multiple-reflections effect activates in the NIR, where the diffuse spectrum carries less weight (Fig. r). Moreover, in the highly metamorphosed rough phase, the NIR spectral albedo lies well below the ≃0.5 balance value that maximizes photon trapping , further limiting its contribution. Therefore, suncup mechanisms cannot produce a larger diffuse signature. The asymmetry instead matches a visible-dominated absorption signature: because the diffuse weight is concentrated in the visible, where LAPs and algae have their strongest influence, any impurity component of the residual is preferentially amplified under diffuse conditions. This explains why the diffuse βwetrough decreases as the hypothesized impurity concentration increases (and approaches 0 in the 2024 algal bloom scenario), and why it requires higher assumed concentrations to approach zero than the direct residual does. Rather than suncup roughness, the larger diffuse residual likely reflects unaccounted-for impurity darkening.

The effective albedo parametrization of (TARTESLH) further supports this interpretation, as it lowers residuals relative to the flat-surface assumption (TARTESsmooth) predominantly under direct illumination, when both suncup mechanisms are active. Under diffuse illumination, it only marginally reduces the spread of residuals or converges towards the flat-surface results (Fig. q), as expected if the diffuse residual is governed by visible-range impurity darkening that a roughness parametrization cannot capture. We note, however, that we cannot attribute the diffuse residual to impurities with full confidence. Under overcast, smooth-wet conditions, TARTES flat-surface albedos exceed the measurements substantially and decrease little through the season, unlike under clear skies, where they track the seasonal decline. Therefore, part of the residual may reflect flat-albedo bias, inaccuracies in the direct–diffuse partitioning rλ or spectral radiation input, or the influence of the surrounding terrain, rather than impurity loading alone.

Under direct illumination, the effective albedo parametrization of brings residuals close to, or slightly below, zero across all years and hypothesized scenarios, with βwetrough averaging -0.02. The mixed, low-to-medium (or black-carbon-dominated) hypothesized concentrations seem the most likely at our site, with the lowest βwetrough=-0.009 across all years. The mixed-high, dust-dominated, and algal-bloom scenarios yield residuals that fall systematically below zero (βwetrough=-0.04 on average). Because βwetrough is averaged over the entire rough period, this suggests these high impurity concentrations are unrealistic as a whole-phase average, though plausible for the late ablation stage, when surface impurities have become most concentrated.

During the rough phase, the measured broadband albedo exhibits a strong logarithmic correlation with the aerodynamic roughness length z0 (ρs=-0.7, p=10-43, Fig. a). Two categories of events deviate from this relationship: spring new-snow layers depositing on pre-existing suncups, which abruptly increase SSA and raise albedo outside the fitted trend, and exceptionally high impurity loadings, such as the 2024 algal bloom. The logarithmic fit also shows that the rate of albedo decay is steepest at low z0 and progressively relaxes as z0 approaches values characteristic of deep suncups.

The shape of the correlation between z0 and the broadband albedo decay suggests a possible link between the evolution of suncup roughness and impurity loading. Webcam acquisitions from our study plot (Fig. ) point to a spatial redistribution of impurities as suncups deepen. Over early-stage suncups (Fig. b), impurities appear homogeneously distributed. At later-stage suncups (Fig. c), despite likely higher total concentrations, impurities appear preferentially located in the hollows (Fig. d and e) and, to a lesser extent, at the ridges (Fig. f and g), leaving the exposed surface apparently cleaner than in the early-stage situation. This redistribution appears to relocate impurities within the suncup topography rather than remove them from the surface. The coupling between suncup growth and impurity loading may therefore be most effective during the early-stage phase, when impurities are more uniformly distributed; as suncups deepen and meltwater concentrates impurities into the hollows, the roughness contribution may grow, while the darkening effect of LAPs, though spatially reorganized, is not necessarily reduced.

Aeolian debris transport also appears to contribute to surface darkening. Figure h (June 03rd, 2023; z0=8.7mm) shows a week when winds blew predominantly from the NW, with a speed-weighted average direction of 305° prevailing for 77 % of the total hours. Dark, rock-like debris is visible on the upwind section of the early-stage suncups. The NW sector of the research field is enclosed by steep rock faces where snow generally disappears well before it does on the field itself, which sits in a slight local terrain depression that allows snow to persist longer. The surface darkening is therefore consistent with rock debris mobilized from the surrounding slopes. Conversely, Fig. i (17 June 2024; z0=8.3mm) corresponds to a week of predominantly SE winds, with a speed-weighted average direction of 147° prevailing for 70 % of the total hours. Although the suncup roughness is nearly identical to that of the previous year, the upwind dark debris is markedly less apparent. Unlike the NW sector, the SE side of the research field opens onto the valley and is not overlooked by steep rock faces, consistent with a reduced local debris supply.