2.1 Incoming and reflected radiation at the surfaces

The incoming light irradiance $E_{\mathit{\lambda }}^{\downarrow }$ on a horizontal surface or on a perfectly levelled upward-looking sensor is the sum of the direct solar irradiance $E_{\mathit{\lambda }}^{\mathrm{sun}}$ weighted by the interception probability cos θ_i (${\mathit{\theta }}_i\in [\mathrm{0},\frac{\mathit{\pi }}{\mathrm{2}}]$) and the diffuse irradiance coming from the sky $E_{\mathit{\lambda }}^{\mathrm{sky}}$:

where θ_i and ϕ_i are the zenith and azimuth solar angles.

A slope is modelled here as a plane defined by its normal N pointing upward with zenith angle θ_n and azimuth angle ϕ_n. This surface receives direct solar radiation with a modified interception probability $\cos {\stackrel{\mathrm{´}}{\mathit{\theta }}}_i$, accounting for the local incidence angle, given by

where the “shadow” function S(x) is 1 for x>0 and 0 for x≤0. This ensures that when the sun is below the slope, the interception probability is null and not negative.

In addition, the slope receives diffuse irradiance from the atmosphere ${\stackrel{\mathrm{´}}{E}}_{\mathit{\lambda }}^{\mathrm{sky}}$, which is lower than that received by the horizontal surface owing to the screening by the slope itself. Assuming isotropic sky radiation, the received diffuse irradiance can be written

where $\cos \stackrel{\mathrm{´}}{\mathit{\theta }}$ is given by Eq. (2) without the i subscript. The integral runs over the sky sector, i.e. the sector of the hemisphere (HS, $\mathit{\theta }\in [\mathrm{0},\frac{\mathit{\pi }}{\mathrm{2}}]$ and $\mathit{\varphi }\in [\mathrm{0},\mathrm{2}\mathit{\pi }]$) above the slope ($\cos \stackrel{\mathrm{´}}{\mathit{\theta }}>\mathrm{0}$, blue shading in Fig. 1b). This integral can be calculated in the reference frame of the slope (Wang et al., 2016; Dumont et al., 2017) involving variable changes and some analytical calculations. An alternative is to write this integral in terms of the flux of the slope normal N through the sky sector:

Let us consider the closed surface including the sky sector (blue shade in Fig. 1b), the unit half disc on the slope above the sensor, and the unit half disc on the horizon plane. According to Gauss' theorem, the total flux of any conservative vector, as N, through this closed surface is null. Thus the flux in Eq. (4) can be deduced from the fluxes through both half discs. The flux through the half disc on the slope – which is perpendicular to N – is $-\frac{\mathit{\pi }}{\mathrm{2}}$ and the flux through the half disc on the horizontal plane is $-\frac{\mathit{\pi }}{\mathrm{2}}\mathit{N}.\mathit{z}=-\frac{\mathit{\pi }}{\mathrm{2}}\cos {\mathit{\theta }}_n$, z being the vertical axis. It follows that

V can be interpreted as the cosine-weighted viewing fraction of the sky seen from the slope.

This reduced contribution from the atmosphere (${\stackrel{\mathrm{´}}{E}}_{\mathit{\lambda }}^{\mathrm{sky}}$) is compensated for by radiation coming from the solid angle under the horizon and above the slope, which impacts both the slope and the downward-looking sensor (the latter is shown as green shading in Fig. 2). This contribution – which we qualify hereinafter by neighbourhood – is written

where we have assumed isotropic radiation and that the surfaces are infinite planes, so that the solid angle of the slope seen from the horizontal surface is the same as the solid angle of the horizontal surface seen from the tilted surface. This implies that the same cosine-weighted viewing fraction 1−V applies to both surfaces as illustrated in Fig. 1b. Summing the direct and the two diffuse contributions, the total incident irradiance on the slope is

and the flux reflected by the slope is

where we have distinguished the direct-hemispherical reflectance ${\stackrel{\mathrm{‾}}{\mathit{\alpha }}}^{\mathrm{dir}}\left(\mathit{\theta }\right)$ (direct albedo) and the hemispherical-hemispherical reflectance ${\stackrel{\mathrm{‾}}{\mathit{\alpha }}}^{\mathrm{diff}}$ (diffuse albedo).

Figure 2The four studied cases depending on the sensor position (mid-slope, top hill) and surface type of the horizontal surface: snow (blue), dark (black). The neighbourhood (i.e. horizontal surface) is seen by the downward-looking sensor in the lower solid angle (green shading, as in Fig. 1). The upper slope is seen by the upward-looking sensor by the upper solid angle (orange shading). Note that the horizontal surface is infinite and the slope is infinite except in the top-hill case.

Download

To continue, the term $E_{\mathit{\lambda }}^{\mathrm{neigh}}$ must be specified. Many scenarios are imaginable depending on the terrain topography and surface type (snow, vegetation, etc.) present in the neighbourhood. For the sake of simplicity, here we consider a horizontal surface and distinguish two cases for the surface type (Fig. 2): (case D) the horizontal surface is dark with reflectivity equal to zero ($E_{\mathit{\lambda }}^{\mathrm{neigh},\mathrm{D}}=\mathrm{0}$) and (case S) the horizontal surface is covered by snow having the same properties as on the slope, i.e. the same reflectivity. Case S is more complex than case D as it depends on the radiation reflected by the slope. The neighbourhood contribution by the snow-covered horizontal surface can indeed be written as

This equation depends on the light flux reflected by the slope (${\stackrel{\mathrm{´}}{E}}_{\mathit{\lambda }}^{\uparrow }({\mathit{\theta }}_i,{\mathit{\varphi }}_i)$), which itself depends on the neighbourhood irradiance (Eq. 9). This interdependence is due to the mutual re-illumination by the two surfaces. We assume that radiation received from the slope is isotropic (as in Eq. 3), meaning that snow is considered to be a Lambertian surface. This very common assumption makes possible the analytical calculations presented in the following. However, we also assume that the direct albedo has an angular dependency (${\stackrel{\mathrm{‾}}{\mathit{\alpha }}}^{\mathrm{dir}}\left({\stackrel{\mathrm{´}}{\mathit{\theta }}}_i\right)$). These two assumptions are physically incompatible, as the albedo of a strictly Lambertian surface has no angular dependence. Nonetheless, the angular dependence of snow albedo has been evidenced by numerous ground observations (e.g. Dumont et al., 2017; Larue et al., 2019) and must be kept. On the other hand, the Lambertian assumption is required to conduct analytical calculations and provide simple formulations. This is the reason why in the following we keep these two somewhat physically incompatible assumptions.

Multiplying Eq. (10) by the mutual re-illumination factor M_λ defined by

and adding the result to Eq. (9) gives the upwelling light flux for the case S:

Similarly, by multiplying Eq. (9) by M_λ, the neighbourhood contribution follows

which is symmetrical to Eq. (12), where the role of the slope and the horizontal surface are permuted.

2.2 Upwelling and downwelling radiation on the sensors

The upward- and downward-looking sensors are considered to be horizontal and to have a perfect cosine response with a 180^∘ field of view. The flux received by the downward-looking sensor placed over a slope comes from the slope and the facing horizontal surface and follows

For the irradiance received by the upward-looking sensor, two cases shall be distinguished (Fig. 2). The first case is when the measurement is taken on the slope far from the horizontal neighbourhood and far from the top of the slope (hereinafter case mid-slope or M). In such a case, the part of the slope above the sensor reflects radiation toward the upward-looking sensor (orange shading in Fig. 2) which adds up to the solar direct and sky radiation. The irradiance in case M is written

where the shadow term $S\left(\cos {\stackrel{\mathrm{´}}{\mathit{\theta }}}_i\right)$ is 1 when the sun directly illuminates the slope and 0 otherwise.

The second case (called T) is when the sensor is above or close to the top of the slope, but still low enough for the downward-looking sensor to mostly view the slope. This is a common case on small slopes. Rigorously, the notion of “top of the slope” is incompatible with the assumption of infinite slope used before, but it is acceptable here as a trade-off between conducting analytical calculations and representing concrete practical situations. In case T, the irradiance received by the upward-looking sensor is given by

From this point, we have all the fluxes required to compute the albedo for four various cases, depending on whether the sensor is located at the top (T) or middle (M) of the slope and on whether the surrounding terrain is dark (D) or covered with snow (S).

Table 1Apparent albedo formulation $\stackrel{\mathrm{´}}{\mathit{\alpha }}\left({\mathit{\theta }}_i\right)=(\mathrm{1}-\stackrel{\mathrm{´}}{r})A^{\mathrm{dir}}\left({\mathit{\theta }}_i\right)+\stackrel{\mathrm{´}}{r}{A}^{\mathrm{diff}}$ for horizontal terrain, small slopes, and for four configurations with significant slopes depending on whether the neighbourhood is covered by snow or by a dark surface, and the albedo and diffuse-to-total ratio measurements are taken mid-slope or at the top of the slope. The slope is not in its own shadow $S\left(\cos {\stackrel{\mathrm{´}}{\mathit{\theta }}}_i\right)>\mathrm{0}$. We have also introduced $M=(\mathrm{1}-V){\stackrel{\mathrm{‾}}{\mathit{\alpha }}}^{\mathrm{diff}}$. The λ dependence is implicit for the sake of simplicity.

Download Print Version | Download XLSX

2.3 Apparent albedo

The measured apparent albedo can now be expressed as a function of the intrinsic direct and diffuse albedo, the diffuse-to-total ratio, and the geometrical parameters. We present here the analytical derivation for the case of “small slopes” and leave the general case of “large slopes” to the Appendix. A summary of all the results is provided at the end of the section.

The small-slope approximation mathematically corresponds to neglecting second-order variations in θ_n (practical upper bounds for θ_n are given in Sect. 4), so that sin θ_n∼θ_n, $\cos {\mathit{\theta }}_n=\mathrm{1}+o\left({\mathit{\theta }}_n\right)$ . It follows that $V=\mathrm{1}+o\left({\mathit{\theta }}_n\right)$ and that for all the cases considered above (cases S or D and T or M), the measured albedo reduces to the same mathematical form:

where we have introduced the geometrical factor

which is the main term carrying the first-order slope effect, and the ratio r_λ between incoming diffuse and total flux far above the slope:

This ratio can be computed with an atmosphere radiative transfer model (e.g. SBDART, Ricchiazzi et al., 1998, 6S Vermote et al., 1997) or can be measured with an upward-looking sensor by obstructing the direct sunlight under the small-slope approximation (as long as V≈1).

For large slopes, the mathematical derivation of the apparent albedo is more complex because V becomes as important as K; i.e. the partial screening of the sky by the slope and the illumination by the neighbourhood have an increasing contribution. The mathematical details are given in the Appendix. The cases DT, DM, ST, and SM lead to different equations. Nevertheless, despite the relatively higher complexity compared to small slopes, a common general form can be found when the diffuse-to-total ratio ${\stackrel{\mathrm{´}}{r}}_{\mathit{\lambda }}$ is measured at the same location as the albedo. This form writes

and it appears to also be suitable for flat surfaces and small slopes as well. Table 1 gives the A^dir and A^diff for all the cases considered in the present paper.

3.1 Albedo correction with known slope parameters

Given the slope parameters (θ_n and ϕ_n), measured ${\stackrel{\mathrm{´}}{\mathit{\alpha }}}_{\mathit{\lambda }}^{\mathrm{mes}}$, and measured r_λ, the goal is to estimate ${\stackrel{\mathrm{‾}}{\mathit{\alpha }}}_{\mathit{\lambda }}^{\mathrm{diff}}$, using the relationship between the direct and diffuse albedos and the small-slope formulation (Eq. 17) given the following equation:

where the only unknown is ${\stackrel{\mathrm{‾}}{\mathit{\alpha }}}_{\mathit{\lambda }}^{\mathrm{diff}}$. It can be solved with any non-linear equation solver. For instance, rearranging this equation to let the difference between the direct and diffuse albedos appear leads to a solution that can be efficiently solved by iterations as follows:

with j the iteration counter, starting at 0 with ${\stackrel{\mathrm{‾}}{\mathit{\alpha }}}_{\mathit{\lambda }}^{\mathrm{diff}}(j=\mathrm{0})=min({\stackrel{\mathrm{´}}{\mathit{\alpha }}}_{\mathit{\lambda }}^{\mathrm{mes}},\mathrm{1})$. The convergence has been tested for a wide range of parameters. Even in the worst case (i.e. the slope opposed to the sun, grazing zenith incidence angle, K=0.2), 10 iterations are sufficient to reach a precision of 0.1 %. In most practical cases (K close to 1), fewer than five iterations are sufficient.

3.2 Albedo correction with unknown slope parameters

The slope parameters are often unavailable or the precision on these parameters is insufficient. If we further assume that surface snow contains negligible amounts of light-absorbing impurities, the intrinsic albedo in the visible wavelengths is nearly constant and close to a value of 1 over a wide range, typically between 400 and 500 nm (Warren and Wiscombe, 1980). This range is also where the slope effect on albedo is the most visible – which is highlighted in Sect. 4 – so that constraining ${\stackrel{\mathrm{‾}}{\mathit{\alpha }}}_{\mathit{\lambda }}^{\mathrm{diff}}={\mathit{\alpha }}_{\mathrm{0}}$ for a range of wavelengths provides a way to estimate K. Indeed, Eq. (24) becomes linear in K if n is calculated by approximating ${\stackrel{\mathrm{´}}{\mathit{\theta }}}_i$ by θ_i, an approximation that is more and more valid as α₀ gets closer to 1. One albedo measurement at one wavelength is in principle sufficient to estimate K, but to improve the reliability we consider here multiple measurements at L different wavelengths λ_l in the range of 400–500 nm. The least-square optimal solution of the linear equation with the unknown K gives the estimate $\stackrel{\mathrm{̃}}{K}$ according to

Here we use α₀=0.98. From $\stackrel{\mathrm{̃}}{K}$, it is possible to estimate $\stackrel{\mathrm{̃}}{n}\left({\stackrel{\mathrm{´}}{\mathit{\theta }}}_i\right)=\frac{\mathrm{3}}{\mathrm{7}}(\mathrm{1}+\mathrm{2}\stackrel{\mathrm{̃}}{K}\cos {\mathit{\theta }}_i)$, which is sufficient to apply the iterative method depicted in Eq. (25) and hence to obtain the diffuse albedo. Note however that knowing K is insufficient to estimate the slope angles θ_n and ϕ_n.

3.3 Unconstrained and constrained correction of the diurnal cycle of albedo with unknown slope parameters

Another practical case is when albedo is measured at different hours during a single day. If we can assume that snow properties have not evolved during that day (e.g. no precipitation, no melt), ${\stackrel{\mathrm{‾}}{\mathit{\alpha }}}_{\mathit{\lambda }}^{\mathrm{diff}}$ is a constant. This assumption leads to the following system of equations:

where ${\stackrel{\mathrm{‾}}{\mathit{\alpha }}}_{\mathit{\lambda }}^{\mathrm{diff}}$ is the unknown along with the two slope parameters θ_n and ϕ_n. Assuming that albedo measurements are taken at N_λ wavelengths, and N_t time steps, the number of equations is N_λN_t and the number of unknowns is N_λ+2. Because of the strong non-linear coupling between the unknowns, no simple analytical method is devisable, but the system can be numerically solved with non-linear least squares using the cost function

where ${\stackrel{\mathrm{´}}{\mathit{\alpha }}}_{{\mathit{\lambda }}_l}^{\mathrm{mod}}\left(t_i\right)$ is given by the right term in Eq. (27). Minimizing the cost function in this study is performed using the Python function scipy.optimize.leastsq implementing the Levenberg–Marquardt algorithm (Levenberg, 1944).

Another method is derived for situations where the snow surface is known to be free of impurities. In such a case, it may be interesting to constrain the albedo value in the blue–green range, ${\stackrel{\mathrm{‾}}{\mathit{\alpha }}}_{\mathit{\lambda }}^{\mathrm{diff}}={\mathit{\alpha }}_{\mathrm{0}},\mathit{\lambda }\in [\mathrm{400}\mathrm{nm},\mathrm{500}\mathrm{nm}]$, as in Sect. 3.2. Adding such a constraint is straightforward using the Python function scipy.optimize.minimize.

3.4 Data

Albedo data were acquired at the Col du Lautaret site (45^∘${\mathrm{2}}^{\prime }{\mathrm{4}}^{\prime \prime }$ N, 6^∘${\mathrm{24}}^{\prime }{\mathrm{18}}^{\prime \prime }$ E) in March and April in the framework of the EBONI campaign running from 2016 to 2019, already presented in Tuzet et al. (2019) and Larue et al. (2019). The albedo data were collected at different locations with various slope configurations in an overall snowy environment, using the manually operated albedometer Solalb, and at a fixed position on a south-east-facing slope using the automatic spectrometer Autosolexs.

Solalb is composed of a single light collector fixed at the tip of a 3 m long arm and connected to a spectrometer (400–1050 nm) (Belke-Brea et al., 2019; Tuzet et al., 2020). Downwelling and upwelling fluxes required to compute the albedo are successively acquired by manually rotating the arm held horizontally, hence pointing the collector upward, then downward. The albedo acquisition is only considered valid if the solar variations stay within 0.1 % during this operation, which takes no more than 30 s. Levelling of the collector, which is critical for the quality of the data, is adjusted and maintained by the operator during the measurements using an electronic inclinometer fitted adjacent to the light collector. The accuracy and stability are usually better than 0.2^∘. Sensor height is about 1 m. The processing of the raw spectra to compute albedo is detailed in Picard et al. (2016b). Repeatability of the measurements in clear-sky conditions is better than 1 %. In addition, the diffuse-to-total ratio is measured by first recording the total downwelling flux, as for the albedo measurement, and second the diffuse downwelling flux by shading the collector from the direct sun using a thin black metallic strip fixed to the arm. The ratio is calculated following the same processing steps as for the albedo. Terrain slope is measured after the Solalb acquisition using a 3 m long and 5 cm wide bar fitted with an electronic inclinometer. In 2018 when Solalb data were obtained, the greatest slope was sought by rotating the bar on the surface until the maximum inclination was found. The azimuth of the bar, giving the aspect of the slope, was measured with a handheld compass. In 2019, at the Autosolexs location, the slope was measured at the end of the season only, by taking two inclination measurements with the bar lying in the north–south and east–west directions. The greatest slope and azimuth were deduced by calculation. Despite the inclinometer intrinsic accuracy (0.1^∘), the precision of the slope and aspect angle obtained with these protocols is relatively mediocre (probably>1^∘) compared to the requirements for the albedo interpretation. It is also worth pointing out that the natural surface is not always a perfect plane, even at the scale of the bar (3 m), but the roughness was not recorded.

Autosolexs has two fixed light collectors pointing upward and downward, which are successively connected to a spectrometer by an optical switch (Picard et al., 2016b) every 12 min. The collectors and the spectrometer have the same specifications as those in Solalb. Sensor height is about 2 m. Data processing, detailed in Picard et al. (2016b), also follows the same steps as for Solalb acquisitions, except that an additional cross-calibration step is required to account for the slightly different responses of the two collectors. In addition to albedo, Autosolexs automatically records the diffuse irradiance using a third light collector shaded by a small rod following the course of the sun. Data used here were acquired on 23 March 2018 in clear-sky conditions.

4.1 Theoretical analysis of the apparent albedo formulations

4.1.1 Quantitative impact of slope on apparent albedo

The impact of the slope is studied considering both 100 % diffuse radiation (overcast conditions) and 100 % direct radiation for various zenith solar angles θ_i. In all cases, the sun is located to the south. For an intrinsic direct albedo of 0.8, Fig. 3 shows the impact of the slope from −35^∘ (north-facing) to 35^∘ (south-facing). With direct illumination, the impact of the slope is considerable and greatly increases with the solar zenith angle. For instance a sun-facing slope of 10^∘ affects the albedo by +0.04, +0.13, and +0.42, at SZAs of 20, 45, and 75^∘ respectively. Neglecting the slope effect leads to detectable albedo errors of 0.01 or larger for slope inclinations larger than 2^∘ at a SZA of 20^∘. This inclination value drops to 0.7^∘ at a SZA of 45^∘ and to 0.3^∘ at a SZA of 75^∘. Such slopes are very small and barely visible to the eye in the field, yet have a detectable effect on the albedo. At SZA=70^∘, which is typical of winter at mid-latitudes, the apparent albedo ranges from 0 to 1.6 over the investigated slope range, a 2-fold variation with respect to the flat albedo value. At SZA=45^∘, which is a typical angle during the melt period in many regions, the apparent albedo ranges between 0.25 and 1.1 for the various formulations. Even when the sun is high (SZA=20^∘), albedo varies significantly from 0.4 to 0.9. Apparent albedo higher than 1.0 for the south-facing slopes may be surprising at a first glance, but it is mainly the consequence of the K factor which accounts for the higher interception probability of the sun beam by these slopes facing the sun compared to horizontal surfaces. It is clear that apparent albedo (being higher or even lower than 1) must not be used for energy budget calculations as is commonly done for a flat terrain.

Figure 3Apparent albedo as a function of slope computed for various formulation and different illumination conditions. The flat albedo is here fixed to 0.8 and the no diffuse radiation is considered.

Download

The differences between the scenarios appear on the graph starting from about a 10^∘ slope (both north- and south-facing). For a slope of 35^∘, the maximum difference is 0.2 for SZA=45^∘ and 0.4 for SZA=70^∘. The smallest apparent albedo in south-facing slopes is for surrounding dark surfaces and the mid-slope position of the sensor, because the downward-looking sensor has a deficit of incoming radiation due to the dark surface, while the upward-looking sensor has an excess due to downward reflections from the upper slope. The opposite is estimated with surrounding snow and a sensor located near the top of the slope, because the downward-looking sensor receives additional illumination from the horizontal area while the upward-looking sensor is not affected by the slope.

The slope effect is largest under direct illumination, which occurs most in the near-infrared domain under clear-sky conditions. Under diffuse radiation (overcast conditions or in the blue and UV ranges) the slope effect is null for the small-slope formulation (bottom right panel in Fig. 3) and very weak when the surrounding surfaces are covered by snow and leads to a decreased albedo for a dark neighbourhood, but this decrease is generally smaller than the impact under direct illumination.

The dependence of the albedo variations with slope on the type of illumination is of practical interest to understand spectral albedo measured in natural conditions, which is addressed in the next section.

4.1.2 Spectral shape of the apparent albedo on a slope

Under clear-sky or partially cloudy conditions, the proportion of direct and diffuse incident radiation varies as a function of the wavelength, and given the contrasted response to the slope between direct and diffuse albedo the shape of measured albedo spectra is distorted over slope. Figure 4 illustrates this distortion by showing theoretical apparent albedo spectra for various slopes under clear-sky conditions. The calculation here assumes pure Rayleigh diffusion, with a given diffuse-to-total ratio modelled as $r_{\mathit{\lambda }}=({\mathit{\lambda }}_{\mathrm{0}}/\mathit{\lambda }{)}^n$ with n=4 and λ₀=350 nm. The snowpack is considered to be homogeneous and infinitely deep. The specific surface area (SSA) is set to 20 m² kg⁻¹, a typical value for near-surface snow in Alpine regions in winter (Dumont et al., 2017) and on the Antarctic Plateau in late summer (Libois et al., 2015). The snow is free of any impurities. To relate the direct and diffuse albedo to snow physical properties, we rely on the asymptotic radiative transfer theory (Eq. 21), where

$$\begin{array}{}\text{(30)}& a_{\mathit{\lambda }}=\mathrm{16}{\displaystyle \frac{\mathrm{4}\mathit{\pi }{n}_{\mathit{\lambda }}^{\prime \prime }}{\mathit{\lambda }}}{\displaystyle \frac{B}{\mathrm{3}(\mathrm{1}-g)}}{\displaystyle \frac{\mathrm{1}}{{\mathit{\rho }}_{\text{ice}}\text{SSA}}},\end{array}$$

with the ice permittivity imaginary part $n_{\mathit{\lambda }}^{\prime \prime }$ taken from Picard et al. (2016a) and the ice density ${\mathit{\rho }}_{\text{ice}}=\mathrm{917}\mathrm{kg}{\mathrm{m}}^{-\mathrm{3}}$ and the shape factors B=1.6 and g=0.845 according to Libois et al. (2014).

Figure 4Spectra of apparent albedo for various slopes under blue-sky conditions (the diffuse irradiance decreases as a power of 4 of the wavelength). The grey curves are calculated with the small-slope approximation.

Download

The flat surface case (second panel in Fig. 4) shows a typical albedo spectrum (Warren and Wiscombe, 1980) with a nearly constant value in the blue and green range (400–550 nm) and a decreasing trend at longer wavelengths attributed to the increasing ice absorption (Warren and Brandt, 2008; Picard et al., 2016a). The absorption feature at 1030 nm is more marked than that at shorter wavelengths (800 and 890 nm) and leads to a local minimum which is often used to infer grain size metrics (Painter et al., 2007; Kokhanovsky et al., 2019). Because we have chosen θ_i=45^∘, the direct and diffuse components (dash grey curves) are similar and as a consequence are overlaid by the apparent albedo spectrum (blue curve) that is a weighted proportion of both components. For other θ_i angles, despite a potentially larger difference between the diffuse and direct albedo, one would observe the same plateau in the blue and green range with a value close to 1 for flat terrain.

When the terrain is not flat, the shape of the albedo spectrum is affected and not only depends on the ice absorption variations but also on the proportion of direct and diffuse illumination. In the red and infrared range (>700 nm), because the illumination is mostly direct, we observe that the apparent albedo is close to the direct albedo and is therefore very sensitive to the slope as shown in the previous section. It is lower for a north-facing slope compared to the albedo over a flat terrain and is larger for a south-facing slope. In the blue range (close to 400 nm) where scattering by the atmosphere is significant, the apparent albedo spectrum lies between the direct and the diffuse albedo curves, and, because the diffuse albedo is always close to 1 for small slopes, the apparent albedo also tends to 1. In the intermediate range, between the blue and the red ranges, the variations in albedo with wavelength are driven by the transition from diffuse-dominant to mostly direct illumination rather than by the variations in the ice absorption. The result is a distortion of the spectra shape that depends on the slope and on the illumination conditions.

The distortion is very different for north- and south-facing slopes. For a north-facing slope, where the direct albedo is reduced by the slope and is thus systematically lower than the diffuse albedo, the apparent albedo shows a marked decrease from the blue to the red ranges (left column in Fig. 4). Conversely, for a south-facing slope, the albedo first increases with wavelength due to the increasing proportion of direct illumination and then decreases because the ice absorption increases rapidly enough to dominate the trend (right column and bottom middle in Fig. 4). The albedo for a south-facing slope thus presents a maximum value in the transition wavelength range, which is always above 1, around 1.008 at 490 nm for a 2^∘ slope, 1.09 at 570 nm, and 1.18 at 600 nm (values taken from the small-slope calculation). These values depend on the illumination conditions and snow characteristics, as lower n, λ₀, θ_i, or SSA lead to a lower maximum. The strongest distortion of the spectra is observed for steep slopes, low sun elevation, low aerosol content, and high altitude. It is noteworthy that all the approximations accounting for the slope effect yield the same general distortion; only small differences appear from about 10^∘ and become significant from about 20^∘, as shown in the previous section and in Fig. 3.

Figure 5Measured, calculated, and corrected albedo for seven acquisitions (rows) taken in different terrain configurations. The first column compares apparent albedo calculated from the theory (small slope and SM case) to measured albedo. The second column compares corrected albedo using measured slope parameters to intrinsic diffuse albedo calculated for a flat surface using measured SSA (grey). The third column is similar to the second column, except that the measured albedo is corrected without using measured slope parameters but by assuming clean snow. The second and third columns also show measured albedo to highlight the change due to the correction.

Download

4.2 Comparison between theoretical calculations and albedo measurements

A set of Solalb measurements acquired in winter 2018 on clean and dry snow has been selected to cover a wide range of slope inclinations and orientations. The comparison with the theoretical apparent albedo using measured slopes, SSA, and diffuse-to-total ratio is shown in Fig. 5 (left column, sorted by increasing slope) for the small-slope approximation and the SM case. The latter was chosen among the four large-slope cases because the surroundings were fully covered by snow and the measurements were taken mid-slope. The agreement is variable, but, in general, the measured (apparent) albedo spectra are clearly affected by the slope as predicted by the theory. All the cases show that the measured albedo is either lower or higher than 1 according to the slope aspect relative to the sun azimuth; that is, for cases 1, 2, 3, 5, and 7 that are opposite to the sun, the albedo is lower than 1 and features a decreasing trend in the blue–yellow range, while for cases 4 and 6 facing the sun, the albedo is higher than 1 and features a maximum in the green–yellow range. The root-mean-square error (RMSE) is in the range of 0.015–0.04, the worst cases being the ones with the largest slope. The bias (calculated minus measured albedo) is variable between about −0.02 and 0.02 except for the largest-slope case (case 7) where it reaches 0.04. Lastly we note that the difference between the two approximations (small slope and SM case) is undetectable in all but case 7 with an 18^∘ slope, and even in this case the difference is much smaller than the discrepancies between the measured and calculated albedo. This indicates that the uncertainties in the albedo measurements and the ancillary data required to perform the calculation (slope, SSA, and diffuse-to-total ratio) are larger than the theoretical formulation differences and therefore that improving the measurement accuracy is a higher priority than developing more advanced theories. As a corollary, using the simple small-slope approximation is probably sufficient for most applications.

Figure 6 shows a similar calculation for the Autosolexs measurements taken on 23 March 2018, a day with continuous clear-sky conditions. Since not all the measured and calculated spectra can be shown (acquisition every 12 min), we selected spectra at a few representative hours (panel a) and albedo time series at a few representative wavelengths (panel b). The slope estimated from a 50 cm resolution digital elevation model (DEM) of the Lautaret area gives θ_n=4.5^∘ and ϕ_n=165^∘. However, the calculation using these angles disagrees with the observations, so that we increased the slope inclination up to θ_n=7.5^∘ to obtain the agreement shown in Fig. 6. A possible explanation is the difference between the snow surface and the DEM that was acquired over bare ground, but we did not perform in situ slope and aspect measurements in the footprint of the sensor, to avoid snow disturbance. This issue emphasizes the difficulty to acquire in situ slope measurements, which are valuable to perform albedo corrections or interpretations. UAV and laser scanners are possible tools to acquire an accurate snow digital surface model in a non-invasive way (Deems et al., 2013; Bühler et al., 2016). In addition we choose SSA=30 m² kg⁻¹ to obtain an agreement in the near-infrared range (1030 nm). This value falls within the range of SSA measured a few days earlier in the surroundings (14 March, Fig. 5).

Figure 6Measured (plain) and simulated (dashed) albedo acquired on 23 March 2018. (a) Albedo spectra acquired at four selected hours during the day; (b) albedo as a function of time for three selected wavelengths.

Download

With the slope being south-south-east, most albedo spectra are affected by sun-facing distortions, i.e. a maximum higher than 1 in the visible range and a concave spectral shape, which is also predicted by the calculation. A greater distortion is observed in the morning when the sun is facing the slope, while the spectrum taken just before the local sunset is almost unaffected. The agreement between the measurements and the theory is overall good even though the adjustment of the slope inclination has a great influence on such a result. The largest discrepancy between measurements and calculations is found at the end of the day, in the near-infrared range, suggesting that snow metamorphism during the day might have decreased the specific surface area. Additional simulations with a SSA of 22 m² kg⁻¹ (dotted line) solve this discrepancy at the end of the day.

4.3 Albedo correction

The correction methods described in Sect. 3 are applied to the measurements used in the previous section, first using the individual measurements taken with Solalb and second the time series of measurements at a single location taken with Autosolexs.

4.3.2 Albedo correction for diurnal cycle measurements with unknown slope parameters

Figure 7 shows Autosolexs albedo spectra corrected using the diurnal time evolution on 23 March 2018 with the unconstrained (blue) and constrained (orange) methods described in Sect. 3.3. All the spectra measured between 08:30 and 16:00 UTC (grey curves) are used as inputs of the methods. Since no in situ SSA were measured near the albedometer it is not possible to compute the expected diffuse albedo that the correction methods are expected to approach. However, fully overcast conditions occurred 4 d later, and, assuming that the near-surface snow did not change during this lapse of time, the albedo measured then (green curves) provides a reference diffuse albedo to assess the performance of the correction.

Figure 7Albedo spectra measured with Autosolexs (grey) and corrected (orange, blue) on 23 March 2018. Measured albedo during overcast conditions on 2 March 2018 are also shown (green).

Download

The albedo spectrum corrected without constraint (blue curve) shows an improvement compared to the measured albedo spectra (grey). The albedo value is more constant in the visible range and closer to 1 in the blue–green range. Nevertheless, it is still higher than 1 and slightly concave, which is a clear sign of insufficient correction. Furthermore, the method produces mixed results depending on the conditions (e.g. taking a subset of the inputs, results not shown). This unstable behaviour comes from the fact that the only constraint to infer the slope parameters is provided by the temporal variations induced by the course of the sun. In turn, any small uncertainty that affects the angular behaviour of the sensor (e.g. imperfect levelling or imperfect cosine response), of the snow surface albedo (e.g. roughness), or of the underlying theory (mainly Eq. 23) may have significant consequences. In our case, the method appears unsuitable for radiative forcing calculations, which require albedo at least lower than 1, but it could be helpful when only wavelength band ratios are necessary (Mary et al., 2013). The case of small slopes may also be more favourable and remains to be explored.

Assuming an impurity-free surface brings a strong constraint on the albedo in the visible, and, if it is true, precisely provides what was missing in the unconstrained method. The constrained method performs well in our case (orange versus green curves, in Fig. 7), especially in the visible range. The slightly higher value in the near-infrared range is not necessarily an error, as it could reflect a decrease in SSA during the 4 d between the clear-sky measurements (23 March) and overcast measurements (27 March). Although the correction method only constrains the albedo value in the visible range, the shape of the spectrum is almost perfectly corrected over the whole spectral range, and the spectrum is smooth (which could have been added as a constraint, but was not). The smoothing comes from the fact that all the measured spectra are effectively used in the estimation process, cancelling the incoherent noise visible in the measurements (grey curves).

Along with the corrected albedo, both methods provide estimates of the two slope angles, which was not possible with the methods using a single acquisition. We obtain a slope of 3^∘ with the unconstrained method. This low value explains why the corrected albedo is still higher than 1 at some wavelengths and still has the traits of a sun-facing slope. In contrast, the constrained method estimates a slope of 7.6^∘ and an aspect of 157^∘ and achieves a good correction. Only the aspect agrees with the DEM (bare ground) of the area.

5.1 Apparent albedo and energy conservation

The theoretical developments and the measured spectra presented throughout this paper make it clear that the “apparent albedo” of a slope is not an albedo sensu stricto and must not be used to compute the energy absorbed by the surface with the usual equation $(\mathrm{1}-\mathit{\alpha })E^{\downarrow }$. This is obvious when the apparent albedo value is higher than 1, because the calculated absorption would then become negative, but this statement holds true even for values lower than 1. The reason is that the apparent albedo is not a purely reflective property of the surface; it also contains the factor converting the energy received on a flat surface to that received by the slope (K), as well as other terms in the case of large slopes. Only the reflective property is subject to energy conservation; K is not. A correct way to calculate the absorption A_λ per unit of flat ground surface area is

which requires more information than just the apparent albedo and the irradiance. It is thereby inapplicable in many situations such as when only the spectral albedo is measured. This equation respects the energy conservation principle even if K>1, because the energy absorbed on the slope is always lower than the irradiance ($E_{\mathit{\lambda }}^{\mathrm{sun}}+E_{\mathit{\lambda }}^{\mathrm{sky}}$), owing to Kcos θ_i<1.

It is worth recalling that measured broadband albedo is similarly subject to the slope effect and that the usual equation $(\mathrm{1}-\mathit{\alpha }){\mathrm{SW}}^{\downarrow }$ is also invalid on a slope when α is measured with horizontal sensors, even though the measured value is lower than 1. Furthermore, the slope effect is impossible to detect from a single acquisition of broadband albedo.

Ideally, the term “albedo” should not be used to refer to the ratio of upwelling and downwelling light fluxes over a slope, but in practice, we recommend at least considering this issue, systematically using the terms “apparent” or “measured”.

5.2 Albedo accuracy and slope effect

All the results presented in Sect. 4 highlight the large and complex sensitivity of the apparent albedo to slope. If one targets an albedo measurement accuracy of 0.01 (Picard et al., 2016b) – for instance to detect small amounts of impurities (Warren, 2013), to detect multiannual trends (Dumont et al., 2014), or to close the surface radiative budget within 4–8 W m⁻² – the error on the slope inclination should remain below 0.6^∘ (considering the sun at a 45^∘ zenith angle and not including aspect error, a very favourable case). This underlines the high level of measurement quality requirements.

Improvement of instruments (equipped with an inclinometer) and protocols (systematic acquisition of diffuse-to-total ratio and slope inclination) is a response to these requirements. Another response is the ad hoc correction of the measured albedo. In a first attempt, it was common in our community to normalize the measured albedo to 1 (or 0.98) in the visible range (e.g. Picard et al., 2016b) because it helped overcome not only the slope effect but also other common problems, such as the cross calibration of the upwelling and downwelling channels or the poor cosine response of the sensors (Zibordi and Bulgarelli, 2007; Picard et al., 2016b), or eliminate operator shadows (Carmagnola et al., 2013). However, this is an error-prone practice, because the spectrum of the scaled albedo on a slope looks similar to the albedo of dirty snow surface in the case of sun-facing slopes, as illustrated in Fig. 8. The convex shape in the visible domain is indeed a common feature of both flat dirty snow and pristine snow slopes. For instance a pristine snow surface with a small slope of 2^∘ has a similar albedo to a flat dirty snow surface with 61 ng g⁻¹ of black carbon, if the apparent albedo measured on the slope is scaled down by 3 % only (blue curve in Fig. 8), so that its maximum is equal to 0.98.

Figure 8Example of scaled spectrum of pristine snow affected by a small slope (2^∘) and comparison to dirty snow spectrum with fitted SSA and black carbon (BC) content.

Download

We thereby recommend not applying rough scaling but instead using a proper correction method as described in this paper. Nevertheless, the risk of over or undercorrections also exists and may result in confusion with dirty snow. For instance overcorrecting measurements taken over pristine snow on a slope facing away from the sun result in a spectrum with a convex shape characteristic of dirty snow. Our conclusion is that small slopes, impurity content and calibration errors, sensor angular response, and operator shadows are interrelated, and only a global assessment of the main error sources and a consistent treatment of these sources can lead to properly corrected albedo spectra.

5.3 The correction methods and their underlying assumptions

The results show that the correction of the slope effect is possible in the domain of validity of small slopes (≲15^∘), but the method to be applied depends on the available information about the actual illumination and slope and/or some assumptions on the snow impurity content. Throughout the paper, we have assumed a known diffuse-to-total ratio, either because it was measured at the same level as the albedo measurements or as a fallback because it was calculated above the mountain ridge with an atmospheric radiative transfer model. Knowing this ratio is absolutely required to correct albedo in the visible domain, except in the very particular case of a solar zenith angle close to the effective angle of the diffuse illumination (in the AART formulation, this occurs when n(θ)≈1 and is usually 45–50^∘). In the near-infrared domain under clear-sky conditions, the diffuse illumination can be neglected, and, if the variation in SSA can also be neglected, a correction without the diffuse-to-total ratio is feasible.

Here we used diffuse-to-total ratio measurements acquired at almost the same time as the albedo acquisitions, which is a very favourable case. Weiser et al. (2016) use nearby measurements of the ratio but considered a constant value for all clear-sky days. The solution to use an atmospheric model is attractive when no measurements are available. Tuzet et al. (2020) obtained consistent results using the SBDART model (Ricchiazzi et al., 1998) parameterized with a generic mid-latitude winter atmosphere and only considering solar zenith angle variations, under the condition that clear-sky conditions are separated from overcast conditions. Dumont et al. (2017) applied the same model but used actual atmospheric profiles obtained from a meteorological analysis and cloud optical depth adjusted using shortwave in situ measurements. The trade-off between costs and benefits of using such a complex setting and more generally the impact of the accuracy of the diffuse-to-total ratio remain to be assessed.

When a single albedo acquisition is available, the most favourable case is when the slope angles (inclination and aspect) or the slope parameter (K or the local solar zenith angle) are precisely known. The correction then requires no assumption on the snow surface (works for clean snow and snow covers with any impurity content). The main required assumption is the angular dependence of the direct albedo (Eq. 23) which depends on the theory used (here AART) and on the smoothness of the surface. However, our results show that the quality of the correction is variable and we suspect the slope parameters' accuracy to be an issue (Fig. 5, second column). Assuming that snow is clean strongly constrains the correction and relaxes the requirement of knowing the slope parameters (Fig. 5, third columns). When the SSA is known with precision, it should also be possible to devise a correction method without the slope parameters (not covered in this paper). At last, the case of dirty snow, unknown SSA, and unknown slope parameters cannot be solved with a single acquisition.

When multiple albedo acquisitions at the same location are available for a wide range of sun positions (e.g. time series of albedo), a correction method exists assuming constant snow properties throughout the acquisitions but without the requirement of assuming clean snow or a known SSA (Eq. 29). The variations in the distortion of the spectrum carry the signature of the slope effect, enabling in principle the recovery of both slope inclination and aspect. The method should best work when the sun azimuth covers a wide range such as in the polar regions during the summer. In mid-latitude regions the range is more limited (less than 180^∘), which may explain the mediocre results we obtained with this method (Fig. 7). Another possible explanation is that SSA was varying at the end of the day as suggested in Sect. 4.2 and Fig. 6. As observed for the single-acquisition case, assuming clean snow provides a strong constraint, and the correction is then excellent (Fig. 7). Although this strong constraint holds most of the time in winter, in spring at the end of the season snow is rarely clean in mountains, and, as discussed in Sect. 5.2, small slopes and small amounts of impurities can have similar signatures.

The correction method proposed by Dumont et al. (2017), applicable to time series of albedo over dirty snow surfaces, makes a stronger assumption to untangle the effect of slope and impurities. It assumes a known analytical formula for the shape of the spectrum, which applies when the type of impurities is perfectly known (e.g. BC or dust of known origin, or algae), and the absorption spectrum can be prescribed in the AART theory yielding a relationship between albedo, SSA, and impurity concentrations. The joint optimization of slope angles, SSA, and impurity concentration is efficient as demonstrated at Col de Porte (Dumont et al., 2017) and Col du Lautaret (Tuzet et al., 2020). If these strong assumptions are acceptable, and depending on the final goal of the correction, this method is certainly the most efficient.

Another important assumption used in the present paper is the perfect levelling of the sensors. It applies well to our measurements because the inclination angles are systematically recorded and proved to be stable within 0.2^∘. Because the effect of the tilt is of the same order as of the slope (in K, Eq. 18, the denominator is calculated with Eq. 2 using the tilt angles instead of the slope angles), this accuracy is sufficient. The method in Weiser et al. (2016) relaxes this assumption and the need for extra measurements. It uses the diurnal cycle of reflected radiation instead of the albedo to infer the slope angles, while the sensor tilt is retrieved using the incident radiation cycle.

The wide range of possible assumptions shows that many methods can be valuable depending on the conditions. Further work should perform a more systematic comparison and exploration of the sensitivity to input uncertainties of each method.