**4. Discussion**

We combined a five-year climatology of the surface albedo, near-surface air temperature, NDSI snow cover and detailed maps of Norwegian land cover and structure to yield a set of empirical models giving estimates of the monthly mean surface albedo as a function of Norwegian land cover classification. Although the endmember fraction and NDSI snow cover alone could explain 55–81% of the variance in the land surface albedo in our model training domain (depending on which albedo and broad band one looks at), the addition of both air temperature and forest structure (volume or biomass) further increased the explained variance to 85–88% (Table 1). We consider this a remarkable outcome given the multiple sources of error inherent in the various land cover products employed in the model training exercise.

For non-forested vegetated land covers, we found strong relationships between air temperature and albedo (all bands), which suggests that tgrowing season phenology was sufficiently synchronized with monthly mean air temperatures. With the exception of croplands ("CRO") and pastures ("PAS"), SW albedo (black-sky) decreased with increasing air temperature, suggesting that it was driven by increased vegetation masking of a higher albedo surface during the growing season. However, for cropland, pastures, and all forest endmembers, the SW albedo increased with increasing air temperature, suggesting either a larger role played by understory vegetation or by increased canopy masking of a lower albedo surface. During the snow season, we found that air temperature and surface albedo were negatively correlated. This is a relationship that is presumably driven by the influence of air temperature on snow metamorphosis and snow physical state [46,87]. When applied outside the training region, we found absolute normalized median errors to be ≤10% for most non-forest endmember models (SW black-sky). The non-forest endmember models on average performed best during spring (MAM) and summer (JJA) where the proportion of the total number of predictions agreeing within 10% of the MCD43A3 retrievals was around 35% and 45%, respectively (Figure S7, Supporting Information). These shares were weighed downward by the persistent positive error for forested peatbogs ("PB-f") and persistent negative error for non-vegetated open areas ("O-nv") whose proportions of the total number of predictions agreeing within 10% of the MCD43A3 retrievals were only 20% and 25% within the validation region, respectively (Figure S7). In general, these two land cover (endmember) types exhibit the largest variation in both geological attributes (exposed mineral composition) and vegetation attributes (vegetation cover fraction). These are two physical attributes important for the surface albedo which were not captured by the models. Furthermore, the model training region contained a disproportionately low share of "O-nv" relative to the validation region (Figure S1), and given the large variation in surface attributes within "O-nv", a larger "O-nv" sample during model fitting would likely have resulted in an improved performance by this endmember model.

Our models were least capable of explaining variance in the NIR albedos, which is unsurprising given the large share of forest cover in the region and the large variation in foliage reflectance in the shortwave infrared (1.5–2.5 μm) band among Fennoscandic tree species [88]. This variation is likely attributed to differences in leaf-level functional traits [19,20,89], and although seasonal trends in air temperature likely explained some of the albedo variance linked to trends in the leaf area [26], such functional controls could not be accounted.

For the forest endmembers, the maximum and minimum albedo predictions under both snow-covered and snow-free conditions (Figure 4) appeared well-constrained when benchmarked to observations reported in other Fennoscandic regions based on satellite remote sensing [32,33,40,41,81]. The dependency of the albedo on forest structure was more difficult to compare quantitatively to the results of other studies since the environmental state conditions (i.e., *SC* and *T*) were often not reported alongside the reported results. However, qualitatively, the general relationship between forest albedo (all endmembers, all bands) and stand volume under snow-free conditions (Figure 4B) appeared be consistent with the regression fits based on the Landsat observations reported in Kuusinen et al. [41]. However, we did not find as strong a dependency as Kuusinen et al. [41] of the VIS broad band albedo on stand total volume (cf. Figure 1 in Kuusinen et al. [41]).

Under snow-covered conditions, the use of an exponential function to model the albedo-structure relationship in forests appeared appropriate when compared to that reported for Finnish forests for age (c.f., Figure 1 in ref. [33]) or biomass (cf. Figure 9A in ref. [40]). However, the shape parameters of our volume (Figure 4A) and biomass models (Figures S2 and S3) seemed to differ from the age-based models of Kuusinen et al. [33] where the asymptotic SW albedo value for DBF (birch) and spruce forests was reached around age 50 in ref. [31]. This is an age that corresponds to a stand volume and biomass in forests of our training domain for which the asymptotic albedo has ye<sup>t</sup> to be realized. The behavior of the SW albedo model for DBF under snow-covered conditions reported in this study appears to be more consistent with the age-based DBF model of Bright et al. [27], which is unsurprising given the partly-overlapping model training domains.
