*2.7. Temperature Data*

The monthly mean temperatures from 2006–2010 were based on 1 × 1 km grids of daily (24-hr) mean air temperatures (2 m) created from three-dimensional spatial interpolation of air temperatures observed at meteorological stations distributed throughout Norway [77]. The accuracy of the gridded daily temperature product (SeNorge 2.0) is ~1 ◦C based on the "leave-one-out" cross validation score [77]. Temperature data were re-projected and downscaled to the nominal resolution of the MODIS products using a nearest neighbor interpolation method. The same temporal synchronization procedure that was applied to the snow cover data was applied to the temperature data prior to monthly aggregation.

Figure 2 provides an overview of the climate predictor space of the model fitting domain associated with the post-processed dependent variable (albedo) dataset.

**Figure 2.** Characteristics of the monthly mean climate predictor dataset used in the model fitting exercise: (**Left**) Statistics; (**Right**) Density scatter. "IQR" = interquartile range; "NDSI" = Normalized Difference Snow Index; and "SC" = Snow cover.

Monthly mean air temperatures rarely fell below −10 ◦C, with a median close to the mean of ~6 ◦C (Figure 2, left panel). Over 75% of the monthly mean snow cover (*SC*) retrievals had values less than 44%, with a median of 0% and mean of ~19%. It is important to note that the highest *SC* values were not necessarily found in the coldest months but in months where the mean *T* was around 2–4 ◦C. These temperatures are characteristic of late winter/early spring when the snowpack is near its deepest (i.e., when a larger amount of short-statured vegetation or other landscape features are buried in snow) [78].

## *2.8. Endmember Data Processing*

Prior to calculating the endmember fractions (*e fi*) corresponding to each MODIS product grid cell, the original forest area of the AR5 land cover product was replaced by the updated area of the SR16 product. It is becoming increasingly understood that the effective spatial resolution of the MCD43A BRDF/Albedo product differs from its nominal resolution of ~500 m [15,68,79]. Recently, Campagnolo et al. [80] applied point spread functions to quantify the effective spatial resolution of various MODIS and other optical satellite remote sensing products. For the v6 MCD43A BRDF/Albedo product, they reported a median effective resolution of 833 m along the east–west transect and 618 m along the north–south transect.

Endmember fractions were computed at both the nominal (500 m × 500 m) and the effective (618 m × 833 m) resolutions before model fitting was executed for both resolutions. Similar to Hovi et al. [81], we apply an elliptical point spread function modeled with a Gaussian distribution, which was defined such that 75% of the signal was assumed to originate from within an ellipse having a diameter of 618 m in the Y (north–south) and 833 m in the X (east–west). Forest structural predictors (Section 2.4) were computed as the weighted means within each pixel using the point spread function as weights, which is illustrated conceptually in Figure S5 of the Supporting Information.

**Figure 3.** Workflow schematic of the linear unmixing model fitting procedure.

The general workflow of the entire linear unmixing (model fitting) procedure is illustrated in Figure 3. Figure S1 of the Supporting Information provides an overview of the distribution of the quality-filtered and temporally-synchronized observational records employed in model fitting and validation by land cover (endmember) type and month.
