**5. Conclusions**

We presented a coherent empirical modeling framework based on linear spectral unmixing of optical satellite retrievals of surface albedo. The unmixing framework allowed us to sample a relatively large area within a finite region of Norway containing multiple land cover features spanning large climate gradients. The unmixing approach may be computationally less expensive than the approaches based on homogenous pixel sampling, which often requires a more extensive spatial domain to be sampled in order to allow for a sufficient number of satellite retrievals. The resulting land (or vegetation) cover-dependent models developed and presented here can be applied to create a spatially-explicit surface albedo dataset (or database) for use in regional land-climate research and in climate-oriented land managemen<sup>t</sup> planning in Norway. The models can also facilitate comparisons to land-cover dependent albedo predictions made by land surface (climate) models, which are difficult to evaluate by direct comparisons to satellite retrievals.

The models required just two environmental state predictors: monthly mean air temperature and monthly mean NDSI snow cover (*SC*), where the latter was also provided by optical satellite remote sensing. Given a normalized error threshold of ≤10%, most of the models proved to be accurate when validated against high quality MODIS albedo retrievals in a region outside of the model training region (based on the medians reported in Figure 9). The exception was the performance for forested peat bogs ("Pb-f") and non-vegetated open areas ("O-nv") whose normalized absolute median prediction errors exceeded 10% in all seasons. Although we reported a large range in upper and lower error quartiles for all models during winter months (DJF; Figures 6–9), these larger errors may be deemed acceptable in most modeling contexts given the lower relevance of surface albedo during periods of low solar energy input (such is the case for Norway during DJF). Finally, although the MODIS NDSI-based *SC* can be higher than the 0.75 constraint shown in Figure 4, the higher values in our domain mostly coincided with the end of the snow season when the monthly mean *T* was also higher. We stress that grea<sup>t</sup> caution should be exercised when extrapolating the model behavior outside the extreme *SC* and *T* ranges presented in Figure 2.

**Supplementary Materials:** The following are available online at http://www.mdpi.com/2072-4292/11/7/871/s1: Table S1: Root mean squared errors of candidate models; Table S2: White-sky model parameters for forests with *volume* as structural predictor; Table S3: Black-sky model parameters for forests with *biomass* as structural predictor; Table S4: White-sky model parameters for forests with *biomass* as structural predictor; Table S5: White-sky model parameters for non-forest endmembers; Figure S1: Distribution of MODIS retrievals used in model fitting and validation by month and land cover type; Figure S2: White-sky model behavior for forests with *volume* as structural predictor; Figure S3: Black-sky model behavior for forests with *biomass* as structural predictor; Figure S4: White-sky model behavior for forests with *biomass* as structural predictor; Figure S5: Weighting scheme applied when fitting models to the *effective* spatial resolution; Figure S6: Distribution of Köppen-Geiger climate zones in Norway and within the study domain; Figure S7: Fraction of total predictions have ≤10% normalized absolute error (SW black-sky) by land cover type and season.

**Author Contributions:** R.M.B. conceptualized and designed the study, compiled and processed all data, carried out all modeling, analysis and validation, produced all figures and wrote the original manuscript draft. R.A. edited and commented on the original draft and R.M.B. and R.A. revised the final manuscript.

**Funding:** This research was funded by the Research Council of Norway (Norges Forskningsråd), gran<sup>t</sup> number 255307/E10.

**Conflicts of Interest:** The authors declare no conflict of interest.
