We found that all random forest models consistently generated RMSE's smaller than one standard deviation of the test year's snow depth. For any given year's predictions, their RMSE's are generally within 10% of each other. That said, these models need more work, perhaps more feature extraction from the elevation map or other types of data such as weather, before they can be considered reliably predictive. In addition, RMSE's vary significantly from year to year, which potentially points to significant variance in model parameters when trained on different datasets.
As for feature importances, we see that their relative importance remains fairly constant across model scales, across snapshots within a year, across years, and when predicting snow-water equivalent (SWE) instead of snow depth. For example, elevation remains the most important feature, even though its importance decreases rapidly through the spring and summer. Other potentially interesting patterns can be seen, such as the importance of daily irradiation peaking during the spring, presumably when the weather is warming up and snow is starting to melt. Figures \ref{335780} to \ref{457812} below compare feature importances across model scales and snapshot dates.