Using repeated patterns in snow distribution modeling: An Arctic example

[1] Snow distribution patterns are similar from one year to the next because they are largely controlled by the interaction of topography, vegetation, and consistent synoptic weather patterns. On a yearly basis none of these controls changes radically. As a consequence, deep and shallow areas of snow tend to be fixed in space, producing depth differences that may vary in absolute, but not relative, amounts from year to year. While this fact is widely known, the use of patterns in modeling snow cover distribution is limited. Here, on the basis of a training set of nine annual snow depth surveys from a small tundra basin in Alaska, we identify the climatological snow distribution pattern (CSDP). Using this and a few depth measurements, the snow distribution for years that were not included in the training set is predicted and mapped with a near-zero bias and RMSE that ranged from 4.4 to 10.4 cm. The accuracy of this strictly empirical approach to modeling the depth distribution is similar to, or better than, the output from a weather-driven physically based snow model. However, in our view a hybrid approach is best. Ingesting the CSDP into SnowModel, a widely used numerical code that simulates snow processes, the accuracy of the model output is improved by up to 60%. This hybrid approach retains the advantages of running a weather-driven numerical code but adds spatial accuracy currently only obtainable from observed snow patterns. The patterns can be captured in several ways, including aerial photography or satellite remote sensing during snowmelt.