Modelling Snow Water Equivalent, A Research Project with GEUS on Greenland Ice Sheet Data

Abstract: Measuring snow water equivalent (SWE) is time-consuming, costly, and at times even simply impossible in remote locations such as on the Greenland ice sheet. This motivates the modelling of SWE from more easily accessible input variables such as snow depth. This study applies both existing modelling and untested machine learning approaches to model SWE using GEUS’ automated weather station data on the Greenland ice sheet. A total of 1,615 observations of SWE and snow depth pairs are selected for specific seasons from 6 ablation sites. The raw data is enhanced and then applied in crossvalidated modelling setups. We find that the out-of-the-box performances of the existing models are comparable to the reported levels in the literature. Additionally, we find that calibration of the existing models improves these accuracies further. The best performing method was found to be the recently introduced Δ-snow model. The machine learning approaches were found to perform just as well as the calibrated existing models. We found support vector regression as the best performing out of the tested machine learning models of support vector regression, XGBoost, and neural network. However, we cannot eliminate the possibility of the other machine learning models performing better when the application is on a larger data set. In general, these findings are encouraging for future modelling of SWE on the Greenland ice sheet. Especially the finding that machine learning models yield comparable performance is encouraging and this suggests it as a viable modelling approach for SWE.