Decision tree-based detection of blowing snow events in the European Alps

摘要

Blowing snow processes are crucial in shaping the strongly heterogeneous spatiotemporal distribution of snow and in regulating subsequent snowpack evolution in mountainous terrain. Although empirical formulae and constant threshold wind speeds have been widely used to estimate the occurrence of blowing snow in regions with sparse observations, the scarcity of in situ observations in mountainous regions contrasts with the demands of models for reliable observations at high spatiotemporal resolution. Therefore, these methods struggle to accurately capture the high local variability of blowing snow. This study investigated the potential capability of the decision tree model?(DTM) to detect blowing snow in the European Alps. The DTMs were constructed based on routine meteorological observations (mean wind speed, maximum wind speed, air temperature and relative humidity) and snow measurements (including in situ snow depth observations and satellite-derived products). Twenty repetitions of a random sub-sampling validation test with an optimal size ratio?(0.8) between the training and validation subsets were applied to train and assess the DTMs. Results show that the maximum wind speed contributes most to the classification accuracy, and the inclusion of more predictor variables improves the overall accuracy. However, the spatiotemporal transferability of the DTM might be limited if the divergent distribution of wind speed exists between stations. Although both the site-specific DTMs and site-independent DTM show great ability in detecting blowing snow occurrence and are superior to commonly used empirical parameterizations, specific assessment indicators varied between stations and surface conditions. Events for which blowing snow and snowfall occurred simultaneously were detected the most reliably. Although models failed to fully reproduce the high frequency of local blowing snow events, they have been demonstrated to be a promising approach requiring limited meteorological variables and have the potential to scale to multiple stations across different regions.