Understanding the snow depth distribution in Alpine terrain using random forests

Abstract

The spatial distribution of snow depths is crucial for addressing various scientific questions in Alpine regions. Information on snow distribution helps to better manage water resources and risks such as snow avalanches or flooding. Snow depth in Switzerland is measured continuously in a network of selected automatic weather stations. In this thesis a Random Forest (RF) algorithm is used to model the snow distribution in the Davos region at different spatial scales. The importance of input parameters such as interpolated snow depth measurements, satellite-based information average snow cover and topographic variables is investigated. The RF-model is trained and validated with drone- and aircraft-based snow depth maps. Variables that provide direct information about the distribution of snow, such as snow cover duration (SCD) or interpolated snow depths from gauging stations, are preferred over topographic variables (elevation, aspect, slope, Topographic Position Index (TPI) and Diurnal Anisotropic Heat (DAH)). To investigate the effect of spatial resolution on RF-model accuracy, different resolutions from 6 m to 100 m were tested. Validation of the RF-model reveals that model accuracy increases with decreasing spatial resolution. The average coefficient of determination (R2) at 6 m and 100 m resolution, being 0.41 and 0.56. In addition, we analyzed the performance of the RF-model on inter-annual and intra-annual time scales. Inter-annual comparison of the snow distribution from the peak of the snow season (in March or April) yields high model accuracy due to similar spatial distribution patterns of snow depth tied to topography. In contrast, an intra-annual comparison results in lower model accuracy due to repeated redistributions during the build-up phase of the snowpack. To test the general validity of the RF-model, we validated the snow model in different regions of Switzerland. The accuracy of the the RF-model varies between a R2 of 0.28 and 0.51 depending on the test region. Similar topographic characteristics of the training and validation region lead to more accurate model results of the snow distribution. The RF approach chosen in this thesis makes a valuable contribution to improved snow distribution modeling. Future area-wide measurements of the evolution of snowpack can contribute significantly to a better understanding and modeling on the intra-annual snow distribution.