Random forests as a tool to understand the snow depth distribution and its evolution in mountain areas

The small scale distribution of the snowpack in mountain areas is highly heterogeneous, and is mainly controlled by the interactions between the atmosphere and local topography. However, the influence of different terrain features in controlling variations in the snow distribution depends on the characteristics of the study area. As this leads to uncertainties in high spatial resolution snowpack simulations, a deeper understanding of the role of terrain features on the small scale distribution of snow depth is required. This study applied random forest algorithms to investigate the temporal evolution of snow depth in complex alpine terrain using as predictors various topographical variables and in situ snow depth observations at a single location. The high spatial resolution (1 m x 1 m) snow depth distribution database used in training and evaluating the random forests was derived from terrestrial laser scanner (TLS) devices at three study sites, in the French Alps (2 sites) and the Spanish Pyrenees (1 site). The results show the major importance of two topographic variables, the topographic position index and the maximum upwind slope parameter. For these variables the search distances and directions depended on the characteristics of each site and the TLS acquisition date, but are consistent across sites and are tightly related to main wind directions. The weight of the different topographic variables on explaining snow distribution evolves while major snow accumulation events still take place and minor changes are observed after reaching the annual snow accumulation peak. Random forests have demonstrated good performance when predicting snow distribution for the sites included in the training set with R2 values ranging from 0.82 to 0.94 and mean absolute errors always below 0.4 m. Oppositely, this algorithm failed when used to predict snow distribution for sites not included in the training set, with mean absolute errors above 0.8 m.

J. Revuelto was supported by a Post-doctoral Fellowship of the AXA research fund (le Post-Doctorant Jesús Revuelto est bénéficiaire d’une bourse postdoctorale du Fonds AXA pour la Recherchem Ref: CNRM 3.2.01/17). IGE and CNRM/CEN are part of Labex OSUG@2020. This work was partly supported by the French national programme LEFE/INSU and an ANR JCJC EBONI grant (ANR-16-CE01-0006). This research was also supported by Lautaret Garden-UMS 3370 (Univ. Grenoble Alpes, CNRS, SAJF, 38000 Grenoble, France), a member of AnaEE-France (ANR-11-INBS-0001AnaEE-Services, Investissements d’Avenir frame) and of the eLTER-Europe network (Univ. Grenoble Alpes, CNRS, LSTER Zone Atelier Alpes, 38000 Grenoble, France). Authors do not have any conflict of interest. We would like to thanks the support of the technical and scientific staff of the CEN (Centre d’Etudes de la Neige) the IGE (Institut de Géosciences de l’Environnement) and the INRAE du Grenoble. Especially thanks to Vincent Vionnet, Yannick Deliot, Gilbert Guyomarc’h and Aymeric Richard for their effort to acquire and process TLS observations at Col du Lac Blanc. Similarly we are very grateful to Laurent Arnaud, Emmanuel Thibert and Ghislain Picard for their support in Col du Lautaret study site. We also thank the two anonymous reviewers and the editor, James P. McNamara, for their kind review of the manuscript.

Peer reviewed