Snow-canopy interactions during rain-on-snow events using Random Forest

Interactions between canopy and snowpack during rain-on-snow (ROS) events are still poorly understood, particularly in complex terrain. To gain further insight into these interactions, we present an inventory of 504 ROS events captured at 69 measurement sites in the Southern Sierra Nevada using seven wireless sensor networks. These ground-based wireless sensor networks were designed so they span a range of topographic and canopy conditions, offering a novel dataset with which to investigate the response of snowpack to rain events. At each measurement location, ground-based snow depth, air temperature, relative humidity, and wind speed data were coupled with forest canopy properties derived from high-resolution remotely sensed Lidar data. We characterized the climatology associated with the observed ROS events as well as analyzed if and how the snow depth response to ROS shifted based on canopy and topographic conditions. We found variations in ROS occurrence and snow depth response across different terrain and vegetation characteristics, which challenges the representativeness of traditional operational measurements because these are generally taken in open, flat sites. Uncertainty in meteorological conditions associated with each event makes establishing simple correlations between physiographic variables and snow depth response challenging. Using climate, topography, and vegetation characteristics as features, we thus trained a Random Forest algorithm to predict change in snow depth during ROS events. The model performed well across all validation events (R² = 0.87, RMSE = 3.91 cm, bias = 2.97 cm, Kling-Gupta Efficiency = 0.72), indicating that concurrent meteorological and physiographic conditions provide a basis for a data-driven approach to predict the change of snow depth to ROS events. ROS events occurring in late winter corresponded to the most uncertainty in model predictions. During this transition period between snow accumulation and ablation, conditions in the snowpack shift between subfreezing and isothermal, creating heterogeneous patterns in cold content that may not be consistent across sites and years. This limits the transferability of information from other locations and seasons, making predication more difficult during this period.