Leveraging Citizen Science, LiDAR, and Machine Learning for Snow Depth Estimation in Complex Terrain Environments

Abstract. The majority of the water supply for many Western US states is derived from seasonal snowmelt in mountainous regions. This study aims to address gaps in basin-scale snowpack modeling by using a multi-step, Gaussian-based machine learning model to generate rapid, high-resolution, snow depth estimates at minimal cost by combining citizen-science snowdepth observations with static LiDAR terrain features collected at a single snow-free date. We focus on reducing personnel danger by modifying the algorithm to minimize the exposure of field sample collectors to avalanche-prone terrain. Using snow observations taken solely within a subbasin (∼9-km²) of a larger basin (∼70-km²), a basin-scale snow depth estimate is modeled for a given date throughout the snow season. Results show that a small number of observations (i.e., 10) within a subbasin can realize snow depth across the greater basin with high accuracy, with a root mean squared error (RMSE) of 0.37 m, and Kling-Gupta efficiency (KGE) of 0.59 when compared to the true snow depth distribution. We test the universality of the algorithm by modeling multiple subbasins of differing spatial characteristics and find similar results. The algorithm shows consistent performance across subbasins with varying spatial characteristics and maintains accuracy even when highrisk avalanche areas are excluded. This method exhibits a potential for citizen-scientist data to safely provide seamless modeled snow depth across different spatial ranges in snow-covered basins.