Airborne Lidar and Machine Learning Reveal Decreased Snow Depth in Burned Forests

Abstract. Wildfires are increasingly burning higher in elevations well into the seasonal snow zone, altering snow accumulation and melt dynamics. However, limited spatially distributed observations throughout the full snow season have constrained our understanding of how these changes vary across space and time. Here, we assess post-fire snow depth changes across nine basins in California’s Sierra Nevada using a machine learning (ML) algorithm, Extreme Gradient Boosting (XGBoost), trained on 50-m resolution airborne lidar. We develop and apply a novel inferential framework, assuming that the ML algorithm trained on each flight captures the effects of fire on snow depth at the time of acquisition. The median cross-validated (5-fold) RMSE across all acquisitions was 0.23 m. During the accumulation season, the trained ML model predicts smaller post-fire snow depth changes than during the ablation season. Across all 115 acquisitions, 77 % of accumulation-season acquisitions and 98 % of ablation-season acquisitions had a lower basin-wide average predicted snow depth in burned areas compared to unburned areas. Lower elevations (<2,500 m) consistently exhibited smaller, near-zero post-fire snow depth changes compared to higher elevations (>3,250 m). South- and east-facing slopes experienced the largest negative post-fire snow depth changes. These results illustrate a new inferential approach to assessing fire impacts on snow using lidar-derived snow depth and provide insights to snowpack dynamics in burned forests that are novel in their spatial extent and resolution, as well as ability to discern fire impacts throughout the snow season.