Machine Learning Based Imputation of Mountain Snowpack Depth within an Operational LiDAR Sampling Framework in Southwest Alberta

Abstract

Airborne LiDAR can support high resolution watershed-scale snow depth mapping that provides the spatial coverage necessary to inform water supply forecasts for mountainous headwaters. This research utilized LIDAR and machine learning to evaluate snow depth drivers and to assess the feasibility of sampling datasets for the spatial imputation of snow depth at the watershed-scale under mid-winter and melt onset conditions. We present a Random Forest based method of extrapolating LiDAR snow depth model values from two flight lines, with insights for future operational use. Models of watershed-scale snow depth developed from spatially constrained flightline training samples correlated with more spatially widespread LiDAR snow depth data but were outperformed by models generated from training data sampled across the entire watershed. Random Forest simulations produced R2 values ranging from 0.41 to 0.61 and RMSE values from 0.7 m to 1.0 m (p < 0.05). By evaluating the performance of snow depth drivers within each simulation, we found that aspect and topographic position index were important drivers regardless of seasonality. LiDAR sampling and machine learning imputation provide a viable framework for substantially reducing the cost of LiDAR-based mountain snowpack monitoring over complex source water regions, but further work is needed to optimize flight path sampling configurations.