Identifying Snow-Covered Areas from Unoccupied Aerial Systems (UAS) Visible Imagery and Lidar: A Comparison of Methods

High-resolution imagery from Unoccupied aerial systems (UAS) offers new opportunities for mapping snow cover at fine spatial scales, particularly in regions with ephemeral and variable snow conditions. This study evaluates a range of classification strategies for generating snow-covered area (SCA) maps from UAS RGB imagery and lidar-derived snow depth collected over open areas in southern New Hampshire, USA and offers practical recommendations for producing UAS-derived SCA maps. We tested machine learning (ML), including Maximum Likelihood Estimation (MLE), Random Forest (RF), and Support Vector Machine (SVM), and threshold-based approaches, exploring the influence of input features and training set composition on classification accuracy and generalization. Results show that classifiers using full red-green-blue inputs consistently yield high performance (accuracy, balanced accuracy and F1 score above 0.96). In contrast, ML and thresholding approaches relying solely on the blue band exhibited lower balanced accuracy (0.83 to 0.86). Training with fewer than 12 orthomosaics reduced the reliability and consistency of snow cover classifiers. When fewer flights are possible, UAV flights that collectively capture the full range of snow-covered area (fSCA) between 20 %–60 %, partial melt, and refrozen surfaces, should be prioritized. We conclude that highly accurate snow mapping from routine UAS optical imagery is possible even in landscapes with variable snow cover and ephemeral, shallow snowpacks. In such regions, lidar-derived estimates of snow-covered area were found to be highly sensitive to the selected snow-depth threshold where shallow snowpacks require low thresholds that may approach measurement uncertainty and increase uncertainty in mapped snow extent.