Title: Improved Snow Distribution Estimates Using a Rapid Response and Photogrammetry System

Abstract: Snow plays a critical role in global hydrology, climate systems, and human activities, particularly in mountainous regions where it is a primary source of freshwater, influences the energy balance, and impacts mobility and commerce. Despite its importance, accurately mapping and predicting snow distribution remains a significant challenge due to the complex spatial and temporal variability of snowpack and the lack of an ideal observation system. This dissertation aims to advance our monitoring capability, understanding of snow distribution, and prediction capacity using Light Detection and Ranging (LiDAR) and photogrammetry techniques.

In chapter 2, I discussed the high-resolution LiDAR-derived datasets of multiple sites in the Western United States that I created, contributing to addressing the scarcity of distributed snow depth data in mountain regions. These datasets, collected during NASA's SnowEx campaigns, provide crucial benchmarks for validating emerging snow monitoring techniques across varied environmental conditions.  In chapter 3, I introduced Ice-road-copters (IRC), an open-source Python toolkit designed to automate the processing of LiDAR and photogrammetry point clouds for snow depth mapping, significantly reducing the manual labor traditionally required for such tasks. This toolkit streamlines noise removal, ground segmentation, DEM coregistration, and raster differencing, facilitating more efficient and consistent production of snow depth maps.

In Chapter 4, I explored the potential of leveraging snow distribution patterns to predict snow depth across diverse mountain environments in the western United States. This involves assessing the repeatability of snow distribution patterns and systematically evaluating how prediction accuracy varies with different pattern types, training data quantities, and spatial scales. Results demonstrate high correlation (r > 0.8) of distribution patterns for snow depths exceeding 0.5 m, while shallow snow conditions during early accumulation or late melt exhibit reduced pattern correlation. Prediction performance is optimized when using temporally consistent patterns—accumulation patterns for pre-peak predictions and ablation patterns for post-peak predictions—yielding mean root mean square errors between 0.2-0.4 m across all sites. Notably, robust predictions can be achieved with as few as 10 observations over a 38 km² area, though prediction confidence improves with increased sampling. Performance degrades with larger spatial extents, with errors approximately doubling when scaling from 38 km² to 3,741 km².

Finally, in Chapter 4, I investigated the integration of LiDAR and photogrammetry for enhanced snow monitoring, demonstrating how combining LiDAR's high accuracy with photogrammetry's cost-efficiency could advance operational snow mapping in mountain watersheds. Results show that both ALS and UAV LiDAR provide similar levels of accuracy when validated against reference measurements, with RMSE values of about 15cm. In contrast, photogrammetry exhibited substantially higher uncertainty that increased with vegetation density—from 27 cm in sparse vegetation to over 1m in dense vegetation—highlighting a critical limitation of this approach for comprehensive watershed monitoring. I proposed an enhanced methodology that combines vegetation masking during photogrammetric processing with gap-filling based on historical LiDAR-derived snow distribution patterns. This integrated approach reduces the RMSE of photogrammetric snow depth maps from 44 cm to 27 cm, demonstrating the potential for synergistic combination of these technologies. 

Advisor: Hans-Peter Marshall

Committee: Shad O'Neel, Ellyn Enderlin, and Rick Forster