Date of Award

Fall 2026

Abstract

Accurate estimation of fractional snow-covered area (fSCA) in forested areas is critical for understanding snowmelt dynamics and meltwater inputs to streams, yet snow cover mapping in forested areas is still challenging using ground-based and satellite-based snow observations. This thesis develops and evaluates multi-scale approaches for monitoring fSCA using field cameras, upscales it to watershed scale using unoccupied aerial systems (UAS), and develops a framework to extract snow cover in forests using Sentinel-2 satellite imagery in mixed deciduous and coniferous forests of New Hampshire, USA.

In the first study, machine learning classifiers’ ability to extract fSCA from field camera RGB imagery were evaluated across 11 open-field sites over four winters. Random Forest (RF) and Support Vector Machine (SVM) classifiers outperformed traditional thresholding methods, achieving classification accuracies of 95% and 94%, respectively, compared to 86% for the best thresholding approach. Image normalization substantially improved cross-camera and cross-site generalization, increasing RF accuracy by up to 41% across different instruments. The RF classifier was applied to field cameras deployed in the forested Hubbard Brook Experimental Forest Watershed 3 (WS3), achieving strong agreement with manually labeled fSCA (MAE = 3.5-9.4%, RMSE = 8.8-17.1%, d1 = 0.80-0.93). Viewshed analyses spatially grounded each camera's observation to the forest ground. UAS-derived snow depth maps from 13 lidar surveys were used to construct a Mean Relative Difference (MRD) spatial weighting that upscaled site-scale camera observations to 10 m watershed-scale fSCA maps at 30-minute temporal resolution, resolving accumulation and melt dynamics unattainable from satellite imagery in this environment.

In the second study, snow cover mapping in mixed deciduous forest was improved for Sentinel-2 using field camera fSCA from 19 cameras at WS3 and Kingman Farm across two winters. Three different snow indices were first tested, and the best performing snow index was corrected for the impact of vegetation. Among the tested indices, NDSI modeled with a hyperbolic tangent function performed best (RMSE = 17.9%, MAE = 10.2%). A vegetation correction using NDVI-weighted NDSI improved accuracy (RMSE = 16.4%, MAE = 8.2%), with pre-winter NDVI recommended for operational use as it avoids contamination from snow-intercepted canopy. Basin-scale fSCA derived from the corrected model agreed closely with camera-based watershed estimates (r² = 0.99, RMSE = 3.3%, MAE = 2.3%), though cloud cover limited usable observations to 12 and 15 dates per winter, constraining satellite observations of rapid snow dynamics.

Together, these results demonstrate a physically grounded, multi-scale framework for monitoring fSCA in forested headwater catchments, combining high-temporal-resolution ground observations with satellite-based mapping to characterize seasonal snow cover dynamics at multiple spatial scales.

Document Type

Master's Thesis

First Advisor

Jennifer M Jacobs

Second Advisor

Jennifer M Jcobs

Third Advisor

Fei Han

Department or Program

Civil and Environmental Engineering

Degree Name

Master of Science

Download

Share

COinS