Date of Award

Fall 2026

Abstract

Background: Understanding the impacts of climate change on the Arctic is critical given the sensitivity of its ecosystems. As global temperatures rise and permafrost collapse, significant changes in land cover, hydrology, and carbon dynamics will occur across these vulnerable regions. Because methane emissions are directly influenced by vegetation communities in permafrost peatlands, accurate land cover classification is important for understanding greenhouse-gas (GHG) dynamics. However, these landscapes are highly heterogeneous at fine spatial scales, requiring high-resolution imagery capable of distinguishing simple and complex vegetation and land cover classes relevant to GHG dynamics. Technological advancements in remote sensing offer modern tools for observing, collecting, and analyzing environmental and climate data. Unpiloted Aerial Systems (UAS) equipped with high-resolution sensors generate detailed spatial data products such as orthomosaics, surface models, and vegetation indices, that allow machine learning algorithms and other analytical tools to monitor environmental change and evaluate climate change impacts. Site: Stordalen Mire (68.35°N, 18.8167°E) is a permafrost peatland near the Abisko Scientific Research Station in northern Sweden. As a long-term research site located above the Arctic Circle, this area provides an important setting for studying the effects of permafrost thaw on biogeochemical cycling and trace-gas emissions. Stordalen Mire is characterized by a gradient of permafrost including intact permafrost, thawing palsa, and fully thawed fens. Simple and complex land cover classifications were developed to better represent transitional sites relevant to biogeochemical cycling. Method: A land cover map of Stordalen Mire was produced using a Random Forest (RF) classifier based on UAS thermal, Light Detection and Ranging (LiDAR) data, and True-Color (RGB) imagery, coupled with vegetation ground-survey observations for model development. Results: Both Random Forest (RF) classification models achieved an overall accuracy of approximately 98.8%, with a mean producer’s accuracy exceeding 95% across land cover classes. Field or plot-level data are critical for remote sensing classification, and additional categorical field data may improve training sample quality and reduce misclassification. Sensor fusion approaches should also be explored to improve class separability. This study helps measure change, reveal spatial patterns, and define critical indicators of climate trends. Collectively, these technologies support the development of a high-resolution land cover classification of Stordalen Mire during the summer of 2025, revealing distinct vegetation patterns across diverse palsa areas, transitional thaw-affected zones, and less diverse fen environments. This classification provides a fine-scale spatial baseline for monitoring biodiversity, permafrost thaw, and climate-driven land cover change.

Document Type

Master's Thesis

First Advisor

MICHAEL W PALACE

Second Advisor

RUTH K VARNER

Third Advisor

FRANKLIN SULLIVAN

Department or Program

Earth Sciences

Degree Name

Master of Science

Download

Share

COinS