FOUR YEARS OF UNMANNED AERIAL SYSTEM IMAGERY REVEALS VEGETATION CHANGE IN A SUB-ARCTIC MIRE DUE TO PERMAFROST THAW

• Main Author: DelGreco, Jessica
• Format: Text
• Language: unknown
• Published: University of New Hampshire Scholars' Repository 2018
• Subjects: climate change; peatland; permafrost thaw; remote sensing; unmanned aerial systems; vegetation change; Geographic information science and geodesy; Environmental science; Arctic; Abisko; Stordalen; Climate change; palsas; permafrost
• Online Access: https://scholars.unh.edu/thesis/1216; https://scholars.unh.edu/cgi/viewcontent.cgi?article=2215&context=thesis

Warming trends in sub-arctic regions have resulted in thawing of permafrost which in turn induces change in vegetation across peatlands both in areal extent and composition. Collapse of palsas (i.e. permafrost plateaus) has also been correlated with increases in methane (CH4) emission to the atmosphere. Vegetation change provides new microenvironments that promote CH4 production and emission, specifically through plant interactions and structure. By quantifying the changes in vegetation at the landscape scale, we will be able to scale the impact of thaw on CH4 emissions in these complex climate-sensitive northern ecosystems. We combine field-based measurements of vegetation composition and Unmanned Aerial Systems (UAS) high resolution (3 cm) imagery to characterize vegetation change in a sub-arctic mire. The objective of this study is to analyze how vegetation from Stordalen Mire, Abisko, Sweden, has changed over time in response to permafrost thaw. At Stordalen Mire, we flew a fixed-wing UAS in July of each of four years, 2014 through 2017, over a 1 km x 0.5 km area. High precision GPS ground control points were used to georeference the imagery. Randomized square-meter plots were measured for vegetation composition and individually classified into one of five vegetation cover types, each representing a different stage of permafrost degradation. Using these training data, each year of imagery was classified by cover type in Google Earth Engine using a Random Forest Classifier. Textural information was extracted from the imagery, which provided additional spatial context information and improved classification accuracy. Twenty five percent of the training data were held back from the classification and used for validation, while the remaining seventy five percent of the training data were used to classify the imagery. The overall classification accuracy for 2014-2017 was 80.6%, 79.1%, 82.0%, and 82.9%, respectively. Percent cover across the landscape was calculated from each classification map and compared ...