• Login
• Admin

Walker, Rachel Z. (2022) A hyperspectral approach to understand the association between PSM (as measured by InSAR data) and vegetation assemblage for a Scottish peatland. MRes thesis, University of Nottingham.

Preview

PDF (Thesis - as examined) - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader Available under Licence Creative Commons Attribution. Download (26MB) | Preview

Abstract

Peatlands are vital ecosystems that store up to a third of terrestrial carbon despite covering 3 % of the land surface; it is therefore important to improve our understanding of these landscapes to enable continuous carbon sequestration as the climate changes. AVIRIS-NG hyperspectral data has the potential to add detail to current understanding from analysis of Peatland Surface Motion (PSM) from InSAR data. PSM is directly linked to vegetation assemblage, erosion rates and land use. Therefore, four sites were chosen; one near natural, two undergoing restoration (starting a decade apart) and one that has been eroded. Machine learning was used to predict Plant Functional Types (PFTs) at each site using fieldwork, satellite imagery and expert knowledge. Exploratory analysis demonstrated that the random forest classifier was better at predicting PFTs in the Flow Country than SVM analysis (using either linear or RBF kernels). The fieldwork was mainly focused on the first restoration site as this site overlapped most with the others and ten PFTs were determined for this location, with an additional three added from fieldwork at the erosion site. Train-test data was created for these 13 PFTs and random forest classifiers applied to the data, the first restoration site underwent additional analysis using the fieldwork-focused specific train-test data to classify the data. The fieldwork focused classification was the most successful with a mean accuracy score of 0.789, with the other mean accuracies ranging from 0.722-0.728, demonstrating the benefits of conducting fieldwork. Within this analysis, the whole dataset was utilised as

well as smaller spectral ranges to determine whether all hyperspectral bands (post pre-processing) need to be used; it was found that the outcomes using the whole dataset were more accurate than the smaller spectral ranges. Additionally, the data was transformed with the original wavelengths, first and second derivatives and continuum removal used to classify the data, with the original and derivative outcomes proving more accurate than continuum removal. Supervised machine learning was much more successful at locating PFTs than the unsupervised k-means cluster analysis; it was concluded that k-means is unsuitable to predict PFT locations. The Peatland Surface Motion (PSM) data was analysed in conjunction with the PFT predictions for the first restoration site using a range of machine learning classification techniques (logistic regression, decision tree, random forest and SVM). Outcomes suggest that there is potential to use the hyperspectral analysis to increase understanding based on PSM outputs, however, further refinement of methods is required to achieve this.

Item Type:	Thesis (University of Nottingham only) (MRes)
Supervisors:	Boyd, Doreen Large, David J.
Keywords:	Peatlands, Hyperspectral, InSAR, Peat Surface Motion
Subjects:	Q Science > QK Botany > QK900 Plant ecology T Technology > TR Photography
Faculties/Schools:	UK Campuses > Faculty of Engineering
Related URLs:	URL URL Type https://github.com/rachelzwalker/Flow_Country_HSI_and_PSM UNSPECIFIED
Item ID:	71860
Depositing User:	Walker, Rachel
Date Deposited:	13 Dec 2022 04:40
Last Modified:	13 Dec 2022 04:40
URI:	https://eprints.nottingham.ac.uk/id/eprint/71860

Actions (Archive Staff Only)

Edit View