Spatial Analysis of Habitat Quality in Semi-Natural Grasslands & Karst Landscapes Using Multi-Spectral Imagery: Case Study in Inis Mor, Ireland

Nicholas Liquori, BA

Master's thesis

Abstract:

Results Based Agri-Environmental Payment Schemes are powerful tools for protecting HNV farmland in Europe. The landscape present on the Aran Islands is one mosaiced by species-rich semi-natural grasslands and exposed limestone bedrock. The presence of rare flora species is owed to long-standing, traditional, low-intensity grazing practices. Caomhnú Árann monitors the farms of those who have enrolled in the program and provides a grade 1-5 ultimately depending on grazing level, presence of indicator species, adequate water provisions for cattle, and scrub maintenance. Unmanned Aerial Vehicle (UAV) acquired multispectral data can aid in the classification of these farmlands and limit the expenditure of time, money, and resources by Caomhnú Árann. Supervised classification by Random Forest Models was used to predict the grazing score of input fields with the highest overall accuracy by a single model of (54.78 %). This accuracy was achieved by the Random Forest Model which used the first 10 dimensions from a Principal Component Analysis as predictors. This model also achieved the highest balanced accuracy of any individual grazing class (class 5 – 84.01%). The variables which underwent the PCA and that built the other random forest models were various zonal statistics of the individual bands from the multispectral imagery, Normalized Difference Vegetation Index (NDVI) and NDVI derived analytical rasters, and a Digital Elevation Model (DEM). This research may have not successfully proven the full employability of a particular classification model as it pertains to the conservation interests of Caomhnú Árann and the Aran Islands, but there is promise in the ability to classify specific grazing scores. Scores 2 and 5 obtained the highest Balanced Accuracies across all 6 Random Forest Models. The results show that a Random Forest Model could be used by Caomhnú Árann in their future research if the collinearity of input data is corrected by performing a Principal Component Analysis, and input data is processed to limit the errors inherent with UAV acquired data.

Keywords

Multi-Spectral Imagery, Semi-Natural Grasslands, Random Forest Classification

Language used: English

Date on which the thesis was submitted / produced: 31. 3. 2022

Thesis defence

• Supervisor: Ing. Vojtěch Barták, Ph.D.

Czech University of Life Sciences Prague

Faculty of Environmental Sciences

Master programme / field:
Landscape Engineering / Landscape Planning