Mapping soil using neural network machine learning and remotely sensed geoscience data; a study in peat
O’Leary, David
O’Leary, David
Loading...
Publication Date
2023-02-17
Type
Thesis
Downloads
Citation
Abstract
Several concepts exist in the discipline of data analysis that are often unknowingly used in geosciences. Specifically, the concepts related to Big Data, Data Assimilation, Digital Soil Mapping, and Exploratory Data Analysis. Recently the availability of large geo-spatial data sets, such as satellite remote sensing, has meant these concepts are appearing more and more in geoscience literature. However, any explanation or explicit understanding of these concepts is often missing. In this thesis, these concepts are highlighted and applied to several large geo-spatial datasets in the context of peatland identification and intra-peatland mapping of physical properties. Peatlands make up ~ 3 % of the land surface globally and account for ~ 10 – 30 % of all soil carbon, making these soils globally important in the carbon cycle. Drained and modified peatlands account for considerable emissions of carbon to the atmosphere. Peatland rehabilitation aims to return these peatlands to their natural state, making them carbon neutral, or even carbon negative in time. Advances in peatland identification and delineation are needed to update national and global peatland inventories and meet national reporting obligations. The geo-spatial datasets used in this thesis are optical satellite and airborne radiometric data. Modern machine learning neural network methods are used to firstly identify peat soil and update peatland boundaries, and secondly to visualise peatland physical property (landcover and soil moisture) variation within a site containing several discontinuous peatlands. The methods developed here follow the principles of data analysis and have applications outside the scope of this thesis.
Funder
Publisher
NUI Galway