Universitetsavisen
Nørregade 10
1165 København K
Tlf: 35 32 28 98 (mon-thurs)
E-mail: uni-avis@adm.ku.dk
PhD thesis defense
PhD thesis defense — Signe Hansen 3 March
Date & Time:
Place:
Aud. C, Department of Geosciences and Natural Resource Management, Øster Voldgade 10, 1350 Copenhagen K
Hosted by:
Geography Section 2, Department of Geosciences and Natural Resource Management, Øster Voldgade 10, 1350 København K
Cost:
Free
Signe Irene Kirstejn Schilling Hansen defends her thesis,
Automatic classification of boulders
Processes and concerns regarding machine learning workflows
Supervisors:
Professor Aart Kroon, IGN
Senior researcher Verner Brandbyge Ernstsen, GEUS
Emeritus Zyad Al-Hamdani, GEUS
Assessment committee:
Professor Christian Winter, Christian-Albrechts-University, Kiel
Scientist Anu Kaskela, Geological Survey of Finland
Professor Thorbjørn J. Andersen (chair), IGN
Summary:
Coastal areas are vulnerable and easily affected by human activity and climate change. To be able to protect these areas, we need mapping of benthic habitats to help identify areas in need of protection. However, these shallow waters are difficult to access by vessel-borne acoustic systems. Shallow water requires smaller vessels, where data collection is both expensive and time-consuming. Topo-bathymetric LiDAR measurements make it possible to cover large areas fast, and this makes it feasible to repeat the measurements for an area and thereby investigate how the area changes over time. Taking advantage of these new possibilities results in frequent and large amounts of data which need to be analysed. To be able to compare climate-related seabed changes from different coastal areas in the world it is important that the data is comparable. The large amount of data, and the requirement of comparability, calls for automated and reproducible data processing and analysis methods.
Stone reefs are among the habitats which need protection. In this thesis topo-bathymetric LiDAR data are used to create machine learning workflows for boulder detection. The overall aim is to improve the understanding of the processes and considerations involved, when using machine learning for mapping boulders in coastal environments using point cloud data.
Random forest machine learning has been carried out in Matlab in combination with feature selection algorithms. Features derived from intensity and height have been used as input data. A training- and test-set has been created from ortho-photos analysis and ground-truthing. The study has resulted in a workflow recommending separated ground-truthed training- and test sets including seabed diversity representative for the area, in which the machine learning model should work. Boulder detection leads to imbalanced data sets, which can be handled by a combination of assigning a weight score to the minority class, and using only a subsample of the majority class. Knowledge about habitats, seabed structures and transition zones between these can lead to a better choice of input data. Knowledge about special characteristics of different habitats can be used to distinguish between them. The establishing of a library is suggested with knowledge about characteristics of different habitats and the relevant discriminators in combination with small ground truthed data sets. To discriminate between boulder and vegetation is very difficult. The work on this thesis led to a conceptual model of the interrelations between the boulder distribution and geomorphology in Rødsand lagoon.
A digital version of the PhD thesis can be obtained from the PhD secretary Mikala Heckscher at mikala@ign.ku.dk