The application of machine learning and 3D photogrammetry for cold-water coral habitat classification in the NE Atlantic

Cold-water coral reefs are complex structural habitats that represent one of the most important deep marine ecosystems. As three-dimensional habitats with high structural complexity, they provide ecosystem services that influence species abundance and biodiversity, being indicators of ecosystem health. These habitats are considered hotspots of biodiversity around the globe, especially in cold and deep waters between 50 and 4000 metres depth. Similar to their tropical counterparts, these habitats are subject to several climate and anthropogenic threats. Over the last two decades, research efforts to identify, map and manage these environments have increased along with the advances in data acquisition. Technologies such as remotely underwater vehicles are equipped with high-resolution sensors that generate gigabytes to terabytes of data. However, data analysis methods are being outpaced by acquisition technologies and there is a latency in the extraction of meaningful information from large datasets. Furthermore, the fine-scale heterogeneity promoted by the three-dimensional scleractinian coral branching structure is often overlooked, being reduced to a two-dimensional scale. This thesis explores methods that can advance seabed mapping to further understand cold-water coral reef habitat features in the deep sea considering their natural, three-dimensional structure and posed data analysis demands given the current technologies. The key aims of the research were to: i) develop an unprecedented 3D imaging classification workflow for CWC habitats of Ireland whilst analysing the suitability and transferability of 2D and 3D data to represent these habitats in high-resolution; ii) quantify facies distribution and spatial variability; iii) link image data to processes driving CWC reef development; iv) develop new forms of visualisation of 3D data of underwater environments; v) derive meaningful information from dense optical datasets. Here, CWC reef habitats in the Porcupine Bank Canyon and the Belgica Mound Province, in the Porcupine Seabight, SW of Ireland were reconstructed in 3D using Structure-from-Motion (SfM) photogrammetry. Point clouds, meshes, orthomosaics and digital elevation models (DEMs) were produced at sub-centimetric resolution. Four different classification workflows were developed and analysed, namely: Multiscale Geometrical Classification (MGC); Colour and Geometrical Classification (CGC); Object-Based Image Classification (OBIA) and; Machine Learning Multiclass classification (MLMC). These first three workflows provided a binary (coral, seabed) classification with accuracy ranging 56 to 74% and provided the analysis of the percentage class distribution for each habitat in 2D and 3D. Results show that there is an impact in mapping CWC in 3D and 2D of at least a tenth of order of magnitude. The MLMC method provided a multiclass (live coral, dead coral, coral rubble, and sediments and dropstones) classification of the 3D point cloud which achieved f1 scores of up to 95.1%. DEMs and classification results were used to assess local and regional CWC patterns in relation to terrain features, facies size and facies distribution. Further investigation revealed that CWC are not randomly distributed within CWC reefs, instead their distribution may be driven by local geomorphometric properties. Aiming to raise awareness and facilitate the interaction of humans with deep-water environments, an application for visualisation of 3D models of CWC in mobile phones was developed. This thesis demonstrates how SfM and machine learning can be used to quantify CWC facies and understand CWC reef habitats.

Grantor dc:publisher

University College Cork

Year dc:date.issued

2023

Author dc:creator

• de Oliveira, Larissa Macêdo Cruz

Advisors dc:contributor.advisor

• Wheeler, Andrew
• Lim, Aaron

• Cold-water corals
• 3D data
• Photogrammetry
• Structure-from-Motion
• Submarine canyons
• Computer vision
• Seabed mapping
• North Atlantic
• Machine learning

Statement dc:rights

• © 2023, Larissa Macêdo Cruz de Oliveira.

Licence dc:rights.uri

• https://creativecommons.org/licenses/by-nc-nd/4.0/

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10468/14482

OAI identifier oai:identifier

oai:cora.ucc.ie:10468/14482