This thesis aimed to develop a methodology using Machine Learning (ML) techniques for the interpretation of deep-sea resources. The deep-sea hosts diverse ecosystems and valuable resources, but potential environmental implications, particularly from mining activities, necessitate effective management strategies. Detailed maps of the sea floor are therefore a necessity, yet such maps have to date only been produced based on manual interpretation which is time consuming and subjective. The study focused on assessing the potential of ML methods to map deep-sea features based on photomosaic and bathymetry data in order to take the first steps in developing an automated, objective, and time-saving technique. This thesis’s method accurately identified and classified features like chimneys at the hydrothermal vent fields, providing insights for resource interpretation and conservation. Integrating ML methods into deep-sea resource management is crucial. The methodology enhances understanding of complex techniques, such as Convolutional Neural Networks (CNN) and Object-Based Image Analysis (OBIA) to overcome a seabed characterization. Simultaneously describing the parameters utilised to achieve a meaningful classification. ML algorithms analyze large data volumes, extract patterns, and predict feature distributions, aiding targeted conservation measures and sustainable resource exploitation. The methodology successfully mapped hydrothermal chimneys in two study areas yet producer accuracies (0,7%) were higher than user accuracies (0,64%), indicating that there were other landforms that shared similar features. The methodology also helps assess potential environmental implications of future mining, supporting informed decision-making and mitigation strategies. It serves also as a foundation for future research to aim at overcoming problems related to incomplete spatial coverage, attempt to better utilize shape and spatial parameters within the OBIA refinement, try to identify more background classes for excluding them from the model, etc. |
