MACHINE LEARNING FOR AUTOMATED RECOGNITION OF SEABED MORPHOLOGIES

The seabed hosts a wide variety of morphological features that record the geological and geomorphological processes shaping continental margins. Their systematic recognition is fundamental for marine geohazard assessment and for the planning of offshore infrastructures. However, conventional mapping relies heavily on manual interpretation of bathymetric data, which is time-consuming, subjective, and difficult to reproduce consistently across large areas. This thesis explores the use of machine learning for the automated recognition of seabed morphological elements from high-resolution bathymetric derivatives. A convolutional architecture based on U-Net is developed and applied to a dataset from the MaGIC Project, covering selected regions of the Italian continental margins. The proposed framework includes dedicated preprocessing, data balancing, and a bidirectional neighborhood-based metric specifically designed to evaluate elongated and discontinuous structures that are poorly captured by standard pixel-wise scores. Three model configurations are tested, multi-class (reduced-class, and binary) to analyze how class granularity affects performance and interpretability. Overall, the proposed framework bridges quantitative image segmentation with geological interpretation, demonstrating that deep learning can effectively support and accelerate seabed mapping, while preserving consistency and reproducibility across extensive marine domains.

The seabed hosts a wide variety of morphological features that record the geological and geomorphological processes shaping continental margins. Their systematic recognition is fundamental for marine geohazard assessment and for the planning of offshore infrastructures. However, conventional mapping relies heavily on manual interpretation of bathymetric data, which is time-consuming, subjective, and difficult to reproduce consistently across large areas. This thesis explores the use of machine learning for the automated recognition of seabed morphological elements from high-resolution bathymetric derivatives. A convolutional architecture based on U-Net is developed and applied to a dataset from the MaGIC Project, covering selected regions of the Italian continental margins. The proposed framework includes dedicated preprocessing, data balancing, and a bidirectional neighborhood-based metric specifically designed to evaluate elongated and discontinuous structures that are poorly captured by standard pixel-wise scores. Three model configurations are tested, multi-class (reduced-class, and binary) to analyze how class granularity affects performance and interpretability. Overall, the proposed framework bridges quantitative image segmentation with geological interpretation, demonstrating that deep learning can effectively support and accelerate seabed mapping, while preserving consistency and reproducibility across extensive marine domains.