Detecting small seamounts in multibeam data using convolutional neural networks

Tobias Ziolkowski^1*Colin Devey¹Agnes Koschmider²

• ¹GEOMAR Helmholtz Center for Ocean Research Kiel, Helmholtz Association of German Research Centres (HZ), Kiel, Germany
• ²University of Bayreuth, Bayreuth, Bavaria, Germany

Seamounts play a crucial role in marine ecosystems, ocean circulation, and plate tectonics, yet most remain unmapped due to limitations in detection methods. While satellite altimetry provides large-scale coverage, its resolution is insufficient for detecting smaller seamounts, necessitating high-resolution multibeam bathymetry. This study introduces a deep-learning-based framework for automated small seamount detection in multibeam bathymetry, combining a CNN-based filtering step with U-Net segmentation to enhance accuracy and efficiency.Using multibeam bathymetric data from the SO305-2 expedition, the proposed approach successfully identified 30 seamounts, many of which were undetectable using satellite altimetry.A hyperparameter optimization study determined the optimal U-Net configuration, achieving a Dice Coefficient of 0.8274 and a Mean IoU of 0.7514. While the model performed well within the training dataset, cross-regional generalization remains challenging, with reduced accuracy observed in areas of highly variable seafloor morphology.The results highlight the limitations of satellite altimetry, as only 14 of the 30 detected seamounts were visible in satellite-derived datasets. This underscores the necessity of high-resolution multibeam surveys for capturing fine-scale seafloor features. In contrast to time-intensive manual annotation-which can require several hours to accurately delineate each individual seamount-the automated U-Net-based segmentation approach analyzed 146,060 km² of multibeam data within seconds, offering substantial time savings and scalability for largescale mapping efforts. Beyond geological mapping, automated seamount detection has broad applications in marine ecology, environmental monitoring, and plate tectonics research. Future work should focus on integrating physical principles and geological constraints, such as typical seamount morphology, size distributions, and tectonic setting, to improve classification accuracy.