Session: 06-17-01 AI Technology for Ocean Engineering - I

Submission Number: 174931

Sparse Point-Guided Fusion of Supervised and Self-Supervised Learning Model for Seaweed Segmentation 

In recent years, the multifaceted utilization of the ocean has attracted increasing attention as a potential solution to global challenges such as climate change and sustainable resource management. The ocean serves as a critical foundation supporting diverse sectors, including energy, environment, ecosystems, and transportation, and plays an indispensable role in achieving sustainable human development. To fully harness its potential, it is essential to accurately capture the state of the ocean and quantitatively understand its variations, thereby enabling the formulation and implementation of optimal measures tailored to specific objectives. Against this background, we aim to realize the Ocean Digital Twin, which digitally represents various oceanic states through data collection and analysis, enabling integrated scenario evaluation and policy planning. As a first step, we focus on the utilization of blue carbon credits derived from seaweed and seagrass meadows in Japanese coastal areas, where accurate classification of underwater vegetation, such as seaweeds and seagrasses, is essential for the quantitative assessment of blue carbon stocks (CO₂ absorption and fixation). However, underwater environments are characterized by turbidity and light attenuation, which significantly degrade image quality and limit the accuracy of conventional classification techniques. Previous studies can be broadly categorized into two approaches: supervised learning, which leverages correspondences between training and test images, and self-supervised learning, which segments regions based on local features within a test image. The former often suffers from performance degradation due to domain gaps between training and test datasets, while the latter can delineate regions but cannot automatically assign class labels, making full automation difficult. Despite their complementary advantages, no unified framework has been established to effectively combine these two paradigms. To address this issue, we propose a novel segmentation method for marine vegetation that integrates supervised and self-supervised learning. Our method enables automatic class labeling that conventional self-supervised learning could not achieve by utilizing the class information obtained from supervised learning. Furthermore, by performing self-supervised learning within each test image after estimating the approximate positions of target species through supervised learning, the proposed framework mitigates the effects of domain gaps and achieves high-precision segmentation. Specifically, the supervised instance segmentation module estimates the approximate positions of each target species as sparse point information, and the self-supervised module refines these regions by automatically segmenting detailed areas for each species. In this study, we evaluate the effectiveness of the proposed method using underwater images captured at an artificial seaweed bed site along the coast of Yamaguchi Prefecture, Japan. Preliminary results and discussions will be presented at the conference. The proposed technology is expected to contribute significantly to improving the accuracy of quantitative blue carbon assessment and enhancing the efficiency of marine ecosystem monitoring.

Presenting Author: Tatsuya Suzuki Ocean Digital Twin CPJ, Converging Technologies Laboratory, Fujitsu Research, Fujitsu Limited

Presenting Author Biography: Tatsuya Suzuki is a Principal Researcher at Fujitsu Research, the R&D division of Fujitsu, a leading IT company in Japan. He specializes in computer vision and has engaged in research on industrial visual inspection for manufacturing, human pose estimation for gymnastics judging support, and generative AI for 3D scene and avatar creation in metaverse applications. Recently, he has been leading cutting-edge research on underwater object recognition, underwater vision enhancement, underwater neural rendering, and marine geophysical data analysis toward the realization of the Ocean Digital Twin.

Authors:

Tatsuya Suzuki Ocean Digital Twin CPJ, Converging Technologies Laboratory, Fujitsu Research, Fujitsu Limited
Kazuya Ijuin Human Digital Twin Division, Global Solutions Business Group, Fujitsu Limited
Hideki Tomimori Ocean Digital Twin CPJ, Converging Technologies Laboratory, Fujitsu Research, Fujitsu Limited
Megumi Chikano Ocean Digital Twin CPJ, Converging Technologies Laboratory, Fujitsu Research, Fujitsu Limited
Katsushi Sakai Ocean Digital Twin CPJ, Converging Technologies Laboratory, Fujitsu Research, Fujitsu Limited