Session: 06-04-02 Marine Engineering and Technology - II

Submission Number: 180348

Field Based Underwater Lot System for Continuous Observation and AI-Based Fish Species Classification 

Underwater video data remain limited because their acquisition is technically challenging due to environmental and operational constraints. Conventional observations have primarily relied on recording-based underwater cameras followed by offline analysis, which makes it difficult to capture and interpret marine conditions in real time. Moreover, the small amount of site-specific data has made it difficult to develop and adapt AI models to local conditions effectively.

This challenge is particularly evident in Japanese coastal waters, where fish species diversity is remarkably high, and environmental factors such as water color, seafloor composition, and current dynamics vary greatly between locations. Because many of these environments are subject to frequent physical disturbances, it has been difficult to maintain continuous observation using conventional equipment. As a result, the accumulation of localized datasets has lagged, preventing effective data-driven model construction for ecological monitoring and fishery management. The delay between data acquisition, manual retrieval, and subsequent analysis has also made it difficult to update AI models quickly enough to reflect seasonal or regional variations in fish assemblages.

To overcome these limitations, this study developed an underwater IoT (UIoT) camera system capable of long-term autonomous operation for continuous visual monitoring of coastal fish communities. The system was designed to operate under typical field constraints, including limited power supply, restricted accessibility, and minimal maintenance capability. It is powered by solar panels and rechargeable batteries, allowing unattended operation for extended periods. The system performs intermittent recording, capturing a one-minute video every 30 minutes, which provides sufficient temporal resolution to observe diel changes and fish behaviors while conserving energy. A mechanical wiper periodically cleans the camera lens to prevent biofouling, maintaining optical transparency throughout multi-week deployments.

Field observations were conducted in Numazu City, Shizuoka Prefecture, Japan, at a shallow coastal coral reef site. The camera was installed at a depth of approximately 3-10 meters, oriented toward a natural reef structure frequented by small to medium-sized reef fish. The deployment continued for several consecutive weeks, during which time the system successfully recorded numerous fish behaviors including schooling, feeding, and territorial interactions. In addition to the UIoT-based recordings, supplementary underwater videos obtained through diver-operated or stationary offline camera setups were also incorporated to enhance data diversity and ensure robustness against potential equipment limitations. The combined dataset represent one of the few long-term, high-frequency video datasets available for shallow coastal ecosystems in Japan.

The recorded underwater videos were used to evaluate an AI-based workflow for automated fish detection and classification. A convolutional neural network (CNN) served as the backbone for visual recognition, and several architectures such as EfficientNet, ResNet, and ConvNeXt were tested to assess model stability under color distortion and contrast variability typical of underwater imagery. The objective was to establish a processing and learning framework capable of robustly classifying fish species across diverse optical and environmental conditions while maintaining sufficient computational efficiency for periodic retraining as new data accumulate.

This framework makes it possible to achieve on-site data collection and adaptive model development, which were previously difficult to realize. Consequently, AI models can now be quickly customized to match the environmental and biological conditions of each region. The proposed approach represents a step toward the practical application of fish species classification AI at an industrial level and contributes to a wide range of marine applications, including coastal monitoring, resource management, and blue carbon assessment.

Presenting Author: Naho Yashiro Graduate School of Frontier Sciences, The University of Tokyo

Presenting Author Biography: Naho Yashiro is currently a Ph.D. student at the Graduate School of Frontier Sciences, The University of Tokyo. She received her M.S. in Mechanical Engineering from Keio University. During her master’s program, she founded MizLinx Inc., a company developing underwater IoT systems, and now serves as its CEO.

Authors:

Naho Yashiro Graduate School of Frontier Sciences, The University of Tokyo
Yuichiro Fujino Graduate School of Agricultural and Life Sciences, The University of Tokyo
Ibuki Igarashi Graduate School of Agricultural and Life Sciences, The University of Tokyo
Mikito Murakami Graduate School of Agricultural and Life Sciences, The University of Tokyo
Sora Ishikawa MizLinx Inc.
Nina Yasuda Graduate School of Agricultural and Life Sciences, The University of Tokyo
Katsunori Mizuno Graduate School of Frontier Sciences, The University of Tokyo