Kimagro Fishfarming Ltd. (Levantina Fish), a Cypriot aquaculture company specializing in the cultivation of Mediterranean species such as gilthead sea bream (Sparus aurata), engaged in a collaborative effort to explore approaches for improving fish weight monitoring through non-invasive and scalable techniques. Conventional practices based on manual sampling can be time-consuming, labor-intensive, and may introduce stress to the fish. Within this context, the collaboration with EuroCC2 Cyprus focused on examining automated methodologies drawing on Artificial Intelligence, computer vision, and High-Performance Computing (HPC).
As part of this work, researchers at the Cyprus Institute investigated the feasibility of a portable stereo-vision setup combined with Deep Learning (DL) methods for fish weight estimation. The experimental system utilizes a stereo camera to capture RGB imagery and depth information of fish during offshore aquaculture activities. These data are processed to identify three anatomical keypoints on each fish—the snout tip, body midpoint, and middle caudal rays—allowing for the reconstruction of fish length in three dimensions. This reconstructed length is then related to weight through a species-specific empirical relationship. Figure 1 outlines the methodological approach examined.
Preliminary results indicate that the approach can achieve a high level of agreement with reference measurements obtained through manual methods, with an observed percentage error of approximately 1.12% under the conditions studied. The work further explored factors influencing performance, including the number of annotated samples and the duration of image acquisition. Findings suggest that, within the experimental setup, consistent estimations can be obtained with a limited dataset (on the order of a few hundred annotated instances) and short recording durations. The inclusion of an additional anatomical reference point—the body midpoint—was found to contribute to improved robustness by partially addressing the effects of fish curvature in underwater imaging.
This work illustrates how HPC resources, such as those available through the Cyclone system at the Cyprus Institute, can support the development and training of data-driven models in applied research settings. The outcomes contribute to ongoing investigations into digital approaches for aquaculture monitoring, with a focus on informing future developments rather than constituting a finalized or production-ready system.
“When individuals with diverse backgrounds come together around a shared objective, opportunities emerge for new lines of inquiry and potential innovation. In a rapidly evolving technological landscape, engaging with emerging tools can contribute to shaping future directions in sustainable aquaculture. We extend our appreciation to the Cyprus Institute for their support throughout this collaborative effort.” — Antonis Kimonides, Owner and Managing Director, Kimagro Fishfarming Ltd.
Figure 1: Overview of the methodological framework for automated fish weight estimation using the proposed DL-based stereo vision system.