An Underwater Image Enhancement Robotic System for Salmon Eggs Monitoring

For centuries, the Pacific Salmon runs in British Columbia were so thick that they were said to "turn the water black." But today, critical numbers from stock assessment tells a different story.

Pacific Salmon Foundation published the State of Salmon Report and says:

Chum salmon: 39% to 89% below long-term average.

Egg-to-fry survival around 26.4% and vary widely.

Fraser sockeye had extremely poor years returns of 855,000 in 2016, 493,000 in 2019, and 291,000 in 2020, which broke record lows.

Salmon populations, despite their critically low numbers in recent years, remain a vital component of Pacific Northwest and Canadian coastal economies, cultures, and ecosystems. They are so deeply embedded in the regional identity and economic fabric—supporting First Nations, commercial, and recreational sectors—that their recovery is considered urgent for future prosperity.

In fisheries management, the annual salmon harvest are allocated based on pre-season predictions and observed in-season strength of the return in a given year. This means that estimated of salmon spawning abundance and survival at each life stage can affect all users of the this critical resource.

Due to limited research on salmon eggs, we developed a robot system for this particular life stage. Instead of conventional box-shaped ROV with propellers, we chose a hexapod-inspired design. A legged robot can navigate uneven terrain more carefully and maintain stability from a low center of mass in complex natural environments, therefore is a more controlled, low-disturbance method for salmon egg observation and monitoring in natural habitat.

The goal of this robot ("Kraken" named after a giant octopus) is to create a non-invasive underwater platform for salmon egg monitoring in natural conditions and it has evolved three iterations over time.

Mechanical Design

The hexapod and associated body, legs and attachment systems were designed and refined in Fusion 360. SimScale was used to perform static force simulations on legs so loads could be estimated and the structure could be reinforced. Computational Fluid Dynamics (CFD) analysis showed that redesigning the front end into a circular dome reduced hydrodynamic drag by 54% compared to the original front geometry. Silicone fins were developed using a 3D printed mold, then fitted onto the legs to support an underwater sine-wave undulation motion in deeper rivers. We used General Purpose Acrylonitrile Butadiene Styrene (GP-ABS) to print and prototype all the parts.

Electrical Design

The robot used a Raspberry Pi and a voltage regulator set to 5.5 V for suitable power distribution. The battery was connected directly to the motor controller, which controls 20 servos in total: 18 underwater servos for locomotion and 2 additional servos to provide 180 degree camera vision control.

Software Design

The software system was developed using Linux Ubuntu and ROS. Programs were run through the Linux environment and then connected to ROS on the Raspberry Pi to control the robotic platform and coordinate its motion.

Integration and Algorithm

The system also incorporated YOLOv11 object detection and the Hybrid Fusion Method (HFM) underwater image enhancement method, both studied from technical literature and online research. These algorithms were used to improve salmon egg detection in difficult underwater images.

Conventional ROV vs Hexapod Platform ROV

Compared with a conventional propeller driven ROV, the hexapod platform offered better stability, more careful movement over uneven terrain, and generated lower disturbance to the surrounding habitat.

This study demonstrated that YOLO object detection in combination with image enhancement improved the detection and enumeration of salmon eggs compared to the original underwater images. Most of the time it is difficult or impossible to detect salmon eggs using only original images due to low visibility, low contrast, colour distortion, suspended particles, overlapping objects or a combination - all of which reduces the reliability of detections. Enhancing underwater images before detection greatly improved the accuracy and consistency of detecting salmon eggs.

The most significant finding of the study was that YOLOv11n performed best after image enhancement among all the popular YOLO methods tested. Specifically, YOLOv11n had the best overall performance across all three measured areas, including perfect recall for detecting the presence of salmon eggs. Perfect recall means that all salmon eggs present in the test data were detected by the model. In this study, recall is especially critical because miss counting salmon eggs, both under and over detections of salmon eggs, can lead to large negative consequences of salmon stock assessment leading to unreliable population estimates.

The results confirm that image enhancement is valuable to the overall detection system, as they all show significantly improved accuracy (higher true positive rate) compared to the original image set. Case studies displayed above provide visual examples of each improvement. For example, case study 1, where the original image contained overlapping salmon eggs, makes it difficult for the YOLO model to delineate (separate/detect) the edges of the eggs because the edges are blended together, resulting in only 15 eggs in the model output. Enhancing the underwater image improves contrast and delineation of egg boundaries, resulting in 21 eggs detected, which is much closer to the true count. This finding was also reflected in the case study 2, the murky water study, where the enhancement improved contrast and visibility, making it easier to detect eggs against the surrounding substrate.

In addition to quantifying the total number of salmon eggs, the detection system can also separate live and dead eggs based on egg colour and condition in an image, providing a estimate of hatching success. As shown in the last picture, dead salmon eggs are identifiable by their opaque white appearance. Using the notation: if live eggs are represented as L, dead eggs are represented as D, and total eggs are represented as T, then:

The live egg percentage is:

The live to dead ratio can be written as L:D. If multiple samples are collected, the average live egg percentage is:

This allows the system to move beyond simple counting and toward estimating overall egg survival rates. This system can then be applied in controlled water tests and field tests, quantifying success across visibility conditions, substrates and water movement in natural stream environments.

This project demonstrates how underwater image enhancement has improved the robotic systems' ability to monitor salmon eggs in challenging underwater environments. Clearer images were shown to significantly improve the detection of salmon eggs, especially when eggs overlap or when water conditions decrease contrast. These findings indicate that image quality is one of the most important factors in detecting salmon eggs using our system and, therefore, should be considered an integral part to evaluate the efficacy of the monitoring system.

Our low-disturbance hexapod-type robotic platform in conjunction with the computer vision detection YOLO system, has many advantages in salmon egg monitoring compared to conventional underwater systems. Specifically, our robotic system is designed to move smoothly through complex underwater environments inducing least disturbance to natural systems, and the computer vision detection system enables faster, more efficient analysis of imagery.

The potential to extend this system's capabilities beyond simply counting salmon eggs is also significant. The system is capable of determining the conditions of those eggs, enabling the estimation of salmon egg survival rates across different locations or years. As such, the system can shed light on the causes for salmon population dynamics and evaluations difference climate and management conditions.

In conclusion, this project has illustrated how engineering can integrate with ecological research on the conservation of critical salmon species for sustainable future in Canada.

After lab testing in controlled environments, we will apply for scientific licences to conduct study in the field. We will expand our dataset to further train the image detection model for more precise enumeration in complex habitats and better detection for egg conditions. The improved abundance and survival estimates can further contribute to population analysis for later stages both in the freshwater and ocean ecosystems.

Future engineering work will focus on further improving the robotic platform and the detection pipeline by using a motorized rig for better depth control and strengthen the structural design by improving fin and leg movement.

We would like to thank Professor Fumin Zhang of the Hong Kong University of Science and Technology for his guidance and support throughout this project. Professor Zhang is a Chair Professor at HKUST and previously served on the faculty at Georgia Tech, where his research focused on marine robotics and autonomous systems. His mentorship helped us better understand forward and inverse kinematics and how to implement these ideas in code. He also demonstrated waterproofing concepts used in his compact underwater robotic systems, which helped informing our design decisions.

We also thank Guo Ran, an engineer at UBC, for helping us with the robot software interface and system integration. His support improved the connection between the hardware and software components of our project. Together, their advices strengthened both the engineering design and practical implementation of our underwater robotic system.

Journal Articles & Reports

An, S., et al. (2024). A hybrid fusion method for underwater image enhancement. Engineering Applications of Artificial Intelligence, 127, 107219. Elsevier.

Kwan, B. (2024). New state of salmon report highlights widespread declines for most salmon in B.C. Pacific Salmon Foundation.

Pacific Salmon Foundation. (2024). State of salmon report. Pacific Salmon Foundation.

Windell, S., et al. (2017). Factors affecting mortality and development success of Pacific salmon eggs and alevins: A review. NOAA Technical Memorandum NMFS.

Software / Tools

Autodesk. (2026). Autodesk Fusion [Software]. Autodesk.

Ultralytics. (2026). Ultralytics YOLO [Software]. GitHub. Retrieved April 19, 2026, from https://github.com/ultralytics

Ultralytics. (2026). Ultralytics YOLO documentation. Retrieved April 19, 2026, from https://docs.ultralytics.com

Webpages

Bena Optics. (2026). The role of ROVs in underwater exploration and the importance of high-precision optics. Retrieved April 19, 2026, from https://www.benaoptics.com

U.S. Geological Survey. (2026). Salmon eggs. Retrieved April 19, 2026, from https://www.usgs.gov

Datasets

salmonlover. (2026). Salmon roe dataset. Roboflow Universe. Retrieved April 19, 2026, from https://universe.roboflow.com

Images

Eiko Jones Photography. (2026). Sockeye salmon [Photograph]. Retrieved April 19, 2026, from https://www.eikojonesphotography.com

National Oceanic and Atmospheric Administration (NOAA). (2021). Overview of the stock assessment process [Image]. Retrieved from https://media.fisheries.noaa.gov/2021-04/Annual%20Summary%20PDF%20508-C3.pdf

AI-Generated Content

OpenAI. (2023). ChatGPT (Mar 14 version) [Large language model]. https://chat.openai.com/chat