RASTECH Magazine

Features Research Technology
Using AI to detect fish mortalities in RAS

Introducing Freshwater Institute’s new MortCAM AI technology.

July 8, 2024  By Kata Sharrer and Rakesh Ranjan, Freshwater Institute

The MortCAM is positioned above the drain plate in the Freshwater Institute’s growout tank. Photos: Freshwater Institute

When raising fish for food, as broodstock, or for conservation purposes, a farm’s biomass or stock is highly valuable. Any sudden increase in mortality can be devastating. Unforeseen losses can significantly impact fish populations and production outcomes, whether occurring during early rearing stages or when fish are ready for harvest. Mortality patterns can serve as early warning signs of fish health and welfare issues or operational failures within aquaculture systems. Failure to recognize these signs promptly can lead to mass mortality events. Therefore, integrating tools for continuous mortality monitoring and alerting operators to unusual or sudden changes is essential for identifying root causes and taking action to prevent escalation.

Well-designed recirculating aquaculture systems (RAS) are engineered to provide optimum rearing conditions for the cultured species. Environmental stressors are minimized by removing solids and CO2, converting ammonia to nitrate, adding oxygen, and maintaining consistent pH for water reuse. However, higher rearing densities associated with RAS production can exacerbate conditions for biotic and abiotic stresses on the fish. As the scale of RAS operations increases, mortality incidents in larger tanks can result in greater losses.

Mortality in RAS is typically monitored through human observation or, in some cases, with the help of monitoring equipment such as an underwater camera. These two common methods are usually limited to when workers are on site, and even then, only intermittently, when attention is on this task. Compounding the issue are the unique characteristics of each farm, which may include deep tanks, low lighting, or cloudy water, making observation challenging. Therefore, continuous monitoring of mortality trends is crucial for effective risk mitigation.


With these challenges in mind, the Precision Aquaculture team at the Conservation Fund’s Freshwater Institute (TCFFI) developed and evaluated MortCam AI, a mortality monitoring tool that utilizes Artificial Intelligence (AI) to provide near real-time mortality tracking. Using the Internet of Things (IoT), the device generates email and text alerts with counts and photo evidence to inform the farm manager and staff of unusual mortality trends.

The MortCAM AI was developed to provide robust mortality detection data and identify trends.

MortCam AI was developed using off-the-shelf components, including a Raspberry Pi 4 single-board computer, RGB camera, underwater enclosure, and power-over-Ethernet (PoE) HAT board. MortCAM development occurred over two phases: 1) image collection and model training and 2) evaluation of the deployed model and notification system. The device was deployed 0.6 m above the bottom drain in TCFFI’s 150m3 semi-commercial scale RAS growout tank, operating with a flow rate of 4,169 liters per minute. During image collection, the tank was stocked with approximately 5,000 Atlantic salmon (Salmo salar) with an average fish weight of 1.63 kg, and a tank density of 50 kg/m3.  

Mortality model
To develop a robust mortality detection model, training data was collected in two tank light conditions: ambient roof lighting and LED-supplemented lighting. MortCam was programmed to capture training images at 15-minute intervals for each light condition over a three-month period. The acquired images were sorted and uploaded to an online AI model training platform for annotation, model training, and visualization. During annotation, each image was manually assessed for the object classes of interest: 1) alive and 2) dead fish in the frame. Bounding boxes were used to select and classify each visible fish instance in the image. One thousand images were used to train the mortality model, and performance was analyzed in terms of mean average precision (mAP) and F1 score, two widely adopted performance metrics. 

Model Performance
The mortality model trained with images captured in ambient light conditions performed best among the models, achieving the highest mAP of 95.5% and an F1-score of 0.92. The performance of the model trained for supplemental light conditions achieved a mAP of 88.7% and an F1-score of 0.86. The glare produced by LED lights on the bottom train plate likely contributed to the degraded model performance in supplemental light conditions. Interestingly, when the performance of the mortality model trained for a certain light condition (e.g., ambient light) was tested in another (e.g., supplemental light) and vice versa, the performance of models declined considerably. This finding suggested that the model wasn’t robust for varying imaging conditions typical in real-world scenarios and may benefit from diversifying the input training data. Therefore, a mixed mortality model was trained by including images from both ambient and supplemental light conditions, and model performance was tested to further explore this idea. As expected, the mixed model was more versatile and better able to accurately predict the classes in varied conditions, achieving a mAP and F1 score of 93.4% and 0.89, respectively. 

MortCam alert system
MortCam AI offers customizable alert settings, allowing users to specify the mortality threshold for triggering email and text notifications. In the validation phase, MortCam was configured to send alerts to authorized users once the threshold of three mortalities was reached. Email and text alerts pinpointed the tank reported with abnormal mortality and provided the mortality count along with the date and time. The alert generated by MortCam attached photo evidence in the email to verify the death count. Any unusual mortality event alerts generated by MortCam can be used to take immediate corrective action to avoid mass mortality events. MortCam also logs daily and cumulative mortality. This enables farm managers to create mortality metrics and benchmarks, which they can use to recognize unusual patterns and identify root causes of fish mortality. 

What’s next?
The developed mortality monitoring system is a Research prototype. Our aim is to eventually commercialize this technology for use by RAS farmers. The mortality model is trained to detect Atlantic salmon mortality within a specific RAS configuration. Generalizing the mortality model to encompass various fish species and RAS configurations would enhance the system’s commercial adaptability. We are extending discussions with our industry partners to explore the commercialization potential and market feasibility of MortCam AI. We believe that the sooner this technology can be used across our industry, the more reliable our production systems become. 

Print this page


Stories continue below