The seafood industry is extremely unpredictable. Weather, sea temperatures, fish stock levels, migration patterns, climate change—all these factors impact fishing operations, trigging gluts and shortages of supply. Quality is unpredictable too, often dependent on environmental factors, catch methods, seasonality, handling at sea, cold chain, and more. Aquaculture also experiences natural variability. Uncontrollable environmental conditions – such as water quality issues caused by weather or pollution and disease outbreaks – affect production and quality in fish and shrimp farms. In short, it’s a risky business.
For these reasons, artificial intelligence (AI) – described as “prediction machines” – holds the promise to make the seafood industry more predictable and therefore more profitable.
“Prediction is the process of filling in missing information,” write the authors of Prediction Machines: The Simple Economics of Artificial Intelligence. “Prediction takes information you have, often called ‘data,’ and uses it to generate information you don’t have. In addition to generating information about the future, prediction can generate information about the present and the past. This happens when prediction classifies credit card transactions as fraudulent, a tumor in an image as malignant, or whether a person holding an iPhone is the owner.”1
The authors—all business professors at the University of Toronto—argue that AI enables us to make better predictions faster and cheaper. That, in short, is driving profits and the adoption of AI across industries: seafood included.
In this article, I provide readers first with an overview of the different types of AI, how they work and their applications. Second, I analyse how AI is currently being used in seafood supply chains, based on a survey of some 365 software applications. Third, I’ll explore how AI can be used in seafood processing, a sector that is currently being under-serviced by AI. I’ll also share cutting-edge research we (ThisFish Inc.) are currently doing to bring AI into the processing sector. Finally, I’ll conclude with some thoughts on how seafood companies can begin to prepare themselves for the AI revolution, ensuring they won’t be left behind.
What is AI and how does it work
Artificial intelligence is often referred to as the Fourth Industrial Revolution. The original revolution was driven by steam power and mechanisation in the 1780s; the second centered on electrification and mass production in the 1870s; and the third began in 1969 with computer electronics. AI is now at the heart of the fourth Industrial Revolution, often called Industry 4.0. Many of the new technologies such as cloud computing, the Internet of Things (IoT), smart sensors and computer chips are enabling AI through advances in data collection, data storage and computational power. AI is itself also enabling many new technologies such as autonomous robots, self-driving cars, augmented reality, and cognitive computing.
What exactly is AI? At its core, AI is about simulating human intelligence in machines and falls within the general field of data science. Machine learning is a subset of AI in which machines make decisions without being programmed. We basically train machines using data. In general, more, and better quality data makes AI smarter. You can “supervise” the learning of AI by labelling data, meaning the input comes with a corresponding output label. For example, you might label an image of a fish (the input) with its proper species (the output). Unsupervised learning involves an algorithm that tries to uncover hidden patterns in data such as detecting anomalies or errors. The US Food and Drug Administration (FDA), for instance, is using AI to see patterns in seafood import data to identify possible illegality. A third type of learning is “reinforced” which occurs when AI interacts with its environment, receiving positive or negative feedback. Chatbots for customer support which ask you to rate its answers is a common example of reinforced learning.
A subset of machine learning is called “deep learning” which uses neural networks inspired by the human brain’s structure and function. This technology is often used to discover hidden patterns in enormously large, complex or multi-dimensional datasets. The most common technologies include computer vision and natural language processing.
Computer vision is a field of AI that enables machines to interpret and make decisions based on visual data from the world. Each pixel in an image is given a number, and then algorithms find patterns that, for example, will differentiate a yellowfin from a bigeye tuna in an image. Computers can classify an image, detect an object, segment an image into components and recognise faces. Computer vision technologies use a type of neural network to achieve high accuracy.
The final technology is generative AI such as the wildly popular ChatGPT or DALL-E for generating images. This technology also uses neural networks to learn patterns from large datasets of images or texts to generate new content. Large language models are trained on millions of pages of text which make them sound human.
How can all this AI technology be applied to the seafood industry? One helpful tool is called the P.A.C Framework, developed by entrepreneur and AI expert Rob May. P.A.C. stands for Predict, Automatic and Classify, which are the core functions of AI. May suggests creating a simple table to see how AI can apply to your business. “To make your first grid, make three columns, one for Predict, one for Automate, and one for Classify. Then on the rows, list key areas of your business. For example, you could list: Customers, Product, and Operations. Then in each box you can figure out how that specific A.I. approach could apply to that area of your business”, May writes.
I have created such a grid (Table 1) looking at different sectors of the seafood value chain including fishing, farming, planning, production and quality control. While not an exhaustive list, it is meant as a brainstorm on how AI might apply to your business. As we’ll see in the next section, AI is now being broadly adopted across the seafood sector.
Table 1: P.A.C. Framework, adapted to the seafood industry
The explosive growth of AI in seafood
AI is now everywhere in our daily lives: in our smartphones, virtual assistants, cars, music recommendations, food deliveries, facial recognition, and so on. The growth has been driven by three factors. First, the volume of digital data has grown in the past 10 years by some 30-fold. AI needs data to learn, and so the growth of digital technologies— especially cloud computing and smartphones—is generating the data to train AI. Second, AI algorithms have become more powerful. For example, ImageNet was a large-scale visual recognition challenge held once a year for the past decade. In 2015, the best AI system beat human performance for the first time. AI can now recognise photos with higher accuracy than humans. And third, computer chips are becoming more powerful and cheaper. They are critical for processing larger amounts of data with more complex algorithms.
As for seafood, I’ve identified a “meta-trend” which helps to explain why AI is being widely adopted in the industry. I call it the Fish n’ Chips Index. Since 1990, the FAO Fish Price Index has increased by 60 percent while the price of semi-conductors or computer chips as declined by more than 50 percent. Today about a third of the software—more than 120 apps—being used in the seafood industry employ AI.
In fact, I’ve analysed more than 365 software apps being used in the seafood industry and published them in an online directory on ThisFish’s website2. There has been an explosion in new tech startups servicing the seafood sector, which peaked in 2018-2019. However, AI is being applied unevenly in the seafood industry.
There are about 100 software apps being used in wild capture fisheries and an additional 100 in aquaculture. Almost 70 percent of aquaculture tech uses AI compared to only 20 percent in fisheries. The reason for the divergence is two-fold: incentives and regulations. In aquaculture, AI is driving down two of the industry’s largest costs: optimisation of feeding and reducing mortalities. Farmers are motivated to adopt AI since it makes business sense. In fisheries, most technology is being driven not by market dynamics, but regulatory compliance to better monitor catches. Fishermen are reluctant to adopt technologies that increase surveillance, with the exception, perhaps, of improved safety at sea. A significant amount of the AI in fisheries is focused on using computer vision to automate the reviewing of video from electronic monitoring systems onboard fishing vessels.
We’ll likely see this gap in AI adoption widen. According to Crunchbase, aquaculture tech companies have raised about USD 632 million in investments, compared to only USD 19 million by fisheries tech companies. In fact, Crunchbase reported that in 2022 alone, 45 aquaculture startups raised USD 292 million.
AI use cases for seafood processing
The seafood processing sector has lagged in both digital transformation and the adoption of AI. Paper record keeping is the predominant method of data collection, causing numerous problems such as human errors, no real-time analytics, slow reporting and poor traceability. There are 29 software products targeting this sector and yet only four feature AI capabilities. For quality testing and verification, there are 13 software products for seafood. More than 50 percent are using AI including computer vision, DNA testing, and chemometrics.
Despite low levels of digitisation and AI adoption, seafood processors are critical nodes in supply chains, aggregating supply, maintaining food safety and enabling global trade. Estimates suggest that about 23 000 companies are involved in seafood processing globally. They represent a bottleneck between the millions of primary producers and billions of consumers.
Despite its critical commercial importance, the processing sector suffers from low profitability. Planet Tracker analysed 89 publicly-listed seafood processing companies to determine that their earnings margin before interest and tax (EBIT) on average was only 3.4 percent.3 Artificial intelligence could potentially improve these margins through automation and optimisation while meeting sustainability goals.
Based on our work digitising seafood processing plants, ThisFish Inc. estimates that an average tuna cannery, processing 100 metric tonnes of raw material per day, generates more than four gigabytes of digital data a year, equal to about 2.7 million pages of text. Fresh-frozen processors of tuna and salmon generate about one gigabyte of data or 680 000 pages. This enormous volume of data could prove to be valuable to the seafood processing sector. Historic data could be used to train machine-learning algorithms to predict quality and production outcomes. In the next section, I’ll share some of the innovative AI research we’ve conducted over the years.
Machine learning
In 2021, data scientists at ThisFish Inc. developed a proof of concept to predict yields in a tuna cannery in Thailand. The data used for analysis came from approximately 22 months of production for skipjack, albacore, and yellowfin tuna. The data consisted of 8 818 unique datapoints for which yields were calculated. A single unit of raw material for which the yields were calculated was a production lot. The yield was calculated as the following:
We divide the variables into two categories: raw material variables and process variables. Raw material variables are those that are inherent properties such as fish species, fish size, harvest method, etc. Process variables are those that are collected as part of the canning process, such as cold storage duration, thawing time, thawing temperatures, cooking time, cooking temperatures, etc. We studied the impact of these variables and tried to determine ways in which these variables can be adjusted to potentially increase yields.
In summary, the primary variables that were found to influence yields were fish size, fish species, time in cold storage, cooking temperature, and cooking time. However, the more interesting finding was the impact of cold storage duration on yields, which impacted recoveries by up to three percent. About 85 percent of the variation in yields was explained by the variables in the model and the AI had a confidence level of 95 percent. Since raw material is the largest cost in production, a yield prediction model could help a cannery in predicting gross margins in production.
Another machine learning project focused on the drained weight in tuna canning. Drained weight is declared on cans and represents the minimum amount of meat weight that remains in a can once its liquid medium—typically oil, brine or spring water—is drained. It’s difficult to predict the drained weight because the tuna in the can absorbs some of the liquid, known as “pick-up.” Many variables can impact the amount of “pick-up” including species, liquid type, meat quality, packing density, can size, flake-to-chunk ratio and others. Most canneries have developed drained weight tables which help production managers determine the fill weight of the cans. However, these tables are simplistic with only a couple variables. They don’t consider all the relevant variables and how these variables interact with each other. As a result, canneries are often surprised when results come back from the drained weight lab.
Furthermore, there are regulatory limits on how many cans are allowed to be underfilled which is called the tolerable negative error. For example, the threshold on a 102-gram drained weight means you can’t have more than 2.5 percent of cans underfilled by 4.5 grams and there’s zero tolerance for extreme underfilling of more than 9 grams. Since the process is unpredictable, canneries purposefully overfill cans from 3–6 grams on average to ensure they don’t break underfilling rules.
In 2023 and 2024, ThisFish Inc. worked with two tuna canneries in Manta, Ecuador to develop a drain weight prediction model. We received more than 100 000 drained weight samples and related variables from Eurofish S.A. and Tri Marine’s SEAFMAN cannery. We trained two separate AI models, one for each cannery, keeping the data separate given confidentiality requirements.
The preliminary results were promising. In general, the AI reduced the standard deviation in the drained weights, pointing to stronger process control. The results also suggest that up to three grams per can could potentially be saved, depending on the benchmark performance of the cannery. One gram of skipjack is equal to about USD 500 000 in raw material costs each year for a 100-metric-tonne-per-day cannery. There are 150 canneries worldwide averaging 100-metric tonnes, meaning AI could potentially save the sector USD 75 million to USD 225 million in raw material costs each year. Once performance is improved, the AI model could then be retrained on the better data, hopefully squeezing out even more savings and strengthening control.
Computer vision
There is a lot of visual inspection in the seafood industry, on raw material, semi-finished and finished products. In general, if humans can see something then we can program AI to see it as well.
At ThisFish Inc., we’ve developed several computer vision algorithms for counting and measuring fish fillets, classifying salmon fillets based on the SalmoFan color scale and detecting five different types of defects on salmon fillets including gaping, trenching, softness, bruising and inoculation scarring (on farmed fish). We’re also working on an algorithm to estimate the weight of a salmon based on an image. Future models could also be trained on whitefish or shrimp .
In the tuna industry, a Japanese tech company has developed Tuna Scope to use computer vision to rate the quality of fresh tuna. ThisFish Inc. has also experimented with using computer vision to estimate the salinity level in frozen skipjack, a research project that is ongoing. Other applications include visual inspection of can seams or detection of dents and other defects on cans. Most seafood processors also do label inspection of finished product, a task that could be automated by a smart camera.
Computer vision could potentially make quality inspection cheaper and better. Currently, most quality inspection involves human inspectors taking random samples to check quality. With computer vision comes automated reducing labour costs and inspection is continuous instead of random sampling. Seafood companies would generate enormous volumes of inspection data which could be used to make better decisions and reduce frivolous claims, since processing plants would have enormous amounts of data to prove they’ve met customer requirements.
Conclusion
In the past, many seafood companies believed a competitive advance could be gained by hiding data and not being transparent about their production process or supply chain. With growing market demands for transparency and responsible sourcing, these attitudes are becoming more anachronistic with each passing day. In the Information Age, lasting competitive advantage will come to those companies that discover hidden insights and patterns in their data through artificial intelligence.
My advice to companies that want to start their AI journey is simple: digitise, prioritise and learn. First, digitise your data. Without data, you’ll have no AI. Second, prioritise by focusing on early wins and cost reductions through automation and optimisations. And third, learn as you go. Take a phased approach in introducing new digital technologies.
While AI can make the industry more profitable, it could also prove valuable for sustainability, especially helping to detect and eliminate seafood fraud in supply chains. Since AI requires clean, comprehensive and complete data for accurate predictions, there will be growing commercial incentives to collect quality data on provenance and traceability.
Stanford professor and AI expert Andrew Ng has stated, “It’s not who has the best algorithm that wins. It’s who has the most data.” One could also say it is those who have the best data that will win.
NOTE: This article was first published in the online InfoFish magazine. Click here for article.
FOOTNOTES
1 Ajay Agrawal, Avi Goldfarb and Joshua Gans Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press, Boston, 2028, pp. 29-30
2 https://this.fish/software-directory
3 François Mosnier, John Willis, Matt McLuckie. Traceable Returns. Planet Tracker. October 2020.
