Can Artificial Intelligence Help Fishermen Find Cod?

It's a cold autumn day in Tromsø and the snow has already settled. Gadus Njord is docked for unloading and getting ready for a fresh fish trip for cod.

On the quay stand two engineers from Catchwise, with their duffel bags full of seasickness tablets and Red Bull. They're going out on the trip to test how well artificial intelligence can support the skipper in finding cod.

This will be Catchwise's third trip to sea, after previous experiences with North Sea herring and mackerel. This time, the honor goes to joining a whitefish trawler, led by skipper Ove Brennskag – an experienced and skilled skipper with many years at sea.

How AI can help find cod

Catchwise's AI models function as a "weather forecast" showing where conditions are best for a specific species.

The model's goal is to predict which areas can provide the best catch, measured in tons per towing hour, based on conditions in the area. The model is based on large amounts of data collected from the entire Barents Sea since 2011, and learns to recognize which conditions are present when there's a good catch.

The model uses several environmental variables in combination, such as bottom temperature, ocean currents, moon phase, and salinity level.

It's important to emphasize that an AI model is created to support the skipper's own assessments, not replace them.

The model helps narrow down the search area. In the same way that the echo sounder can give you an indication of whether to shoot off, a model gives an indication of which areas you should search.

How did it go?

Despite generally poor catch conditions, the model managed to provide good estimates for expected catch per towing hour in the field. We saw a good correlation between the times we went to areas that the model had rated as good, both on the echo sounder and in the catch.

Although the model is still in a "break-in" phase, it manages to capture which factors affect catch efficiency on a general basis. On the trip, we also experienced that good hauls were made in areas the model hadn't picked up. This is useful learning that we'll take with us to the next version.

We received good feedback that the predictions were easy to understand and use, and that they provided a good indication of expected catch. The skipper and mate provided good suggestions for improvements along the way and interpretations of what the model showed based on their own experience.

In the image below, you can see a clip from the application as we drove into a good field. A darker color corresponds to an area that the model thinks is good compared to the lighter areas around it.

As we hit the field, we could quickly see that there was quite a bit of activity on the echo sounder and the trawl bag filled up faster. The images from the echo sounder are from before and after we arrived at the good field.

The way forward

We learned an incredible amount on the trip and have seen that the model manages to capture where the good fishing areas are. It's a great start! Now we'll take all the learning from the trip back and improve the model so it can become an even better tool for fishermen.

A big thank you to skipper Ove Brennskag, mate Kristian Krøvel, and the rest of the crew on Gadus Njord for an educational experience, and to Ronny Vaagsholm for making this collaboration possible. We look forward to the continuation!

See pictures from the trip below: