THIS ARTICLE/PRESS RELEASE IS PAID FOR AND PRESENTED BY the Institute of Marine Research - read more

The Institute of Marine Research is working across disciplines to adopt artificial intelligence. Here computer scientist Endre Moen is at his desk.

Artificial intelligence has been taught to 'read' salmon scales

Determining the age of salmon from their scales is a task for trained experts. Now artificial intelligence may give them a hand.

Communications Adviser
Presented by: Institute of Marine Research
Published

Like trees have growth rings, fish have information stored in their scales. In the case of salmon, experts can interpret this information to ascertain the age of the fish, how long it has lived in the sea, how many times it has spawned and whether it comes from a fish farm. However, this can be a time-consuming process.

That’s why scientists at the Institute of Marine Research (IMR) have taught a type of artificial intelligence known as a neural network how to interpret images of salmon scales seen through a magnifying glass.

Easy to distinguish wild and farmed salmon, but harder to determine their age

About salmon scales

The images of the salmon scales come from the Institute of Marine Research and Rådgivende Biologer AS. The scales were collected from salmon in rivers along Norway’s west coast.

The images, which were analysed by several experts in advance, contained the following information:

  • Origin of the fish (wild or farmed salmon)
  • Whether or not it had spawned
  • Number of years it had lived in a river
  • Number of years it had lived in the ocean

Photos with missing information were not used.

“The neural network had a 97 per cent hit rate when asked to distinguish between farmed and wild salmon. But this is a relatively simple task that experts can also perform quickly,” says Endre Moen, a computer scientist at the IMR.

“The most interesting questions are how long the salmon have lived in the ocean and river respectively,” he continues.

To measure how well the neural network performed these tasks, the scientists looked at how often it agreed with six human experts.

Second best at ascertaining sea age

“The network agreed with the experts 94 per cent of the time when estimating the sea age of the fish. In fact it was second best – only beaten by one human expert who was 97 per cent in agreement with the others,” says Moen.

When estimating the river age of the salmon, the network scored 66 per cent. In this case the experts scored between 75 and 85 per cent, but they had access to additional information as well as the image. For example, which river the fish came from.

This matters because salmon have local habits in terms of the age at which they leave rivers as smolts. It can be anywhere from two to six years, depending on the local conditions, such as the water temperature and size of the river.

In any case, the river age was hardest to estimate for both artificial and human intelligence.

Here is one of several thousand scales used in the study. The dark grey section at the bottom of the picture is the part of the scale that is visible on the outside of the salmon. The age has been determined by an expert.

The quality of the model answer matters

“Interpreting the river age is a more difficult task, and the one where there is most disagreement between the experts. This is reflected in the neural network,” explains Endre Moen.

To train the neural network, the researchers fed in over 6,000 images of salmon scales, along with the experts’ interpretations. Afterwards, it took an 'exam' by interpreting a further 150 pictures by itself.

This is called supervised machine learning.

“If the experts are uncertain or disagree, it makes the artificial intelligence more uncertain when it starts operating independently.”

Neural networks (artificial intelligence)

  • A computer program that uses methods that mimic the ones found in the human brain.
  • Neural networks can be trained in various ways (machine learning).
  • Supervised machine learning is when they are fed a number of examples of correct answers.
  • The aim is to enable the network to find the correct answers to cases that it has never seen before.
  • However, errors in the teaching material can cause the network to repeat mistakes.
  • The massive increase in computing power in recent years has breathed fresh life into the field of artificial intelligence. At the IMR, it is an interdisciplinary focus involving several specific tasks that would benefit from automation. For example, recognising fish species from echo sounder images

Not fully understood how it thinks

We cannot be sure that the neural network notices the same characteristics as the humans. The researchers are not entirely sure it reaches the right answer.

“We know that it first looks for simple characteristics in the images, such as short horizontal and vertical lines. It may be that the network gives importance to both the shape and size of the scales, or it might be completely different things,” says Moen.

The type of neural network they used was originally developed to recognise objects in images, such as dogs, cats, bikes and houses, but has been modified to be applicable to ageing of fish scales.

AI can join the expert panel

The initial purpose of the study is to demonstrate the possibilities offered by the new technology.

“For the moment, it is not safe to leave the whole analysis of the salmon scales to a neural network. But as a cautious first step, the network can be included as one of several experts on a panel. This could help to free up human resources or increase reliability,” says Endre Moen.

Reference:

Vabø et al. Automatic interpretation of salmon scales using deep learning, Ecological Informatics, 63, 2021. DOI: 10.1016/j.ecoinf.2021.101322

See more content from the Institute of Marine Research:

artificial intelligence institute-marine-research salmon partner aquaculture
Powered by Labrador CMS