Aquaculture Europe 2023

September 18 - 21, 2023


Add To Calendar 19/09/2023 15:15:0019/09/2023 15:30:00Europe/ViennaAquaculture Europe 2023DEEP LEARNING-BASED METHOD FOR FISH BEHAVIOURAL CHANGE QUANTIFICATIONStolz 2The European Aquaculture Societywebmaster@aquaeas.orgfalseDD/MM/YYYYaaVZHLXMfzTRLzDrHmAi181982


Qin Zhang1,*, Eleni Kelasidi2, Martin Føre1, Biao Su2

1Department of Engineering Cybernetics, NTNU, Norway

2Department of Aquaculture Technology, SINTEF Ocean, Norway




In this abstract, results are presented from the study of the behavioural change of Atlantic salmon ( Salmo Salar ) in sea cages when exposed to dynamically changing environments.  Several field trials have been conducted in  an industrial- scale fish farm to observe and study  the  behavioural responses  of fish towards different influence factors,  especially to the structures (obstacles) of various shapes, sizes, and colours.  Ping360 sonars (BlueRobotics - Ping360 Scanning Image Sonar)  were used to collect  relevant fish behaviour data in industrial-scale sea cages . A d eep  learning-based method  was applied to  the 360-degree sonar data and  used to identify fish swimming patterns around  the structures and quantify the  minimum  distance  the fish kept from the structures . C hanges in the behaviour of  the fish to wards structures with different appearance were studied.

This work was financed by the Research Council of Norway through the project: CHANGE   ̶  An underwater robotics concept for dynamically changing environments [1] and RACE   ̶   Fish Machine Interaction [2] .

Materials and methods

T he structure was  produced  in six different  versions that varied in shape, size, and colour, and equipped with a stereo camera  and two  Ping360 sonars: one on the top and one on the bottom , as shown in Figure 1. The Ping360 sonars were set  with a range of 5 meters.  The structure was placed in a water depth of 8 m in each of the tests conducted in June 2021, October 2021 and September 2022 in an industrial-scale fish farm of SINTEF ACE [3].

 The sonar data provide  images of a 360-degree view of its surroundings, with the areas of fish having greater intensity.  The circular swimming behaviour of fish around structures makes them appear as rings in sonar images. After creating a training dataset via manually labelling fish swimming patterns on a few hundred randomly selected sonar images, a deep learning semantic segmentation UNet++ model [4] was trained to identify fish swimming patterns around structures. T he distances between fish and structures can then be estimated by averaging the distance from each pixel on the  edge of the  parts  identified  as fish swimming patterns to the structure centre , as shown by the red curves in Figure 2.


The deep learning-based method has been applied to the sonar image of fish accumulated around structures over 1-, 5-, and 10-minute time periods, Figure 2 shows examples of the accumulated fish presence around the  cylindrical  structures over a 5-minute period and the corresponding fish behaviour quantitation results. W hen the structures were big cylinders,  fish with an average size of 1 kg  seemed to stay  approx.  1.5 m from the yellow structure  and 0.8 m from the white structure. Similar fish-to-structure distances were obtained when the structures were big cubes. W hen  interacting with  small cylinders, the same fish  maintained approx. 0.9 m  and 0.6 m from the yellow structure and the white structure, respectively. Based on our sonar data, f ish always stayed closer to the white structures than to the yellow struc tures if the structures were  of the same shape and size.

In September 2022, the fish were of  an average size of 1 kg and  a  population  size of 172,563 individuals  and  stayed approx. 1.5 m away from a yellow cube structure. In June and October 2021, average fish size was 2.5 kg and 5 kg,  while  population  size was 195,832 and 99,243, respectively. T he distance to the same  yellow cube  structure  then increased to  approx. 2 m and 2.5 m, respectively. In all three cases, the biggest-sized fish had the longest distance to the structure, while the smallest-sized fish had the closest distance to the same structure.  This indicates that there seemed to be a  relationship between  fish  size and the distance the fish maintain from the structure.

Conclusion and future work

T he colour of structures  seems  to be  affecting fish behaviour, and fish keep shorter distances from white structures compared to the yellow ones. O ur results present n o evidence that structure shape has an impact on fish , but structure size  may  have an effect as f ish were closer to smaller structures than to the big ones.  In addition,  fish stayed farther away from the same structure as they got bigger, suggesting that the size of fish themselves  can  also a factor influencing their behaviour. More studies  are needed to confirm  these findings  and explore further relationships  between  fish  size and fish-to-structure distance.


[1]                      CHANGE  ̶ An Underwater Robotics Concept for Dynamically Changing Environments.

[2]                      RACE  ̶  Fish-Machine Interaction.

[3]                       SINTEF ACE. 

[4]                      Zhou et al., 2020. UNet++: Redesigning skip connections to exploit multiscale features in image segmentation . IEEE Transactions on Medical Imaging 39, 1856-1867.