Introduction
Norway is the largest producer of Atlantic salmon (Salmo salar), with more than 300 million salmon put out to sea each year (Sommerset et al., 2020). In addition, in 2020 1,701,347 Atlantic salmon were reported used for research in Norway, which accounted for 74.5 percent of the total number of reported research animals (Kristiansen et al., 2021; Mattilsynet, 2021). In both research projects and commercial salmon farming, the welfare of the fish in a production unit is monitored by welfare indicators (WIs) that are either operational (OWIs) or laboratory-based (LABWIs). Morphological OWIs are for the most part manually assessed on a sub-sample of individual fish from the production unit. Examples of morphological OWIs are fin damage, eye damage, skin damage and gill damage, among others. Each OWI can be scored in different ways but an increasingly common approach is to score the injuries on a 0-3 scale, where 0 is basically undamaged and 3 can be classified as a severe injury (Noble et al., 2018). The process of manually assigning these welfare scores is time consuming and prone to both intra- and inter-observer error due to the subjective nature of the scoring method and the motivation and skill of the observer, amongst other factors. This paper discusses the feasibility of using hyperspectral imaging as a rapid and objective method of documenting and quantifying a selection of morphological OWIs.
Material and methods
We sampled n=1834 fish in total, across four experiments, and developed methods for image analysis based on machine learning algorithms using spectral data in the images. Automatic scores from the image analyses were compared to manual scores of WIs for each fish and this was used to evaluate the agreement between the two methods. In all experiments, hyperspectral images were taken on both the left and right side of each fish. The fish were presented to the camera on a conveyor belt moving at a speed between 10 and 20 cm/s. Morphological OWIs were recorded by trained manual observers for the same fish and used as reference data. For each WI the correlation between the hyperspectral index and the manual WI score was calculated. In trial 1, the dorsal fin was analyzed and compared with the manual OWI. The interrater agreement between 2 observers for dorsal injuries was calculated to be 0.66 in percentage agreement with a Spearman’s correlation of 0.56 and a Cohen’s kappa of 0.40. In trial 2, hyperspectral images of 120 fish, sampled before, during and after smoltification, were analyzed and correlated with smoltification status, assessed by quantifying plasma chloride levels after a 24-hour seawater challenge test. In trial 3, the use of hyperspectral imaging to evaluate eye haemorrhaging was tested. In trial 4, the potential of hyperspectral imaging to detect the degree of sea lice infection was assessed. The reference data in Trial 2 was a LABWI, namely blood serum chloride ion concentration after a 24-hour sea water challenge (Noble, et al., 2018) unlike the other trials which used OWIs as reference data.
Results and discussion
In trial 1 and 3 we observed the lowest correlation between the subjectively human scored OWIs and the hyperspectral scores (table 1). This can be due to a phenomenon called ‘attenuation of error’ where correlations are generally lower between two variables if one or both is measured with error (Adolph & Hardin, 2007). Part of the error can be attributed to the OWIs scored by human observers. Despite a standardized scoring scheme, different scorers weighted different aspects of an injury differently. Nevertheless, the correlation between the HSI scores and manual OWI scores in trial 1 is on par with the interrater agreement between two human scorers, which demonstrates that the hyperspectral system agrees almost as well with a human observer as two human observers agree with each other. The higher agreement for the LABWI supports this hypothesis. Counting lice is arguably less prone to subjective variation in scoring, although observer fatigue and thoroughness can influence the process, thus possibly resulting in a higher correlation in trial 4.
Conclusions
Together these results demonstrate the feasibility for using hyperspectral imaging as a technique for automatically monitoring welfare in Atlantic salmon and quantifying traits such as fin and eye injuries, smoltification status and lice infection levels.
References
Adolph, S. C., & Hardin, J. S. (2007). Estimating Phenotypic Correlations: Correcting for Bias Due to Intraindividual Variability. Functional Ecology, 21(1), 178-184. http://www.jstor.org/stable/4139400
Kristiansen, T. S., Bleie, H., Bæverfjord, G., Engh, E., Hansen, K. A. E., Lybæk, S., Mattison, J., & Ryeng, K. A. (2021). Årsrapport 2020. https://www.forsoksdyrkomiteen.no/wp-content/uploads/2021/06/Arsrapport-2020-ENDELIG.pdf
Mattilsynet. (2021). Bruk av dyr i forsøk 2020.
Noble, C., Gismervik, K., Iversen, M. H., Kolarevic, J., Nilsson, J., Stien, L. H., & Turnbull, J. F. (Eds.). (2018). Welfare indicators for farmed Atlantic salmon: tools for assessing welfare. https://nofima.no/wp-content/uploads/2021/05/FISHWELL-Welfare-indicators-for-farmed-Atlantic-salmon-November-2018.pdf.
Sommerset, I., Walde, C. S., Jensen, B. B., Bornø, G., Haukaas, A., & Brun, E. (2020). Fiskehelserapporten 2019.