Introduction
Accurate estimation of salmon lice (Lepeophtheirus salmonis) levels is essential for fish welfare, farm management, and regulatory compliance in aquaculture. While traditional manual counting methods are still widely used, they rely on small sample sizes and may not always provide representative estimates (Jeong et al., 2021). Recent years have seen the development of camera-based solutions that identify salmon lice using dedicated machine learning models. Such image-based monitoring systems offer a non-invasive alternative, enabling larger sample sizes and more frequent observations (Rabe et al., 2024). However, image-based identification remains challenging due to the small size and varying life stages of salmon lice and the variation in image quality that occurs under commercial conditions. Moreover, detector training still depends on human annotation, which can vary between individual analysts. Such inter-observer variability may affect the performance of automated detection systems trained on these data (Quinn et al., 2023). This study evaluates the consistency of lice counting among trained analysts to better understand how this variability may influence image-based monitoring.
Materials and methods
Image sequences of Atlantic salmon (Salmo salar) were sampled from a representative selection of farming conditions using a commercial underwater camera system (Stingray Marine Solutions AS). A total of 2,160 sequences were initially assessed for image quality by two independent analysts based on visibility, lighting, and sharpness. Only sequences with agreement classified as Good or Fair quality were included for further analysis. Salmon lice were then counted and classified independently by three experienced analysts using standardized guidelines, distinguishing between adult female and mobile lice (adult males and pre-adult stages I and II of both sexes). During lice counting, analyst uncertainty was also recorded. Inter-observer agreement was evaluated using Cohen's kappa for quality assessment and the intraclass correlation coefficient (ICC) and exact agreement for lice counts.
Results
Agreement between analysts on image quality classification was moderate, with 63% of sequences consistently classified. Salmon lice count showed high inter-observer reliability across both lice stages and image quality levels (ICC > 0.93). Exact agreement between analysts was high overall, but lower for mobile lice and Fair-quality sequences compared to adult females and Good-quality sequences. Uncertainty in lice detection and stage classification was significantly higher in Fair-quality sequences.
Discussion
The results demonstrate that trained analysts can produce highly consistent salmon lice counts from image sequences, with excellent inter-rater reliability across lice stage groups and image quality levels. Although exact agreement was lower, particularly for mobile lice and Fair-quality images, most differences between analysts were small, indicating that variability was largely random rather than systematic.
Image sequence quality was a key driver of variability. Reduced quality increased uncertainty in both salmon lice detection and stage classification, highlighting its impact on reliable interpretation. Despite this, overall counting patterns remained stable, suggesting that experienced analysts can maintain consistency even under less optimal conditions.
These findings are important for automated detection and monitoring under commercial farming operations. Consistent human annotations provide a strong foundation for machine learning, but lower image quality can reduce sensitivity, especially for the smaller mobile lice stages and under challenging visibility conditions. Overall, image-based monitoring can provide robust and scalable estimates based on large sample sizes when image quality and inter-observer variability are taken into consideration.
References
Jeong, J., Stormoen, M., Thakur, K.K., Revie, C.W. (2021). Imperfect estimation of Lepeophtheirus salmonis abundance and its impact on salmon lice treatment on Atlantic Salmon farms. Frontiers in Marine Science. 8:763206. https://doi.org/10.3389/fmars.2021.763206
Quinn, L., Tryposkiadis, K., Deeks, J., De Vet, H.C.W., Mallett, S., Mokkink, L.B., Takwoingi, Y., Taylor-Phillips, S., Sitch, A. (2023). Interobserver variability studies in diagnostic imaging: a methodological systematic review. British Journal of Radiology. 96(1148) 20220972. https://doi.org/10.1259/bjr.20220972
Rabe, B., Murray, A.G., Ives, S.C., Moriarty, M., Morris, D.J. (2024). Searching for sea lice: surveillance to assess environmental infection pressures to model and inform sea lice infestation management. Aquaculture Environment Interactions. 16 241-262. https://doi.org/10.3354/aei