An image synthesis framework for enhanced salmon louse larvae (Lepeophtheirus Salmonis) detection in complex seawater conditions

Salmon lice (Lepeophtheirus salmonis) infections pose significant threats to farmed and wild salmon populations, making early detection during pre-infective and infective stages essential for sustainable management. Traditional methods, such as PCR, are costly and unsuitable for large-scale deployment, while existing machine-learning approaches are limited by the lack of annotated data from complex seawater environments and are largely restricted to simplified laboratory conditions. To address these challenges, this study presents an automated synthetic data generation method based on the Segment Anything Model (SAM), specifically designed to improve the detection of multi-stage salmon louse larvae in real-world and complex seawater environments. A total of 54 orthogonal experiments were conducted to optimize the factor configuration for synthetic data generation, resulting in a high-quality dataset comprising 120,864 images. To maximize the utility of synthetic data and enhance model performance, we evaluated YOLO series models and the transformer-based RT-DETR-L model. Our experiments revealed that pretraining on synthetic data before fine-tuning (or hybrid training) consistently improved model performance across multiple evaluation metrics. The YOLOv8n’s F1-score increased from 70.6% to 87.7%, a notable relative improvement of 24.22%. In real-world seawater tests, models trained with synthetic data improved recall by 11.0% to 87.0% compared to models trained exclusively with the original data and outperformed well-trained biologists, further validating the effectiveness and practicality of the synthetic approach. This study provides a scalable solution for monitoring salmon louse larvae in large-scale seawater environments, improving the welfare of farmed and wild salmon. The source code is publicly available at https://github.com/jay-zc/synthetic-dataset (as of July 2025).