Artificial Intelligence (AI) and Machine Learning (ML) can assist producers to better manage recirculating aquaculture systems (RASs). ML is a data-intensive process, and model performance primarily depends on the quality of training data. Relatively higher fish density and water turbidity in intensive RAS culture produce major challenges in acquiring high-quality underwater image data. Additionally, the manual image annotation involved in model training can be subjective, time-consuming, and labor-intensive. Therefore, the presented study aimed to simulate fish schooling behavior for RAS conditions and investigate the feasibility of using computer-simulated virtual images to train a robust fish detection model. Additionally, to expedite the model training and automate the virtual image annotation, a process flow was developed. The ‘virtual model’ performances were compared with models trained on real-world images and combinations of real and virtual images. The results of the study indicate that the virtual model trained solely with computer-simulated images could not perform satisfactorily (mAP = 62.8%, F1 score = 0.61) to detect fish in a real RAS environment; however, replacing a small number of the virtual images with real images in the training dataset significantly improved the model’s performance. The M6 mixed model trained with 630 virtual and 70 real images (virtual-to-real image ratio: 90:10) achieved mAP and F1 scores of 91.8% and 0.87, respectively. Furthermore, the training time cost for the M6 model was seven times shorter than that for the ‘real model’. Overall, the virtual simulation approach exhibited great promise in rapidly training a reliable fish detection model for RAS operations.

The You Only Look Once (YOLO) has been widely adopted in aquaculture monitoring and management due to its real-time performance and deployment flexibility. The recently introduced YOLO26 architecture incorporates Non-Maximum Suppression (NMS)-free end-to-end inference and is optimized for deployment on resource-constrained CPU-based devices, making it particularly relevant for edge deployment in commercial aquaculture applications. Nevertheless, its performance, operational efficiency, and deployment suitability compared with previous YOLO generations remain largely unvalidated in aquaculture-specific scenarios. This study benchmarks YOLO26 against three Ultralytics predecessors (YOLOv5u, YOLOv8, and YOLO11) across nano, small, and medium model scales for the detection of fish mortality, a critical indicator of fish population health and welfare, in recirculating aquaculture systems (RAS). Twelve model variants were evaluated for detection accuracy, training efficiency across seven dataset sizes, and inference performance on both high-performance NVIDIA A100 GPUs and the resource-constrained, CPU-only Raspberry Pi 5 edge device. All models achieved comparable performance on the full dataset, with mAP50 varying by only 1.25 percentage points across three independent training runs, indicating minimal influence of architectural generation on final mortality detection accuracy when sufficient training data are available. However, notable differences emerged in data efficiency and deployment performance. YOLOv8 demonstrated the strongest training efficiency, achieving 90% mAP50 with only 400 training images, whereas YOLO26 nano and small variants required 1000 images to reach comparable accuracy. In contrast, YOLO26 exhibited advantages during edge deployment, with YOLO26n achieving the highest inference speed on the Raspberry Pi 5 at 7.84 ± 0.13 FPS across three benchmark sessions, while YOLOv5mu outperformed all contemporary medium-scale architectures on CPU-based hardware. These results demonstrate that architectural novelty alone is an insufficient criterion for model selection. The findings support a deployment-oriented framework in which training data availability, target hardware, and inference requirements collectively inform model selection for aquaculture applications.