Mamba-based super-resolution and semi-supervised YOLOv10 for freshwater mussel detection using acoustic video camera: A case study at Lake Izunuma, Japan

Freshwater mussels are essential for maintaining the balance and ecological integrity of aquatic environments. Continuous monitoring of mussel populations provides critical insights into ecosystem dynamics, and supports conservation and benthic resource management. However, traditional monitoring methods face challenges, such as water clarity issues, labour-intensive surveys, outdated technologies, incomplete datasets, and limited spatial coverage. In a field survey at Lake Izunuma, Japan, a new mussel-monitoring approach was developed for turbid underwater environments using a custom-designed Sonar Speedy Scanner (SonarSS) system integrated with Mamba-based image reconstruction and an improved YOLOv10 identification model (Mussel-YOLO). The MambaIR model reconstructed super-resolution images that were processed using the Mussel-YOLO model for precise mussel detection. Among seven evaluated super-resolution reconstruction methods, MambaIR demonstrated superior performance, achieving a signal-to-noise ratio (PSNR) of 32.33 dB and a structural similarity index metric (SSIM) of 74.29 %, delivering excellent image quality. Mussel detection using MambaIR-generated images achieved the highest accuracy with an mAP50 of 91.4 % at optimal magnification. This research has significant implications for the biodiversity conservation, resource management, and environmental monitoring in low-transparency waters. This method supports sustainable and cost-effective conservation practices by enhancing underwater species detection accuracy and reducing manual labour. In addition, it can be applied to other areas, such as monitoring invasive species or assessing aquatic ecosystem health in response to climate change. • High-resolution ARIS sonar successfully detects the mussel status in turbid waters. • Offers a low-cost, non-invasive acoustic method to supplement manual surveys. • Semi-supervised learning cuts needed labels for training deep learning models. • Integrates SRR and detection model to boost long-range monitoring accuracy.