Using real-time object detection to identify morphologically similar invasive and indigenous mussels in South AfricaNovel Approaches

Yi-Fei Gu¹

¹The University of Hong Kong, China

Invasive species often compress the niche of indigenous species, altering community structure and ecosystem function. Efficient and accurate detection of invasive species is, therefore, critical for monitoring invasion status and ecosystem health. Morphological similarities between invasive and indigenous species can lead to misclassification under field conditions and hinder assessments of invasion threat. A novel approach, using tailored species annotation strategies and state-of-the-art YOLO models, was developed to detect the invasive Mediterranean blue mussel, Mytilus galloprovincialis, which is often misidentified as the indigenous brown mussel, Perna perna, along the South African coast due to colour polymorphism and endolithic erosion. This approach separated the morphologically similar invasive species into high confidence “invasive” and low confidence “unsure” classes. The YOLOv11 model in its large-scale configuration outperformed other YOLO variants, achieving an average precision of 70.7% in invasive species detection. A mobile phone app designed for detecting M. galloprovincialis is available at app stores. The proposed methods and mobile phone app can be adapted for other species detection tasks, improving invasive species monitoring and management in various ecosystems. Identifying the low confidence “unsure” class remains, however, challenging and continued research into how to separate morphologically similar species is needed. Our study demonstrates the effectiveness of combining ecological expertise with advanced deep learning techniques for invasive species detection. Future research should focus on refining methods for classifying morphologically similar samples to enhance the accuracy and reliability of species detection in field conditions.