Smart detection of biofouling: A deep learning framework for water infrastructure surveillance

Golden mussel (Limnoperna fortunei) is a significant freshwater biofouling species, responsible for approximately 40% of hydropower disruptions in South America, and is increasingly threatening water infrastructures worldwide through rapid spread facilitated by global shipping. Accurate lifecycle detection is essential for effective control of golden mussel biofouling. However, traditional microscopic identification methods are time-consuming and labor-intensive, while also suffering from reduced accuracy due to morphological ambiguity, complex backgrounds, and sparse discriminative features. In this study, we propose MusselDet, a novel end-to-end detection and instance segmentation framework based on an enhanced Co-DETR architecture, to track the key life stages of the golden mussel. MusselDet consists of three key modules: (1) a Class-Aware Convolutional Block Attention Module (CA-CBAM) to enhance feature discrimination in sparse data environments, (2) a Learning Dynamic Graph Convolutional Network (LDGCN) that recovers missing larval features through temporal-spatial graph learning, and (3) a Mussel Location-Sensing Transformer (MLST) to achieve fine-grained instance segmentation and assess the density of clustered adult mussels. Evaluated on a custom-built Golden Mussel Dataset (GMD), MusselDet outperforms baseline and state-of-the-art methods in precision, recall, and inference speed, achieving a mean average precision (mAP₅₀) of 90.3% and 52.4 frames per second (FPS). This study introduces a comprehensive and scalable framework for lifecycle monitoring of golden mussel biofouling, providing a foundation for real-time and cost-effective surveillance of water infrastructure. Beyond detection, the framework offers a practical tool to facilitate the prevention and management of invasive biofouling species.