Automatic Detection and Assessment of Corals in Shallow Sea Regions Based on Deep Learning Models

In order to improve the efficiency of coral monitoring in shallow sea areas, we propose a deep learning-driven process for coral automatic detection and assessment using attention-embedded Yolov8 model and DeepSort target matching and tracking algorithm. Image datasets are firstly collected based on videos taken by underwater robots, and labeled according to expert guidance. An attention-embedded Yolov8 model has been trained using the datasets to achieve the purpose of intelligently identifying different types of corals. Thereafter Deep Sort algorithm is applied for matching and tracking individual corals in consecutive frames and re-identifying of corals of different types. And the assessment, including automatic coral counting and pixel-based coverage estimation, has been seamlessly integrated into the process. Experiments have been successfully conducted with underwater video datasets from shallow waters in the South China Sea, achieving over 90% accuracy in the automatic identification and counting of various coral types. The process proposed in this study represent a significant step towards the development of reliable and automated techniques for coral reef detection, and can be applied to provide efficient support for coral ecosystem surveys.