Marine Species and Scientific Name Classification Using YOLOv8

In marine biodiversity monitoring and conservation efforts underwater species detection plays a pivotal role. This study surveys the implementation of YOLOv8, a state-of-the-art object detection model, fine-tuned using a pretrained FishInv.pt model for real underwater species detection. The dataset consists of 25 marine species, which was manually annotated and split in to training (70%), validation (20%), test (10%). The model was trained on google colab pro with Tesla T4 Gpu, using 100 epochs, a batch size of 16 and to address underwater image challenges such as low contrast and light scattering. Various data augmentation methods and techniques are used. Experimental results show that YOLOv8 was able to achieve a detection speed of 29.3 FPS making the model suitable for monitoring. The model was able to achieve a mean average precision (mAP@50) of 0.812, recall of 0.61 with a precision of 0.94. A comparative analysis with faster R-CNN and YOLOv5 highlights how the superior trade-of between speed and accuracy of YOLOv8 in challenging under water environments helps in the effective detection of species. However, dataset imbalance led to misclassification, particularly among species which are morphologically similar and related. Future work of the project aims to enhance recall with the help of hyperparameter tuning, advanced dataset expansion, and model optimization for improved computational efficiency. Development of real-time monitoring system, AI powered marine monitoring system was developed supporting ecological research and conservation efforts.