Contrastive Learning for Detecting Invasive Aquatic Species: An Analysis of SimCLR and Self-Supervised Methods with Unlabeled Video Data

Invasive aquatic species pose significant ecological risks to US waterways, necessitating improved detection methods. Traditional models struggle with large, unlabeled datasets, underscoring the need for enhanced automated recognition systems. This study explores the effectiveness of self-supervised learning methods on unlabeled video data in developing a robust and autonomous aquatic invasive species recognition framework. It explores three main questions: SimCLR's performance on unlabeled image data compared to other self-supervised learning methods, the impact of the size and diversity of the unlabeled image dataset, and SimCLR's impact on downstream tasks. This study utilizes a specialized dataset comprising invasive and non-invasive mussel larvae from water sample videos to investigate the efficacy of each method in classifying aquatic species. The dataset diversity, size, image scale, and model performance were evaluated using benchmark datasets. The research findings notably highlighted the performance of SimCLR, which utilized a ResNet50 architecture, achieving remarkable accuracy rates of 91.46% on the aquatic species dataset. This demonstrates the potential and challenges of self-supervised learning in classifying aquatic species from unlabeled datasets.