This study explores the application of unsupervised machine learning (ML) to advance sustainable fisheries management and scientific analysis by automating the categorization of fish images and minimizing the reliance on manual data labeling. The research presents a novel pipeline that utilizes unsupervised learning to identify developmental stages and species without needing extensive labeled datasets.
The approach integrates two neural networks for foreground detection and feature extraction, with pre-trained models like VGG16, ResNet, and EfficientNet, ensuring consistent and reliable results. In addition to the preliminary dataset containing tropical anemone fish, the developed algorithms were tested on additional datasets containing various species, demonstrating their versatility and robustness.
Dimensionality reduction techniques, including Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP), were compared, with UMAP excelling in detecting intricate correlations and preserving high-dimensional structures. The preliminary classification achieved a 74% accuracy rate in classifying the developmental stages of A. Ocellaris using PCA with k-means clustering. Moreover, UMAP’s ability to accurately classify species highlights the potential of unsupervised ML in fisheries science, offering a pathway to more sustainable management practices through predictive modeling without the need for large labeled datasets.