Testing the feasibility of deep learning approaches to enhance monitoring of marine macroinvertebrates: Insights from a case study using the gastropod Peringia ulvae
摘要
Coastal ecosystems, while vital for their ecosystem services, face growing threats from climate change and human activities. Traditional methods for monitoring these systems, such as morphological species identification and laboratory analysis, are time-consuming and resource-intensive, restricting their scalability. Recent advances in Deep Learning have demonstrated significant potential for automating organism identification and counting from images, streamlining ecological research efforts. Using a common marine gastropod, Peringia ulvae, as a model species, we evaluate the potential for state-of-the-art computer vision methods incorporating self-supervised and few-shot learning techniques, to quantify this marine macroinvertebrate. We tested five machine learning models (CounTR, Segment Anything, Training-Free Object Counting, Grounding DINO, DeepDataSpace) across imaging devices, organism aggregation levels, and compared the performance across models and with human counts to determine accuracy and robustness. Using high-resolution images from a DSLR camera, we found that CounTR and Deep Data Space excelled in scenarios where P. ulvae were aggregated as opposed to spaced. However, model accuracy plateaued at