Practical application of artificial intelligence for ecological image analysis: Trialling different levels of taxonomic classification to promote convolutional neural network performance

The integration of artificial intelligence (AI), particularly convolutional neural networks (CNNs), into ecological research presents new opportunities for the automated analysis of image-based data. This study explores the practical application of CNNs for ecological image analysis by trialling annotation to different levels of taxonomic classification to determine their impact on model performance. We systematically compare various annotation strategies, evaluating their effects on the accuracy of CNNs in ecological contexts; as well as considering the feasibility of manually annotating training data to different levels. We demonstrate that variation in annotations groupings (animal, phylum or morphology) has little impact on model performance, despite large differences in class numbers. Consequently, the decision for annotators should hinge on whether to invest effort in detailed annotation at the beginning of a project or to perform finer sorting of model predictions at the end. These findings provide practical guidance for optimizing the workflow in AI-driven ecological studies, offering flexibility without compromising model performance. • Manual annotation of images is a significant bottleneck slowing ecological data use. • Assigning coarser taxonomy to individuals would quicken training data compilation. • Using three different levels of taxonomy had minor effects on model performance. • Suggests annotation strategy can be flexible without impacting CNN results. • Optimal AI-driven ecological workflows with minimal annotation effort identified.