Using machine learning effectively in marine benthic ecology: a question-driven toolkit

Manual annotation of marine image datasets is time-consuming, driving interest in machine learning (ML) to automate image analysis. However, benthic ecologists often lack the expertise needed to determine when and how ML can be applied effectively to their datasets. To explore the practical use of ML in benthic ecology, user perceptions and model training experiences were collected during a workshop. RootPainter, an interactive, corrective-annotation ML software, was used to introduce ML concepts and workflows. Participants developed models to segment benthic organisms from time-lapse and remotely operated vehicle imagery. Real-time feedback during training improved efficiency and understanding of model behaviour. Pre- and post-project qualitative surveys identified ecological and technical factors influencing model performance. From these findings, we propose five questions to guide efficient, robust annotation of marine benthic images for ML model development: (i) ‘do you have expert taxonomic and machine learning knowledge?’; (ii) ‘what is your ecological research question?’; (iii) ‘is the species sufficiently abundant?’; (iv) ‘are you consistent in your selection of training data?’; and (v) ‘have you anticipated the strengths and limitations of your imaging tool?’. Responses to these questions can provide practical guidance to help marine ecologists make informed choices, accelerate model training and improve performance.