An interpretable approach to automating the assessment of biofouling in video footage

Biofouling-communities of organisms that grow on submerged hard surfaces-creates pathways for the spread of invasive marine species and diseases. To manage this risk, international vessels are increasingly required to demonstrate effective biofouling management, which in turn necessitates underwater inspections and the analysis of large volumes of hull imagery to verify biofouling status. Automated assessment with computer vision can streamline this process. This work shows how the interpretable Component Features (ComFe) approach, combined with a DINOv2 Vision Transformer (ViT) foundation model, can address this challenge efficiently and effectively. ComFe achieves competitive performance in comparison to previous non-interpretable CNN methods, with fewer parameters and greater transparency-highlighting which image regions and training examples drive classifications. All code, data, and model weights are publicly released.