Applying computer vision and machine learning tools to improve the diagnostic accuracy of in-situ and digital image-based assessments of coral health and disease

The prevalence of disease is steadily increasing in coral reefs around the world, and there is a need for efficient methods of monitoring and tracking the health of corals. Coral health assessments are primarily conducted using in-situ survey visual observations of disease in the field. Recent technological advancements in the field of computer vision allow researchers to collect high-resolution imagery of benthic habitats and reconstruct the images into 3D models and 2D orthomosaics. However, little is known about the relative efficacy or diagnostic accuracy of these two methods. This study contrasts the diagnostic performance of in-situ and digital methodologies for detecting diseases affecting coral reefs on Hawaii Island. We surveyed multiple study plots on coral reefs located on both the windward and leeward side of Hawaii Island. For each plot, an in-situ visual analysis of coral health is conducted by the diver, and images are collected and later compiled to create high-resolution orthomosaics for digital analysis. Both methods assess the same coral colonies, resulting in paired health diagnoses for a variety of diseases. Lacking a gold-standard diagnosis of disease, a latent class model is used to estimate the sensitivity and specificity of both methods. We find that in-situ assessments of coral health have a higher sensitivity and lower specificity to detecting diseases and conditions of reduced health when compared to digital analyses using the orthomosaics. However, the effect size is relatively modest, indicating that while in-situ provides a more sensitive diagnostic approach, the techniques are of comparable accuracy and should both be considered viable methods of characterizing and monitoring coral health. In order to explore potential improvements to this approach, we applied artificial intelligence (AI) tools to automate the identification of disease states in the digital analyses. Precision-Recall curves indicate that AI has the capacity to improve the efficacy of digital analyses by assisting humans in identifying diseases from digital imagery. This study provides promising results for integrating digital tools to improve monitoring of these important marine ecosystems.