Automated identification of benthic epifauna with computer vision

Benthic ecosystems are chronically undersampled, particularly in environments >50 m depth. Yet a rising level of anthropogenic threats makes data collection ever more urgent. Currently, modern underwater sampling tools, particularly autonomous underwater vehicles (AUV5), are able to collect vast image data, but cannot bypass the bottleneck formed by manual image annotation. Computer vision (CV) offers a faster, more consistent, cost effective and sharable alternative to manual annotation. We used TensorFlow to evaluate the performance of the Inception V3 model with different numbers of training images, as well as assessing how many different classes (taxa) it could distinguish. Classifiers (models) were trained with increasing amounts of data (20 to 1000 images of each taxa) and increasing numbers of taxa (7 to 52). Maximum performance (0.78 sensitivity, 0.75 precision) was achieved using the maximum number of training images but little was gained in performance beyond 200 training images. Performance was also highest with the least classes in training. None of the classifiers had average performances high enough to be a suitable alternative to manual annotation. However, some classifiers performed well for individual taxa (0.95 sensitivity, 0.94 precision). Our results suggest this technology is currently best applied to specific taxa that can be reliably identified and where 200 training images offers a good compromise between performance and annotation effort. This demonstrates that CV could be routinely employed as a tool to study benthic ecology by non-specialists, which could lead to a major increase in data availability for conservation research and biodiversity management.