Taxonomic resolution of coral image classification with Convolutional Neural Network

Coral reefs are the most complex, diverse, and sensitive marine ecosystems, which are globally undergoing drastic changes. Changes in coral coverage, abundance, and diversity are difficult to track at adequate taxonomic resolution in a fast and efficient way. Deep Learning-enabled image recognition can help to increase the accuracy and can add to automating the entire survey process. However, the extent to which coral phylogenetic relationships within the higher taxonomic ranks are reflected by shared visual traits of the constituent species is an unresolved research subject. As a consequence, it is even more questionable whether the taxonomy of coral reefs at these levels can be identified from images using Machine Learning techniques. In this study, we analyzed the performance of Convolutional Neural Networks (CNN) through different taxonomic ranks to classify the underwater images. We have used 1,15,296 images from the CoralNet database comprising 104 species. A classifier was developed by fine-tuning the pre-trained ResNet34. The developed CNN classified 87.5% of 34,543 test images correctly to species level and 91.78% to genus level. The average classification recall on species level was 83.99%. The ability to classify coral photographs at the species level can significantly boost the amount of occurrence data collected from survey sites. Even anticipating a higher taxonomic level will be a great place to start for additional research on a completely unknown specimen. Hence, the new technique will be a very useful tool for forecasting and understanding ecological responses to the environmental changes.

Automated summary

This study analyzed the performance of Convolutional Neural Networks (CNN) in classifying underwater images at different taxonomic ranks. The results showed that deep learning techniques can accurately classify images to species and genus levels. The application of this technique can enhance data collection efficiency and quantity, facilitating the prediction and understanding of ecological responses to environmental changes.