Classifying coral shore regions, a deep learning approach

Abstract

Coral reef environments are important marine ecosystem and they are in decline due to climate change. Monitoring coral reefs is crucial but has its challenges since coral reefs are dynamic and data is sensitive to weather conditions. This often leads to difficulties for acquiring accurate data. Coral reef surveys are normally done during ideal conditions which are not always available when researchers are present. That is only one of several challenging aspects in coral reef classification. This study explores the possibilities of applying deep learning to classify drone imagery from coral reefs as a promising new approach for coral reef classification, but also as a new approach for producing geophysical output. This study also explores if it is possible to classify non ideal shore imagery. The possibility to work with non ideal imagery would provide more options for the analysis of coral reef surveys. The images are classified using a segmentation approach with a convolutional neural network. The binary classifications achieve high accuracies of 94% and 96% and IoU values of 77% and 71%. From the segmented output a coral density map is derived. The production of the coral density map is the one of the final steps towards producing promising geophysical output. The classification of non ideal imagery led to some difficulties, but it should be possible to fully correctly classify those as well, which can be discussed in further research