Combining Drones and Deep Learning to Automate Coral Reef Assessment with RGB Imagery

Abstract

Coral reefs and their associated marine communities are increasingly threatened by anthropogenic climate change. A key step in the management of climate threats is an efficient and accurate end-to-end system of coral monitoring that can be generally applied to shallow water reefs. Here, we used RGB drone-based imagery and a deep learning algorithm to develop a system of classifying bleached and unbleached corals. Imagery was collected five times across one year, between November 2018 and November 2019, to assess coral bleaching and potential recovery around Lord Howe Island, Australia, using object-based image analysis. This training mask was used to develop a large training dataset, and an mRES-uNet architecture was chosen for automated segmentation. Unbleached coral classifications achieved a precision of 0.96, a recall of 0.92, and a Jaccard index of 0.89, while bleached corals achieved 0.28 precision, 0.58 recall, and a 0.23 Jaccard index score. Subsequently, methods were further refined by creating bleached coral objects (>16 pixels total) using the neural network classifications of bleached coral pixels, to minimize pixel error and count bleached coral colonies. This method achieved a prediction precision of 0.76 in imagery regions with >2000 bleached corals present, and 0.58 when run on an entire orthomosaic image. Bleached corals accounted for the largest percentage of the study area in September 2019 (6.98%), and were also significantly present in March (2.21%). Unbleached corals were the least dominant in March (28.24%), but generally accounted for ~50% of imagery across other months. Overall, we demonstrate that drone-based RGB imagery, combined with artificial intelligence, is an effective method of coral reef monitoring, providing accurate and high-resolution information on shallow reef environments in a cost-effective manner.