Development of Coral-Coverage Estimation Method Using Deep Learning and Sea Trial: at Kujuku-Shima Islands

Comprehensive and effective survey methods of coral distribution are indispensable for environmental conservation in the sea. Observation methods by divers, autonomous underwater vehicles (AUVs), and aerial imagery have been investigated for decades. However, effective methods in turbid water have not been developed sufficiently. In this paper, we propose a practical coral-coverage estimation method by combining an effective survey system (SSS: Speedy Sea Scanner) and a deep-learning based estimation method. We tested the performance of the proposed method in Kujuku-Shima islands, Nagasaki, Japan. Experimental results showed that corals can be distinguished with accuracy of about 80% in places with relatively high transparency, and the error of coverage estimation is 10% or less.