UECNN: an Underwater-Enhanced Convolutional Neural Network For Coral Reef Health Classification

Abstract

Coral reef ecosystems are dealing with an unprecedented drop right now because of climate change, ocean acidification, and human driven pressures, all stacking at once. So there really is a need for scalable and fairly meticulous automated systems to continuously monitor the health of coral reef. The current deep learning possibilities do not handle that well with the optical realities of underwater imagery, especially that depth dependent blue-green colour distortion, the low contrast that comes from water turbidity. Here we introduce UECNN (Underwater-Enhanced Convolutional Neural Network), a new architecture focused on splitting coral reef health into three simple groups: Bleached, Diseased, and Healthy. When we evaluated it on a dataset with 6,400 coral reef images, UECNN got a mean test accuracy of 90.95% ± 0.31% over three distinct seeds 42, 123, 456. Finally, the Grad-CAM visualizations align with the whole narrative.