Framework for Automatic Coral Reef Extraction Using Sentinel-2 Image Time Series

Abstract

Abstract Using supervised and unsupervised classification on a single image to extract coral reef extent results in missing data and wrong extraction results. To improve the accuracy of coral reef extraction, this study proposes a novel technical framework for automatic coral reef extraction based on an image filtering strategy and spatiotemporal similarity measurements of pixel-level Sentinel-2 image time series. This method was applied to the Anda Reef, Daxian Reef, and Nanhua Reef, China, using 1464 Sentinel-2 images obtained from 2015–2020. Sentinel-2 images were automatically selected considering space, time, cloud cover, and image entropy after atmospheric correction. With the binary classification measurement standard using the digitization coral reef results of the Sentinel-2 images as the true value, the time series established by the modified normalized difference water index demonstrated high robustness and accuracy. Analyzing the time series curves of the coral reef and deep water verified that the spatiotemporal similarity measurement of this framework can stably extract the boundaries of the coral reef.