A survey on underwater coral image segmentation based on deep learning

Abstract

ABSTRACT Image-based coral reef survey technologies have revolutionized the monitoring of coral reefs by offering a cost-effective and noninvasive method for collecting data across large spatial scales and extended periods. Among these technologies, underwater videography has emerged as a well-established and reliable tool for remote sensing in coral research. Automatic segmentation of coral images represents a forward-looking and fundamental research area in underwater remote sensing. It aims to address a major challenge that limits traditional in situ underwater coral survey research: the difficulty of automatically generating accurate and reproducible high-resolution maps of the underlying coral reef ecosystems. Understanding recent achievements and their relevance to coral ecology monitoring needs is crucial for future planning. This paper presents a literature review on underwater coral image segmentation, focusing on the deep learning implementation pipeline. Furthermore, we introduce a new densely annotated dataset specifically designed for the semantic segmentation of underwater coral images. We systematically evaluate State-of-the-Art (SOTA) methodologies and novel techniques not previously applied to coral image semantic segmentation using the proposed dataset. We then discuss their feasibility in this context. Our goal for this review is to spark innovative ideas and directions for future research in underwater coral image segmentation and to provide readers with an accessible overview of some of the most significant advancements in this field over the past decade. By accomplishing these objectives, we hope to advance research in underwater coral image segmentation and support the development of effective monitoring and conservation strategies for coral reef ecosystems.