A High-Quality Underwater 3D Reconstruction Solution for Coral Reef Environments Leveraging Advanced Photogrammetric Computer Vision Techniques

Coral reefs, with their intricate 3D structures, are vital marine ecosystems increasingly threatened by environmental stressors. Accurate 3D reconstruction of these underwater structures is essential for scientific research and conservation management. While underwater photogrammetry has emerged as a promising tool for this purpose, technical challenges persist in capturing fine-scale features under underwater conditions, particularly the intricate morphology of coral reefs and the highly complex textures that hinder high-fidelity reconstruction. Recent breakthroughs in computer vision and deep learning have introduced new opportunities for underwater photogrammetry. This study begins by outlining a photogrammetric workflow that can integrate current advanced technologies, followed by summarizing cutting-edge methods used in the key stages, i. e., sparse and dense reconstruction. Building on previous research, we propose a hierarchical reconstruction strategy for accurate and efficient dense modeling. Our approach first performs an efficient global coarse-grained reconstruction to capture the overall scene structure, followed by fine-scale modeling in key regions of interest. Using image data collected from a coral reef site at Moorea Island, we compare and evaluate various techniques, analyzing their respective strengths and limitations. In sparse reconstruction, the classical feature method scale-invariant feature transform demonstrates competitive performance. Deep learning–based methods, such as ALIKED feature and the SuperGlue matching network, achieve superior results on certain metrics. For dense reconstruction, Neural Radiance Fields and 3D Gaussian Splatting–based methods yield high-quality reconstructions but are computationally intensive. In contrast, the deep learning–based multi-view stereo approach achieves comparable reconstruction quality with greater efficiency. Experimental results on reconstruction result fusion further validate that our approach offers a scalable and practical solution for coral reef monitoring, advancing conservation science and ecosystem management practices.