An integrated pipeline to quantify benthic community composition and structural complexity on coral reefs by combining structure-from-motion, photogrammetry and machine learning.
Supervisor: Renato Morais (Univ. Perpignan)
This research project aimed to develop an integrated pipeline for analysing coral reef ecosystems, combining advanced techniques in machine learning, photogrammetry, and structure-from-motion. The primary objective was to extract comprehensive data on community composition and structural complexity, with a focus on demonstrating the pipeline’s efficacy for large-scale coral reef studies. The study addressed a significant gap in coral reef research by incorporating the fundamental geometric relationship between three key complexity metrics: surface rugosity, fractal dimension, and height range. This approach offers a more comprehensive 3D characterization of structural complexity than previous studies. A protocol was developed for processing extensive datasets of benthic images using photogrammetry techniques. Additionally, a custom R routine was created to digitally quantify the three metrics representing the 3D characterization of structural complexity on coral reefs. Due to time constraints and the volume of data, the study focused on a single location: the Dongsha Atoll located in Taiwan. This site is part of a broader project, including three other locations, investigating seascape drivers of high and low fish biomass production on coral reefs. In this study, 10 Digital Elevation Models and ortho-photo-mosaics were generated, enabling the quantification of benthic community composition and structural complexity metrics. The analysis indicated that outer rim habitats exhibited higher rugosity and fractal dimension compared to lagoon habitats. These results support the notion that both biotic and abiotic factors significantly influence the structural complexity of reef systems within an atoll. This observed variation in structural metrics highlights the complex interplay between environmental conditions and biological communities in shaping reef topography. For benthic annotation analysis, the semi-automatic annotation tool from TagLab was employed, significantly reducing annotation time, although its use was limited due to time constraints. This thesis represents a novel approach in coral reef research, combining Structure from Motion (SfM) photogrammetry with AI-based interactive image segmentation software. It establishes a pipeline capable of monitoring structural complexity variations and benthic composition at scale. However, further investigations are necessary to refine and optimize these innovative techniques. The study’s findings contribute to the growing body of knowledge on coral reef structural complexity and offer promising methodologies for future large-scale reef monitoring and analysis projects.