Coral classification at scale

computer vision
machine learning
conservation
Image classifiers for coral growth forms and benthic attributes, trained on 500,000+ images with 14M+ annotations.

Role: built the models at Sparkgeo, with the CoralNet research initiative, for MERMAID

The problem

Reef surveys produce far more photos than skilled annotators can label. CoralNet holds many images and annotations, but the labels are inconsistent across projects. MERMAID wanted consistent, high-quality labels and models that classify coral growth forms and benthic attributes from transect survey imagery.

What I did

  • Built computer-vision models in Python with scikit-learn, PyTorch, DuckDB and Dask on S3 and EC2.

  • Trained them on 500,000+ images with 14M+ annotations.

  • Contributed to the grant writing that won a $2M Bezos Earth Fund award (2026) to move coral-reef deep-learning models from research into production for MERMAID. This grant was used for MermaidSeg to provide a segmentation model for users from images and video.

  • Contributed to the application that won the AWS IMAGINE Grant for Nonprofits (2024 to 2025). ## Result

  • 75%+ accuracy, and substantially less manual effort for skilled coral annotators.

  • Added to the MERMAID platform for users to classify coral.

Related: MermaidSeg and mermaid-classifier.