2024 GBR in-water colony level bleaching severity dataset and AIMS Bleaching ViT Model

Updated: 2026-09-01

This dataset was developed to support the training and validation of colony-level artificial intelligence (AI) models for assessing coral bleaching severity on the Great Barrier Reef (GBR). The research is important because it enables automated, standardized estimation of bleaching impacts while maintaining compatibility with historic in-water coral bleaching survey methods. The data were collected to create a high-quality annotated benchmark for detecting and classifying coral colonies into five ordinal bleaching severity categories, ranging from unbleached to recent mortality. The dataset was derived from photo-quadrat images collected during February and March 2024 as part of the Australian Institute of Marine Science (AIMS) 2024-2025 bleaching response program. Images were obtained from standardized photo-transect surveys comprising three replicate 50 × 1 m transects per site, with each image representing a 1 × 1 m nadir-view photo-quadrat. Data collection involved manual segmentation and annotation of coral colonies. Expert coral reef ecologists assigned polygon masks, taxonomic classifications, and one of five bleaching severity categories to each colony. A total of 344 images containing 4,576 colony-level annotations were generated, with all annotations independently quality checked to minimize observer bias. Additional annotated coral segments were included to augment the training dataset. The collected data are analysed using AI-based image analysis techniques to train and validate models that estimate colony-level bleaching severity. Analyses focus on predicting ordinal bleaching categories for dominant hard-coral morphologies while ensuring robust performance across varying coral communities and bleaching intensities observed during the 2024 GBR bleaching event. For more information refer to the manuscript Expert validated colony-level coral bleaching AI model for image-based coral reef monitoring programs (Remmers et al, 2026)[Manuscript in preparation].