TsinghuaCorals/reefnet_species_images

ReefNet Species Images

Full-resolution coral reef images with species-level annotations from the ReefNet dataset.

Dataset Description

This dataset contains 51,080 images from 54 CoralNet sources with 365,030 annotation patches across 245 species.

Each annotation in metadata.parquet specifies a patch center (Row, Column) within a full-resolution image. Crop a 224×224 patch centered at these coordinates for classification tasks.

Splits

The curated 92-class subset used for training has the following splits:

Split	Patches	Description
train	48,313	Training set (max 1000/class)
image_val	32,792	Image-level validation (10%)
image_test	33,090	Image-level test (10%)
source_val	8,074	Source-level holdout (WAPA Interns)
null	485,522	Not in curated split (other species/annotations)

Metadata Columns

Column	Description
file_name	Relative path: {source}/{image_name}
split	Split assignment (null if not in curated set)
in_curated_92class	True if in the 92-class training split
label_taxonomy	Full taxonomy string
label_coralnet	Original CoralNet label
source	CoralNet source name
Row, Column	Patch center coordinates
Ecoregion, Province, Realm	Geographic classification
Latitude, Longitude	GPS coordinates

Citation

@article{reefnet2024,
  title={ReefNet: A Large-Scale Coral Reef Annotation Dataset},
  year={2024},
  note={https://huggingface.co/datasets/ReefNet/ReefNet-1.0}
}