NMFS-OSI/noaa-pacific-benthic-cover-t3

Dataset Card for NOAA Pacific Benthic Image Classification Dataset Tier 3 (Fine-Grained)

Dataset Details

Dataset Description

This dataset contains benthic reef image crops for Tier-3 coral reef classification, built from CoralNet-derived annotations associated with NOAA Pacific Islands Fisheries Science Center (PIFSC) Ecosystem Sciences Division (ESD) monitoring programs.

Images are 224×224 crops associated with one label per crop and split into train/validation/test. Tier-3 labels are fine-grained functional and taxonomic classes (e.g., individual coral genera/morphologies, algae types) compared to the broad Tier-1 groupings. Metadata supports both direct cloud URL access and Hugging Face datasets streaming workflows.

• License: Public Domain (NOAA Open Data)
• Shared by: NOAA Open Data
• Curated by: Ecosystem Sciences Division (ESD) | Pacific Islands Fisheries Science Center (PIFSC)
• Image Size: 224 x 224 pixels (patch-based)

Dataset Sources

• Cloud Repository: Public Google Cloud Storage source bucket path: gs://nmfs_odp_pifsc/PIFSC/ESD/ARP/pifsc-ai-data-repository/coralnet_mirror/pacific_t3_splits

Dataset Structure

The dataset is split into:

• train: 265,046
• validation: 56,770
• test: 56,864

Total: 378,680 image crops across 67 classes.

Features (Parquet/HF metadata)

• image_url: HTTPS object URL in public GCS
• gcs_uri: canonical gs:// URI
• label: integer class index
• label_name: class short code
• label_desc: class description
• file_name: image file basename
• split: train/val/test marker

Classes

T3 contains 67 classes. The top 10 classes by image count (all splits combined) are shown below.
See class_list_full.csv for the full class inventory and class_counts_full.csv for per-split counts.

Rank	Class	Count (all splits)
1	SAND	29,691
2	TURFH	29,691
3	CCAH	29,691
4	PESP	29,691
5	MOEN	28,911
6	POMA	25,908
7	EMA	21,930
8	HALI	20,411
9	TURFR	15,409
10	POCS	12,539

Uses

Direct Use

Suitable uses:

• Training/evaluating fine-grained benthic cover classifiers (Tier-3)
• Transfer learning for coral-reef benthic imagery
• Benchmarking class-imbalance handling methods for many-class problems
• Building marine habitat monitoring model baselines at genus/morphology level

Out-of-Scope Use

• Species-level ecological inference beyond Tier-3 label semantics
• Regulatory or policy decisions without domain expert review
• Habitat trend inference without survey design/statistical context
• Identification or any application not related to benthic habitats

Dataset Creation

Curation Rationale

Goal: provide a reusable machine-learning-ready Tier-3 classification dataset aligned with NCRMP/PIFSC benthic image workflows, with cloud-native hosting and reproducible splits for model development at fine-grained resolution.

Source Data (See Below)

Source records and context are drawn from NOAA InPort metadata for:

• Stratified Random Survey (StRS) benthic images and annotated benthic cover products
• Climate-station/fixed site benthic images and annotated benthic cover products

Data Collection and Processing

High-level workflow:

1. Photoquadrat surveys collected by divers under NCRMP methods.
2. Benthic images annotated using CPCe (historically) and CoralNet (more recent periods).
3. Tier-3 labels retained at the fine-grained functional/taxonomic level.
4. Crops generated and assembled into class folders for train/val/test.
5. Metadata manifests generated for HF streaming and analysis.

Who are the source data producers?

Primary producers are NOAA PIFSC ESD teams and collaborators conducting NCRMP field missions and annotation operations in U.S. Pacific Islands regions.

Annotations

Annotation process

Annotation points were produced from benthic image analysis pipelines described in NOAA metadata. Human analysts assign fine-grained benthic category labels at points using Tier-3 label sets. This dataset retains those Tier-3 labels directly.

Who are the annotators?

Trained NOAA/PIFSC-associated analysts and partner contributors using program SOPs and calibration/quality-control procedures described in NCRMP metadata.

Bias, Risks, and Limitations

• Geographic Bias: Data is primarily from the Pacific Region; performance may differ in other regions.
• Environmental Bias: Imagery collected under certain lighting and seasonal conditions.
• Temporal variability and differences in camera conditions across years/missions
• Label uncertainty from difficult imagery and fine-grained taxonomic ambiguity
• Significant class imbalance across 67 classes; many rare classes have very few examples
• Not intended for species-level ecological claims beyond label semantics

Recommendations

• Use class-balanced sampling or loss weighting given strong imbalance across 67 classes
• Keep train/val/test separation strict to reduce leakage risk
• Consider hierarchical or Tier-1 auxiliary supervision to handle rare Tier-3 classes
• Validate externally on independent sites/years before deployment
• Consult expert guidance for ecological interpretation

Personal and Sensitive Information

• No personal, sensitive, or identifiable information is present.
• Images are of marine environments only.

---

Glossary

• NCRMP: National Coral Reef Monitoring Program
• PIFSC ESD: Pacific Islands Fisheries Science Center, Ecosystem Sciences Division
• StRS: Stratified Random Survey
• Tier-3: Fine-grained benthic functional/taxonomic classes

Metadata / Citation

Citation:
Pacific Islands Fisheries Science Center (2025). Ecosystem Sciences Division (ESD);

Related Metadata:

• Benthic cover from StRS annotations
• Benthic cover from climate-station annotations
• Benthic images from StRS Sites
• Benthic images from Fixed climate stations

Dataset Card Contact

For questions or inquiries, contact:
Michael Akridge – Michael.Akridge@noaa.gov

---

Related Datasets

• Broad-class T1 curated split dataset (8 classes, train/validation/test):
  https://huggingface.co/datasets/NMFS-OSI/noaa-pacific-benthic-cover-t1
• All-annotations unsplit T3 dataset (96 classes):
  https://huggingface.co/datasets/NMFS-OSI/noaa-pacific-benthic-cover-t3-all

---

Disclaimer

This repository is a scientific product and is not official communication of the National Oceanic and Atmospheric Administration, or the United States Department of Commerce. All NOAA project content is provided on an 'as is' basis and the user assumes responsibility for its use. Any claims against the Department of Commerce or Department of Commerce bureaus stemming from the use of this project will be governed by all applicable Federal law. Any reference to specific commercial products, processes, or services by service mark, trademark, manufacturer, or otherwise, does not constitute or imply their endorsement, recommendation or favoring by the Department of Commerce. The Department of Commerce seal and logo, or the seal and logo of a DOC bureau, shall not be used in any manner to imply endorsement of any commercial product or activity by DOC or the United States Government.