CoralNet: Curated Coral Reef Image Datasets

• CoralNet sources are a diverse collection of coral reef image datasets, annotated and curated globally to support ecological research and machine learning applications.
• The datasets undergo rigorous multi-stage curation including filtering, expert re-annotation, and taxonomic mapping with WoRMS, achieving over 90% agreement in refined subsets.
• Their integration into benchmarks like ReefNet highlights challenges in domain adaptation and class imbalance while enabling robust cross-source evaluation.

CoralNet sources are a diverse collection of annotated coral reef image datasets, contributed by a broad network of researchers and institutions, and curated on the CoralNet platform. These sources underpin major advances in automated coral monitoring, robust machine learning model development, and cross-domain benchmarking. Through rigorous curation, taxonomic standardization, and expert verification, CoralNet sources have become foundational for high-fidelity, large-scale coral classification datasets—enabling ecological research and conservation at global scale.

1. Data Aggregation and Curation from CoralNet

CoralNet, as an open-access annotation and analytics platform, hosts more than 1,300 individually contributed sources, each representing collections of images and point-based labels from coral monitoring programs worldwide. ReefNet (Battach et al., 19 Oct 2025) exemplifies the systematic aggregation of this resource, applying a multi-stage filtering pipeline to select sources with high ecological and technical relevance, sufficient image and annotation counts, and appropriate focus on reef-building hard corals. Non-pertinent sources (e.g., calibration sets or datasets lacking adequate coverage) are excluded.

The remaining sources are subjected to re-annotation protocols whereby marine biologists refine and verify point labels according to consensus taxonomic standards and map them to the World Register of Marine Species (WoRMS). This process reduces label ambiguity and achieves expert agreement rates exceeding 90% in some subsets, as compared to the initial 73% typical of raw, crowd-annotated data.

2. Taxonomic Mapping and Standardization

A persistent challenge in integrating CoralNet sources into machine learning workflows is heterogeneity in label granularity and nomenclature. To mitigate this, sources included in benchmarking datasets such as ReefNet are re-mapped to a unified taxonomic schema at the genus level (with Fungiidae maintained at family due to verification difficulty). All annotation labels are reconciled with WoRMS, ensuring global consistency and up-to-date systematic classification.

This taxonomic normalization enables fine-grained morphological studies, robust ecological comparison across geographic regions, and downstream compatibility with other global biodiversity databases.

3. Benchmark Datasets and Evaluation Protocols

By leveraging curated CoralNet sources, large-scale datasets such as ReefNet (Battach et al., 19 Oct 2025) have been constructed, comprising approximately 925,000 genus-level point annotations spanning 76 sites and covering diverse marine provinces (e.g., Indo-Pacific, Red Sea). Each source’s imagery is supplemented with metadata (location, camera settings, etc.).

ReefNet defines two primary evaluation settings:

Benchmark Type	Training/Testing Strategy	Purpose
Within-Source	Train/validate/test splits within each CoralNet source	Measures local domain accuracy
Cross-Source	Train on multiple sources, test on entirely withheld CoralNet sources	Measures generalization

These stratified protocols are designed to prevent leakage and ensure that cross-source evaluations genuinely probe domain shift, reflecting deployment realism where models face previously unseen reefs, imaging conditions, or annotation practices.

4. Ecological and Technical Significance

The breadth and annotation quality of CoralNet sources aggregated in ReefNet allow for detailed ecological inquiry:

• Fine-grained taxonomic discrimination, including rare or morphologically similar genera, across variable reef environments.
• Spatial and temporal variability can be assessed, supported by rich metadata embedded during curation.
• A rigorously verified label set enables robust model comparison, as all entries share common annotation protocols and taxonomy.

Technically, the diversity of imagery, sampling locations, and annotation origins introduces significant covariate shifts, presenting a substantive challenge for both supervised and zero-shot learning approaches. Models typically display a marked drop in performance in cross-domain settings, highlighting the necessity of domain-adaptive algorithms.

5. Advances in Machine Learning and Conservation

ReefNet’s integration of CoralNet sources provides a publicly available, ML-ready platform that catalyzes research across several axes:

• Facilitates the development of domain generalization methods, zero-shot classifiers, and robust taxonomic recognition pipelines.
• Benchmarks modern approaches in terms of Macro Recall and Micro-accuracy, and explores advanced techniques such as focal loss for class imbalance.
• Inspires broader data standardization and expert-in-the-loop curation practices globally.

The platform supports both ecological analysis—by providing consistent, expert-verified annotations for population and community structure—and machine learning innovation by serving as a challenging and realistic testbed.

6. Limitations and Future Directions

The transition from raw CoralNet sources to a benchmark like ReefNet is nontrivial: it requires significant manual re-annotation and curation efforts to ensure ecological fidelity and taxonomic consistency. Persistent challenges include:

• Handling cases where ground-truth verification is difficult (e.g., maintaining family-level resolution for some taxa).
• Managing class imbalance, as rare genera remain under-represented even in large-scale compilations.
• Domain adaptation, evidenced by substantial performance gaps between within-source and cross-source benchmarks.

A plausible implication is that future work will benefit from expanding annotation coverage, developing semi-automated expert verification tools, and improving model architectures for domain-robust classification.

7. Summary Table: CoralNet Source Utilization in ReefNet

Process	Purpose	Outcome
Filtering & Curation	Select high-quality, relevant sources	76 expert-verified, globally distributed
Taxonomic Mapping to WoRMS	Standardize genus/family assignments	Unified label set for ML compatibility
Metadata Integration	Support technical/biogeographic analysis	Enhanced contextual information
Benchmark Partitioning	Prevent overfitting, test generalization	Within- and cross-source benchmarks
Expert Re-Annotation	Reduce annotation noise/ambiguity	Expert agreement >90% in some subsets

The aggregation, verification, and standardization of CoralNet sources set a new standard for large-scale, taxonomically consistent coral monitoring datasets, serving both the ecological and computational communities. ReefNet’s rigorous use of these sources underscores their ongoing centrality in global coral reef research and machine learning benchmarking.