ReefNet: A Large-Scale Dataset and Benchmark for Fine-Grained Coral Reef Recognition
Authors: Abdulwahab Felemban, Yahia Battach, Faizan Farooq Khan, Yuqian Fu, Xuhui Liu, Yesmeen M. Khattab, Yousef A. Radwan, Xiang Li, Natalie Dunn, Susanne Bähr, Tullia Isotta Terraneo, Fabio Marchese, Sara Beery, Luis Silva, Rhonda Suka, Karla Gonzalez Martinez, Marta Ezeta Watts, Federica Barreca, Nicolas Oury, Burton H. Jones, Francesca Benzoni, Mohamed Elhoseiny
Organizations: King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia · Massachusetts Institute of Technology (MIT), Cambridge, MA, USA
Coral reef monitoring requires scalable recognition of visually similar taxa across heterogeneous underwater surveys. We introduce ReefNet, a public dataset that aligns approximately 925K hard-coral point annotations from 76 CoralNet sources and an independently collected Al-Wajh dataset with the World Register of Marine Species taxonomy. A stratified expert audit of 8,962 patches guides filtering to a 479,027-annotation benchmark with 39 labels: 38 genera and the family Fungiidae. Expert agreement is 92% on the retained audit subset, rather than an exhaustive accuracy estimate. We evaluate pretrained models through zero-shot recognition, few-shot adaptation, and full fine-tuning, each on within-source ReefNet and held-out-source Al-Wajh tests. Biological vision-language pretraining provides the strongest zero-shot results; supervised adaptation substantially improves recognition, with DINOv2 reaching 65.79% macro recall on Al-Wajh. Matched text-description experiments improve all four evaluated MLLMs on Al-Wajh, while a GPT-5.4 reasoning ablation does not improve over direct classification. Long-tail recognition, label uncertainty, and geographic imbalance remain challenges. ReefNet provides a taxonomically grounded resource for studying these challenges; the Al-Wajh test is a regional stress test rather than evidence of global robustness. Public data and resource links are documented in the supplement.
Figures & tables
Figure 1: ReefNet benchmark overview. All crops from one original image share a split. Adaptation uses ReefNet training points and validation-based selection; zero-shot models use no ReefNet training examples. Every regime is evaluated on both test sets. Al-Wajh never enters training or selection; its metric averages over 13 present genera. The standard protocol retains 39 output labels; Claude/GPT API prompts instead list 13 Al-Wajh candidates. Photos illustrate the protocol and do not identify actual split membership.
Number of annotations
Dataset
Geographic coverage
Annotation type
Lowest taxonomic level
Mapped to WoRMS
Number of images
Hard corals
Genus-level annotations
CoralNet Beijbom et al. (2015)
World
Sparse points
Species
No
4,524,792*
N/A
No
CoralSCOP Zheng et al. (2024)
World
Masks
Order
No
41,297
330,144 ‡
No
CoralVOS Ziqiang et al. (2023)
17 sites (South China Sea)
Masks
Order
No
60,456
N/A
No
MosaicsUCSD Edwards et al. (2017)
Palmyra
Masks
Species
No
16 †
44,008
Yes
Eilat Raphael et al. (2020)
Eilat
Sparse points
Genus
No
212
∼ 12,000
Yes
Table 1: Comparison with existing coral classification datasets. *Image counts reflect availability as of May 13, 2025. Hard coral/genus-level annotations are not specified for CoralNet because its labels are not standardized. † MosaicsUCSD includes 16 annotated orthomosaics, each from ∼ 1.5K images. ‡ CoralSCOP Zheng et al. (2024) contains 330,144 binary coral/non-coral masks.
Figure 2: ReefNet data construction pipeline. We apply source filtering, WoRMS taxonomy alignment, semantic label consolidation, and expert verification to construct ReefNet, yielding ∼ 925K point annotations across 39 hard-coral label classes together with a new Red Sea test set from Al-Wajh Lagoon.
Figure 3: Distributions of Dataset Annotations. (a) The 445,985 benchmark training samples across 39 labels (38 genera and Fungiidae), on a log scale. Dashed separators mark the fixed head/body/tail groups; annotations give observed training-count ranges. (b) Full-release geographic distribution of annotations; nearby sources are grouped, and marine ecoregions are shown in light blue Spalding et al. (2007) .
Split
ReefNet
Al-Wajh
Annotations
Classes
# Sources
Annotations
Classes
# Sources
Train
445,985
39
68
–
–
–
Val
9,999
38
63
–
–
–
Test
23,043
39
66
4624
13
1
Table 2: Dataset Split Statistics. ReefNet provides train/val/test splits covering 39 coral classes (38 in validation). Al-Wajh is used only for cross-source testing and contains 4,624 annotations from 13 classes overlapping with the ReefNet label set.
Name only
Name + description
Model
ReefNet
Al-Wajh
ReefNet
Al-Wajh
InternVL-3.5 Wang et al. (2025)
4.0
9.8
3.2
10.7
LLaVA-NeXT Li et al. (2024b)
2.6
8.3
–
–
LLaVA-OneVision Li et al. (2024a)
3.3
9.3
–
–
MiniCPM-V 4.5 Yao et al. (2024)
3.6
9.5
4.4
12.5
Qwen3-VL Bai et al. (2025)
5.1
15.2
5.9
16.8
Table 3: Zero-shot macro recall (%). Best and second-best available scores are bold and underlined within each column , separately for each text condition. Dashes indicate unavailable description results. Both text conditions use the same test sets. † Claude/GPT API prompts list 39 candidates on ReefNet and 13 on Al-Wajh, unlike the standard 39-label protocol; Al-Wajh comparisons across these candidate sets are not matched. API access: 26–27 July 2026. Appendix F states the documentation scope for the open-model description runs.
Model
Within-Source Test (ReefNet)
Al-Wajh Test
1-shot
5-shot
10-shot
20-shot
1-shot
5-shot
10-shot
20-shot
Qwen3-VL Bai et al. (2025)
5.56
9.15
17.24
27.22
10.93
13.77
25.87
31.27
MiniCPM-V 4.5 Yao et al. (2024)
2.85
7.32
14.95
21.95
7.68
8.52
12.77
20.88
InternVL-3.5 Wang et al. (2025)
2.00
2.87
10.44
18.34
10.31
9.01
12.80
12.87
CLIP Radford et al. (2021)
8.10
23.02
38.41
45.71
23.16
29.38
47.64
57.30
OpenCLIP Ilharco et al. (2021)
9.70
21.45
39.61
45.95
18.40
23.27
45.84
46.51
Table 4: Few-Shot Hard Coral Classification Macro Recall (%). Single-run models are fine-tuned with k examples per label from ReefNet. Bold and underline denote the best and second-best result in each column. Results are reported for the within-source ReefNet test set and the cross-source Al-Wajh test set.
Category
Model
ReefNet Test
Al-Wajh Test
Vision
DINOv2 ViT-L/14 Oquab et al. (2024)
79.27
65.79
DINOv3 ViT-L/16 Siméoni et al. (2025)
77.42
59.76
SigLIP 2 ViT-L/16 (visual) Tschannen et al. (2025)
74.21
56.71
Public MAE ViT-L/16 He et al. (2022)
61.24
49.74
ViT-B/16 Dosovitskiy et al. (2021)
79.84
54.84
EfficientNet-B3 Tan and Le (2019)
71.29
52.58
Table 5: Full-data fine-tuning macro recall (%). Scores are single-run point estimates; best and second-best across all listed models in each column are bold and underlined. Vision/VLM training uses 20 epochs; MLLM LoRA uses two, without compute matching. ViT and BioCLIP2 seed variability is reported separately in Appendix C.5 .
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: ReefNet Curation Pipeline. The diagram illustrates the sequential filtering and exclusion steps applied to 1,366 CoralNet image sources to arrive at the final ReefNet dataset of 76 sources. Sources were excluded due to an insufficient number of human-verified annotations, ecological or privacy concerns, and calibration-related issues.
Figure 5: Screenshot of the ReefNet manual verification platform. Experts assess genus-level annotations by reviewing image patches alongside full-image context and assign structured labels through a dropdown interface. The tool supports efficient, large-scale quality control of coral annotations.
Figure 6: Distribution of annotations, images, and sources across marine ecoregions. Each bar group represents one of 26 marine ecoregions covered in the ReefNet dataset, as defined by the MEOW classification. The plot shows the number of annotations, distinct images, and image sources per ecoregion (log scale). Geographic coverage is notably high in regions like Hawaii, Samoa, and the Mariana Islands, but more limited in some Atlantic and Western Indian Ocean regions.
Figure 7: Map showing the origin of the imagery in the Al-Wajh test dataset. Panel a shows the location of panel b and c in the Red Sea. Panel b shows the Al-Wajh lagoon with the specific locations of the annotations in the test set. Panel c shows the location of a limited set of images taken outside the Al-Wajh lagoon. Map contains Natural Earth data and Bing satellite imagery.
Figure 8: Log-scale Distribution of Annotations Across 13 Hard Coral Genera in the Al-Wajh Lagoon Dataset. The distribution shows a bias toward some classes, comparable to the distribution in ReefNet.
Observed labels per crop
Crops
Fraction (%)
One
449,314
93.80
Two
26,991
5.63
Three
2,431
0.51
Four or more
291
0.06
Appendix
Table 6: Label composition of all 479,027 benchmark crops, measured using observed point annotations.
Set
Pairs
Any overlap (%)
Mean overlap (%)
All pairs
Overlapping pairs
ReefNet benchmark
2,891,942
43.75
13.77
31.47
ReefNet test
111,895
42.83
13.70
31.98
Al-Wajh
8,473
23.57
7.39
31.34
Appendix
Table 7: Spatial overlap among same-image crop pairs. The ReefNet benchmark row includes training, validation, and test crops.
Table 8: Pretrained initialization and preprocessing for the four added visual backbones. Input dimensions and resize lengths are in pixels. Each model uses a 39-label classifier, 20 fine-tuning epochs, and checkpoint selection on ReefNet validation. ImageNet normalization uses the mean and standard deviation listed in General Settings; ReefNet denotes the dataset-specific mean/std used for those runs. SigLIP 2 uses its visual encoder, and MAE uses the public pretrained checkpoint.
Model
Test
Condition
Empty
Off-list
Claude Opus 5
ReefNet
Direct, names
0
2
Claude Opus 5
Al-Wajh
Direct, names
0
0
GPT-5.4
ReefNet
Direct, names
0
6
GPT-5.4
ReefNet
Reasoning, names
1
0
GPT-5.4
ReefNet
Direct, descriptions
0
4
GPT-5.4
Al-Wajh
Direct, names
0
7
Appendix
Table 9: Invalid API predictions retained as errors. Off-list counts exclude empty predictions. Each ReefNet run has 23,043 samples; each Al-Wajh run has 4,624.
Model
Training
Test
Run 1
Run 2
Run 3
Mean ± SD
ViT-B/16
Full
ReefNet
79.84
79.90
79.01
79.58±0.50
ViT-B/16
Full
Al-Wajh
54.84
55.37
55.62
55.28±0.40
BioCLIP2
Full
ReefNet
79.23
78.96
78.39
78.86±0.43
BioCLIP2
Full
Al-Wajh
58.28
57.83
57.80
57.97±0.27
Qwen3-VL
Capped
ReefNet
50.96
50.45
50.16
50.52±0.41
Qwen3-VL
Capped
Al-Wajh
34.72
34.68
32.66
34.02±1.18
Appendix
Table 10: Macro recall (%) across three independent seeds per study, with mean and sample standard deviation. Run indices do not denote numeric seed values. Qwen3-VL uses a reduced, class-capped training set and is not a full-data comparison.
Figure 9: Head, body, and tail macro recall. Selected original-suite models on the ReefNet within-source test set (39 labels, 23,043 samples). Each panel separates full fine-tuning, few-shot adaptation, and zero-shot recognition. The lower few-shot segment shows k=5 recall and the total bar height shows k=20 recall. Groups contain 4 head, 23 body, and 12 tail labels, defined by training frequency. The ViT series uses the auxiliary per-class record described below.
Figure 10: Per-class recall of ViT-B/16 on the ReefNet within-source test set. Classes are sorted by recall (ascending). Dark blue bars indicate the four head genera (collectively 78.4% of training samples); medium blue bars are body genera; red bars identify classes with fewer than 200 test samples; this test-count criterion is distinct from the training-frequency groups above. The dashed line shows the macro recall (79.8%).
Figure 11: Normalized confusion matrix for ViT-B/16 on the ReefNet within-source test set. Each row is normalized by the true class count (row sum = 100%), so diagonal entries equal per-class recall. Notable off-diagonal clusters include the abundant genera ( Acropora , Pocillopora , Montipora ), the compact-corallite group ( Acanthastrea , Dipsastraea , Favites ), and pairs such as Platygyra ↔ Favites and Astreopora → Montipora .
Model
Macro Recall
Macro Precision
Macro F1
Accuracy
Vision Models
ResNet-50
45.6
55.9
48.3
80.7
EfficientNet-B3
71.3
75.7
72.4
90.2
ViT-B/16
79.8
83.0
80.7
94.6
DeiT-B/16
76.2
80.8
77.6
93.7
BEiT-B/16
66.7
73.7
71.0
89.6
Appendix
Table 11: Additional metrics on the ReefNet within-source test set (39 labels, 23,043 samples). Macro recall—the primary metric reported in the main paper—is included for reference. Precision and F1 may be affected by classes absent from model predictions; a caveat noted in Section D.3 . Bold indicates the best result within each model group.
Ground truth
BioCLIP2
Confidence
Qwen3-VL
BioCLIP2 confident but wrong; Qwen3-VL correct (273 cases)
Cyphastrea
Plesiastrea
1.00
Cyphastrea
Porites
Leptoria
0.99
Porites
Pavona
Montipora
0.98
Pavona
Both predict the same wrong genus; BioCLIP2 confident (280 cases)
Favites
Goniastrea
1.00
Goniastrea
Appendix
Table 12: Reported examples of complementary and shared classification failures. Confidence values belong to BioCLIP2; the 273 and 280 counts describe the corresponding groups, with three examples shown from each.
Figure 12: Historical ReefNet/CoralNet qualitative examples from the 2025 cross-source split 00 (34,040 test crops), using an MAE-pretrained ViT. These are not from the current within-source benchmark. GT: ground truth; Pred: prediction. Green/red prediction text indicates correct/incorrect classifications.
Figure 13: Historical Al-Wajh qualitative examples from the 2025 MAE-pretrained ViT experiment (4,606 test crops, 512×512 pixels), distinct from the current 4,624-sample benchmark. GT: ground truth; Pred: prediction. Green/red prediction text indicates correct/incorrect classifications.
GPT-5.4 response condition
Reference text
ReefNet
Al-Wajh
Direct classification
Names
14.18
26.39
Reasoning before label
Names
12.33
22.91
Direct classification
Names + descriptions
14.92
32.22
Reasoning before label
Names + descriptions
Not run
30.39
Appendix
Table 13: GPT-5.4 reasoning ablation: macro recall (%). GPT prompts use 39 candidates on ReefNet and 13 on Al-Wajh. The within-source reasoning-plus-description condition was not run.
Label
Description
Acanthastrea
The Acanthastrea coral is characterized by its large, fleshy polyps, which are circular to irregular in outline, often showcasing a mottled palette of dull browns, greens, and occasional vivid hues. Its colony is encrusting and broad, resembling a flat, stone-like surface, accentuated by spine-like structures and gentle, folded textures that convey a spiny, rich complexity.
Acropora
Acropora is a vibrant, table-like coral with porous skeletal structures, exhibiting bright blue, green, yellow, or brown hues. It features flattened branches with ridges and unique tubular corallites at tips, growing over 2 meters wide, forming dynamic underwater habitats.
Agaricia
Agaricia corals are massive and plate-like, featuring tan to yellow-brown colors with deep valleys and irregular fronds. They form complex horizontal plates, characterized by clear septocostae and a substantial, textured surface.
Astreopora
Astreopora is a massive, hemispherical coral with rough, spine-covered surfaces in cream, beige, or brown, displaying crater-like corallites. These distinct, irregularly spaced structures reveal vibrant polyp tissue in hues of white, bright green, or yellow.
Colpophyllia
Large, domed colonies resembling a human brain, with meandering brown ridges and valleys in whitish, green, or tan. Can grow up to 2 meters, robust, with retracted polyps by day, color variations enhancing its distinctive appearance.
Cyphastrea
The Cyphastrea coral genus features massive, encrusting to columnar colonies with smooth or hillocky surfaces. Color varies from mottled grey to vibrant hues, with rounded corallites exhibiting 10-12 primary septa, creating a visually dense and intricate structure.
Appendix
Table 14: Exact 39-label description texts used in GPT-5.4 prompts. Al-Wajh prompts include only the 13 candidates listed in Appendix C.4 .
Figure 14: Density plots per CoralNet source of the average Red, Green, and Blue values of each image.
Figure 15: Scatter plot showing the resolution of each annotated image per source.
CoralNet source
Listed contributors
Listed affiliations
2002
Zuhairah Dindar
University of New South Wales
2016
Zuhairah Dindar
University of New South Wales
2Maldives_2Katie
Katie Lubarsky, Hugh Runyan
University of California San Diego
403 He’eia Restoration
Cynthia Hunter, Dorian Brunzelle, Kaitlyn Jacobs, Alana Minato, Cody Powers, Connor Antonis, Devon Stapleton, Dylan Rich, Jeany Robledo, Jacob Nygaard, Jordan Pounds-Crihfield, Jacquelyn Simpson, Kevin Christensen, Kaitlin Hooper, Madeline Payne, Renee Wold, Shayna Arakaki, Zachary Clark
Hawaii Institute of Marine Biology, University of Hawaii at Manoa
403 He’eia Restoration ’21
Cynthia Hunter, Dorian Brunzelle, Kaitlyn Jacobs, Keisha Bahr, Brooklyn Bennett, Marisa Bhao-Intr, Brittany Kernodle, Corey Ling, Dan Zhuo, Desiree Shaw, Johann Vollrath, Kenzie Vierra, Lena Marinkovich, Lauryn Pisciotto, Mellisa Gajardo, Madeleine Perez, Sophia Hanscom, Samantha Thomas, Toranosuke Degawa, Zada Boyce-Quentin
Hawaii Institute of Marine Biology, University of Hawaii at Manoa
Ain Sukhna Sea Urchin and Coral Cover
Omar Attum
Indiana University Southeast
Appendix
Table 15: Contributors and affiliations for the 85 CoralNet sources retained at the pre-final curation stage (76 sources in the curated release; 68 in the classification benchmark). This list is based on publicly available information from the CoralNet platform, supplemented by our in-depth online research into each source and outreach to their original contributors.
In order to design conservation and restoration strategies to counter the global decline of coral reefs, ecological monitoring of reefs needs to be scaled up dramatically. Computer vision methods are increasingly used to tackle the vast amount of data: as the paradigm of data collection in reefs shifts from highly standardized and constrained survey images to unconstrained imagery on scalable platforms, it is necessary to design machine learning methods that help to get a fine-grained understanding of reefs from general-purpose reef imagery. This paper provides CoralscapesV2, an extension of the Coralscapes dataset for general-purpose visual scene understanding in reefs. CoralscapesV2 increases the dataset size, scope, label completeness and quality for semantic segmentation, and extends the number of classes from 39 to 95 fine-grained visual categories. Furthermore, CoralscapesV2 provides 65k exhaustive fish instance mask annotations, meticulously annotated to completeness by using the video, revealing that annotation of fish based on only static images is insufficient. CoralscapesV2 is the first dataset for panoptic segmentation in coral reefs, capturing a wide range of scenarios in the wild, posing a challenging benchmark for contemporary semantic segmentation and instance segmentation models. CoralscapesV2 is an important step towards general-purpose panoptic segmentation in coral reefs, which has substantial implications for scaling up coral reef monitoring, as it can be employed in a wide range of applications from benthic cover mapping from robot or handheld videos to designing methods for automated quantification and understanding of fish behavior and fish-reef interactions.
Jonathan Sauder, Thomas Ruckli, Gabrielė Strodomskytė +17
Automated image recognition is increasingly used to scale ecological monitoring beyond manual annotation, yet ecologists lack evidence-based guidance on how much labelling effort reliable deployment at new sites requires. We present a decision framework quantifying the trade-off between labelling effort and recognition accuracy when transferring vision systems across marine habitats. The benchmark spans five datasets, three oceans, and three taxonomic groups (fish, corals, invertebrates), from tropical reefs in the Great Barrier Reef and French Polynesia to a temperate Danish fjord. We evaluated four recognition models (DINOv2, CLIP, ResNet-50, EfficientNet-B4) under four adaptation strategies (linear probing, LoRA, Visual Prompt Tuning, full fine-tuning) across three protocols: within-habitat transfer across 20 reef sites (240 runs), cross-dataset geographic transfer along a difficulty gradient (40 runs), and few-shot adaptation curves with 0-100 labelled samples per class (648 runs). Frozen self-supervised foundation features (DINOv2 + linear classifier, 1,538 trainable parameters) generalised to unseen reef sites at least as well as fully fine-tuned convolutional baselines four orders of magnitude larger; they learned species-diagnostic, habitat-invariant representations, whereas baselines encoded habitat-specific shortcuts that fail at new sites. As few as 10-20 labelled images per species sufficed to deploy reliable recognition at a new site, cutting annotation effort by roughly an order of magnitude. Solution. Programmes expanding to new sites can deploy reliable recognition by pairing a frozen, open foundation model (DINOv2) with a simple linear classifier and annotating only 10-20 images per species - roughly 1-4 hours per site. The framework lets programmes budget labelling effort against expected accuracy across sites, ecosystems, and platforms.
Alzayat Saleh, Mostafa Rahimi Azghadi
College of Science and Engineering, James Cook University, Townsville, QLD, Australia · Centre for AI and Data Science Innovation, James Cook University, Townsville, QLD, Australia
Reliable visual recognition of shellfish is important for ecological monitoring and aquaculture, but real-world imagery exhibits substantial variation in acquisition conditions, appearance, and posture. Existing marine vision resources cover diverse taxa and imaging settings, but fine-grained mollusc recognition remains underexplored in a group-disjoint benchmark that jointly evaluates class prediction, representation-space retrieval, and condition sensitivity. To address this gap, we introduce ShellfishNet, a provenance-aware, group-disjoint benchmark of 40,130 images from 100 marine mollusc classes, organized into 30,879 acquisition or observation groups. Rather than treating recognition as a single clean-accuracy leaderboard, ShellfishNet evaluates model configurations across classification, image-to-image retrieval, taxonomic error structure, and controlled condition sensitivity. We benchmark representative convolutional, transformer-based, hybrid and efficient, self-supervised, and vision-language models on classification and retrieval. Separately, we evaluate open image-captioning and vision-language models on the captioning subset and assess selected fixed checkpoints under controlled generic and marine-inspired synthetic transformations. Across the benchmark, image-level and class-aware classification largely agree, whereas retrieval provides complementary model rankings and taxonomy projection reveals that a non-negligible subset of exact-class errors remains correct at higher taxonomic ranks. Selected fixed checkpoints with similar clean accuracy can also exhibit distinct sensitivity profiles under controlled transformations. These results show that model comparison on shellfish imagery should account for both task and image condition, rather than rely on a single clean classification score. The dataset and evaluation protocols will be publicly released.
Ziheng Zhou, Yang Wang, Nan Wang +2
IT College, Shanghai Ocean University · Fudan University · DP Technology