CNN and Transformer-Based Deep Learning Models for Coral Health Classification

Abstract

Coral reefs are among the most important marine ecosystems, providing habitat for nearly 32% of marine species and supporting the livelihoods and food security of more than 500 million people worldwide. In addition to sustaining marine biodiversity, coral reefs protect coastlines and contribute to economic and cultural activities in many regions. Despite their significance, coral reefs are rapidly declining due to climate change, ocean warming, and, in particular, coral bleaching. The most severe global coral bleaching event to date began in 2023 and is still ongoing, affecting approximately 84% of the world’s coral reefs, highlighting the urgent need for reliable methods to detect bleaching at an early stage. In response to this challenge, this paper proposes deep learning-based models for coral health classification using underwater images. The approach compares the performance of three convolutional neural network architectures, ResNet50, EfficientNetB3, and ConvNeXt-Tiny, with a transformer-based model, Swin Transformer-Tiny, for classifying corals as healthy or bleached. Transfer learning is applied to all models, and their performance is evaluated using a publicly available dataset containing 923 labeled coral images. The results show that all models achieve effective classification performance, with ConvNeXt-Tiny and Swin Transformer-Tiny attaining the highest accuracy of 86.33%, outperforming ResNet50 and EfficientNetB3. These findings provide insight into the advantages of newer CNN and transformer-based architectures for learning complex visual patterns in underwater coral images. The results further demonstrate their suitability for practical coral reef monitoring systems, supporting reliable and early detection of bleaching in real-world conservation and reef management applications.

Introduction

Coral reefs, vital marine ecosystems supporting 32% of marine species and protecting shorelines, are increasingly threatened by climate change, ocean acidification, and coral bleaching. Coral bleaching—caused by the loss of symbiotic algae—has escalated in frequency and severity, with the 2023 event affecting 84% of global reefs. Early detection of bleaching is critical for conservation and preserving biodiversity.

This study proposes using deep learning models for automated coral classification from underwater images. Four architectures were compared: ResNet50, EfficientNetB3, ConvNeXt-Tiny, and Swin Transformer-Tiny. A dataset of 923 images (healthy and bleached corals) underwent extensive preprocessing and augmentation to enhance model robustness. Models were fine-tuned via transfer learning, using pre-trained ImageNet weights and custom classification heads.

Results:

• All models effectively distinguished healthy and bleached corals.

• Test accuracy: ResNet50 (84.17%), EfficientNetB3 (83.45%), ConvNeXt-Tiny (86.33%), Swin Transformer-Tiny (86.33%).

• ConvNeXt-Tiny and Swin Transformer-Tiny outperformed others due to their ability to capture richer spatial and contextual features.

The study demonstrates that modern CNN and transformer-based models can improve coral bleaching detection, enabling timely conservation interventions and more accurate reef monitoring.

Conclusion

This paper presented a comparative analysis of CNN and Transformer-based models, ResNet50, EfficientNetB3, ConvNeXt-Tiny and Swin Transformer-Tiny for effective coral health classification. Among the models, ConvNeXt and Swin Transformer achieved the highest test accuracy of 86.33%, outperforming ResNet50 (84.17%) and EfficientNetB3 (83.45%). By comparing these four models, we gain validated insights into their capabilities in coral health classification. Specifically, we observe that while all models are effective at distinguishing healthy and bleached corals, the more recent architectures, ConvNeXt and Swin Transformer, better capture complex spatial and contextual features, leading to higher classification performance. ResNet50 and EfficientNetB3, though effective, are comparatively less efficient on this dataset, highlighting the significance of architecture choice for tasks involving limited and complex image data. Overall, the results demonstrate the potential of deep learning models for automated coral reef monitoring, even when trained on relatively small datasets.

References

[1] National Ocean Service, “What is a coral reef made of?,” Noaa.gov, 2019. https://oceanservice.noaa.gov/facts/coralmadeof.html [2] “The coral - Monaco Oceanographic Institute, Albert I Foundation,” Institut océanographique, Jul. 19, 2021. https://www.oceano.org/en/thematic-pages/the-coral/ (accessed Jan. 02, 2026). [3] Y. K. Wong, Z. Zheng, M. Zhang, D. J. Suggett, and S.-K. Yeung, “CoralSCOP-LAT: Labeling and analyzing tool for coral reef images with dense semantic mask,” Ecological Informatics, vol. 91, p. 103402, Nov. 2025, doi: https://doi.org/10.1016/j.ecoinf.2025.103402. [4] D. O. Obura et al., “Coral Reef Monitoring, Reef Assessment Technologies, and Ecosystem-Based Management,” Frontiers in Marine Science, vol. 6, no. 580, Sep. 2019, doi: https://doi.org/10.3389/fmars.2019.00580. [5] D. N. Handiani, N. S. Ningsih, and E. Beliyana, “Coral bleaching occurrence and its relation to marine heatwave events in the Southwestern waters of South Sulawesi, Indonesia, as part of the Coral Triangle region,” Journal of Marine Systems, vol. 252, p. 104136, Sep. 2025, doi: https://doi.org/10.1016/j.jmarsys.2025.104136. [6] T. N. F. Roach, J. Dilworth, C. M. H, A. D. Jones, R. A. Quinn, and C. Drury, “Metabolomic signatures of coral bleaching history,” Nature Ecology & Evolution, pp. 1–9, Feb. 2021, doi: https://doi.org/10.1038/s41559-020-01388-7. [7] C. M. Eakin, H. P. A. Sweatman, and R. E. Brainard, “The 2014–2017 global-scale Coral Bleaching event: Insights and Impacts,” Coral Reefs, vol. 38, no. 4, pp. 539–545, Jul. 2019, doi: https://doi.org/10.1007/s00338-019-01844-2. [8] Margaux Monfared, “84% of the world’s coral reefs impacted in the most intense global coral bleaching event ever | ICRI,” ICRI, Apr. 23, 2025. https://icriforum.org/4gbe-2025/ [9] A. Raphael, Z. Dubinsky, D. Iluz, and N. S. N, “Neural Network Recognition of Marine Benthos and Corals,” Diversity, vol. 12, no. 1, p. 29, Jan. 2020, doi: https://doi.org/10.3390/d12010029. [10] Chowdhury, M. Jahan, Shahriar Kaisar, M. E. Khoda, A. Karim, and R. Naha, “Coral Reef Surveillance with Machine Learning: A Review of Datasets, Techniques, and Challenges,” Electronics, vol. 13, no. 24, pp. 5027–5027, Dec. 2024, doi: https://doi.org/10.3390/electronics13245027. [11] S. Jamil, M. Rahman, and A. Haider, “Bag of Features (BoF) Based Deep Learning Framework for Bleached Corals Detection,” Big Data and Cognitive Computing, vol. 5, no. 4, p. 53, Oct. 2021, doi: https://doi.org/10.3390/bdcc5040053. [12] Ajay, Akanksh M, and Mamatha Balipa, “Enhancing Coral Health Monitoring with a Hybrid CNN-ViT Model for Bleaching Prediction,” pp. 74–79, May 2025, doi: https://doi.org/10.1109/icaiss61471.2025.11041783. [13] A. Kaur, K. S. Gill, M. Kumar, and R. Rawat, “VGG19’s Role in Safeguarding Coral Reefs using a CNN-Based Monitoring Model,” 2024 4th Asian Conference on Innovation in Technology (ASIANCON), pp. 1–5, Aug. 2024, doi: https://doi.org/10.1109/asiancon62057.2024.10838114. [14] M. Thamarai and S. P. Aruna, “Stressed Coral Reef Identification Using Deep Learning CNN Techniques,” Journal of Electronic & Information Systems, vol. 5, no. 2, pp. 1–9, Sep. 2023, doi: https://doi.org/10.30564/jeis.v5i2.5808. [15] G. Xin, H. Xie, S. Kang, Y. Chen, and Y. Jiang, “Improved research on coral bleaching detection model based on FCOS model,” Marine Environmental Research, vol. 200, p. 106644, Sep. 2024, doi: https://doi.org/10.1016/j.marenvres.2024.106644. [16] Y. Shuang et al., “Classification of pain expression images in elderly with hip fractures based on improved ResNet50 network,” Frontiers in Medicine, vol. 11, Jul. 2024, doi: https://doi.org/10.3389/fmed.2024.1421800. [17] L. Mahin, A. Shaikh, Aniket Sandupatla, Rushikesh Pudale, A. Bakare, and Prof. Mallesh Chavan, “Classification of Simple CNN Model and ResNet50,” International journal for research in applied science and engineering technology, vol. 12, no. 4, pp. 4606–4610, Apr. 2024, doi: https://doi.org/10.22214/ijraset.2024.60677. [18] Deep Ajabani, Z. A. Shaikh, A. Yousef, K. Ali, and M. A. Albahar, “Enhancing skin lesion classification: a CNN approach with human baseline comparison,” PeerJ Computer Science, vol. 11, pp. e2795–e2795, Apr. 2025, doi: https://doi.org/10.7717/peerj-cs.2795. [19] G V Kartheek and S. Rani, “Robust Multi-Class Classification for Real-Time Agricultural Applications Using Efficient and Adaptive Deep Learning,” International Journal of Scientific Research in Science and Technology, vol. 11, no. 6, pp. 90–99, Nov. 2024, doi: https://doi.org/10.32628/ijsrst2411490.

Copyright

Copyright © 2026 Hajara Sabnam Kareem Navaz. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.