1. Abstract

According to the World Wide Fund for Nature, 90 percent of coral reefs are estimated to be lost by 2050 if immediate action is not taken to address the climate issue. Therefore, this proposal introduces a comprehensive pipeline for real-time monitoring of coral reefs. Our framework uses onboard sensors in BlueROV2 to send images of corals to be processed by an enhanced YOLO-ViTDet for coral  identification and condition classification. Then, the bounding box prediction will be passed to the Swin Transformer U-Net for pixel-level severity analysis. If over 30 percent of the scanned coral exhibits a high severity of unhealthy conditions, an immediate alert will be sent to the operator. The machine learning integration will be optimized using state-of-the-art GPUs and quantization. The entire framework is built upon the cutting-edge advancements in machine learning and robotics, with the goal of assisting marine biologists in preventing the deterioration of coral conditions.

 

2. Introduction

With the rapidly increasing ocean temperature and changing ocean chemistry, coral reef ecosystems are suffering from prevalent diseases, bleaching, and even death. Scientists around the world have warned that the existence of coral reefs is in great danger unless immediate actions are taken to safeguard them (Rahman et al., 2023; Andrello et al., 2021; Nama et al., 2023). Traditional coral monitoring practice requires an experienced and knowledgeable diver to survey the coral reefs in person, noting down any anomalies and reporting for data storage. This approach may easily disturb corals and make regular real-time checkups challenging. Another major limitation is that this practice is inefficient for mass monitoring, where divers must survey various dangerous locations within a short period of time. Recent advances in automated coral monitoring with machine learning and robotics have offered a promising solution (Piñeros et al., 2024; Teague et al., 2022). This proposal builds upon the state-of-the-art technologies to develop a robust system that enables mass monitoring of coral reefs in real time.

 

3. Literature Review

Increasingly, deep machine learning frameworks have been applied to monitor coral conditions in underwater settings. Parsons et al. (2018) combined aerial data with in-water imagery to improve coral bleaching classification accuracy. Lu et al. (2024) introduce SCoralDet, a coral detection model based on YOLO that uses a Multi-Path Fusion Block and an Adaptive Power Transformation label strategy to enhance detection accuracy. Raber and Schill (2019) designed Reef Rover, an unmanned surface vehicle that maps coral reefs using open-source Global Positioning System drone technology. Llewellyn and Bainbridge (2015) present a detailed analysis of the traditional coral monitoring system and a more efficient automated robotics system. Boonnam et al. (2022) experiment with both supervised and unsupervised machine learning techniques to improve coral bleaching prediction accuracy. Modasshir et al. (2018) employ a CNN and an AUV for accurate coral identification and counting. Cardenas et al. (2024) present a comprehensive overview of the interconnected roles machine learning and robotics play in practical coral monitoring systems.

​

Though there have been numerous studies on implementing machine learning and robotics models to locate and classify coral conditions, there is a severe lack of comprehensive pipelinesfor real-time detection and pixel-level analysis. This proposal therefore investigates how a two-stage machine learning pipeline can enhance real-time coral identification and severity analysis on edge devices embedded within autonomous underwater vehicles.

 

4. Methodology

Data Acquisition and Augmentation

This study will employ a dataset of 3,049 labeled high-quality coral images under five health conditions: healthy coral, bleached disease, band disease, white pox disease, and dead coral (Rajan & Damodaran, 2023). To further increase the number of training samples, we will employ StyleGAN3 to generate an additional 6,000 labeled images, which expert marine biologists will verify. Then, we will employ data augmentation techniques such as Gaussian noise addition and elastic transformations for better generalizability. Lastly, we will employ a Denoising Diffusion Probabilistic Model adapted for image restoration, namely Palette (Saharia et al., 2021), to enhance the data samples. The entire training framework can be seen in Figure 1.

 

Deep Learning Models

YOLO-ViTDet

We will use a novel hybrid model that combines two state-of-the-art architectures, YOLOv11 and ViT, to both draw bounding boxes around detected corals and classify their current condition. This combination allows the model to rely on the YOLO component for fast real-time classification and ViT’s transformer architecture for more nuanced analysis. This model will employ CSPDarkNet as the backbone, utilizing its split-and-merge strategy for enhanced computational efficiency. The model will be measured against Complete IoU loss and cross-entropy loss during training. The final performance will be evaluated with mAP and precision-recall curves. After training the model on our preprocessed dataset, it will be benchmarked against Faster R-CNN and EfficientDet.

​

Swin Transformer U-Net

Next, bounding box crops extracted from YOLO-ViTDet outputs will be resized and fed into Swin Transformer U-Net for a pixel-level analysis. The multi-scale transformer’s design allows it to extract local features while significantly reducing computation requirements. The U-Net encoder-decoder structure enables detailed feature analysis while reconstructing spatial information. This model will be trained with Dice and Focal loss and evaluated with the Dice coefficient, IoU, and per-class pixel accuracy. This model will be trained on detailed masks, created by expert marine biologists, that categorize coral disease severity into three levels: minimal, moderate, and severe.

​

Real-Time Deployment and Optimization

The entire machine learning framework will be optimized for deployment on the NVIDIA Jetson Xavier NX, which will be integrated directly into the BlueROV2 underwater vehicle. The Palette diffusion model will be integrated as a real-time data augmentor to improve prediction accuracy; YOLO-ViTDet will be used for coral identification and classification; and the Swin Transformer U-Net will be employed for pixel-level analysis. Each of these models will be quantized with NVIDIA’s TensorRT framework to reduce computational burden. Additional strategies, such as lower input resolution and model pruning, may be implemented to optimize the system further. The entire real-time deployment framework can be seen in Figure 2.

​​​

​

Bibliography

Andrello, M., Darling, E. S., Wenger, A., Suárez‐Castro, A. F., Gelfand, S., & Ahmadia, G. N. (2021). A global map of human pressures on tropical coral reefs. Conservation Letters, 15(1).

https://doi.org/10.1111/conl.12858

​

Boonnam, N., Udomchaipitak, T., Puttinaovarat, S., Chaichana, T., Boonjing, V ., & Muangprathub, J. (2022). Coral Reef Bleaching under Climate Change: Prediction Modeling and Machine Learning. Sustainability, 14(10), 6161.

https://doi.org/10.3390/su14106161

​

Cardenas, J. A., Samadikhoshkho, Z., Rehman, A. U., Valle-Pérez, A. U., de León, E. H.-P., Hauser, C. A. E., Feron, E. M., & Ahmad, R. (2024). A systematic review of robotic efficacy in coral reef monitoring techniques. Marine Pollution Bulletin, 202, 116273.

https://doi.org/10.1016/j.marpolbul.2024.116273

​

Llewellyn, L. E., & Bainbridge, S. J. (2015). Getting up close and personal: The need to immerse autonomous vehicles in coral reefs. https://doi.org/10.23919/oceans.2015.7401909 Lu, Z., Liao, L., Xie, X., & Yuan, H. (2024). SCoralDet: Efficient real-time underwater soft coral detection with YOLO. Ecological Informatics, 85, 102937.

https://doi.org/10.1016/j.ecoinf.2024.102937

​

Modasshir, M., Rahman, S., Youngquist, O., & Rekleitis, I. (2018, December 1). Coral Identification and Counting with an Autonomous Underwater Vehicle. IEEE Xplore.

https://doi.org/10.1109/ROBIO.2018.8664785

​

Nama, S., Shanmughan, A., Nayak, B. B., Bhushan, S., & Ramteke, K. (2023). Impacts of marine debris on coral reef ecosystem: A review for conservation and ecological monitoring of the coral reef ecosystem. Marine Pollution Bulletin, 189(189), 114755.

https://doi.org/10.1016/j.marpolbul.2023.114755

​

Parsons, M., Bratanov, D., Gaston, K., & Gonzalez, F. (2018). UA Vs, Hyperspectral Remote Sensing, and Machine Learning Revolutionizing Reef Monitoring. Sensors, 18(7), 2026.

https://doi.org/10.3390/s18072026

​

Piñeros, V . J., Reveles-Espinoza, A. M., & Monroy, J. A. (2024). From Remote Sensing to Artificial Intelligence in Coral Reef Monitoring. Machines, 12(10), 693.

https://doi.org/10.3390/machines12100693

​

Raber, & Schill. (2019). Reef Rover: A Low-Cost Small Autonomous Unmanned Surface Vehicle (USV) for Mapping and Monitoring Coral Reefs. Drones, 3(2), 38.

https://doi.org/10.3390/drones3020038

​

Rahman, Md. N., Shozib, S. H., Akter, Mst. Y ., Islam, A. R. Md. T., Islam, Md. S., Sohel, Md. S., Kamaraj, C., Rakib, Md. R. J., Idris, A. M., Sarker, A., & Malafaia, G. (2023). Microplastic as an invisible threat to the coral reefs: Sources, toxicity mechanisms, policy intervention, and the way forward. Journal of Hazardous Materials, 454, 131522.

https://doi.org/10.1016/j.jhazmat.2023.131522

 

Rajan, S. K. S., & Damodaran, N. (2023). MAFFN_YOLOv5: Multi-Scale Attention Feature Fusion Network on the YOLOv5 Model for the Health Detection of Coral-Reefs Using a Built-In Benchmark Dataset. Analytics, 2(1), 77–104.

https://doi.org/10.3390/analytics2010006

​

Saharia, C., Chan, W., Chang, H., Lee, C. A., Ho, J., Salimans, T., Fleet, D. J., & Norouzi, M. (2021). Palette: Image-to-Image Diffusion Models. ArXiv:2111.05826 [Cs].

https://arxiv.org/abs/2111.05826

​

Teague, J., Megson-Smith, D. A., Allen, M. J., Day, J. C. C., & Scott, T. B. (2022). A Review of Current and New Optical Techniques for Coral Monitoring. Oceans, 3(1), 30–45.

https://doi.org/10.3390/oceans3010003

​​

Appendix​​​​​​​​​​​​​​​​​​​​​​​​