Enhancing Coral Reef Restoration Assessment Using Machine Learning in Gili Asahan and Kecinan Bay, Lombok, Indonesia

Methods

Summary

Our project aims to establish a near-real-time monitoring system for a coral reef restoration site, optimizing resource use and minimizing the need for extensive fieldwork, particularly SCUBA diving. This system will provide critical data for early monitoring of coral growth and early detection of disturbances such as storms or bleaching events, allowing for timely intervention by stakeholders and policy makers.

1. Underwater monitoring system:
   • Equipment: We will install a pan-tilt-zoom (PTZ) CCTV camera in a custom-built underwater housing. This housing will include an automatic wiper mechanism to prevent algae build-up on the camera lens. We also planned to install a UV light near the camera to monitor coral spawning, which will be programmed to light up at specific times each night, this will enhance the coral spawning data and occurrence.
   • Power Supply: The camera system will be powered primarily by solar panels, ensuring sustainability and reducing the need for frequent maintenance.
   • Data Logger: We will install a data logger with a buoy that will be equipped with sensors to collect pH, dissolved oxygen, temperature and salinity data surrounding the reef sites area, this data will be crucial to for early growth coral monitoring.
   • Connectivity: The live video feed will be accessible via PCs and smartphones, allowing remote monitoring of the reef site, which we can also use to control the camera's movement. The data logger will send data through PCs and smartphones every 30 minutes.
2. Data Collection and Utilization:
   • Machine Learning Integration: We will use YOLOv10 algorithm for image classification and analysis. The model will be trained on a dataset of 10,000 annotated images per site (two sites in total).
   • Annotation Process: Ten marine biologists will be employed to annotate the images. A comprehensive training program will be conducted for these annotators to ensure high quality annotations.
   • Model deployment: After training, the YOLOv10 model will be used for benthic classification and segmentation of photogrammetric data. This will increase the efficiency and accuracy of our analysis, and act as a visual aid for project progress. 
3. Outcome and Impact:
   • Monitoring and response: The system will enable continuous monitoring of coral health and early detection of disturbances, allowing rapid action to mitigate potential damage.
   • Stakeholder engagement: Real-time data and analysis will be made available to stakeholders and policy makers, supporting informed decision-making and fostering collaborative conservation efforts.

• Reproducibility: Detailed protocols for installation, maintenance, and operation of the underwater monitoring system will be documented. A comprehensive guide to the machine learning model training and deployment process will be provided.

  Workshops and training on data annotation and system operation will be conducted to ensure that similar projects can be effectively replicated.

• Visual Aids:

  Successfully Installed Underwater Camera Footage
  

Camera Setup: 

Solar Panel and Recorder Integration:

Data Logger Setup:

System Illustration:

Image Annotation Process:

Challenges

The challenges we anticipate are bad weather and low tidal conditions during the installation process, which could delay setup and affect how effective are we on installing the equipment. In addition, there is a risk of imperfect annotation, which could compromise the quality of our training data. Data collection and model training are time-consuming processes that require significant effort and coordination. In addition, we may initially encounter problems with low model accuracy. To address these challenges, we will schedule installations during favorable weather windows and provide robust equipment security. We will provide extensive training to our annotators to ensure high quality data and implement rigorous validation procedures to refine our model. Regular evaluations and iterative improvements will be made to improve model accuracy and ensure the reliability and effectiveness of the system.

Pre Analysis Plan

Our pre-analysis plan includes several steps to ensure a rigorous and comprehensive analysis of our project data and results. Our primary hypothesis is that real-time monitoring using PTZ CCTV cameras and machine learning can significantly improve early-stage coral growth tracking and disturbance detection compared to traditional methods.

We will use the YOLOv10 algorithm for image classification and analysis, training the model with annotated images from our two reef sites. To handle multiple outcomes, such as variations in coral health, growth rates, and disturbances, we will use a multi-class classification approach. This will allow us to accurately categorize different coral types and disturbances.

To deal with the potential variance in our data, we will use statistical techniques such as cross-validation and bootstrapping. This will help us assess the robustness and reliability of our model. We will also perform sensitivity analyses to understand how different factors, such as environmental conditions and annotation quality, affect the performance of our model.

We will use both quantitative and qualitative methods to interpret the data. Quantitatively, we will analyze the frequency and severity of disturbances, changes in coral cover, and growth rates. Qualitatively, we will review annotated images and model predictions to ensure that they are consistent with observed patterns and expert assessments.

Periodic interim analyses will be conducted to monitor progress and make necessary adjustments. This iterative process will help us refine our methods and improve the accuracy and effectiveness of our monitoring system. Finally, we will comprehensively document our findings and methods to facilitate replication and scalability of the project.