AI-driven Dispensing of Coral Reseeding Devices for Broad-scale Restoration of the Great Barrier Reef

Coral reefs are facing imminent collapse due to climate change and pollution, making automated, large-scale restoration efforts crucial. The proposed work presents an AI pipeline for real-time deployment of coral reseeding devices to enhance the efficiency and scalability of reef restoration.

The gist: A highly configurable AI pipeline is presented for the real-time deployment of coral reseeding devices, achieving 77.8% deployment accuracy across five Great Barrier Reef sites by integrating image labeling, automated classification, and a decision-making module that determines whether to deploy based on classifier analysis.

Pipeline Architecture

The proposed pipeline consists of three core components designed for real-time operation: (i) the image labeling scheme, which addresses data availability; (ii) the classifier, which performs automated analysis of underwater imagery at the image or patch-level while enabling quantitative coral coverage estimation; and (iii) the decision-making module that determines whether deployment should occur based on the classifier’s analysis. The system is designed to be highly configurable to accommodate diverse operational constraints.

Image Labeling Scheme

The choice of labeling scheme depends primarily on available annotation resources, including cost and access to domain experts, as well as on the level of information and accuracy required from the pipeline. Three schemes are discussed:

1. Human Expert Labeling: This is described as the most expensive scheme, requiring both financial resources and expert marine scientists to annotate the data. It can be performed at two levels: ‘Image-Level’, where each image is assigned a single label, or ‘Patch-Level’, where each image is divided into smaller crops that are individually labeled. Image-level labeling has a lower cost but generally yields lower accuracy for the overall pipeline compared to patch-level labeling.

2. Patch-wise Pseudo-labeling with CLIP: This scheme leverages human expert image-level labels and then adopts the pseudo-labeling scheme from [18] using the large vision-language model CLIP. Patches matching the image-level label “Coral” are retained as “Coral,” while patches identified as “Sand” or “Water” are labeled as “No Deploy.”

3. Unsupervised Patch-wise Labeling with ChatGPT-4o: This scheme uses ChatGPT-4o for completely unsupervised pseudo-labeling of seafloor patches, leveraging the visual understanding capabilities of large multimodal models to incorporate new data with minimal human input.

Classifier Module

The classifier module is tasked with performing image analysis, and two types are proposed:

1. Image-Level Classifier: This method directly maps input images from the camera system to one of the output classes. It is fast but its accuracy is limited because images often contain mixed-class elements, and expert definitions of a suitable location vary, introducing inconsistency.

2. Patch-Level Classifier: This version partitions the original image into a regular grid of smaller patches, with the model generating labels for each patch. This has the advantage that processing smaller patches rather than resizing the entire image preserves more detail, which significantly improves classification accuracy. It also facilitates multi-camera integration.

Decision-Making Module

The decision-making module depends on the selected classifier, either ‘Image-Level’ or ‘Patch-Level’. For the ‘ImageLevel’ configuration, deployment occurs if the predicted class is “Deploy.” For the ‘Patch-Level’ configuration, two strategies are proposed:

1. Thresholding with Patches: This strategy emulates marine scientist heuristics by computing the ratio of the “Deploy” patches relative to the combined number of “No-Deploy” and “Coral” patches. Deployment occurs if this ratio exceeds a calibrated threshold α.

2. Spatial Patch Aggregation: This leverages a lightweight Convolutional Neural Network [34] to account for spatial relationships between neighboring patches, outputting a deployment probability, and the device is released when this probability surpasses the threshold α.

Experimental Validation

The pipeline was validated at five sites across the Great Barrier Reef, benchmarking performance against annotations from expert marine scientists. The system achieves 77.8% accuracy for deployment and 89.1% accuracy for sub-image patch classification, with real-time model inference achieved at 5.5 frames per second on a Jetson Orin. Ablation studies show that the ‘Patch-Level Classifier’ outperforms the ‘Image-Level Classifier’ by 15.2% overall F1 score (Table III). Furthermore, testing with CLIP supervision still achieves a high overall F1 score of 73.32%, demonstrating that large vision-language foundation models can supplement human labels in annotation-constrained settings. The final results show the pipeline's accuracy is comparable to an ecologist, despite challenges posed by the "limited size of the training dataset and inherent label noise.

Here are specific improvements to existing AI systems based on this research, categorized by the component they enhance:

) 1. Enhance Data Efficiency through Knowledge-Augmented Pseudo-Labeling:

The current reliance on expensive human expert labeling (especially for patch-level data) is a major bottleneck.

The improved system will integrate a tiered pseudo-labeling pipeline using Large Vision-Language Models (LVLMs). Specifically, the system will utilize the ChatGPT-4o model to generate initial, unsupervised pseudo-labels for seafloor patches. This reduces the required human annotation burden by leveraging the model's visual understanding of common elements (sand, water) and providing a baseline for training.

) 2. Implement Flexible Multi-Tiered Classification Architectures:

The current choice between Image-Level and Patch-Level classifiers is rigid based on operational needs.

The improved system will feature a dynamically configurable classifier selection mechanism that allows operators to select the optimal analysis granularity (Image-Level vs. Patch-Level) based on real-time constraints (e.g., required interpretability for an ecologist, or speed for rapid vessel transit). This flexibility is crucial for adapting the AI to diverse operational scenarios without requiring a complete model retraining.

) 3. Optimize Decision Making via Adaptive Aggregation Strategies:

The current decision modules are fixed (simple thresholding or a lightweight CNN).

The improved system will incorporate an adaptive decision module that dynamically switches between two strategies:

a) A refined Thresholding with Patches strategy, where the threshold parameter α is adjusted based on real-time operational parameters (e.g., available coral devices in the vessel).

b) A Spatial Patch Aggregation strategy, which uses a lightweight CNN to predict deployment probability, allowing for more nuanced decisions that account for spatial relationships between adjacent patches rather than just a simple ratio count.

) 4. Ensure Robustness Against Class Imbalance via Advanced Loss Functions:

The current use of the multi-class Focal Loss is good but can be further refined.

The improved system will utilize an advanced, dynamically weighted focal loss formulation that not only addresses class imbalance (predominance of 'No-Deploy') but also incorporates a weighting factor based on the uncertainty or rarity of specific environmental conditions detected in the imagery, ensuring that hard-to-classify novel substrate types receive disproportionately high attention during training.

) 5. Achieve Real-Time Edge Deployment via Model Optimization:

The current system runs at 5.5 FPS on a Jetson Orin.

The improved system will employ continuous model optimization and quantization techniques (e.g., leveraging MobileNetV3-small backbone variants) to maximize inference speed while maintaining acceptable accuracy (targeting >80% F1 score). This ensures that deployment decisions are made with minimal latency, directly enabling higher vessel velocity and operational efficiency in dynamic marine environments.

The improved AI system will be a highly autonomous, real-time Reef Guidance System capable of:

1. Detecting and classifying reef substrate (Coral vs. No Deploy/Substrate) from high-resolution underwater imagery in real-time at depths up to 10m.

2. Determining whether to deploy coral reseeding devices based on a flexible, user-configurable pipeline that balances expert oversight with AI automation.

3. Providing interpretable feedback to marine scientists by visualizing the specific patches and image features driving the deployment decision (e.g., showing which specific patch ratio triggered a 'Deploy' command).

4. Operating efficiently across various vessel configurations and hardware setups by selecting the most appropriate model backbone and decision strategy for the current operational context.