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

Abstract

Coral reefs are on the brink of collapse, with climate change, ocean acidification, and pollution leading to a projected 70-90% loss of coral species within the next decade. Reef restoration is crucial, but its success hinges on introducing automation to upscale efforts. In this work, we present a highly configurable AI pipeline for the real-time deployment of coral reseeding devices. The pipeline consists of three core components: (i) the image labeling scheme, designed to address data availability and reduce the cost of expert labeling; (ii) the classifier which performs automated analysis of underwater imagery, at the image or patch-level, while also enabling quantitative coral coverage estimation; and (iii) the decision-making module that determines whether deployment should occur based on the classifier's analysis. By reducing reliance on manual experts, our proposed pipeline increases operational range and efficiency of reef restoration. We validate the proposed pipeline at five sites across the Great Barrier Reef, benchmarking its performance against annotations from expert marine scientists. The pipeline achieves 77.8% deployment accuracy, 89.1% accuracy for sub-image patch classification, and real-time model inference at 5.5 frames per second on a Jetson Orin. To address the limited availability of labeled data in this domain and encourage further research, we publicly release a comprehensive, annotated dataset of substrate imagery from the surveyed sites.

Figures (5)

• Figure 1: The Reef Guidance System incorporates a camera system which provides top-down, wide field of view, high resolution imagery of the seafloor; on-board compute to run the Artificial Intelligence (AI) model in real-time; a weakly supervised model training framework; and automated deployment of coral reseeding devices based on model decisions.
• Figure 2: Proposed Pipeline Schematic. We propose a flexible pipeline where different variations of our approach can be selected to best suit the operational requirements of the reef restoration program. Left: the user can select an Image Labeling Scheme from three options based on whether a domain expert is available to annotate data examples of new locations and environmental characteristics. Each scheme has differing levels of human expert input. Center: we propose two classifiers, image-level and patch-level, to provide flexibility in terms of the level of interpretability of model outputs as well as the configuration of the boat during coral reseeding (i.e. whether the devices can be deployed from one side or both sides of the vessel simultaneously; the location of the deployment hardware on the vessel). Right: we propose three alternatives for the decision-making module and evaluate the performance in terms of accuracy and inference speed.
• Figure 3: Input prompt for the ChatGPT pseudo-labeling scheme.
• Figure 4: (a) Precision–Recall curve showing the different deployment approaches. The threshold $\alpha$ can be tuned based on operational requirements. A higher $\alpha$ increases recall for the Deploy class, meaning that images labeled as No Deploy are less frequently misclassified as Deploy, but more Deploy images are classified as No Deploy. This corresponds to a more conservative deployment strategy. (b) Example confusion matrix for $\alpha=0.3$.
• Figure 5: Real-world deployment results for our proposed Reef Guidance System, visualized as GPS maps comparing an ecologist's deployment decisions (left) alongside the outputs of our proposed pipeline (right). Green markers indicate 'Deploy' frames, where coral devices can be released; red markers indicate 'No Deploy' frames where the corals should not be released. Where the red and green markers are the same between the left and right images, this means both the ecologist and the pipeline are in agreement. We highlight an example image for both 'Deploy' and 'No Deploy' on the left image, and also show incorrect decisions for our pipeline in the right image. These examples also show the difficulty of the images and semantic similarity between 'Deploy' and 'No Deploy' sample frames. Overall, our pipeline (right) matches closely with the ecologist's decisions.