Accelerating coral rubble instability assessments with machine learning: insights from the Great Barrier Reef

Abstract

The stability of coral rubble is crucial for coral reef recovery since a stable coral rubble base is essential for coral regeneration. To enhance the effectiveness of coral reef restoration and to model the recovery and connectivity of coral reef ecosystems, it is crucial to predict coral rubble stability over extensive areas and long timeframes. The assessment of rubble instability on individual small-scale coral reefs can be calculated using physical-based models. However, these models are time consuming, limiting their ability to predict coral rubble instability over large spatial scales with high resolution. Here, the Random Forest machine learning algorithm is used to efficiently process large databases containing information on coral reef hydrodynamics and related coral rubble instability and provide estimates on the importance of specific variables in the classification of rubble as stable or unstable. Adopting a Random Forest model provides a significant advantage in addressing the non-linear problems inherent in risk assessment and significantly reducing the model calculation time. In this study, physics-based and Random Forest-based assessment models were adopted to evaluate coral rubble instability risk across the Great Barrier Reef (GBR), with the performance of the trained Random Forest model evaluated against the physics-based model. Seven hydrodynamic characteristics were selected, and 1.2 million physics-based model samples were utilised in the Random Forest model for training (70 %), validation (20 %) and testing (10 %). These samples include the southern, central, and northern reef areas of the Great Barrier Reef, ensuring that the model is comprehensively trained and capable of producing more accurate and reasonable prediction results. Results show that the Random Forest model reduces the prediction time by several order of magnitude compared to the physics-based model: assessing rubble instability in the entire GBR takes only two minutes with the Random Forest model. Additionally, the trained Random Forest model eliminates the resolution limitation: it can rapidly assess new cases irrespective of data resolution changes, without the need for recalculations, unlike physics-based models. Moreover, the Random Forest model enables broader application for assessing coral rubble instability beyond just the GBR. This is because the Random Forest model only requires relevant key factor data to assess coral rubble instability. Among all seven factors, mean depth has the greatest impact on the prediction results, with a Gini decrease index of 34 %. The Gini decrease index measures the importance of a factor by indicating how much it decreases the impurity in the data split; a higher value means greater importance. Other factors have indices around 10 %. Therefore, even in regions with limited data, rubble instability can be effectively assessed with only depth measurements and wave climate information. This study demonstrates a novel and highly successful approach to predicting rubble instability on large scales, offering value guidance for coral reef recovery efforts in the GBR and significantly reducing the required field-based data.