Monitoring through many eyes: Integrating disparate datasets to improve monitoring of the Great Barrier Reef

Erin E. Peterson, Edgar Santos-Fernández, Carla C. M. Chen, Samuel Clifford, Julie Vercelloni, Alan Pearse, Ross Brown, Bryce Christensen, Allan R. James, Ken Anthony, Jennifer Loder, Manuel González-Rivero, Chris M. Roelfsema, M. Julian Caley, Camille Mellin, Tomasz Bednarz, Kerrie L. Mengersen

Introduction

The Great Barrier Reef (GBR) is an ecosystem of immense scale, covering an area roughly the size of Italy. This vastness presents a significant logistical challenge: no single organization, regardless of resources, can monitor the entire reef comprehensively. While the Australian Institute of Marine Science (AIMS) maintains "gold standard" long-term datasets, their surveys are necessarily restricted to specific, accessible reefs.

In parallel, a wealth of data is collected by universities, NGOs, and citizen scientists using diverse methods, cameras, and expertise levels. Historically, these disparate datasets have rarely been analyzed together due to significant variations in quality, scale, and reliability. This research presents a novel framework to bridge this gap, integrating high-precision professional data with high-volume, variable-quality data to produce a more accurate picture of reef health.

Methodology: The Weighted Bayesian Approach

To synthesize these fragmented data sources, the researchers employed a weighted spatio-temporal Bayesian Beta regression model.

The Model Framework

The study utilized a Stochastic Partial Differential Equation (SPDE) approach via the r-INLA package in R. This method is specifically designed to handle large spatial datasets that are typically computationally prohibitive for standard Gaussian processes. By using SPDE, the model can efficiently interpolate coral cover across the vast spatial extent of the GBR.

Mechanistic Weighting

The core innovation of this study is its "Secret Sauce": a mechanistic weighting scheme. Rather than treating all observations as equal, the model assigns weights based on three critical factors:

  1. Image Extent: The physical area of the reef captured in a single photograph.
  2. Sampling Intensity: The number of points classified per image.
  3. Classification Accuracy: For non-professional sources, accuracy was determined by having contributors classify validation images previously vetted by experts.

By accounting for these factors, the model can leverage the volume of citizen science and AI data without allowing lower-quality observations to degrade the overall predictive power.

Data Integration: Five Primary Sources

The pilot study integrated five distinct sources of image-based hard-coral data:

  • AIMS LTMP & MMP: Professional long-term and marine monitoring programs (the "gold standard").
  • XL Catlin Seaview Survey: Large-scale surveys utilizing automated AI classification (CoralNet).
  • UQ-RSRC: Specialized surveys from the University of Queensland Remote Sensing Research Centre.
  • Reef Check Australia: Data contributed by trained citizen scientist divers.
  • Environmental Covariates: Spatial layers including cyclone exposure, thermal bleaching events, sea surface temperature anomalies, and bathymetry.

Key Findings and Insights

Predictive Power

The integration of disparate data led to a dramatic improvement in model performance. The "All Data" model was 42.76% more accurate at predicting coral cover (based on Root Mean Square Prediction Error) than the model relying solely on official monitoring data. Furthermore, the correlation between observed and predicted coral cover jumped from nearly zero (-0.056) in the professional-only model to 0.7431 in the integrated model.

Uncertainty as a Metric

Interestingly, the addition of more data did not always lead to a decrease in uncertainty. In some regions, the inclusion of new data points revealed high spatial variability that the professional-only model had previously smoothed over. This provides a more honest and scientifically useful estimate of our actual knowledge regarding specific reef locations.

The Value of "Imperfect" Data

The study demonstrates that the "perfect" should not be the enemy of the "good." Even though citizen science data accounted for only ~0.3% of the total observations in this pilot, the framework proved that integrating non-gold-standard data fills critical gaps where no professional data exists. Simulations suggested that scaling up community engagement to a "Zooniverse-style" level (e.g., 1,000+ classifications per image) could allow citizen data to provide certainty levels comparable to professional surveys.

Conclusion and Future Directions

This research confirms that the impact of cyclones and thermal bleaching remains the primary driver of coral loss, and the inclusion of "messier" datasets did not obscure these critical signals.

Future iterations of this work aim to transition from "mechanistic" (fixed) weights to "stochastic" weights, allowing the model to estimate biases internally. As AI-driven classification becomes more prevalent, this weighting framework will be essential for integrating various algorithms with different accuracy signatures. Ultimately, this "many eyes" approach provides a scalable, robust pathway for monitoring the world's most vital and threatened marine ecosystems.

Summary updated

Abstract

Numerous organisations collect data in the Great Barrier Reef (GBR), but they\nare rarely analysed together due to different program objectives, methods, and\ndata quality. We developed a weighted spatiotemporal Bayesian model and used it\nto integrate image based hard coral data collected by professional and citizen\nscientists, who captured and or classified underwater images. We used the model\nto predict coral cover across the GBR with estimates of uncertainty; thus\nfilling gaps in space and time where no data exist. Additional data increased\nthe models predictive ability by 43 percent, but did not affect model\ninferences about pressures (e.g. bleaching and cyclone damage). Thus, effective\nintegration of professional and high-volume citizen data could enhance the\ncapacity and cost efficiency of monitoring programs. This general approach is\nequally viable for other variables collected in the marine environment or other\necosystems; opening up new opportunities to integrate data and provide pathways\nfor community engagement and stewardship.\n
BibTeX
@article{peterson2020monitoring,
  author = {Erin E. Peterson and Edgar Santos-Fernández and Carla C. M. Chen and Samuel Clifford and Julie Vercelloni and Alan Pearse and Ross Brown and Bryce Christensen and Allan R. James and Ken Anthony and Jennifer Loder and Manuel González-Rivero and Chris M. Roelfsema and M. Julian Caley and Camille Mellin and Tomasz Bednarz and Kerrie L. Mengersen},
  title = {{Monitoring through many eyes: Integrating disparate datasets to improve monitoring of the Great Barrier Reef}},
  journal = {Environ. Model. Softw.},
  year = {2020},
  doi = {10.1016/j.envsoft.2019.104557},
}
Figure