An Integrated Bayesian Network approach to Lyngbya majuscula bloom initiation

Sandra Johnson, Fiona Fielding, Grant Hamilton, Kerrie MengersenMarine 2009 · DOI

TL;DR

This paper introduces an Integrated Bayesian Network (IBN) approach to model and understand the complex factors leading to Lyngbya majuscula blooms. By combining scientific and management perspectives within a unified framework, this IBN offers a powerful tool for scenario modeling and decision-making in environmental management.

What Problem Does This Paper Address?

Lyngbya majuscula blooms are a recurring and increasing problem in coastal areas worldwide, posing significant environmental and health risks. Understanding the intricate web of scientific and management factors that trigger these blooms is crucial for effective mitigation. However, this knowledge is often fragmented across different research groups and management bodies. This paper addresses the challenge of integrating this diverse knowledge into a cohesive and actionable model.

Methods

The authors propose an Integrated Bayesian Network (IBN) approach. This involves:

  • Bayesian Networks (BNs): Used to represent probabilistic relationships between variables (factors) and their influence on bloom initiation.
  • Object-Oriented Bayesian Networks (OOBNs): Allow for modularity and reusability by defining network components as objects, facilitating the construction of complex models from smaller, manageable parts.
  • Dynamic Object-Oriented Bayesian Networks (DOOBNs): Extend OOBNs by incorporating temporal dynamics, allowing the model to capture how factors change over time and influence sequential events, crucial for understanding bloom initiation which is time-dependent.
  • Integration of Models: The core of the IBN is the integration of multiple BNs (Science Network, Management Network) and a separate water catchment simulation model (E2 model). This allows for a holistic view, where management actions (modeled in the Management Network) influence nutrient loads, which are then simulated through the catchment model and fed into the Science Network to predict bloom initiation.

Datasets

Not explicitly detailed as training datasets. The models were populated using:

  • Expert elicitation
  • Data from simulation models
  • Statistical models
  • Data from monitoring sites and government agencies
  • Meta-data compiled in a Healthy Waterways report.

Key Insights and Ideas

  • Holistic Modeling is Essential: No single factor drives Lyngbya blooms; an integrated approach considering both scientific and management factors is necessary.
  • Bayesian Networks as a Unifying Framework: BNs, particularly OOBNs and DOOBNs, provide a flexible and powerful way to represent complex, uncertain relationships and integrate diverse knowledge sources.
  • Temporal Dynamics Matter: Bloom initiation is a process that unfolds over time, making dynamic models (DOOBNs) more insightful than static ones for prediction and understanding.
  • Scenario Modeling for Decision Support: The IBN framework is highly effective for exploring "what-if" scenarios, allowing managers to test the potential impact of different management strategies.
  • Stakeholder Collaboration is Key: The development of the networks involved close collaboration with various stakeholders (scientists, managers, local experts), ensuring the model reflects real-world knowledge and concerns.

Key Results

  • Sensitivity Analysis: Identified the "available nutrient pool (dissolved)" as the most influential factor affecting bloom probability, with changes in nutrient availability causing a 77% variation. Other significant factors include bottom current climate, sediment nutrients, dissolved iron, dissolved phosphorus, light, and temperature.
  • Scenario Testing: Demonstrated that management actions (e.g., upgrading point sources) and climate events (e.g., severe summer storms) can significantly alter bloom initiation probabilities. For instance, a summer rainfall event increased bloom probability to 52% compared to 42% in the original static model.
  • DOOBN Performance: The dynamic model (DOOBN) improved prediction and interpretability by incorporating time-specific dynamics, aligning better with the conceptual understanding of Lyngbya behavior.
  • IBN Effectiveness: The integrated approach successfully combined scientific understanding with management inputs to provide a cohesive tool for assessing bloom initiation risk.

Limitations and Future Work

  • Data Limitations: The paper notes that for some observed outbreaks, data for all nodes in the BN model were not available, limiting direct outcomes comparison.
  • Validation Scope: The paper acknowledges that a full comparison against classical multivariate techniques or a rigorous test of the BNs' output reliability on unseen data was beyond the scope of this paper.
  • Future Work:
    • The Science BN has been adopted by Healthy Waterways and will be validated with future data.
  • The authors suggest the IBN approach can be extended to investigate other aspects of the organism, such as growth, biomass, and decay, by modifying the Science Network.
  • The integrated network approach is conceptually suitable for investigating other environmental outcomes influenced by nutrient outputs and water movement in catchments.

Summary updated

BibTeX
@inproceedings{johnson2009integrated,
  author = {Sandra Johnson and Fiona Fielding and Grant Hamilton and Kerrie Mengersen},
  title = {{An Integrated Bayesian Network approach to Lyngbya majuscula bloom initiation}},
  booktitle = {Marine},
  year = {2009},
  doi = {10.1016/j.marenvres.2009.07.004},
}
Figure