Varela Gonzalez, Patricia Y (2017-12). PROBABILISTIC RISK MAPPING COUPLING BAYESIAN NETWORKS AND GIS, AND BAYESIAN MODEL CALIBRATION OF SUBMARINE LANDSLIDES.. Doctoral Dissertation. Thesis uri icon

abstract

  • A spatial and causal probabilistic methodology is introduced for risk assessment based on the coupling of a conceptual Bayesian Network (BN) model and GIS to generate risk maps. The proposed integration of these spatial events is referred to as BN+GIS, which features forward and inverse modeling, denoted in this work as spatial prognosis and spatial diagnosis, respectively. This approach is illustrated through two case studies: (1) environmental risk associated to oil and gas site developments implemented in the Barnett Shale Play in Texas, and (2) landslide susceptibility in the Elliott State Forest in the Oregon Coastal Range. This approach will equip stakeholders, such as land owners, operators, regulators, government officials, and other related organizations with a platform that can help them improve the assessment of future potential risk scenarios, and to identify likely consequences that would lead to undesirable states of environmental risks ahead of time. A sensitivity analysis was performed on BN+GIS to study the influence of some of the user-defined parameters on the model's results, such as sample size, spatial interval of the systematic sampling methodology, and the prescribed diagnosis distribution used for decision making purposes. As an additional effort to portray the potential application of the Bayesian paradigm on risk assessment, a parameter estimation methodology is implemented using bathymetry data and CPT logs. This approach is illustrated through a study case, where information was mined from existent landslides to perform a Bayesian calibration on an infinite slope model. This approach allowed to estimate posterior probability distributions of physical parameters given a prescribed factor of safety, to assess the most likely depth of failure, and to identify the optimum amount of samples required to maximize the reliability of the inferences. This work focusses on providing a substantial contribution to improved policymaking and management through the use of integrated sources of evidence such as real data, model predictions and experts educated beliefs.
  • A spatial and causal probabilistic methodology is introduced for risk assessment based
    on the coupling of a conceptual Bayesian Network (BN) model and GIS to generate risk
    maps. The proposed integration of these spatial events is referred to as BN+GIS, which
    features forward and inverse modeling, denoted in this work as spatial prognosis and
    spatial diagnosis, respectively. This approach is illustrated through two case studies: (1)
    environmental risk associated to oil and gas site developments implemented in the
    Barnett Shale Play in Texas, and (2) landslide susceptibility in the Elliott State Forest in
    the Oregon Coastal Range. This approach will equip stakeholders, such as land owners,
    operators, regulators, government officials, and other related organizations with a
    platform that can help them improve the assessment of future potential risk scenarios,
    and to identify likely consequences that would lead to undesirable states of environmental
    risks ahead of time. A sensitivity analysis was performed on BN+GIS to study the
    influence of some of the user-defined parameters on the model's results, such as sample
    size, spatial interval of the systematic sampling methodology, and the prescribed
    diagnosis distribution used for decision making purposes. As an additional effort to
    portray the potential application of the Bayesian paradigm on risk assessment, a
    parameter estimation methodology is implemented using bathymetry data and CPT logs.
    This approach is illustrated through a study case, where information was mined from
    existent landslides to perform a Bayesian calibration on an infinite slope model. This
    approach allowed to estimate posterior probability distributions of physical parameters
    given a prescribed factor of safety, to assess the most likely depth of failure, and to
    identify the optimum amount of samples required to maximize the reliability of the
    inferences. This work focusses on providing a substantial contribution to improved
    policymaking and management through the use of integrated sources of evidence such as
    real data, model predictions and experts educated beliefs.

publication date

  • December 2017