Using a Bayesian Network Framework to Predict Permafrost Thaw in the Arctic

Abstract

Rising global temperatures are a threat to Arctic ecosystems. Thawing permafrost is expected to expose previously frozen carbon to microbial decomposition, an action that will promote further warming and have consequences for both the natural environment and human communities. However, there is a critical gap in the ability of current permafrost models to simulate permafrost thaw under future projected climate conditions. A model based on Bayesian methods may help address existing limitations in the representation of physically complex processes and availability of observational data. A particular strength of Bayesian methods over more traditional methods is the ability to integrate various types of evidence (e.g., observations, model outputs, or expert assessments) into a single model through probability and statistics. This ability is particularly helpful in regions such as the Arctic that have sparse or no data. Here, I outline a new modeling framework using a Bayesian network (PermaBN) to simulate permafrost thaw in the continuous permafrost region of the Arctic. The PermaBN model development process involves: (1) identifying variables relevant to permafrost thaw via extensive literature review and collaboration with experts at Texas A&M University, (2) pre-validating and validating the model via expert assessment, and (3) evaluating the model with physical observations from a local case study. Pre-validation and expert assessment validation results show that, as expected, increases in thaw depth are expected to be low under initial conditions favoring lower temperatures, increased soil moisture conditions, and high active layer ice content while changes are expected to be high under initial conditions favoring higher temperatures, decreased soil moisture conditions, and low active layer ice content. Model evaluation shows that performance of PermaBN is enhanced when system conditions are known. Future work includes refining the model probabilities, calibrating the model, and evaluating the model performance using a pan-Arctic case study. Results from this study are expected to provide better predictions of permafrost thaw that can then be applied to carbon modeling studies, infrastructure hazard assessments, and policy decisions aimed at mitigation of and adaptation to permafrost thaw.