Using Hybrid AI for Multiscale Wildfire Risk Assessment and Mitigation

The risk of wildfires has increased significantly in recent years and touched communities not previously at high risk (Burke et al., 2021; Radeloff et al., 2018). Effective mitigation of wildfire risk is essential to reduce the potential for catastrophic losses. Accurate assessment of the wildfire risk at a property level will enable fire departments and homeowners to take appropriate mitigation measures and reduce losses. We developed a Wildfire Integrated Modeling, Prediction, and Learning Environment (WIMPLE), a hybrid AI (HAI) tool for wildfire risk assessment under a NASA-funded project. WIMPLE is based on our Scruff HAI framework (Pfeffer & Lynn, 2018), which provides integration of different types of AI models, sharing and composition of models, and spatiotemporal flexibility.

WIMPLE combines models of different spatiotemporal resolutions from regional to property level to generate a stochastic risk assessment of wildfire occurrence and impact on a property. The user provides the WIMPLE system with the location and layout of their property and WIMPLE provides the user with information on the risk posed by wildfires to their property. These risk metrics include risk of a wildfire in surrounding region, risk of a wildfire in the property, and expected damage to the property (USD). These values are calculated using several distinct models all integrated under the Scruff HAI framework.

The foundational component of the HAI model is the environmental condition model. The environmental condition model generates a probability distribution of likely future environmental conditions from marginalizing the NASA Earth Exchange Global Daily Downscaled Projections (Thrasher et al., 2022) by year and season. The advantage of using this dataset is that it incorporates different climate change trajectories in the estimations of future environmental conditions. This allows the underlying predicted environmental conditions for other model components to reflect the progress of climate change.

The distributions of environmental conditions generated by the environmental condition model are used in the regional fire likelihood, regional fire spread, and property fire spread models. The regional fire likelihood model uses a Random Forest model trained on data from past fires and climatic conditions in a region to determine the likelihood of fires occurring in that region in the future. It generates the probability of a fire occurring in a region based on the predicted environmental conditions from the environmental condition model. The regional fire spread model uses the set of environmental conditions from the environmental condition model in combination with the USFS regional wildfire simulation tool, FlamMap (Finney, 2006) to generate a likely area of effect of a fire impacting the target region. Finally, the property fire spread model uses the conditions from the environmental condition model and a layout of the property supplied by the user to simulate a fire spreading across the property with Fire Dynamics Simulator (FDS) (Kevin McGrattan, et al., 2013). This component simulates the damage and trajectory of a fire through the target property under different initial ignition points and uses the results to determine the expected damage to the property in USD.

The goal of WIMPLE is to provide homeowners with information on the risk wildfires pose to their property and suggest mitigation approaches to lessen that risk. WIMPLE uses explainable AI (XAI) techniques to visualize the outputs of the model components in the user interface at the regional and property level scales. Each of these scales informs different characteristics of wildfire risk. The regional risk visualization conveys the risk of wildfires around the property, potentially motivating the end user to consider alternative house locations. The property level visualization shows the simulation of different paths that a fire could take across a property, highlighting potential hazards that property owners should address.

In future development, we hope to extend the property level simulation to include counterfactual simulations to highlight the important role of mitigating fire hazards on a homeowner’s property. For instance, a counterfactual could read: “Under the current property layout, the expected damage from wildfires is $300,000. However, if you remove the trees at points A and B, the expected damage would decrease by 30% to $210,000.” This statement would be accompanied by simulations of a fire in the property without the trees mentioned in the mitigation suggestion above. This allows the end user to see the visual and statistical impacts of wildfire mitigation measures, hopefully motivating them to take action to protect their property.

References

Burke, M., Driscoll, A., Heft-Neal, S., Xue, J., Burney, J., & Wara, M. (2021). The changing risk and burden of wildfire in the United States. Proceedings of the National Academy of Sciences, 118(2), e2011048118. http://doi.org/10.1073/pnas.2011048118

Finney, M. A. (2006). An Overview of FlamMap Fire Modeling Capabilities.

Kevin McGrattan, Randall McDermott, Craig Weinschenk, & Glenn Forney. (2013). Fire Dynamics Simulator, Technical Reference Guide, Sixth Edition. http://doi.org/10.6028/NIST.sp.1018

Pfeffer, A., & Lynn, S. K. (2018). Scruff: A deep probabilistic cognitive architecture for predictive processing. Presented at the Conference on Biologically Inspired Cognitive Architectures, Prague, Czech Republic.

Radeloff, V. C., Helmers, D. P., Kramer, H. A., Mockrin, M. H., Alexandre, P. M., Bar-Massada, A., ... Stewart, S. I. (2018). Rapid growth of the US wildland-urban interface raises wildfire risk. Proceedings of the National Academy of Sciences, 115(13), 3314–3319. http://doi.org/10.1073/pnas.1718850115

Thrasher, B., Wang, W., Michaelis, A., Melton, F., Lee, T., & Nemani, R. (2022). NASA Global Daily Downscaled Projections, CMIP6. Scientific Data, 9(1), 262. http://doi.org/10.1038/s41597-022-01393-4