PhD Student Washington State University
A Transferable, Cause-Specific Deep Learning Framework for Wildfire Ignition Probability Estimation
Co-Author: Ji Yun Lee (Washington State University)
Abstract: The effectiveness of wildfire mitigation actions depends on knowing not only where ignitions are likely to occur, but also why they occur and how that risk changes over time and space. Thus, accurate ignition prediction is essential for identifying priority areas and designing cause-specific, efficient wildfire mitigation strategies. However, most ignition probability estimation models have several methodological challenges. Their static, site-specific nature fails to capture dynamic changes in landscape and climatic conditions, while requiring re-training for new regions due to a lack of transferability. Moreover, they often use simple binary classification algorithms, making strong assumptions about non-ignition data and providing little insight into statistical ignition causes.
To address these limitations, our study develops a transferable, cause-specific deep learning (DL) framework for estimating wildfire ignition probability. The framework starts with a generalized model, which is trained on multi-region datasets to learn transferable relationships between key drivers and observed ignitions. To relax the assumptions in binary classification, the generalized model utilizes presence-only methods (e.g., Maximum Entropy) within a DL architecture to better represent ignition probability distributions. This model is further designed to generate cause-specific ignition predictions, which can support targeted mitigation strategies. Finally, transfer learning is used to fine-tune the generalized model for local conditions, aimed at producing accurate site-specific sub-models without much additional data. By combining generalized underlying ignition patterns, cause-specific prediction, and site-level adaptation, the proposed framework enables accurate ignition risk assessment, reduces the need for redeveloping region-by-region models, and directly links probabilistic forecasts to targeted, problem-oriented wildfire mitigation decisions.
Researcher University of Washington, NHERI RAPID; University of California, Berkeley
From Imagery to Engineering Insight: Automating Post-Disaster Fire Damage Prediction with AI Using RAPID Imagery
Co-Authors: Karen Dedinsky (University of Washington), Jeffrey Berman (University of Washington), and Joseph Wartman (University of Washington)
Abstract: As wildfire and urban fire events grow in frequency and severity, engineers need methods to rapidly convert large volumes of post-disaster imagery into actionable information on structural performance and damage patterns. For this purpose, tools that enable automated prediction of fire damage from imagery are becoming increasingly critical, as they offer a rapid, consistent, and scalable approach for post-disaster engineering assessment.
We present an AI-enabled workflow for automated fire-damage classification from post-disaster imagery, developed by the RAPID facility at the University of Washington. The workflow is implemented in the open-source rAPIdtools Python package, which streamlines image ingestion, building-level image extraction, AI-based damage prediction, and dataset generation into a unified and extensible pipeline.
Leveraging RAPID post-disaster imagery, rAPIdtools enables repeatable, end-to-end automation that can be deployed across large spatial extents with minimal manual intervention. We evaluate the performance of this approach using RAPID datasets for the 2025 Palisades and Eaton fires, comparing automated classifications against manually curated labels. The results indicate that AI-based image analysis can create accurate building-level damage maps at large-scale and substantially reduce the time, cost, and safety risks associated with traditional post-fire reconnaissance.
By standardizing damage classification and enabling rapid creation of large, labeled datasets, this work supports the development of fragility functions, performance-based fire engineering, and community-scale resilience analysis for fire-borne hazards. The rAPIdtools framework lowers barriers to incorporating AI into post-disaster workflows, promotes reproducible research, and advances data-driven improvements in fire-resilient design, planning, and mitigation.
Practicing engineer Exponent
Assessing Structure-to-Structure Fire Vulnerability in WUI Neighborhoods
Co-Author: Liana Wong (Stanford University)
Abstract: The increasing frequency and severity of Wildland–Urban Interface (WUI) fires demand scalable, data-driven approaches to quantify structure-to-structure fire spread vulnerability. While many existing risk models emphasize wildfire behavior such as fuels, weather, and topography, post-ignition losses in dense residential neighborhoods are often governed by spatial vulnerability, particularly building separation distance and the resulting radiant heat exposure. This study presents a computational framework that advances toward a WUI structural risk assessment index by focusing explicitly on vulnerability driven by building proximity. BRAILS++ is used to automatically extract granular building attributes including footprints, heights, and areas from multi-source imagery. The framework is demonstrated using the Pacific Palisades area in Los Angeles, which was severely impacted by the January 2025 Palisades Fire. By integrating BRAILS++ feature extraction with the enclosure fire dynamics program CFAST, an automated workflow is developed to quantify building- and cluster-level vulnerability. The framework evaluates how structural arrangement influences potential structure-to-structure radiative exposure, rather than relying on post-fire damage classification. While many existing studies prioritize construction materials, this study emphasizes building separation distance coupled with CFAST-based radiative heat flux as the primary vulnerability metric. Radiative heat flux estimates depend on simplified building geometry and assumed ventilation and fuel load conditions defining post-ignition fire scenarios. Secondary factors including defensible space and WUI wildfire-resistant construction compliance are also considered. Building-level and cluster-level vulnerability indices are computed by weighting the calculated metrics. The results enable data-driven identification of high-risk spatial hotspots and support targeted mitigation and resilience planning in WUI communities.
PhD Student University of Nevada, Reno
Probabilistic Framework for Uncertainty Propagation in Wildfire Risk Assessment
Co-Authors: Dani Or (University of Nevada, Reno) and Hamed Ebrahimian (University of Nevada, Reno)
Abstract: We present a Probabilistic Wildfire Risk Assessment (PWRA) framework that advances wildfire risk analysis by defining risk as a spatial probability density function of loss while systematically propagating uncertainty across the full hazard-to-loss chain. The PWRA employs a deterministic uncertainty-propagation strategy based on the Generalized Unscented Transform that greatly reduce computation burden of prevailing approaches that depend on large ensembles of stochastic fire simulations. This enables efficient propagation of uncertainty in ignition conditions, fire weather, and fuel properties that require only a small fraction of the simulations typically needed by Monte Carlo–based methods. The PWRA framework is modular, supporting consistent coupling of a range of hazard and system models with downstream damage and loss components, and accommodating diverse uncertainty sources, performance metrics, and stakeholder-oriented objectives. The framework is demonstrated through an application to the 2018 Camp Fire domain. Results show that PWRA reproduces spatial patterns of burn probability consistent with historical observations and generates exceedance-rate and return-period maps for multiple fire behavior metrics—including flame length, fireline intensity, and heat per unit area. By providing a unified, flexible, computationally efficient, and uncertainty-aware formulation for wildfire risk analysis, the PWRA framework helps bridge the gap between wildfire hazard modeling and fire safety engineering, supporting the development of performance-based metrics that inform mitigation strategies and design decisions in wildfire-prone environments.
Associate Professor Unniversity of California, Los Angeles
Wildfire Exposure and Household Financial Distress: Evidence from Mortgage Default Data
Abstract: Wildfires are an increasingly frequent and severe natural hazard, with cascading impacts that extend beyond physical damage. While prior research has examined wildfire effects on housing markets and insurance, less is known about how wildfire damage affects financial distress and mortgage performance. This study investigates the causal effect of wildfire exposure on mortgage default using a high-resolution longitudinal dataset that links loan-level performance records with wildfire damage information.
We employ a difference-in-differences (DiD) framework to compare changes in probable mortgage default outcomes for loans located in wildfire-affected ZIP codes to those in unaffected areas before and after major wildfire events. To address covariate imbalance, heterogeneous exposure, and potential model misspecification, we apply modern causal inference methods, including doubly robust DiD estimators that combine outcome regression with propensity score weighting. This approach improves the credibility and efficiency of estimated treatment effects while allowing flexible control for borrower, loan, and neighborhood characteristics.
Our findings provide new empirical evidence on the financial consequences of wildfire damage for households, demonstrating how disaster exposure translates into elevated mortgage distress and default risk over time. The results reveal meaningful spatial and temporal heterogeneity in impacts, with implications for mortgage markets, disaster recovery policy, and climate-related financial risk assessment. More broadly, this work illustrates how integrating hazard data with large-scale financial datasets and advanced causal inference techniques can improve understanding of disaster impacts within complex socio-technical systems. The study contributes to ongoing NHERI SimCenter efforts to develop data-driven, interdisciplinary frameworks for modeling and managing disaster risk.
Assistant Professor University of Waterloo
An agent-based model for couple housing and household recovery simulation
Co-Authors: Ali Nejat (Texas Tech University), Sara Hamideh (Stony Brook University), and Elaina Sutley (The Univesity of Kansas)
Abstract: As climate change increases the frequency and severity of disasters, proactive planning for post-disaster housing recovery is essential to mitigate long-term social and economic disruption. Computational models can support this planning by simulating potential recovery trajectories, yet many existing approaches are limited by overwhelming data requirements or narrow applicability to past events. Moreover, existing models tend to either focus on the progress of repairs to the housing unit (housing recovery) or shelter status of the impacted households (household recovery).
This research introduces a novel agent-based model for post-disaster recovery simulation that couples housing and household recovery. In our model, individual households, insurers, and contractors are agents governed by empirical behavior rules, and incorporates modifiable system-level constraints, such as contractor availability, to reflect context-specific recovery dynamics. The model simulates how housing damage forces households into temporary housing arrangements (e.g., stay in a hotel or rental unit) until the housing unit is rebuilt, and how displacement incurs additional living costs to households, which may translate into delays in housing recovery decisions. Thus, the model captures correlations between household and housing recovery processes.
We demonstrate the model’s utility by hindcasting two California wildfires—the 2017 Tubbs Fire and the 2018 Camp Fire. Our model reproduces temporal and spatial patterns of recovery observed in building permit and construction data. By balancing generalizability with data realism, our model provides a flexible and transferable tool for post-disaster recovery planning, supporting more effective decision-making under uncertainty and enhancing community resilience in the face of escalating climate risks.