📄 earth science

Predicting Fire-Following-Earthquake Ignition Probabilities in California: Validation Across Earthquake Events

This study develops and validates a census-tract-level logistic regression model using historical California earthquake data to effectively predict and prioritize spatial fire-following-earthquake ignition risks, demonstrating strong performance in both cross-validation and application to the USGS HayWired earthquake scenario.

Original authors: Guochuan Wei, wei Wang, Donghui Ma

Published 2026-10-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guochuan Wei, wei Wang, Donghui Ma

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When the ground shakes in a city, the immediate danger is often the collapse of buildings and the rupture of infrastructure. Yet, history shows that the most devastating consequence of a major earthquake is frequently the fire that follows. In the 1906 San Francisco earthquake, flames destroyed far more of the city than the shaking itself did, consuming thousands of structures over a vast area. This secondary hazard, known as fire following earthquake, arises when the violent motion displaces contents, breaks gas lines, or topples water heaters, igniting fires just as the emergency response system is overwhelmed by the initial disaster. With water supplies cut and roads blocked, these individual blazes can merge into uncontrollable infernos. For city planners and emergency managers, the critical challenge is not just predicting that fires will happen, but knowing exactly where they are most likely to start. This knowledge allows them to position resources strategically before the ground even begins to move, potentially saving lives and limiting destruction.

A team of researchers from the Beijing University of Technology has tackled this problem by building a new way to map the risk of these fires across California. They focused on a specific question: if an earthquake strikes a particular neighborhood, how likely is it that a fire will start there? To answer this, they looked back at seven major earthquakes that occurred in California between 1983 and 2014. They gathered data on every census tract—a small, defined neighborhood used for population counts—within the affected areas. In total, they analyzed over 7,000 observations, noting which neighborhoods experienced at least one fire and which did not. They then cross-referenced these outcomes with a wide range of factors, including how hard the ground shook, the age and type of buildings, how many people lived there, and even the time of year the earthquake happened.

The researchers faced a significant hurdle in their work: how to ensure their predictions would hold up if a brand-new earthquake struck tomorrow. Many previous studies tested their models by splitting data from the same earthquake into training and testing groups, which can create an illusion of accuracy. To avoid this, the team used a rigorous method where they built their prediction tool using data from six earthquakes and then tested it entirely on the seventh, a process they repeated for every single event in their dataset. This approach, known as nested cross-validation, acts as a strict stress test, ensuring the model learns general rules about fire risk rather than just memorizing the details of past events. The results were promising. When the model was tested against these completely unseen earthquakes, it successfully ranked the neighborhoods, placing the areas that actually caught fire near the top of the list. Specifically, when the researchers looked at the top 20 percent of neighborhoods predicted to be at highest risk, they found that these areas contained nearly 72 percent of all the actual fires that had occurred across the seven historical events.

The study also identified the specific ingredients that make a neighborhood more prone to ignition. The model found that the intensity of the shaking was a primary driver, but so were the characteristics of the built environment. Older housing stock, particularly homes built before 1970, and a higher density of buildings heated by natural gas significantly increased the likelihood of a fire starting. The researchers combined these factors into a final predictive tool that assigns a probability score to every neighborhood. They tested this tool on a hypothetical future scenario: a magnitude 7.0 earthquake on the Hayward Fault, known as the HayWired scenario. The model predicted that the highest risks would be concentrated in Alameda, Contra Costa, and Santa Clara counties. This finding aligned closely with independent estimates of where fires would likely occur in that same scenario, suggesting the model captures the real-world dynamics of fire spread effectively.

The researchers did not stop at a single method; they tested whether their results depended on how they measured the shaking or how they split the data. They found that while different ways of measuring ground motion changed the specific ranking of some neighborhoods, the overall ability of the model to identify high-risk areas remained strong. The most consistent results came when using a specific measure of shaking intensity that correlates well with how humans feel the earthquake and how structures respond. The study concludes that by focusing on the right combination of ground motion, building age, and fuel sources, it is possible to create a reliable map of fire risk. This map does not predict the exact moment a fire will start, but it provides a clear, data-driven guide for where emergency crews should focus their attention. In a world where resources are always limited, knowing where to look first is a powerful tool for reducing the devastation that follows the shaking.

Technical Summary: Predicting Fire-Following-Earthquake Ignition Probabilities in California

Problem Statement
Fire following earthquake (FFE) is a critical secondary hazard that can cause damage exceeding the initial seismic event, as historically demonstrated in the 1906 San Francisco and 1995 Great Hanshin earthquakes. Effective emergency response requires rapid spatial prioritization of resources, yet existing predictive models often lack rigorous validation for earthquakes excluded from the training data. Previous studies frequently partitioned data from the same earthquake event between training and testing sets, leading to potential overestimation of predictive performance and a failure to characterize how well models rank ignition risks for entirely new, held-out seismic events.

Methodology
This study developed a census-tract-level logistic regression model to predict the probability of at least one recorded FFE ignition. The research utilized a dataset of 7,074 earthquake-census-tract observations derived from seven historical California earthquakes (1983–2014), containing 155 recorded ignitions.

Data and Variables: The model incorporated five categories of explanatory variables: ground-motion intensity measures (PGA, PGV, MMI), exposure variables (total building floor area, population density), building vulnerability proxies (pre-1970 housing share, wood/URM/mobile-home floor-area shares), potential ignition source proxies (gas-heated housing density, building occupancy shares), and environmental variables (mean air temperature). Variables were transformed (e.g., logarithmic, square-root) to address skewness.
Modeling Framework: A nested cross-validation design with leave-one-event-out (LOEO) splitting was employed.
- Outer Loop: One entire earthquake event was held out as the test set, while the remaining six events formed the training set. This process was repeated for all seven events.
- Inner Loop: Within the outer training set, a second LOEO split was used to select the optimal model structure from 211 candidate structures (combinations of a base model plus up to three optional variables).
- Assessment: Performance was evaluated on the completely held-out outer test events using Area Under the Receiver Operating Characteristic Curve (AUROC), Average Precision (AP), and Recall@Top-k% (where $k \in \{5, 10, 20\}$ ).
Final Application: The final model structure was selected based on aggregated inner-loop performance, refitted using all 7,074 historical observations, and applied to the USGS HayWired earthquake scenario to identify priority screening sets.

Key Contributions

Rigorous Validation Protocol: The study implements a nested LOEO cross-validation framework that strictly separates model selection from assessment and ensures that entire earthquake events are held out during testing. This addresses the limitation of previous studies where data from the same event was split across training and testing, providing a more realistic estimate of a model's ability to generalize to future, unseen seismic events.
Priority Screening Performance: The study quantifies the effectiveness of the model in identifying high-risk census tracts. It demonstrates that the Top20% of census tracts ranked by predicted probability capture a significant majority of actual ignition locations across diverse historical events.
Scenario Application: The validated model is applied to the HayWired scenario, offering a spatially explicit method for prioritizing resources that aligns with established regional ignition estimates.

Results

Predictive Performance: Across the seven outer test events, the model achieved an event-mean AUROC of 0.833 and an event-mean AP of 0.267.
Recall Metrics: The Top20% priority screening sets contained 71.9% (97 out of 135) of all census tracts with at least one recorded ignition. In every individual outer test event, the Top20% set included at least half of the observed ignition tracts.
Model Structure: The final selected model structure included four variables: ShakeMap instrumental intensity (MMI), total building floor area (ln(1+TFA)), pre-1970 housing-unit share (Pre70), and utility-gas-heated occupied housing unit density (square-root transformed). All coefficients were positive, indicating that higher values in these variables correlate with increased ignition probability.
Calibration: Calibration analysis revealed that predicted probabilities were slightly overestimated overall (intercept = -0.638, slope = 0.824). However, in the low-probability range where most observations reside, the predictions closely matched observed ignition proportions.
Sensitivity: Sensitivity analyses comparing different ground-motion measures (MMI vs. ln(PGA) vs. ln(PGV)) showed that while rankings varied for specific events, the MMI model generally provided superior overall ranking discrimination and recall. Comparisons with standard stratified five-fold cross-validation indicated that LOEO results were slightly more conservative but more appropriate for assessing performance on completely new events.
HayWired Scenario: When applied to the HayWired scenario, the model identified priority screening sets concentrated in Alameda, Contra Costa, and Santa Clara counties. These three counties accounted for 93.1% to 95.4% of the Top5% to Top20% screening sets, closely matching the 83.5% share of ignition estimates reported in the original HayWired study for the same counties.

Significance
The paper claims that its results characterize the ranking performance of FFE ignition models across historical events and support scenario-based priority screening under specific data and modeling assumptions. By demonstrating that a census-tract-level model can effectively identify high-risk areas in held-out events, the study provides a methodological basis for improving urban FFE risk reduction and emergency response planning. The findings suggest that focusing response resources on the top-ranked 20% of census tracts can capture the majority of potential ignition sites, offering a practical tool for decision-makers when resources are constrained.

Technical Summary: Predicting Fire-Following-Earthquake Ignition Probabilities in California

More like this