Characterizing and Predicting Wildfire Evacuation Behavior: A Dual-Stage ML Approach

This study utilizes a dual-stage machine learning approach on a large-scale survey of residents in California, Colorado, and Oregon to identify latent behavioral typologies based on household resources and predict evacuation outcomes, revealing that while transportation mode is highly predictable from static characteristics, evacuation timing remains difficult to classify due to its dependence on dynamic fire conditions.

The Gist

Imagine a massive, chaotic dance floor during a wildfire. Everyone is trying to leave the building, but they aren't all moving the same way. Some are sprinting out the door immediately; others are stopping to grab their pets, pack a suitcase, or wait to see if the fire is actually coming. Some have cars, some have bikes, and some are stuck because they don't have a ride.

This paper is like a team of detectives trying to figure out who is on that dance floor, how they are moving, and why they are moving the way they are. They used a giant survey (like a digital questionnaire) sent to people in California, Colorado, and Oregon to gather clues.

Here is the breakdown of their investigation using simple analogies:

1. The Detective Work: Two Types of Tools

The researchers didn't just look at the data; they used two different "superpowers" (Machine Learning) to understand it.

• Superpower A: The "Grouping" Glasses (Unsupervised Learning)
  Imagine you walk into a room full of strangers and you have to sort them into teams without asking them anything. You just look at their shoes, their backpacks, and how they are standing.

  • What they did: They used tools called K-Modes and Latent Class Analysis to sort the 853 survey respondents into 6 distinct "tribes" based on their habits and resources.
  • The 6 Tribes they found:
    1. The "Stuck" Group: People with no cars, no plans, and no tech. They are the most vulnerable.
    2. The "Stable" Group: Long-time residents with multiple cars and a solid plan. They know exactly what to do.
    3. The "Newcomers": Renters or recent movers who might be confused or lack resources.
    4. The "Super-Prepared": People with cars, GPS, smartphones, and a written plan. They are ready to go.
    5. The "Pet Parents": People whose evacuation is complicated by animals (pets or livestock). They have extra hurdles to jump over.
    6. The "Mixed Bag": People who didn't fit neatly into any category; their decisions were all over the place.

• Superpower B: The "Crystal Ball" (Supervised Learning)
  Now, imagine you want to predict the future. You have a list of facts about a person (do they own a car? do they have a plan?) and you want to guess what they will do next.

  • What they did: They built computer models to predict two specific things: When will they leave? and How will they leave?

2. The Big Surprise: What the Crystal Ball Could (and Couldn't) See

The researchers tested their "Crystal Ball" on two questions, and the results were very different.

Question 1: "How will you get out?" (Transportation Mode)

• The Result: The Crystal Ball was amazingly accurate (about 89% right).
• The Analogy: This is like predicting if someone will drive a car or ride a bike based on whether they own a car. If you own a car and have a plan, you almost certainly drive. If you don't, you might walk or take a bus.
• Why it matters: Emergency planners can use this to know exactly how many cars will be on the road and how many people will need rides. They can prepare the right amount of buses or shelters.

Question 2: "When will you leave?" (Evacuation Timing)

• The Result: The Crystal Ball was pretty bad (only about 60% right).
• The Analogy: This is like trying to predict exactly when a person will decide to jump off a diving board just by looking at their swimsuit. You can't do it!
• Why it failed: The time someone leaves depends on things that happen right now—like seeing smoke, hearing a siren, or getting a text message. These are "real-time" events that a survey filled out months ago cannot capture. It's too chaotic and changes too fast to predict with a static list of facts.

3. The Takeaway for Real Life

So, what does this mean for saving lives?

• We can plan the "How": Because we know exactly who has cars and who doesn't, emergency managers can be very smart about traffic control and helping people without vehicles. We can say, "Okay, this group needs a bus; that group needs a ride."
• We can't predict the "When" easily: We can't just look at a person's age or income and say, "They will leave at 4:00 PM." Because the timing is so unpredictable, we need better, faster warning systems. We need to keep people informed in real-time (like live smoke maps or instant alerts) rather than relying on old data.

In a nutshell:
This study tells us that while we can easily predict how people will escape a wildfire based on their resources (cars, pets, plans), we cannot easily predict when they will leave because that depends on the chaotic, changing situation of the fire itself. The best defense is to help people prepare their "How" now, so they are ready when the "When" finally happens.

Technical Summary

1. Problem Statement

Wildfire evacuations in the Western United States (specifically California, Colorado, and Oregon) present unique challenges due to the rapid, unpredictable nature of fire spread and the heterogeneity of resident decision-making. Existing research often relies on post-disaster interviews or simulation models that fail to capture real-time decision complexity. Furthermore, traditional statistical methods often assume linear relationships or predefined behavioral categories, limiting their ability to model the high-dimensional, interdependent factors influencing evacuation behavior (e.g., demographics, resource access, risk perception, and pet ownership).

The core problem addressed is the lack of a comprehensive framework that simultaneously:

1. Discovers latent behavioral typologies within large-scale survey data without prior assumptions.
2. Predicts critical evacuation outcomes (timing and transportation mode) using static household characteristics to inform emergency planning.

2. Methodology

The study employs a Dual-Stage Machine Learning Approach combining unsupervised learning for pattern discovery and supervised learning for predictive modeling.

A. Data Collection and Preprocessing

• Source: A large-scale survey conducted via Amazon Mechanical Turk (MTurk) in April–May 2022.
• Population: Residents of wildfire-prone areas in CA, CO, and OR.
• Sample: Initial 1,312 responses reduced to 853 valid responses after rigorous three-stage quality control (attention checks, consistency verification, and logical household composition rules).
• Variables: Sociodemographics, mobility (vehicle count, GPS access), risk perception, past experience, pet ownership, and behavioral outcomes (timing, mode, route, destination).

B. Unsupervised Learning (Behavioral Clustering)
To identify latent subgroups, the authors applied three complementary techniques:

1. Multiple Correspondence Analysis (MCA): Used to visualize the high-dimensional categorical data and reduce it to latent dimensions, revealing a continuous behavioral spectrum rather than discrete clusters.
2. K-Modes Clustering: Applied to categorical data to partition households into distinct groups based on shared characteristics. The optimal solution identified 6 clusters.
3. Latent Class Analysis (LCA): A probabilistic approach used to validate the K-Modes results, confirming the stability of the identified behavioral typologies.

C. Supervised Learning (Predictive Modeling)
To predict key outcomes, three algorithms were trained on a 75/25 train-test split:

• Algorithms: Logistic Regression, Random Forest, and XGBoost.
• Target Variables:
  1. Evacuation Timing: Binary classification (Early vs. Late).
  2. Transportation Mode: Multi-class classification (Private vehicle vs. Non-private/Other).
• Evaluation Metrics: Accuracy, Weighted F1-Score, Cohen's Kappa, and ROC-AUC.

3. Key Contributions

• Hybrid ML Framework: First study to integrate unsupervised behavioral clustering (MCA, K-Modes, LCA) with supervised classification (RF, XGBoost) specifically for wildfire evacuation survey data.
• Identification of Latent Typologies: The study moves beyond broad demographic correlations to define specific, actionable behavioral archetypes based on resource availability and preparedness.
• Differentiation of Predictability: The research explicitly distinguishes between outcomes that are structurally determined by household attributes (transportation mode) and those driven by dynamic, real-time situational factors (evacuation timing).
• Scalable Data Collection Validation: Demonstrates the viability of MTurk for behavioral disaster research when paired with strict quality control, despite known sampling biases.

4. Key Results

A. Unsupervised Clustering Findings
The analysis revealed six distinct behavioral clusters:

• Cluster 0 (Resource-Limited): Limited vehicle access, minimal planning, low tech support. High risk of evacuation barriers.
• Cluster 1 (Stable/Resident): Long-term residents, multiple vehicles, family-based destinations.
• Cluster 2 (Transient): Short-term renters/recent movers with constrained mobility and weak preparedness.
• Cluster 3 (Highly Prepared): Multiple vehicles, high tech access (GPS/Smartphones), written evacuation plans. Fastest/most deliberate evacuees.
• Cluster 4 (Pet/Livestock Owners): Decisions heavily constrained by animal logistics.
• Cluster 5 (Heterogeneous/Uncertain): Ambiguous response patterns, lacking a structured decision process.

B. Supervised Learning Performance

• Transportation Mode Prediction (High Success):
  • Models achieved high accuracy (~86–89%) and F1 scores.
  • Conclusion: Transportation mode is strongly determined by static household features (vehicle ownership, income, preparedness). It is highly predictable.
• Evacuation Timing Prediction (Low Success):
  • Models showed modest performance (Accuracy ~53–62%, ROC-AUC ~0.61).
  • Conclusion: Timing is difficult to predict using static data alone. It is heavily influenced by dynamic, unobserved variables such as real-time smoke visibility, fire proximity, and community alerts.

5. Significance and Implications

• For Emergency Management:
  • Resource Allocation: Since transportation mode is predictable, planners can accurately forecast traffic loads and allocate resources (e.g., buses for non-vehicle households) based on demographic profiles.
  • Targeted Messaging: Different clusters require different communication strategies. For example, "Cluster 0" needs assistance with mobility, while "Cluster 4" needs specific guidance on pet evacuation.
• For Policy and Planning:
  • The low predictability of timing suggests that static demographic models are insufficient for predicting when people will leave. Emergency systems must rely on dynamic warning systems and real-time situational intelligence rather than just pre-event demographic profiling.
• Methodological Advancement: The study validates a dual-stage ML pipeline that can uncover hidden behavioral structures in complex disaster data, offering a template for future research in other hazard contexts (e.g., hurricanes, floods).

Limitations & Future Work:
The study acknowledges MTurk sampling biases (younger, more educated, tech-savvy). Future work should integrate real-time hazard data (fire spread models, smoke sensors) and longitudinal data to improve timing predictions and address equity gaps for digitally marginalized populations.