Evac-Cast: Interpretable Evacuation Forecasting

• Evac-Cast is an interpretable, survey-free machine learning framework designed to forecast tract-level evacuation rates by leveraging hazard intensity, vulnerability, readiness, and built environment data.
• It employs hazard-specific XGBoost regressors alongside SHAP for transparent, non-linear interaction modeling, enabling clear feature attribution across hurricanes and wildfires.
• The framework achieves mean absolute errors of about 4.5% for hurricanes and 3.5% for wildfires, supporting targeted emergency management and resource allocation.

Evac-Cast is an interpretable, survey-free machine-learning framework for tract-level evacuation forecasting across hurricanes and wildfires. It was introduced to address a core operational challenge in emergency management: producing timely, accurate, and transparent predictions of evacuation rates at fine spatial resolution before and during hazard events. The framework integrates over 20 features from four dimensions—hazard intensity, community vulnerability, evacuation readiness, and built environment—then trains hazard-specific XGBoost regressors and interprets them with SHAP. In the reported evaluation, Evac-Cast was applied to two hurricanes in Florida and two wildfires in California, achieving mean absolute errors of approximately $4.5\%$ for hurricanes and $3.5\%$ for wildfires while revealing a stable cross-hazard feature hierarchy led by hazard intensity (Li et al., 1 Aug 2025).

1. Operational problem and conceptual scope

Evacuation forecasting is operationally consequential because accurate forecasts enable targeted messaging, pre-staging of resources such as fuel and contra-flow operations, appropriate shelter scaling, and informed assignment of search-and-rescue assets to neighborhoods anticipated to exhibit low departure rates (Li et al., 1 Aug 2025). The problem is difficult because behavioral drivers are multi-factorial, highly non-linear, and context-dependent across hazards; traditional predictors such as risk perception, social ties, and efficacy beliefs are often obtained through post-event surveys that are costly, slow, and prone to recall bias; many analyses assume linear relationships or rely on black-box models with limited transparency; and most studies focus on single hazards, limiting generalizability.

Evac-Cast addresses these constraints by replacing time-consuming micro-level behavioral surveys with a tract-scale model built from hazard intensity, vulnerability, readiness, and built environment. Its core design choice is to use multi-source, high-availability datasets that can be updated near real time and aligned to census tracts. Separate gradient-boosted tree regressors are trained for hurricanes and wildfires, allowing the model to capture non-linear interactions without sacrificing interpretability, while SHAP provides both global and local attribution of feature effects (Li et al., 1 Aug 2025).

A common misconception is to treat Evac-Cast as a generic all-hazards predictor. The reported system is not a single universal model; it uses separate models for hurricanes and wildfires. The generalizable claim is more limited and more precise: the dimensional ordering of feature importance is reported as stable across these hazard classes, even though the underlying regressors are hazard-specific.

2. Events, geography, and evacuation-rate construction

Evac-Cast was evaluated on four events spanning two hazard types: Hurricane Ian (2022), Hurricane Milton (2024), the Kincade Fire (2019), and the Palisades–Eaton fires (January 2025) (Li et al., 1 Aug 2025). The spatial unit is the US census tract within counties that issued official evacuation orders for each event. The temporal windows differ by hazard: for hurricanes, the forecast period is the three days leading up to landfall; for wildfires, the window begins on the ignition date, and the peak evacuation rate within the evolving window is used as the proxy.

Event	Geography and window	Evacuation-rate source
Ian (2022)	Florida; three days leading up to landfall	Cuebiq mobility data with Resilitix AI
Milton (2024)	Florida; three days leading up to landfall	Facebook “Population During Crisis”
Kincade (2019)	California; ignition date to peak evacuation window	Cuebiq mobility data with Resilitix AI
Palisades–Eaton (2025)	California; ignition date to peak evacuation window	Facebook “Population During Crisis”

The outcome construction differs by data source. For Ian and Kincade, evacuation rates were computed from Cuebiq mobility data using Resilitix AI techniques to infer stable pre-event home locations and detect evacuation and return transitions via five-day rolling windows. A device was labeled “evacuated” when it remained outside its pre-event home tract for at least three consecutive days, and “returned” when presence was re-established for at least three consecutive days; device-level trajectories were then aggregated to tract-level daily evacuation rates. For Milton and Palisades–Eaton, evacuation rates were computed from Facebook “Population During Crisis” by aggregating Bing level-14 tiles to tracts via tile centroids, averaging 8-hour counts to daily values, and calculating daily evacuation rates versus baseline (Li et al., 1 Aug 2025).

This measurement strategy is central to the framework’s survey-free character. A plausible implication is that Evac-Cast operationalizes evacuation behavior through mobility-derived tract aggregates rather than through respondent-level intention or perception data.

3. Feature dimensions and geospatial-temporal engineering

Evac-Cast integrates over 20 features across four dimensions: hazard intensity, community vulnerability, evacuation readiness, and built environment (Li et al., 1 Aug 2025). Hazard intensity includes proximity to the hazard, official risk designations, and meteorological variables. For hurricanes, proximity is the minimum distance from tract centroid to best-track center across the forecast window using the NOAA NHC GIS archive; for wildfires, it is the distance from tract boundary to the nearest CAL FIRE perimeter. Official designations include Florida evacuation zones $A$–$E$ and CAL FIRE Fire Hazard Severity Zones. Meteorology is queried at tract centroids from the Open-Meteo Historical Forecast API and aggregated over event-specific windows.

Community vulnerability variables are tract-level ACS measures: median income, poverty rate, youth ratio, elderly ratio, high school attainment, living alone ratio, household size, disabled ratio, and non-white ratio. Evacuation readiness comprises renter share and households without personal vehicle. Built-environment variables include building density from Bing Maps building footprints, population density from ACS and TIGER/Line, road density from TIGER/Line, and mobile-home share (Li et al., 1 Aug 2025).

Geospatial alignment is implemented through tract centroids for meteorology, spatial joins for zones and hazards, and polygon-boundary geometry for distance calculations. For categorical designations such as evacuation zones and fire severity, the paper states that features are aligned to tracts via overlay, but encoding details are not explicitly specified. The paper also does not explicitly detail missing-data handling or normalization. It notes that XGBoost does not require feature scaling and that missing values are typically handled internally by learned default directions in splits; one-hot or ordinal mapping would be appropriate where categorical encodings are used. These are identified as operational assumptions following common practice in tree-based modeling rather than as fully specified parts of the published implementation (Li et al., 1 Aug 2025).

The feature design embodies the paper’s central claim about behavioral proxies. Hazard intensity is interpreted as conveying perceived risk; socioeconomic and demographic attributes are treated as reflecting capacity constraints and likely compliance; and built-environment indicators are treated as capturing feasibility and exposure. Because no ablation study is reported, this interpretation remains inferential, but it is explicitly supported by the reported performance and SHAP hierarchy.

4. Model architecture, objective, and training protocol

Evac-Cast trains hazard-specific XGBoost regressors, selecting gradient-boosted decision trees for their ability to capture non-linearities and interactions, robustness to mixed feature types, and support for clear post-hoc interpretability via SHAP (Li et al., 1 Aug 2025). The additive model is written as

$f(x) = \sum_{k=1}^{K} f_k(x),$

with training objective

$$L(\theta) = \sum_{i=1}^{N} \ell(y_i,\hat{y}_i) + \sum_{k=1}^{K} \Omega(f_k),$$

where

$\hat{y}_i = \sum_{k=1}^{K} f_k(x_i).$

For regression, the paper gives squared error loss,

$\ell(y,\hat{y}) = (y-\hat{y})^2,$

and the usual XGBoost regularization,

$$\Omega(f) = \gamma T + \frac{1}{2}\lambda \lVert w \rVert^2,$$

with optional $L_1$ regularization $3.5\%$0 to encourage sparsity.

The reported training protocol uses a random $3.5\%$1 train/test split per hazard type. To address imbalanced targets, the training data are oversampled with SMOGN. Hyperparameters are selected through random search with 10-fold cross-validation on the training data to minimize mean squared error; the explored parameters include max_depth, learning_rate, gamma, and min_child_weight. The paper does not disclose the final tuned values for n_estimators, max_depth, learning_rate, subsample, colsample_bytree, reg_alpha, or reg_lambda, stating that these are event-specific and selected via cross-validation (Li et al., 1 Aug 2025).

The end-to-end pipeline begins with defining the study area as tracts in counties with official evacuation orders, constructing mobility-derived evacuation outcomes, assembling hazard, ACS, TIGER, and Bing context features, aggregating them to event windows, splitting the data, applying SMOGN, tuning the XGBoost model, evaluating with standard regression metrics, and generating SHAP explanations. For deployment, the paper states that features can be computed from real-time feeds for a new event, tracts can be scored, and both predictions and SHAP explanations can be published (Li et al., 1 Aug 2025).

A methodological caveat is that the study reports random splits at the tract level and hazard-specific modeling, but no additional spatial or temporal leakage controls are described. The paper recommends grouped cross-validation by county or event date for future operationalization.

5. Performance, error metrics, and SHAP-based interpretation

Performance is reported on held-out test sets. For hurricanes, the paper gives $3.5\%$2, $3.5\%$3, $3.5\%$4, and $3.5\%$5. For wildfires, it gives $3.5\%$6, $3.5\%$7, $3.5\%$8, and $3.5\%$9 (Li et al., 1 Aug 2025). The associated formulas are

$A$0

$A$1

and

$A$2

The headline interpretation in the paper is mean absolute errors of approximately $A$3 for hurricanes and $A$4 for wildfires, with $A$5 values around $A$6–$A$7. Event-level breakdowns, calibration curves, statistical significance tests, and confidence intervals are not reported.

Interpretability is provided through SHAP, which decomposes predictions as

$A$8

where $A$9 are feature-level Shapley values (Li et al., 1 Aug 2025). For hurricanes, global SHAP importance is led by maximum wind speed and hurricane proximity, with evacuation zone designation, wind gusts, and precipitation also acting as strong positive contributors. The paper further reports positive contributions from renter ratio, mobile-home share, road density, median income, living alone ratio, youth percentage, and high school attainment, while population density contributes negatively. The local explanations indicate that high maximum wind speed and proximity increase tract-level predicted evacuation, higher renter and mobile-home shares increase evacuation, and dense tracts tend to show lower predicted evacuation, reflecting urban shelter-in-place feasibility.

For wildfires, global SHAP importance is led by fire proximity and fire hazard severity zone, followed by meteorological variables such as mean temperature, minimum humidity, maximum wind, and gusts. High school attainment is described as notably high impact; median income and non-white ratio are also influential. Local explanations show that tracts closer to the fire or in higher severity zones have higher predicted evacuation, and higher educational attainment and income correlate with greater predicted evacuation (Li et al., 1 Aug 2025).

The cross-hazard result is especially important: the dimensional hierarchy is stable across hazards, with hazard intensity dominating, followed by community vulnerability, built environment, and evacuation readiness. The paper reports that the gap between hazard intensity and vulnerability is large for hurricanes and narrower for wildfires, which it interprets as reflecting different decision regimes for slow-onset widely forecast threats versus rapid-onset localized threats.

6. Operational use, limitations, and term ambiguity

The intended operational uses include forecasting tract-level evacuation rates one to three days before hurricane landfall and continuously during wildfire progression; targeting public messaging to tracts with low predicted evacuation rates; tailoring message content based on SHAP-driven feature profiles; allocating resources such as fuel, contra-flow timing, and shelter capacity; and prioritizing search-and-rescue assets for tracts with low predicted departure under high hazard exposure (Li et al., 1 Aug 2025). The reported input streams at forecast time are the latest hazard intensities and proximity measures, current ACS/TIGER/Bing context, and mobility-derived baselines and near-real-time evacuation-rate estimates for monitoring and calibration. Meteorological and hazard feeds can be updated hourly to daily, and predictions can be refreshed daily or sub-daily depending on feed availability.

The framework has clear limitations. Network topology is not explicitly modeled; hazard scope is limited to hurricanes and wildfires; hyperparameter values and leakage controls are not fully detailed; calibration and uncertainty quantification are not reported; and robustness checks beyond the held-out test sets are not described (Li et al., 1 Aug 2025). The paper recommends grouped cross-validation, event-level leave-one-out validation, uncertainty methods such as bootstrapped ensembles and conformal prediction, and extension to additional hazards such as floods and earthquakes. It also notes ethical concerns: mobility data coverage may be uneven across demographics, ACS features reflect structural inequalities, and geospatial aggregation can misalign irregular tracts. Fair deployment is therefore framed as using predictions to expand support rather than reduce resources for low-evacuation areas.

The name itself is context-sensitive. In the disaster-forecasting literature, Evac-Cast denotes the interpretable evacuation-prediction framework just described (Li et al., 1 Aug 2025). In unrelated literatures, the same string is used or mapped differently: it corresponds to vacuum operation of the CERN Axion Solar Telescope in low-mass axion searches (Collaboration et al., 2017, Iguaz, 2011); to solvent-less acrylic-coating concepts for the DEAP detector, especially hybrid in-situ polymerization and, to some extent, photoinitiated CVD (Kuźniak, 2012); to a cooling-enhanced rapid-curing fiberglass cast with passive daytime radiative cooling (Zhang et al., 27 Jan 2026); and to an evacuated, ultra-high-vacuum-compatible RF rod-casting process for intermetallic feedstock preparation (Bauer et al., 2016). This suggests that disciplinary context is essential when interpreting the term.

Within emergency management and computational hazard response, however, Evac-Cast most specifically refers to a tract-scale, interpretable, survey-free forecasting framework in which macro-level proxies for behavior, combined with XGBoost and SHAP, are used to predict and explain evacuation across hurricanes and wildfires (Li et al., 1 Aug 2025).