Source-linked AI summary
The spatial anatomy of urban wildfire vulnerability: a spatially validated GeoAI framework reveals the roles of building density and vegetation moisture in structure loss during the 2025 Palisades Fire
Parastoo Farajpoor, Mohammadreza Narimani
TL;DR
Urban wildfire models need to establish whether predictive relationships transfer across neighborhoods rather than relying only on random validation. This study uses spatially validated pre-fire data and separate post-fire analyses for the Palisades Fire, finding that apparent ROC-AUC 0.92 declined to approximately 0.75 under 1 km spatial blocking.
Problem
Urban wildfire models can appear overly accurate when random validation ignores spatial dependence, leaving transferability of structure-loss relationships across neighborhoods insufficiently tested.
Method
The study links pre-fire environmental, topographic, and built-environment data to inspected structures, compares logistic regression and gradient boosting under spatial validation, and separates post-fire analyses.
Results
ROC-AUC fell from 0.92 under random cross-validation to approximately 0.75 under 1 km spatial blocking, while logistic regression performed equally well and was better calibrated.
Takeaways & Limitations
The supported application is neighborhood-scale screening and moisture-aware vegetation management, not parcel-level prediction, with uncertainty communicated explicitly.
Takeaways & Limitations
The single-event analysis is conditional on CAL FIRE-inspected structures and does not estimate how changing wind, ignition location, or suppression would alter losses across events.
Abstract
from arXiv · showhide
Urban wildfire resilience depends on interactions among built form, vegetation condition, and extreme fire weather, yet city-scale risk models often overlook whether predictive skill transfers across neighborhoods. We developed a spatially validated GeoAI workflow for the January 2025 Palisades Fire, linking 12,081 CAL FIRE damage inspections to pre-fire Sentinel-2 vegetation indices, Landsat surface temperature, LANDFIRE fuels, terrain, and OpenStreetMap buildings and roads. Among 9,883 inspected residential structures, 5,566 were destroyed. Random cross-validation yielded ROC-AUC 0.92 for the integrated XGBoost model, but 1 km spatial block validation reduced performance to 0.75; logistic regression performed similarly and was better calibrated. Building count within 100 m was the strongest predictor, with destruction odds increasing 4.12-fold per standard deviation. Vegetation moisture and greenness showed opposing conditional associations: NDMI at 100-300 m was protective (OR 0.52), whereas NDVI at 30-100 m was positively associated with destruction after accounting for moisture (OR 1.74). Predictive information was concentrated at the 100-300 m neighborhood scale. A separate post-fire track mapped burn severity and vegetation recovery without leakage. The results support neighborhood-scale susceptibility screening, moisture-aware vegetation management, and spatial block validation as a minimum standard for single-event urban wildfire modeling.
1. Introduction
Urban wildfire vulnerability emerges from interactions among built form, vegetation condition, and fire exposure across neighboring parcels, while expanding wildland–urban interfaces increase hazard. This study examines which pre-fire conditions were associated with structure destruction in the Palisades Fire and whether those relationships transfer across space.
- Urban wildfire as a coupled system: Urban-penetrating fires become coupled urban-system failures because ignited buildings generate fuel and embers that can propagate through the built fabric.Consequences depend on the arrangement and condition of buildings, vegetation, roads, and services.
- Motivation: Exposure is increasing as development extends into fire-prone landscapes and the wildland–urban interface spreads across biomes and continents.Within communities, structure loss is rarely explained by a single factor; housing arrangement, local structure density, and location are repeatedly identified as important.
- Built environment and spatial scale: Community-scale morphology matters alongside the home-ignition zone because radiant heat, ember exchange, and repeated ignition opportunities operate across neighboring parcels.Structure density and location can rival or exceed coarse vegetation-cover measures.
- Vegetation condition: Vegetation attributes are not interchangeable: greenness, canopy cover, moisture, fuel type, and spatial arrangement can either support ecological functions or increase combustible exposure.Moisture-sensitive optical indices have been linked to live fuel moisture in southern California shrublands.
- Validation motivation: Random train-test splits can overstate predictive accuracy when spatially autocorrelated observations share landscape, neighborhood, and fire-exposure histories.For planning, discrimination alone is insufficient because a well-ranked map may still assign unreliable probabilities.
- Study objective: The study asks which pre-fire environmental, topographic, and built-environment conditions were associated with structure destruction and how well those relationships transfer across space.Its design compares interpretable logistic regression and gradient boosting under random and spatially blocked validation while distinguishing vegetation amount from moisture.
2. Materials and methods
The study used an inspection-level, spatially supported GeoAI workflow with separate pre-fire susceptibility and post-fire impact tracks. Data were harmonized across temporally distinct remote-sensing composites, terrain, fuels, built-environment variables, and descriptive community context while preventing outcome leakage.
- Analysis unit and spatial support: The individual CAL FIRE-inspected structure was the primary analysis unit, with environmental and neighborhood variables summarized across 0-30 m, 30-100 m, and 100-300 m supports.These supports represented immediate structure-adjacent, intermediate neighborhood, and broader community contexts.
- Analytical design: Two complementary tracks separated pre-fire residential-destruction modeling from post-fire event-impact and community-context analysis, preventing post-fire indicators from entering susceptibility predictors.The separation was declared before modeling to avoid leakage from burn severity, post-fire spectral change, observed damage, and recovery.
- Damage-inspection data: 12,081 classified structures remained after excluding 56 inaccessible records from the 12,137-record CAL FIRE incident service.Residential structures were then identified from the classified inspection population, while post-fire roof, eave, vent-screen, siding, and window fields were excluded because their observability depended on the outcome.
- Temporal design: Three temporal windows kept immediate pre-fire predictors, season-matched baselines, and post-fire observations distinct.The immediate pre-fire period was 1 October 2024-6 January 2025, the seasonal baseline covered October-December 2022-2024, and post-fire observations were restricted to the impact track.
- Predictor construction: Predictors combined Sentinel-2 NDVI and NDMI, Landsat surface temperature, LANDFIRE fuels, 10 m terrain derivatives, and date-scoped OpenStreetMap buildings and roads.Sentinel-2 processing used s2cloudless probability below 40%, with 16 immediate pre-fire scenes and 42 seasonal-baseline scenes; raster predictors were summarized as means over concentric rings.
- Community context and limitations: Community context was descriptive because only twelve tracts contained inspected structures, and network distance measured baseline fire-station proximity rather than evacuation performance or response time.The analysis joined tract-level residential destruction fractions to the CDC/ATSDR Social Vulnerability Index and median pre-fire network distance.
3. Results
The 2025 Palisades Fire destroyed 5,566 of 9,883 inspected residential structures, with losses clustered in dense urban areas. Spatial validation substantially reduced apparent model performance and identified neighborhood building arrangement, vegetation condition, and spatial scale as central results.
- Observed destruction: 5,566 of 9,883 inspected residential structures were destroyed, and losses were strongly clustered with Moran’s I = 0.59.Losses concentrated in the contiguous urban fabric of Pacific Palisades and along the Highway 1 corridor.
- Model validation: 0.168 ROC-AUC separated random from 1 km spatial validation for integrated XGBoost, which scored 0.921 +/- 0.007 versus 0.753 +/- 0.060.Integrated logistic regression achieved spatial ROC-AUC 0.756 +/- 0.092, essentially matching XGBoost.
- Model validation: 0.757 ROC-AUC was achieved by spatially validated XGBoost for M2, exceeding M0 at 0.697, M1 at 0.722, and M3 at 0.753.Logistic probabilities were reasonably calibrated with a mean recalibration slope of 1.02, while XGBoost was overconfident at 0.69.
- Post-fire interpretation: 71% of the 68 km2 burned perimeter was classified as high severity by study-derived dNBR, while 10 m spectral change did not resolve structure-scale exposure.Mean dNBR in the 30-100 m ring increased from 159 for structures with no recorded damage to 218 for Affected and 236 for Destroyed structures, with broad class overlap.
- Social vulnerability: 0.29 Spearman rho with p = 0.35 indicated no detectable association between tract destruction fractions and SVI in the narrow, low-vulnerability study population.The twelve census tracts occupied SVI overall percentiles of 0.00-0.26, within the least-vulnerable national quartile.
4. Discussion
The discussion identifies neighborhood building concentration and vegetation moisture as the most transferable spatial signals of structure destruction, while showing that spatial validation substantially reduces apparent model skill. It frames the workflow as a neighborhood-screening baseline with explicit scale, calibration, uncertainty, and single-event limitations.
- Urban morphology: Neighborhood building concentration carried more transferable destruction information than any single environmental variable, with a nonlinear response above roughly 50-60 buildings within 100 m.The observational inflection should guide closer neighborhood assessment, not serve as a zoning cutoff or causal density rule.
- Spatial scale: Building counts across 100-300 m distinguished urban fabric with different loss rates, whereas nearest-neighbor distance was short almost everywhere.This neighborhood support captures cumulative exposure from many possible ignition pathways.
- Vegetation condition: Vegetation was not uniformly protective or hazardous: at fixed vegetation amount, greater moisture was strongly protective, while at fixed moisture, greater vegetation amount was associated with higher destruction odds.The discussion attributes this contrast to combustible biomass increasing exposure while moisture limits ignitability and fire spread.
- Validation and calibration: The 0.17 AUC optimism gap separated an apparently excellent model from a moderately informative neighborhood-screening model, and gradient boosting added little transferable value beyond logistic regression under spatial blocking.XGBoost was overconfident outside local blocks, while the logistic model remained closer to observed frequencies.
- Validation and calibration: Credible evaluation should resemble deployment: random withholding tests interpolation within familiar neighborhoods, whereas spatial blocking better approximates prioritization in another part of the city.Residual autocorrelation means even block validation cannot reproduce omitted fire dynamics, so out-of-fold surfaces require cautious interpretation.
- Implications and limitations: The workflow supports coordinated adjacent-building hardening, moisture-aware vegetation management, and indicators paired with their valid scale and uncertainty.Future work should replicate the pipeline across DINS events, add hardening data, incorporate fire progression and wind exposure, and test stability across urban forms and social contexts.
5. Conclusion
The analysis identifies neighborhood structure and vegetation moisture as the most informative pre-fire correlates of residential destruction. Its predictive relationships depended on spatially respectful evaluation, which reduced apparent model performance and supported neighborhood-scale screening.
- Conclusion: 12,081 damage inspections showed neighborhood structure and vegetation moisture were the most informative pre-fire correlates of residential destruction.Building count within 100 m produced the largest association, while moisture-sensitive vegetation information was protective and greenness alone was not.
- Conclusion: Building count within 100 m produced the largest association and a marked nonlinear model response.
- Conclusion: ROC-AUC 0.92 under random cross-validation fell to approximately 0.75 under 1 km spatial blocking.Under spatial blocking, a simpler logistic model performed equally well and was better calibrated.
- Conclusion: The practical outcome is a neighborhood-scale screening framework grounded in spatially validated relationships.
Data availability statement
Public source products, metadata, analysis-ready data, model outputs, and derived products are openly available through a Zenodo repository record.
- Data availability statement: Zenodo provides the analysis-ready structure-level feature table, study boundary, manuscript tables, model diagnostics, out-of-fold predictions, SHAP sample, final figures, manifests, and related derived products.The repository record is available at https://doi.org/10.5281/zenodo.22061862.
- Data availability statement: Table 1 lists all public source products and collection identifiers.Persistent source links, retrieval metadata, and checksums are documented in the repository manifest.
- Data availability statement: The same Zenodo record provides additional clipped source data products.The supplied passage identifies a clipped Sen… product, but its full name is truncated.
Ethics statement
The study used publicly available data and involved no human participants, identifiable personal information, animals, or intervention; ethics approval and informed consent were therefore not required.
- Publicly available geospatial, environmental, infrastructure, census, and structure-inspection data were used.
- No human participants, identifiable personal information, animals, or intervention were involved, so institutional ethics approval and informed consent were not required.
Generative AI statement
The authors used ChatGPT for language refinement and visualization enhancement, then reviewed and edited all AI-generated revisions for relevance and accuracy.
- Generative AI statement: ChatGPT improved grammatical accuracy, sentence structure, and visualizations, while the authors reviewed and edited every AI-generated revision.The authors checked the revisions for relevance and accuracy.
Abbreviations
The paper uses abbreviations for evaluation metrics, inspection data, remote-sensing indices, geospatial platforms and data sources, explainability, vulnerability, and wildland–urban interface concepts.
- AUC denotes area under the curve, while DINS denotes Damage Inspection.
- NDMI and NDVI denote the Normalized Difference Moisture Index and Normalized Difference Vegetation Index, respectively.
- GEE and OSM denote Google Earth Engine and OpenStreetMap, respectively.
- dNBR denotes differenced Normalized Burn Ratio, SHAP denotes Shapley additive explanations, and SVI denotes Social Vulnerability Index.
- WUI denotes wildland-urban interface.