Explainable GeoAI for urban wildfire structure loss: vegetation and built environment in the 2025 palisades fire

1 Introduction

Wind-driven fires that penetrate dense residential districts are not simply larger wildfires. Once one or more buildings ignite, structures become fuel, local ember production intensifies, and fire can propagate through the built fabric after the initial wildland flame front has passed (Cohen, 2000; Calkin et al., 2023; Metz et al., 2024). These events are therefore urban-system failures as well as ecological disturbances. Their consequences depend on the arrangement and condition of buildings, vegetation, roads, and services, and on whether an urban system can resist loss, sustain essential functions, and recover (Meerow et al., 2016; McWethy et al., 2019).

Exposure to this hazard is increasing as the wildland–urban interface (WUI) expands. In the United States, development has continued to extend into fire-prone landscapes, while global mapping shows that the interface between settlements and wildland vegetation is now widespread across biomes and continents (Radeloff et al., 2018; Carlson et al., 2022; Schug et al., 2023). Within affected communities, structure loss is rarely explained by a single factor. California studies repeatedly identify housing arrangement, local structure density, and location as major correlates of loss, often with effects that rival or exceed coarse measures of vegetation cover (Syphard et al., 2012; Kramer et al., 2019; Syphard and Keeley, 2019). The home-ignition zone remains important, but community-scale morphology also matters because radiant heat, ember exchange, and repeated ignition opportunities operate across neighboring parcels (Cohen, 2000; Calkin et al., 2014; Syphard et al., 2014; Knapp et al., 2021; Syphard et al., 2021). Evidence beyond California spans complementary approaches: building-risk indices in Lebanon and Greece integrate building and surrounding-site characteristics (Mhawej et al., 2017; Papathoma-Köhle et al., 2022), post-fire surveys in Portugal examine ignition pathways and structural damage (Ribeiro et al., 2020), and Mediterranean vulnerability mapping integrates exposure, sensitivity, and coping capacity (Oliveira et al., 2018). Optical-microwave mapping in Rhodes and severity/exposure assessment in southwestern Türkiye further demonstrate why settlement context and land cover matter when interpreting remotely sensed fire impacts (Makineci, 2024; Zengin and Acar, 2025).

Vegetation is an equally necessary but easily oversimplified component of this system. Green space can moderate heat, stabilize soils, and support ecological functions, yet dry, continuous biomass can increase exposure to combustible materials. Greenness, canopy cover, vegetation moisture, fuel type, and spatial arrangement are therefore not interchangeable. Moisture-sensitive optical indices have been linked to live fuel moisture in southern California shrublands (Dennison et al., 2005), and recent California analyses show that the relationship between urban vegetation and structure loss varies with vegetation type, maintenance, spatial scale, and built context (Escobedo et al., 2025; Kenny et al., 2026). This distinction matters for planning: a greenness-only map can conflate actively transpiring vegetation with desiccated fuel.

Urban Informatics provides a useful frame for integrating these relationships without reducing them to a stack of disconnected maps. Following recent syntheses, we treat Urban Informatics as the coupling of five interconnected dimensions—urban sensing, urban data and data infrastructure, urban computing, urban science, and urban systems and applications—built on the disciplinary foundations of urban science, informatics/computing, and Geographic Information Science (GIScience)/geomatics (Batty, 2013; Shi et al., 2022). In this study, urban science frames the relationships among built form, vegetation, and structure loss; informatics/computing implements the predictive and diagnostic models; and GIScience/geomatics aligns observations across space, time, and spatial support. Open Earth observation products, administrative inspections, and volunteered geographic information can enable rapid post-event analysis, but decision readiness depends on transparent provenance, scale-aware modeling, reproducible evaluation, and a clear path from diagnostics to intervention (Pradeep, 2026). A related optical geospatial artificial intelligence (GeoAI) study provides updatable neighborhood-scale urban canopy screening with explicit spatial diagnostics (Narimani et al., 2026). Data abundance alone does not guarantee reliable planning evidence. Here, GeoAI denotes geospatial artificial intelligence, including machine learning with spatial predictors and spatially explicit evaluation; the term is not restricted to deep learning.

The January 2025 Los Angeles fires have already generated important empirical study. Kenny et al. (2026) found structure density to be the strongest predictor of home loss across the Eaton and Palisades fires, with smaller and less consistent urban-canopy effects. Norlen et al. (2026) likewise showed that urban morphology was more informative than vegetation cover and that relationships changed across community, neighborhood, and parcel scales. Building on those findings, we emphasize the joint reporting of temporally defined predictors, within-event spatial validation, and probability-calibration diagnostics. The Palisades event therefore still offers an opportunity to ask not only which conditions were associated with loss but also how their performance changes when predictions are evaluated in spatially held-out parts of the same fire.

This distinction is consequential because spatially autocorrelated predictors and outcomes can make random train-test splits appear far more accurate than they are. Nearby observations share landscape, neighborhood, and fire-exposure histories; placing them on both sides of a random split allows a model to learn local signatures rather than transferable relationships (Roberts et al., 2017; Valavi et al., 2019; Ploton et al., 2020). For planning applications, discrimination alone is also insufficient. A map that ranks neighborhoods well but assigns overconfident probabilities can misdirect resources and invite parcel-level interpretations unsupported by the data.

This study asks which pre-fire environmental, topographic, and built-environment conditions were associated with structure destruction among California Department of Forestry and Fire Protection (CAL FIRE)-inspected structures in the Palisades Fire, and how well the models discriminate destruction in spatially held-out parts of the event. We assemble a fully open predictor stack with strict temporal separation; compare interpretable logistic regression and gradient boosting under random and spatially blocked validation; distinguish vegetation amount from vegetation moisture across three spatial scales; and maintain a separate impact-and-context track for spectral impact, relative greenness, social vulnerability, and service accessibility. We do not collapse these dimensions into an arbitrary resilience index. Instead, we treat structure resistance, post-fire spectral response, and community context as related but non-interchangeable parts of urban wildfire resilience. The objective is a reproducible neighborhood-screening framework that makes both observed associations and their interpretive scale visible.

2 Materials and methods

2.1 Study design and urban-informatics architecture

The workflow was organized as a data-to-decision system rather than a conventional layer-overlay exercise (Figure 1). Stages 1–5 map onto the five Urban Informatics dimensions: urban sensing and observations; urban data and infrastructure; urban computing/GeoAI; urban science and interpretation; and urban systems and applications. Stage 6 summarizes the proposed decision-support pathway from screening to professional assessment and planning action. Within computing, a primary prediction track estimates associations with observed structure destruction using pre-fire information only, while a separate impact track characterizes spectral impact, relative greenness, and community context. This separation was declared before modeling so that spectral impact, post-fire spectral change, observed damage, and relative-greenness summaries could not leak into the susceptibility predictors. Model-based neighborhood screening is distinct from planning decisions and is intended to precede professional assessment, field verification, and intervention.

The individual CAL FIRE-inspected structure is the primary analysis unit. All metric operations were performed in North American Datum 1983 (NAD83)/Universal Transverse Mercator (UTM) zone 11 N (EPSG: 26911). Environmental variables used three concentric supports: a 0–30 m disk and 30–100 and 100–300 m annuli. Building-count supports are defined separately in Section 2.4. These supports represent an immediate structure-adjacent zone, an intermediate neighborhood zone, and a broader community context. Because 10–30 m satellite pixels cannot resolve individual shrubs, fences, or yard-maintenance details, the innermost support is interpreted as a local remote-sensing context rather than a parcel-level defensible-space inspection.

2.2 Study area and event context

The Palisades Fire ignited on 7 January 2025 in the Santa Monica Mountains above Pacific Palisades, Los Angeles, California. It burned approximately 23,448 acres (about 95 km²) before containment at the end of January (CAL FIRE, 2025b). The study area comprises the final interagency fire perimeter plus a 2 km buffer, covering approximately 218 km² (National Interagency Fire Center, 2025) (Figure 2). The buffer retains inspected structures near the mapped perimeter and provides environmental and accessibility context without treating uninspected buildings as undamaged.

The event followed exceptional antecedent dryness and coincided with strong Santa Ana conditions. Study-area Gridded Surface Meteorological (gridMET) precipitation for the period between October 2024 and December 2024 totaled 5.5 mm, compared with a median of 111 mm for the period 1980–2024. One-hundred-hour dead fuel moisture reached its lowest values of the preceding year while vapor-pressure deficit remained elevated. ERA5-Land area means for 7–9 January reached 7.8 m s − 1 for 10 m wind speed and 9.6% for minimum relative humidity (Muñoz-Sabater et al., 2021). These coarse meteorological data describe the common forcing under which structures were exposed; they were not used as structure-level predictors because their 3–9 km grids provide little credible spatial contrast across the fire footprint. The event-context series is shown in Supplementary Figure S1. The event is interpreted as a wind-dominated fire transitioning from chaparral into dense urban fabric (Guzman-Morales and Gershunov, 2019; Keeley and Syphard, 2019).

2.3 Damage observations and analysis population

The primary outcome derives from the CAL FIRE Damage Inspection program, which classifies inspected structures as No Damage, Affected (>0–10%), Minor (10–25%), Major (25–50%), Destroyed (>50%), or Inaccessible (CAL FIRE, 2025a). The authoritative Palisades incident service contained 12,137 records when retrieved on 12 August 2026. Fifty-six inaccessible records were excluded, leaving 12,081 classified structures. Residential structures—Single Residence, Multiple Residence, and Mixed Commercial/Residential—formed the primary modeling population (n = 9,883), of which 5,566 (56.3%) were destroyed. The binary outcome was therefore Destroyed versus Not Destroyed (No Damage through Major). A predeclared sensitivity analysis grouped Major with Destroyed. Intermediate damage classes were retained for descriptive figures but were too sparse for stable ordinal modeling (Table 1B).

Table 1. Public data architecture and audited analysis population. Section A lists the data sources; Section B reports DINS damage classes by structure category.

Inspections were treated as the sampling frame. Structures outside the inspected set were considered unknown rather than undamaged. Drip Edge, Insulation, Nailing, and Sheathing (DINS) also records roof, eave, vent-screen, siding, and window characteristics observed after the fire. Those fields were excluded because observability depended on the outcome: for example, eave type was unknown for 69.2% of destroyed structures and 0.3% of undamaged structures. Including such variables would allow the model to infer destruction from the absence of inspectable material, a form of outcome leakage also noted in post-fire engineering assessments (Insurance Institute for Business and Home Safety, 2025). Assessor-derived year built and improved value were retained because they existed independently of the post-fire inspection and were 100 and 96.5% complete, respectively. Relative to the final fire perimeter, 8,555 of 9,883 modeled residential points (86.6%) fell inside the perimeter polygon, and 5,360 (54.2%) lay within 500 m of the perimeter edge (Supplementary Table S1). The perimeter-edge subset is a sensitivity population, not evidence of complete inspection coverage. Odds ratios (ORs) describe conditional associations within the DINS-inspected set. The one-way footprint match described below assesses geolocation support; it does not estimate the proportion of the residential building inventory inspected.

2.4 Open data architecture and feature engineering

Table 1A summarizes the public data architecture, Table 1B audits the DINS analysis population, and Table 2 lists the complete predictor dictionary. Three temporal windows were kept distinct: an immediate pre-fire period (from 1 October 2024 to 6 January 2025), a season-matched baseline (October–December phase of the period 2022–2024), and post-fire observations used only in the impact track. The use of temporally consistent composites follows standard remote-sensing practice: rich spectral inputs do not ensure transferability when acquisition windows, ground reference, and spatial validation are poorly aligned (Roberts et al., 2017; Valavi et al., 2019). True- and shortwave-infrared false-color scenes provide pre-fire, active-fire, and post-fire visual context (Supplementary Figure S2). In contrast, pre-fire predictor maps and post-fire impact maps are presented separately in Figures 3, 4, respectively.

Table 2. Complete predictor dictionary, source, and spatial support.

Sentinel-2 Level-2A surface reflectance was processed in Google Earth Engine (Gorelick et al., 2017) using s2cloudless probability below 40% and scene-classification masking. The immediate pre-fire composite used 16 scenes, and the seasonal baseline used 42 scenes. The Normalized Difference Vegetation Index (NDVI; Tucker, 1979) and Normalized Difference Moisture Index (NDMI; Gao, 1996) were calculated using Equations 1, 2:

NDVI = ρ NIR − ρ Red ρ NIR + ρ Red (1)

NDMI = ρ NIR − ρ SWIR ρ NIR + ρ SWIR (2)

where ρ NIR, ρ Red, and ρ SWIR denote surface reflectance in the near-infrared, red, and shortwave-infrared bands, respectively. NDVI characterizes vegetation greenness, whereas NDMI was used as a moisture-sensitive indicator of canopy condition rather than as a direct measurement of fuel moisture (Dennison et al., 2005). Index anomalies were calculated as the immediate pre-fire value minus the season-matched 2022–2024 baseline.

Landsat 8/9 Collection 2 Level-2 surface temperature, quality masked and composited across 21 scenes, provided pre-fire land-surface temperature. Landscape Fire and Resource Management Planning Tool (LANDFIRE) 2024 supplied Scott and Burgan fire-behavior fuel models and forest canopy cover at 30 m (Scott and Burgan, 2005; Rollins, 2009; LANDFIRE, 2025). The 2024 release was selected because later products may encode the fire scar. Elevation, slope, northness, eastness, 300 m topographic position, and terrain ruggedness were derived from the United States Geological Survey (USGS) Three-Dimensional Elevation Program (3DEP) 10 m digital elevation model.

Built-environment variables were derived from a date-scoped OpenStreetMap (OSM) snapshot dated 1 January 2025, 6 days before ignition (OpenStreetMap contributors, 2025). The extract contained 28,302 building polygons and 4,904 drivable road edges; 99.1% of residential DINS points matched a building footprint within 20 m. Derived variables included building footprint area, edge-to-edge distance to the nearest neighboring building, building counts within 30, 100, and 300-m-radius disks, distance to the nearest drivable road and dead-end node, and road-length density within 300 m. Fire station and hospital locations supported the contextual accessibility analysis. Census-tract geometry came from Topologically Integrated Geographic Encoding and Referencing (TIGER)/Line and social context from the 2022 Centers for Disease Control and Prevention and the Agency for Toxic Substances and Disease Registry (CDC/ATSDR) Social Vulnerability Index (Flanagan et al., 2011; Centers for Disease Control and Prevention/Agency for Toxic Substances and Disease Registry, 2024). For building counts, eligible OSM building centroids falling within circular buffers of radius 30, 100, and 300 m around each structure were counted using a centroid-in-disk spatial predicate. Axis-aligned bounding boxes were used only as an internal candidate search and do not define the support. The target’s matched footprint is included when its centroid falls inside the disk. All mapped building polygons exceeding 10 m², including garages and sheds when separately represented, were eligible. These counts describe the historical OSM representation of neighborhood building concentration rather than a complete residential census. Environmental predictors use the concentric supports defined above, and road-length density uses a 300 m-radius disk.

Raster predictors were summarized as valid-pixel means over the 0–30 m disk and 30–100 and 100–300 m annuli. Fuel-model classes were collapsed into interpretable group fractions, and Euclidean distance to the nearest wildland-fuel pixel was calculated using LANDFIRE data. Collinearity was assessed using Spearman correlations and variance inflation factors (VIFs). Perfectly dependent fuel subfractions were removed, and the interpretable model used a pruned predictor set in which pairs with an absolute correlation above 0.85 were resolved using process relevance and data support. The final integrated set contained 37 predictors organized into predictor blocks: M0, terrain and distance to wildland fuel; M1, M0 plus vegetation, moisture, thermal, and fuel variables; M2, M0 plus built-environment variables; and M3, all predictor families. The pruned logistic specification therefore excludes several collinear terms retained by eXtreme Gradient Boosting (XGBoost), including building count within the 300 m-radius disk. XGBoost uses all 37 predictors, whereas the logistic model retains 31. Table 2 identifies model membership and Table 3 reports explicit support. Supplementary Tables S4–S6 provide the vegetation correlation matrix, candidate-matrix VIF diagnostics, and single-index sensitivity fits. M1 and M2 are alternative extensions of M0, whereas M3 combines the environmental and built-environment blocks. The complete model-specific inventory in Table 3 identifies predictors excluded from the logistic refit as pruned.

Table 3. Model-specific logistic-refit associations and XGBoost importance for the complete 37-variable M3 dictionary.

Spatial support is reported explicitly. Coefficient estimates apply only to the 31-variable pruned logistic refit; all 37 variables belong to the XGBoost dictionary. Mean |SHAP| and ranks correspond to the final fitted M3 XGBoost model.

Odds ratios are conditional on the DINS-inspected residential sample. The displayed confidence intervals and p-values are classical model-based summaries and do not adjust for residual spatial dependence. For building count within the 100 m-radius disk, the spatial block-bootstrap 95% interval (1 km blocks; 1,000 successful replicates) is (1.784, 6.006), compared with the model-based interval (3.271, 4.265). “Not estimated (pruned)” denotes exclusion from the pruned logistic refit while XGBoost mean |SHAP| and rank remain defined. Building-count supports are circular centroid-in-disk buffers.

2.5 Models, spatial validation, and interpretation

Two model families were used to separate interpretability from non-linear flexibility. Logistic regression modeled the probability that structure i was destroyed as shown in Equation 3:

logit [ P ( Y i = 1 ) ] = β 0 + ∑ j = 1 p β j x ij (3)

where P ( Y i = 1 ) is the probability that structure i was destroyed, β 0 is the intercept, β j is the coefficient for standardized predictor x ij, and p is the number of predictors. The logistic model used L2 regularization and balanced class weights. Standardized odds ratios and model-based 95% confidence intervals (CIs) were obtained from an unpenalized refit of the pruned specification and are interpreted as conditional associations rather than causal effects. These model-based intervals and p-values do not account for residual spatial dependence in the out-of-fold residuals. They are descriptive summaries of model-based uncertainty rather than spatially robust tests.

Table 3 and Figure 5C report conventional model-based 95% confidence intervals for the unpenalized logistic refit. For the building count within a 100 m-radius disk, a spatial block-bootstrap 95% interval (1 km blocks; 1,000 successful replicates) is additionally reported in Section 3.4 and the Table 3 note as an assessment of un under block resampling.

Gradient-boosted trees were fitted with XGBoost and tuned over tree depth, learning rate, and number of estimators (Chen and Guestrin, 2016). Predictor contributions were summarized with SHAP values, which are interpreted as model attributions rather than causal effects (Lundberg and Lee, 2017).

The primary evaluation used 5-fold spatial block cross-validation. Structures in each 1 km grid cell were assigned together to a fold; block allocation was seeded, and cells assigned to different folds could be adjacent. Median imputation, scaling, and inner three-fold hyperparameter tuning were fitted within each outer training set. Inner GroupKFold kept observations from the same spatial cell together, but did not impose separation between adjacent cells. Consequently, this design prevents splitting a cell between training and testing but does not guarantee a minimum geographic gap or non-overlapping predictor supports. The 500 m and 2 km analyses assess sensitivity to block size; no exclusion buffer was applied. Random five-fold cross-validation was retained as an interpolation benchmark (Roberts et al., 2017; Valavi et al., 2019; Ploton et al., 2020).

Performance was evaluated using ROC-AUC, average precision (reported as Area Under the Precision-Recall Curve [PR-AUC]), balanced accuracy, F1 score, and Brier score (Saito and Rehmsmeier, 2015). Balanced accuracy and F1 used a probability threshold of 0.5. Reliability curves and a diagnostic logistic regression of held-out outcomes on the logit of held-out predictions quantified calibration; ideal intercept and slope are 0 and 1. These diagnostic regressions did not recalibrate or modify the out-of-fold probabilities. This distinction matters because balanced class weights alter the fitting target. Fold-level calibration estimates are reported in Supplementary Table S7. Means and standard deviations summarize outer-fold performance, and residual Moran’s I assesses remaining spatial structure. Fixed-settings XGBoost sensitivity analyses varied scale, block size, outcome, and population; these are distinguished from the nested-tuning models in Table 4. Paired fold-level AUC differences are provided descriptively in Supplementary Table S8.

Table 4. Out-of-fold performance and sensitivity analyses.

(A) Nested-tuning models; parentheses show sample SD across 5 folds (Delta Degrees of Freedom [ddof] = 1). (B) Fixed-settings XGBoost sensitivity models; SD is calculated across the five folds using ddof = 0. PR-AUC denotes average precision. The all-scales row in B is not the M3 logistic model in A.

2.6 Impact, relative greenness, and community context

The impact track was analytically separated from the pre-fire models. Spectral impact was characterized from pre- and post-fire Sentinel-2 Normalized Burn Ratio (NBR) composites. NBR, differenced NBR (dNBR), and relative dNBR (RdNBR) were calculated using Equations 4–6 (Key and Benson, 2006; Miller and Thode, 2007):

NBR = ρ NIR − ρ SWIR 2 ρ NIR + ρ SWIR 2 (4)

dNBR = 1000 ( NB R pre − NB R post ) (5)

RdNBR = dNBR max ( ∣ NBR pre ∣ , 0.001 ) (6)

where ρ NIR and ρ SWIR 2 are surface reflectance in the near-infrared and longer-wave shortwave-infrared bands; NB R pre and NB R post are the pre- and post-fire NBR composites; dNBR is their difference multiplied by 1,000; and RdNBR is obtained by dividing that scaled dNBR by the square root of max(|NBR_pre|, 0.001), where NBR_pre in the denominator is the fractional (unscaled) pre-fire NBR and 0.001 is a numerical floor (Miller and Thode, 2007), to reduce dependence on initial vegetation condition.

Scaled dNBR thresholds of 100, 270, and 440 define descriptive spectral-change bins: dNBR ≤ 100, 100 < dNBR ≤ 270, 270 < dNBR ≤ 440, and dNBR > 440. These bins organize the imagery and relative-greenness series; they are not a locally calibrated severity classification for developed land. Structural loss, ash, exposed soil, and debris clearance can contribute to urban dNBR. We therefore report continuous dNBR around inspected structures and interpret the mapped classes as spectral impact (Key and Benson, 2006; Makineci, 2024; Zengin and Acar, 2025).

Post-fire relative greenness was tracked monthly from February 2025 through July 2026. To account for seasonal phenology, the aggregate NDVI ratio for time t and fixed spectral-impact class s was calculated using Equation 7:

R t , s = mean ( NDV I t , s ) B m ( t ) , s (7)

Here, R t , s is the ratio of the mean NDVI in month t and spectral-impact class s to B m ( t ) , s, the spatial mean of NDVI from the month-matched 2022–2024 reference composite for class s. The reference composite pools cloud-masked observations from the corresponding calendar month across the three baseline years. NBR and NDVI are calculated from reflectance composites before spatial averaging. A ratio of 1 denotes the class-level seasonal reference, not restoration of vegetation structure or ecosystem function. Valid-pixel support may vary among months because of masking. The series combines developed and wildland surfaces and is interpreted as an aggregate measure of relative greenness; in developed areas it also reflects demolition, debris clearance, and landscaping. The three-year reference controls seasonal comparison over a short baseline; it does not quantify long-term baseline uncertainty.

Community context joined tract-level residential destruction fractions to the CDC/ATSDR Social Vulnerability Index and to median pre-fire network distance from inspected structures to the nearest fire station. The accessibility measure describes baseline service proximity rather than evacuation performance or emergency response time. The tract analysis is descriptive because only 12 tracts contained inspected structures.

2.7 Reproducibility and quality assurance

The workflow was implemented in Python using geopandas, rasterio, scikit-learn, statsmodels, XGBoost, SHAP, osmnx, libpysal, and esda, with Google Earth Engine used for satellite compositing. Every source was recorded in a manifest that included the provider, version, spatial and temporal support, access parameters, retrieval time, license, and checksum. Deterministic seeds fixed fold assignments and model tuning. Automated assertions tested coordinate reference systems, unique structure identifiers, expected index ranges, temporal precedence, and the exclusion of post-fire or observability-biased variables from the predictor matrix. Analysis-ready data and derived products are archived in Zenodo (Farajpoor and Narimani, 2026), and the public replication code is available at https://github.com/MohammadrezaNarimaniUCDavis/Palisades_Urban_Wildfire_GeoAI.

3 Results

3.1 Event setting and spatial pattern of loss

The fire occurred after the second-driest October–December period in the 45-year study-area record and during a sharp combination of low relative humidity, low dead-fuel moisture, and strong winds (Supplementary Figure S1). Within the inspected residential population, 5,566 of 9,883 structures were destroyed. Loss was strongly clustered (Moran’s I = 0.59) and concentrated in the contiguous urban fabric of Pacific Palisades and along the Highway 1 corridor. Structures in the northern and western canyon-edge subdivisions and many inspected structures near the buffered perimeter experienced markedly lower destruction rates (Figure 2).

3.2 Pre-fire environmental and built-environment contrasts

The pre-fire environmental maps show a shrub-dominated landscape with substantial local variation in vegetation moisture and thermal condition (Figure 3). In contrast, the post-fire maps show pronounced changes in greenness, moisture, spectral impact (dNBR), and surface temperature (Figure 4). The built-environment maps show the highest structure densities in the southeastern urban fabric, where many losses occurred (Figure 6). Across original DINS classes, destroyed structures occupied denser neighborhoods, were surrounded by lower NDMI, were closer to dead-end road nodes, occurred at higher local topographic positions, and tended to be older than structures with no recorded damage (Figure 7). Nearest-neighbor building distances were short throughout the developed area (median 2.6 m), so neighborhood building count separated the classes more clearly than nearest-neighbor distance alone.

3.3 Spatially held-out performance and calibration

The validation design changed the apparent quality of the models more than the choice of learner did. Under random cross-validation, the integrated XGBoost model achieved an ROC-AUC of 0.921 ± 0.007, a precision-recall AUC of 0.938, and a Brier score of 0.112. Under 1 km spatial blocking, the same specification fell to ROC-AUC 0.753 ± 0.060 and Brier score 0.213. The integrated logistic model achieved spatial ROC-AUC 0.756 ± 0.092, similar to XGBoost in mean discrimination despite its simpler form (Figure 8; Table 4A). The approximately 0.17 AUC difference between random and spatial validation therefore represents spatial optimism rather than independent-event validation.

The built-environment block had a higher mean spatial-CV AUC than the vegetation block. XGBoost achieved 0.697 for M0, 0.722 for M1, 0.757 for M2, and 0.753 for M3. The paired M2-minus-M1 differences were positive in all five folds (mean 0.034, standard deviation (SD) 0.024; range 0.013–0.064), but the small number of folds and overlapping training samples support an ordered tendency rather than an independent replication test (Supplementary Table S8). For the integrated logistic model, calibration intercept and slope were 0.210 ± 0.515 and 1.020 ± 0.460 across folds; corresponding XGBoost values were 0.344 ± 0.583 and 0.690 ± 0.155. Logistic calibration was closer to ideal on average, while both models varied substantially between held-out areas (Supplementary Table S7). Residual Moran’s I remained 0.48–0.55. This residual structure is compatible with omitted event exposure and other spatial processes, but does not identify their separate contributions or establish a spatial independence range. Table 4A reports nested tuning. Table 4B reports fixed-settings XGBoost, not the 31-predictor logistic model; its standard deviations use the population convention across the five folds, whereas Table 4A uses the sample convention.

3.4 Dominant predictors and scale dependence

Both model families identified neighborhood building arrangement as the dominant signal (Figure 5; Table 3). In the logistic model, a one-standard-deviation increase in building count within the 100 m-radius disk was associated with a 3.735-fold increase in the conditional odds ratio of destruction among inspected residential structures (model-based 95% CI: 3.271–4.265; spatial block-bootstrap 95% CI: 1.784–6.006). The 100 and 300 m building-count variables also dominated XGBoost SHAP attributions on the full predictor set; the 300 m count is pruned from the logistic specification for collinearity with the 100 m count and road-length density, but retains SHAP rank 1 in the full XGBoost fit (Table 3). The SHAP response for building count within the 100 m-radius disk was non-linear, with a pronounced increase above approximately 50–60 buildings. This value is a model-derived inflection within the Palisades data, not a universal density threshold.

In the joint pruned model, NDMI at 100–300 m was associated with lower destruction odds (OR: 0.53, model-based 95% CI: 0.48–0.59), as was NDMI at 30–100 m (OR: 0.72). NDVI at 30–100 m was positively associated in that joint specification (OR: 1.76, 1.51–2.05), but its sign reversed in the single-index model with otherwise matched controls (OR: 0.833, 0.775–0.895). By comparison, the separately entered NDMI at 100–300 m remained negatively associated (OR: 0.554, 0.515–0.596). Thus, the moisture association is more stable across these specifications than the positive greenness coefficient that reverses under single-index models. The joint coefficient is conditional on joint adjustment and is not evidence that greater greenness independently increases loss. Correlations and candidate-matrix VIFs document substantial predictor dependence (Supplementary Tables S4–S6). The positive NDMI-anomaly coefficient alongside a negative NDMI-level coefficient is consistent with suppression among correlated predictors and should not be interpreted independently.

Logistic coefficients and SHAP rankings summarize different aspects of the fitted models. Road-length density within 300 m had an OR of 1.109 (model-based p = 0.097) yet SHAP rank 11; distance to the nearest road had an OR of 1.119 (p < 0.001) and rank 31; and immediate (0–30 m) canopy cover had an OR of 0.887 (p = 0.002) and rank 36. A conditional linear coefficient and a global attribution magnitude need not agree. Non-linearity, interactions, correlated predictors, and the different predictor sets can all contribute; the ranking differences alone do not isolate a mechanism.

Terrain, access, and construction-era effects were smaller but coherent. Destruction odds ratio increased at higher local topographic positions and on east-facing aspects and decreased with elevation, distance from wildland fuels, and distance from dead-end nodes. Newer construction and greater assessed value were protective conditional on the other predictors. Restricting vegetation and building-count variables to matched nominal scales showed a tendency for spatial-CV discrimination to increase with scale: ROC-AUC was 0.721 at 0–30 m, 0.744 at 30–100 m, and 0.753 at 100–300 m, compared with 0.756 for the full multi-scale model (Figure 9); these overlapping fold-level summaries are reported as ordered tendencies, without an independent test of scale differences. The broadest support-scale group also carried the largest aggregate SHAP contribution. Mean discrimination was similar when Major damage was grouped with Destroyed, when all structure categories were included, and when 2 km blocks were used. Five-hundred-meter blocks yielded a higher AUC (0.809) than 1 and 2 km blocks (0.756 and 0.765). This pattern is consistent with proximity-related optimism, although changes in fold composition also contribute; it does not demonstrate that 1 km removes spatial dependence (Figure 9; Table 4B). As a conditioning sensitivity, refitting the pruned logistic model on the 5,360 residential structures within 500 m of the final perimeter edge retained the directions of the leading conditional associations (building count within the 100 m-radius disk – OR: 3.917; NDVI 30–100 m – OR: 1.593; NDMI 100–300 m – OR: 0.389; Supplementary Table S2).

3.5 Spectral impact and post-fire relative greenness

Spectral impact varied across the analyzed perimeter (Figure 10). The dNBR classes describe spectral change rather than locally calibrated ecological severity over urban land. Mean dNBR in the 30–100 m ring was 159 for structures with no recorded damage, 218 for Affected structures, and 236 for Destroyed structures. The broad class overlap underscores the distinction between neighborhood spectral change and an individual structure’s fire exposure.

The aggregate relative-greenness series was strongly seasonal and ordered by the spectral-impact classes. At the first reported post-fire observation, NDVI was approximately 54, 40, and 29% of the month-matched reference in the low-, moderate-, and high-impact bins. Class means approached or briefly exceeded the reference during winter 2025–2026 and were 64–77% of that reference by July 2026 (Figure 10). Because these summaries combine developed and wildland land cover, they describe changes in regional greenness rather than ecological recovery. They do not separate vegetation response from changes associated with clearing and re-landscaping in developed areas.

3.6 Community context

The 12 census tracts containing inspected structures occupied a narrow, low range of social vulnerability range: Social Vulnerability Index (SVI) overall percentiles were 0.00–0.26, near the least-vulnerable quarter of the national distribution. Tract destruction fractions ranged from 0 to 0.90, with a descriptive Spearman’s rho of 0.29 (p = 0.35; Figure 11). Median network distance to the nearest fire station ranged from 0.7 to 7.4 km, with no consistent destruction gradient. This restricted-range comparison does not imply that social vulnerability is unimportant to wildfire resilience; it shows that this single, affluent study area contains too little social variation to estimate that relationship reliably (Davies et al., 2018; Norlen et al., 2026).

4 Discussion

4.1 Neighborhood morphology and the spatial pattern of loss

The clearest result is that urban morphology, especially the concentration of neighborhood buildings, was the strongest conditional correlate of destruction within the inspected residential set. This is consistent with the building-density findings of Kenny et al. (2026) and the multi-scale urban-morphology results of Norlen et al. (2026), while extending them in three ways: the association is estimated conditional on a pre-fire environmental stack, predictive models are evaluated under spatial blocking, and their probability calibration is assessed. The non-linear response above roughly 50–60 buildings within the 100 m-radius disk is compatible with a change in propagation opportunity as repeated ember and radiant-heat exposure accumulates across dense fabric. Because the analysis is observational, the inflection may motivate further neighborhood assessment but is not a zoning cutoff or causal density rule. Building concentration may also encode where the fire traveled, the timing of exposure, and neighborhood-specific suppression conditions. The current analysis cannot separate these pathways. We therefore interpret building count as a within-event screening correlate rather than an isolated causal vulnerability mechanism.

The result also clarifies the scale at which open urban data are most useful. Nearest-neighbor distance was short almost everywhere, whereas building counts within 100 and 300 m-radius disks distinguished urban fabric with very different loss rates. That support summarizes neighborhood concentration and is compatible with multiple ignition pathways, a pattern consistent with prior California structure-loss studies and the Camp Fire analysis (Syphard et al., 2012; Knapp et al., 2021; Syphard et al., 2021). The protective year-built association is compatible with progressive improvements in construction practice, but year built also encodes neighborhood age, layout, and redevelopment history. It should therefore guide screening for retrofit needs rather than be interpreted as a direct building-code effect.

4.2 Vegetation condition reframes the greenness-risk relationship

The environmental findings argue against treating urban vegetation as uniformly protective or hazardous. Neighborhood moisture was negatively associated with loss in both joint and single-index specifications. The positive joint-model NDVI coefficient, however, became negative when NDVI was entered separately, so an independent harmful effect of greenness is not supported. In the exceptionally dry months preceding the fire, moisture-sensitive information remains physically relevant. Still, correlated spectral indices cannot uniquely distinguish combustible biomass, vegetation condition, and their surrounding built context. The immediate-ring canopy effect was small and partly protective, further weakening any interpretation that simple canopy removal would improve resilience.

For urban greening and defensible-space policy, the actionable distinction is therefore among the amount of vegetation, moisture conditions, arrangement, and continuity. NDVI alone cannot make that distinction. Moisture-sensitive indices, irrigation context, fuel type, and maintenance need to be interpreted together, especially at the 100–300 m neighborhood scale where the observed model summaries were most informative (Dennison et al., 2005; Escobedo et al., 2025). Higher-resolution canopy and parcel data could refine local implementation; recent open optical GeoAI study demonstrates how updateable canopy screening can complement structural inventories while retaining spatial uncertainty (Narimani et al., 2026).

4.3 Spatial validation is a substantive urban-informatics result

The 0.17 AUC optimism gap is not a technical footnote. It is the difference between an apparently excellent model and a moderately informative neighborhood-screening model. Gradient boosting benefited most from random spatial mixing, yet it did not outperform logistic regression once neighborhoods were held out. The more flexible learner therefore offered little improvement in mean spatially held-out discrimination in this single-event setting. Calibration moved in the same direction: XGBoost was overconfident outside local blocks, while the logistic model was closer to ideal calibration on average, with substantial fold-to-fold variation.

This result has direct implications for Urban Informatics: model evaluation must reflect the intended spatial decision context. The substantial gap between random and spatial cross-validation is therefore a substantive finding for neighborhood screening, not only a technical diagnostic. A data-to-decision system is credible only if evaluation resembles the decision it will support. Randomly withholding nearby structures evaluates interpolation within familiar neighborhoods; spatial blocking more closely approximates prioritization in a different part of the city. The higher AUC under 500 m blocks is consistent with proximity-related optimism. Similar AUCs at 1 and 2 km do not establish independence: block width is not a minimum train-test distance, and residual Moran’s I does not identify a correlation range. Accordingly, the out-of-fold surface should be interpreted as a calibration-assessed, within-event screening surface, not a deterministic parcel forecast. Spatially structured validation, calibration, and block-size sensitivity should be minimum reporting elements for single-event urban hazard models (Roberts et al., 2017; Ploton et al., 2020).

4.4 Translating diagnostics into planning and design

The findings motivate a clear pathway from computation to action that keeps model-based screening distinct from planning decisions. Potential users include municipal planning and building departments, fire authorities, and rebuilding or retrofit programs. The intended decision scale is neighborhood-to-block screening rather than automated parcel-level decisions. First, computational outputs can flag areas of high building concentration for coordinated professional assessment alongside independent structure-hardening inspections. The relevant assessment unit is the adjacent-building interface and surrounding block: ember-resistant vents and roofs, ignition-resistant fencing, reduced combustible continuity, and coordinated treatment of the closest structures. Second, vegetation programs can replace greenness-only screening with moisture-aware assessments that preserve the cooling and ecological value of urban canopy while targeting dry, continuous fuels. Third, any intervention should follow field verification; the spatially evaluated surface indicates where further assessment may be useful and does not automate enforcement, insurance, or regulatory decisions.

The key design principle is to couple every indicator with its valid scale and uncertainty along the pathway screening → agency/professional assessment → field verification → planning action. This aligns with decision-oriented Urban Informatics, in which sensing, data infrastructure, computing, urban science, and systems applications are connected through an explicit implementation pathway rather than ending with a map (Shi et al., 2022; Pradeep, 2026). Because the acquisition and analysis are scripted, the same workflow can be rerun shortly after damage-inspection data are released, creating comparable evidence across events and supporting a future multi-city benchmark. Such benchmarking is essential before any Palisades-specific threshold is generalized.

4.5 Scope, transferability, and unresolved processes

The design intentionally holds one extreme event fixed and asks why inspected structures within that event experienced different outcomes. This strengthens the analysis of spatial contrasts but does not estimate how changing wind, ignition location, or suppression would alter loss across events. Inference is conditional on DINS-inspected structures, and structures outside that frame remain unknown. Building-envelope characteristics were not modeled because their post-fire observability was strongly outcome dependent; excluding them protects internal validity while leaving a known component of home ignitability unresolved (Zamanialaei et al., 2025).

Sensor support defines a second boundary. Ten- to 30 m products characterize local and neighborhood contexts but cannot resolve individual yard treatments, fence materials, vents, or fine-fuel continuity. The neighborhood-scale tendency should therefore be understood as a property of the present predictors and validation design, not as evidence that parcel-level conditions are irrelevant. Similarly, dNBR combines vegetation and structural change in urban areas and is retained only as an impact descriptor. The SVI result is limited to 12 low-vulnerability tracts and should motivate cross-event comparisons with more diverse communities rather than a general conclusion about equity.

These boundaries point directly to the next analytical step: replicate the pipeline across DINS events, add independent structure-hardening data, incorporate time-resolved fire progression and wind exposure, and evaluate whether neighborhood effects remain stable across urban forms and social contexts. The present study supplies an open, single-event baseline for that comparison, alongside growing Mediterranean and European evidence on WUI exposure, burn mapping, and building vulnerability (Mhawej et al., 2017; Oliveira et al., 2018; Ribeiro et al., 2020; Papathoma-Köhle et al., 2022; Makineci, 2024; Zengin and Acar, 2025). Its contribution is an auditable single-event benchmark that connects neighborhood-scale associations with their evaluation conditions. Comparable studies across urban forms, construction practices, and fire regimes are needed to test which relationships persist.

5 Conclusion

A reproducible urban-informatics analysis of 12,081 Palisades Fire damage inspections shows that neighborhood structure and vegetation moisture were the most informative pre-fire correlates of residential destruction. Within the DINS-inspected residential frame, building count within the 100 m-radius disk produced the largest conditional association and a marked non-linear model response. At the same time, higher neighborhood NDMI was consistently associated with lower loss across the examined specifications. The NDVI coefficient was specification-dependent. The validation comparison is itself an Urban Informatics contribution: because evaluation must match the intended spatial decision context, the drop from ROC-AUC 0.92 under random cross-validation to approximately 0.75 under 1 km spatial blocking is a substantive result for neighborhood screening, not only a technical caveat. Under spatial blocking, a simpler logistic model had similar mean discrimination and a calibration slope closer to one. The practical outcome is a neighborhood-scale screening framework, not a parcel-level prediction system. It motivates the coordinated assessment of the dense urban fabric and vegetation conditions, supported by field observations and explicit evaluation of uncertainty. The separated spectral-impact, relative-greenness, and community analyses broaden the resilience interpretation without contaminating the pre-fire model. By combining open data, temporal leakage control, spatial validation, transparent model interpretation, and a direct planning translation, the workflow provides a reproducible protocol for comparative urban wildfire resilience studies; coefficients and density thresholds from this single event should not be treated as operationally transferable until tested across additional fires and urban forms.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. During the preparation of this work, the authors used ChatGPT to improve grammatical accuracy, refine sentence structure, and enhance visualizations. All AI-generated revisions were thoroughly reviewed and edited by the authors to ensure relevance and accuracy.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frsus.2026.1975724/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Charithmali Chethika Abenayake, University of Moratuwa, Sri Lanka

Reviewed by: Recep Uğur Acar, Dumlupinar University, Türkiye

Pavithra Jayasinghe, University of Sri Jayewardenepura, Sri Lanka

Enes Zengin, Istanbul Technical University, Türkiye

Footnotes

AUC, area under the curve; DINS, Damage Inspection; dNBR, differenced Normalized Burn Ratio; GEE, Google Earth Engine; NDMI, Normalized Difference Moisture Index; NDVI, Normalized Difference Vegetation Index; OSM, OpenStreetMap; SHAP, Shapley additive explanations; SVI, Social Vulnerability Index; WUI, wildland–urban interface.