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

Abstract

Urban wildfire resilience depends on how built form, vegetation condition, and extreme fire weather interact. Thus far, city-scale decision tools often combine spatial data without testing whether their apparent predictive skill persists when evaluated in spatially separated settings. We developed a spatially validated geospatial artificial intelligence (GeoAI) and urban-informatics workflow for the January 2025 Palisades Fire, linking 12,081 California Department of Forestry and Fire Protection (CAL FIRE) damage inspections to pre-fire Sentinel-2 vegetation amount and moisture, Landsat surface temperature, Landscape Fire and Resource Management Planning Tool (LANDFIRE) fuels, 10 m terrain, and a dated OpenStreetMap (OSM) representation of buildings and roads. Residential destruction (5,566 of 9,883 inspected residential structures) was modeled using logistic and gradient-boosting models with nested tuning under random and 1 km spatial-block cross-validation. Random cross-validation suggested excellent discrimination for the integrated gradient-boosting model (ROC-AUC 0.92), but spatial validation reduced performance to 0.75; an interpretable logistic model achieved a similar mean Receiver Operating Characteristic and Area Under the Curve (ROC-AUC) under spatial blocking and had a calibration slope closer to one. Within the inspected residential set, neighborhood building count was the strongest predictor: conditional destruction odds ratio (OR) increased 3.74-fold per standard deviation in building count within a 100 m-radius disk. Higher neighborhood Normalized Difference Moisture Index (NDMI) was associated with lower destruction odds ratio, whereas the positive Normalized Difference Vegetation Index (NDVI) coefficient in the joint model reversed when NDVI was entered separately; the opposing vegetation signals were therefore specification-dependent. Neighborhood-scale predictor sets tended to outperform immediate-surrounding sets, although their fold-level variability overlapped. A separate impact track described spectral change, relative greenness, and community context without introducing post-fire predictors. The results position open geospatial modeling as a neighborhood-scale screening tool rather than a parcel-level prediction system. The findings support neighborhood-focused assessment of built form and vegetation condition, with spatial validation and probability diagnostics reported alongside prediction maps. The reproducible workflow provides a basis for cross-event comparison; applying its coefficients or density ranges operationally requires independent-event validation.