Wildfire-SSR: Geolocation-Aware Multimodal Fusion for Structure-Level Wildfire Risk Assessment
Authors: Yu-Hsuan Ho, Yiming Xiao, Hao-En Lai, Ali Mostafavi, Samuel D. Brody
Organizations: Urban Resilience.AI Lab, Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX · Artie McFerrin Department of Chemical Engineering, Texas A&M University, College Station, TX · Institute for Disaster Resilient Texas, Texas A&M University, Houston, TX
Structure-level wildfire vulnerability assessment is critical for reducing losses in wildland-urban interface communities, yet existing methods often fail to combine parcel-scale ignition pathways with community-scale geographic coverage. This study introduces Wildfire-SSR (Satellite and Street-view for wildfire Risk), a geolocation-aware multimodal framework that integrates macro-scale satellite context with micro-scale street-view evidence. A Vision-Language Model extracts wildfire-relevant building and site attributes from pre-event street-view panoramas and converts them into structured semantic text. These features are spatially aligned with satellite imagery through a decoupled cross-attention fusion decoder, while a frozen satellite-pretrained DINOv3 encoder captures overhead visual context. Alternative geolocation embedding strategies are evaluated to anchor text-derived features within each satellite patch. The framework is tested on approximately 37,000 structures affected by the 2025 Eaton and Palisades fires in Los Angeles, using wildfire risk labels derived from relative post-event damage patterns to better represent inherent structural vulnerability. Rotary Positional Embedding provides the strongest spatial grounding, achieving a building localization F1 score of 0.6735 and a harmonic mean F1 score of 0.2476 for five-class pixel-level risk classification, outperforming unimodal baselines and an unanchored multimodal configuration. The model also shows a conservative bias in higher-risk classes, reducing severe-risk underestimation, and completes end-to-end assessment of the study area in 37 minutes on a single GPU. Wildfire-SSR reframes cross-view fusion as localized semantic alignment and provides a scalable tool for mitigation prioritization, resilience planning, and insurance-oriented wildfire risk assessment.
Figures & tables
Figure 1: Overview of the proposed geolocation-aware multimodal framework for structure-level wildfire risk assessment. The pipeline addresses the limitations of isolated unimodal approaches by integrating macro-scale satellite imagery with micro-scale street-view data. A Vision-Language Model first extracts semantic building vulnerabilities from street-level images into structured text descriptions. To bridge the spatial disconnect between modalities, the geographical coordinates of the street-view captures are mapped to local relative coordinates and injected into CLIP-encoded text embeddings. A decoupled cross-attention multimodal fusion decoder then aligns these geolocation-aware text features with overhead visual features extracted via DINOv3. The proposed model processes these fused representations into pixel-wise risk labels, which are then transformed into a detailed, actionable community-scale wildfire risk map.
Figure 2: Overview of the model architecture. Ground-level street-view images are first processed by a Vision-Language Model to extract structured semantic attributes, which are combined with their geographic coordinates to form spatially anchored text embeddings. A multimodal fusion decoder then integrates these location-aware micro-level features with macro-level visual features extracted from pre-event satellite imagery via a frozen satellite-pretrained vision encoder to generate precise, structure-level wildfire risk maps.
Figure 3: Structure of the two-way cross-attention block utilizing a decoupled attention mechanism. Geolocation embeddings (GEs) are injected specifically into the queries ( Q ) and keys ( K ) to facilitate spatial routing, while values ( V ) remain purely semantic to preserve micro-level building attributes. The block consists of three sequential steps: sparse text self-attention to capture geometric dependencies between structural, sparse-to-dense cross-attention to update textual context with overhead visual data, and dense-to-sparse cross-attention to infuse the macro-level visual map with ground-sourced semantic details.
Figure 4: Schematic representation of the geometric localization algorithm. To preserve view-angle perspective, orthogonal normal vectors ( ns ) are cast from the street-view camera’s heading direction. The intersection of these street vectors with the heading vectors ( nb ) originating from adjacent building centroids determines the precise spatial anchor point on the observed facade.
Figure 5: Study area: final perimeters of the 2025 Eaton and Palisades wildfires in the Los Angeles region, showing the wildland-urban interface neighborhoods used for structure-level vulnerability evaluation. Perimeter data are sourced from the City of Los Angeles GeoHub ( City of Los Angeles, 2025c ) .
Geolocation Embedding
Trainable Param # (M)
Building Localization F1 Score
Building Wildfire Risk Classification F1 Score
Minimal
Low
Moderate
High
Extreme
Harmonic Mean
Rotary Position Embedding
100
0.6735
0.2330
0.2147
0.1747
0.6842
0.2462
0.2476
Random Fourier Features
100
0.6699
0.3588
0.2296
0.1048
0.5320
0.2385
0.2198
Spatial Grid Mapping
349
0.2284
0.3916
0.4012
0.0000
0.4027
0.2267
0.0000
Table 1: Performance comparison of geolocation embedding strategies. Best results are bolded, second-best are underlined.
Figure 6: Pixel-level wildfire vulnerability predictions from the proposed RoPE-based multimodal framework (right) compared against the ground-truth building-level risk labels (left) for representative tiles in the Eaton and Palisades study areas. Colors denote the five ordinal risk classes from minimal to extreme; the comparison illustrates the model’s capacity to capture broad spatial patterns of structural vulnerability while highlighting residual confusion between adjacent low and moderate tiers.
Metric
Minimal
Low
Moderate
High
Extreme
Harmonic Mean
Pixel-level F1
0.2330
0.2147
0.1747
0.6842
0.2462
0.2476
Building-level F1
0.1561
0.1993
0.2241
0.7481
0.2173
0.2594
Building-level Adjacent F1
0.2177
0.5058
0.2717
0.7705
0.2937
0.3861
Building-level Upper Adjacent F1
0.1645
0.2146
0.2628
0.7670
0.2884
0.3022
Table 2: Quantitative evaluation of building wildfire risk classification across pixel-level and building-level aggregated metrics. Best results are bolded, second-best are underlined.
Model
Trainable Param # (M)
Building Localization F1 Score
Building Wildfire Risk Classification F1 Score
Minimal
Low
Moderate
High
Extreme
Harmonic Mean
Satellite Imagery
58
0.6527
0.0595
0.4181
0.1362
0.6976
0.1529
0.1450
Street-view Text Descriptions
100
0.0884
0.0445
0.0763
0.0517
0.5168
0.1159
0.0764
(w/o Geolocation Embedding)
Street-view Text Descriptions
100
0.3296
0.2441
0.2361
0.2267
0.4376
0.0958
0.1964
(Geolocation Embedding: RoPE)
Table 3: Ablation study results on input modalities and geolocation embeddings. Best results are bolded, second-best are underlined.
Model
Building Localization F1 Score
Building Wildfire Risk Classification F1 Score
Minimal
Low
Moderate
High
Extreme
Harmonic Mean
Satellite Imagery + Risk Labels as Text
0.6806
0.4748
0.3165
0.7516
0.7878
0.1516
0.3458
(w/o Geolocation Embedding)
Satellite Imagery + Risk Labels as Text
0.6618
0.6584
0.1976
0.4360
0.9187
0.8082
0.4466
(Geolocation Embedding: RoPE)
Satellite Imagery + Risk Labels as Text
0.6420
0.6092
0.3309
0.4261
0.9267
0.7727
0.5329
Table 4: Results of the label reconstruction diagnostic experiment utilizing explicit risk levels as text input. Best results are bolded, second-best are underlined.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1: Example street-view image used for VLM-based semantic feature extraction.
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.
Parastoo Farajpoor, Mohammadreza Narimani
Department of Biological and Agricultural Engineering, University of California, Davis, Davis, CA, 95616, USA
Wildland urban interface (WUI) fires increasingly threaten communities and infrastructure, motivating the need for rapid visual assessment. Recent research has focused primarily on detecting fire and assessing post-event damage. However, structure-level fire threat identification during active WUI fire events remains underexplored. This study evaluates Qwen3-VL-8B-Instruct for this active-fire decision using two separately scored tracks: 1,824 synthetic images from 491 designed WUI scenarios and 45 frames from a controlled-burn experiment conducted in collaboration with Columbus Fire & EMS Department. Under zero-shot prompting, the model returned valid target fields for 1,808 synthetic images and 39 controlled-burn frames. Accuracy among valid outputs was 94.0% for the synthetic track and 69.2% for the controlled-burn track, compared with a 16.0% random-selection baseline for the latter. On the synthetic track, the model generally identified the intended target across varied designed WUI scenarios when the relevant fire-spread cues were visually distinct. Controlled-burn performance was well above random selection, but errors showed difficulty interpreting spatial relations in complex scenes. The dual-track protocol provides scalable stress testing in designed scenarios and a controlled-burn reference for assessing model behavior under physical fire conditions. Future work can use the synthetic data for VLM post-training and then evaluate the adapted models on larger and more diverse real WUI fire datasets.
Zhangding Liu, Neda Mohammadi, John E. Taylor
School of Computational Science and Engineering, Georgia Institute of Technology, Atlanta, GA, USA · School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, GA, USA · The University of Sydney, Sydney, NSW, Australia
Satellite and street-view imagery see different parts of a damaged building: roofs from above, facades and debris from the ground. Average accuracy hides whether fusion methods use this complementary evidence. We study the conflict cases, the samples on which independently trained single-view models disagree, as a diagnostic subset. On paired collections from the 2025 Eaton wildfire and Hurricanes Ian and Milton, conflicts make up 9.6-33.2% of the test data. Fusion improves conflict accuracy over either view, yet an oracle that selects the correct single view per conflict remains 0.19-0.32 above the best evaluated fusion method. A linear fusion gate over building-visibility features, calibrated confidences, and disagreement recovers part of this headroom where ground photographs are aimed at the building: on Eaton it beats end-to-end fusion by +0.072 conflict accuracy p < 10^-4 and calibrated averaging by +0.051 p = 0.0001; on the panoramic hurricane datasets it is indistinguishable from end-to-end fusion. Cropping panoramas toward the building raises the conflict gain of fusion, whereas random crops of the same size do not. The spatial density of conflicts, computed without labels, correlates with tile-level wildfire damage Spearman r = 0.615, 25 tiles. Conflict cases thus show how much complementary evidence fusion uses, which view it carries under each hazard and capture setting, and where current methods fall short.
Yifan Yang, Lei Zou
Department of Geography, Texas A&M University, College Station, Texas, USA