Wildfire risk assessment goes geolocation-aware: RoPE anchors street-view semantics into satellite tiles, reaching 0.2476 harmonic F1 over five risk classes
Synopsis
Structure-level wildfire vulnerability assessment has struggled to combine parcel-scale ignition pathways with community-scale geographic coverage. Wildfire-SSR has a vision-language model turn pre-event street-view panoramas into structured text, then uses latitude and longitude to anchor each description at a relative pixel position inside its satellite tile before fusing it with satellite features through a decoupled cross-attention decoder. On roughly 37,000 structures affected by the 2025 Eaton and Palisades fires in Los Angeles, the configuration using Rotary Position Embedding reached a building localization F1 of 0.6735 and a five-class pixel-level risk harmonic-mean F1 of 0.2476, above both unimodal baselines and an unanchored multimodal configuration.
Interpretation
Structure-level wildfire risk can now be assessed from pre-event satellite imagery and street views alone, producing a footprint and a five-tier risk map for each building instead of requiring manual inspection building by building. Manual inspection gives rich detail but does not scale, overhead remote sensing covers large areas but misses the facade conditions that decide whether a structure ignites, and one-to-one satellite-patch-to-street-view cross-view fusion is computationally impractical at community scale. On roughly 37,000 structures affected by the 2025 Eaton and Palisades fires, the full multimodal configuration reached a building localization F1 of 0.6735 and a five-class risk harmonic-mean F1 of 0.2476, above satellite-only at 0.6527 and 0.1450, street-view-text-only at 0.0884 and 0.0764, and unanchored multimodal at 0.6453 and 0.2115.
The method compresses ground evidence into text and places it back into the satellite tile by coordinate, so fusion happens on semantic-to-location correspondence rather than pixel-to-pixel alignment of two viewpoints. The vision-language model GPT-4o extracts attributes such as roof material, siding type and vegetation-structure contact from pre-event street-view panoramas into structured key-value text, and XGBoost recursive feature elimination cuts 25 attributes to 16. A frozen DINOv3 encoder pretrained on satellite imagery supplies overhead features and a frozen CLIP text encoder encodes the descriptions; decoupled cross-attention injects positional embeddings only into queries and keys while values stay purely semantic. Among three geolocation strategies, Rotary Position Embedding was strongest on real data, Random Fourier Features reached a localization F1 of 0.6699 but a harmonic-mean F1 of 0.2198, and Spatial Grid Mapping collapsed the harmonic mean to zero because its moderate-risk F1 was zero.
The whole pipeline runs over the entire study area on a single GPU, giving pre-fire screening an operational throughput. The end-to-end run executes data loading, model inference and georeferenced risk-map generation sequentially at batch size 1 without parallel optimization. On a single NVIDIA RTX A6000 it assessed about 36,978 structures in roughly 37 minutes, with pure inference at 3.79 seconds per pixel tile.
Perspective
The demonstrated scope is the wildland-urban interface of the 2025 Eaton and Palisades fires in Los Angeles, with inputs of pre-event WorldView-3 imagery at about 0.30 m ground sampling distance, Google street-view panoramas from 2020 to 2024 and CAL FIRE post-event damage inspection records, and outputs of pixel-level and building-level risk maps; the 37-minute and 3.79-second-per-tile timings come from sequential execution at batch size 1 on a single RTX A6000, which the authors describe as an unoptimized performance floor.
Street-view coverage is spatially uneven, historical captures may be temporally misaligned with pre-event satellite imagery, and trees, walls or parked vehicles can occlude facades; risk labels are built from deviations of post-event damage from the neighborhood majority class and still inherit noise from fire behavior, suppression dynamics and neighborhood effects; semantic cues for the adjacent low and moderate classes are subtle and harder to distinguish. The authors propose incorporating richer pre-fire inspection data and complementary modalities such as parcel records or fuels data, and testing transferability across regions, building typologies and fire regimes.
