Aerial damage assessment: Eaton Fire

Contents

Aerial damage assessment: Eaton Fire#

The January 2025 Eaton Fire destroyed thousands of structures in Altadena, California. The RAPID Facility flew the area in February 2025; this example grades the buildings in one orthomosaic patch on the Cal Fire / CHS damage scale with a local Gemma 4 model, entirely offline.

Source: examples/fire_damage_from_aerial.py.

from pathlib import Path
from rapidtools import (
    AerialImageryExtractor,
    AssetAnalyzer,
    PhysicalAssetCollection,
    Pipeline,
    download_dataset,
)
from rapidtools.models import load

# 1. Sample data from the rAPIdtools registry.
raster_path, footprint_path, prompt_path = download_dataset(
    ['eaton_patch2', 'altadena_sample_buildings', 'aerial_chs_prompts']
)

# 2. The building inventory.
buildings = PhysicalAssetCollection.from_geojson(footprint_path)

# 3. Crop each footprint and outline it for the model.
extractor = AerialImageryExtractor(
    dataset=raster_path,
    save_directory=Path('eaton_fire_aerial_feb25/overlaid_imagery'),
    overlay_asset_outline=True,
    outline_shape='rotated_bbox',
    outline_buffer='2 m',
    outline_width='1%',
    image_prefix='eaton_trinity_25',
    keep_multiple_copies=True,
)

# 4. Grade every crop with the prompt file (a JSON schema for CHS levels).
analyzer = AssetAnalyzer(
    load('gemma4', model_id='google/gemma-4-E2B-it'),
    prompt=prompt_path,
    batch_size=8,
)

# 5. Run and export.
graded = Pipeline([extractor, analyzer]).run(buildings)
graded = graded.filter_empty()
graded.to_geojson(
    'eaton_footprints_CHS_with_gemma4.geojson', ignore_properties=['image_assets']
)

Notes#

  • keep_multiple_copies=True appends a counter when two assets share a centroid, which happens with stacked or duplicated footprints.

  • The prompt file asks for a JSON object with a chs_level key, so the grade lands in gemma4_chs_level. Open the GeoJSON in QGIS and colour by that attribute to see the damage map.

  • Swap load('gemma4', ...) for load('gemini', api_key=...) to grade with a hosted model; nothing else changes. The script examples/fire_damage_from_aerial_gemini.py does exactly that.

  • examples/chs_from_shapefile_example.py shows the same pipeline reading a Shapefile inventory and writing one back with to_shapefile().

Contents