Street-level damage and recovery

Contents

Street-level damage and recovery#

Street-level imagery captured at two dates tells a longitudinal story: what was destroyed, and what has been cleared or rebuilt since. This example fetches Mapillary panoramas from February and September 2025 for the same buildings, grades the initial damage, then re-evaluates only the damaged buildings for recovery progress.

Source: examples/fire_recovery_analysis_from_street.py.

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

footprints, damage_prompt, recovery_prompt, token_file = download_dataset(
    ['eaton_patch1_bing_buildings', 'street_chs_prompts',
     'street_recovery_prompts', 'mapillary_token']
)
buildings = PhysicalAssetCollection.from_geojson(footprints)
token = token_file.read_text().strip()

# 1. Two surveys, two folders. Date windows select the survey.
for folder, start, end in (
    ('output/street_feb', '2025-01-01', '2025-05-01'),
    ('output/street_sep', '2025-06-01', '2025-12-01'),
):
    buildings = MapillaryImageExtractor(
        access_token=token, save_directory=folder,
        start_date=start, end_date=end, cast_corner_rays=False,
    )(buildings)

# 2. Initial damage from the February views only.
model = load('gemma4', model_id='google/gemma-4-E2B-it')
damage = AssetAnalyzer(
    model, prompt=Path(damage_prompt).read_text(),
    batch_size=4, max_images_per_asset=2,
    image_filter=lambda img: 'street_feb' in str(img.path),
)
buildings = damage(buildings)

# 3. Keep the damaged buildings.
damaged = buildings.filter_by_attribute('gemma4_chs_level', '0', operator='!=')

# 4. Recovery: both surveys, more images per asset.
recovery = AssetAnalyzer(
    model, prompt=Path(recovery_prompt).read_text(),
    batch_size=2, max_images_per_asset=4,
)
damaged = recovery(damaged)
damaged.filter_empty().to_geojson(
    'damaged_buildings_recovery_status.geojson', ignore_properties=['image_assets']
)

Notes#

  • Both extractor runs attach images to the same assets; the analyzer’s image_filter decides which survey each pass sees.

  • The same model instance serves both passes, so the weights load once.

  • Vehicles along a street-level survey builds the inventory from the imagery itself instead of starting from footprints.

Contents