Hummingbird Data Schemas

Tornado — buildings_damage_{event_id}.geojson

One feature per building within the modelled track corridor. Geometries are Overture Maps polygon footprints.

geometry: Polygon

Field	Type	Description
id	string	Overture Maps building UUID
state	string	US state abbreviation
wind_speed_mph	float	Peak wind speed at building centroid (mph)
r_miles	float	Distance from track centerline (miles)
track_side	int	Side of track: 0 = left, 1 = right
along_track_frac	float	Fractional position along track (0.0 = start, 1.0 = end)
area_m2	float	Building footprint area (m²)
construction_class	string	Structural class: W1 (light wood frame), W2 (heavy wood), C1–C3 (concrete), S1–S5 (steel), MH (manufactured housing), etc.
si_occtype	string	Occupancy type from Structure Inventory: RES1 (single-family), RES2 (manufactured home), RES3 (multi-family), COM1–COM10 (commercial), IND1–IND6 (industrial)
si_roof_material	string	Roof covering material from SI (asphalt_shingle, metal, tile, etc.)
si_med_yr_blt	float	Median year built for the census block group
damage_ratio	float	Structural damage ratio, 0.0 (no damage) to 1.0 (total destruction)
damage_index	float	Composite damage severity index including SI modifiers
damage_state	string	Categorical state: none / slight / moderate / extensive / complete
p_destruction	float	Probability of total destruction (damage_ratio ≥ 0.5)
in_debris_zone	bool	Whether building is within the debris impact corridor
rcv_structure	float	Replacement cost value — structural component (USD)
rcv_contents	float	RCV — contents (USD)
rcv_ale	float	RCV — additional living expense (USD)
rcv_total	float	Total ground-up RCV (structure + contents + ALE) (USD)
rcv_insuredloss	float	Estimated insured loss after take-up rate and coverage limits (USD)

Hail — structures_hail_{date}.geojson

One feature per building within the MESH swath. Multiple files per event represent different processing dates or state/county batches; the latest-dated file is canonical.

geometry: Polygon

Field	Type	Description
fd_id	string	FEMA / HAZUS flood-district building identifier
lon	float	Building centroid longitude
lat	float	Building centroid latitude
occtype	string	Occupancy type (RES1, COM1, etc.)
bldgtype	string	Building construction type (W1, C3, etc.)
val_struct	float	Structure replacement value (USD)
footprint_sqft	float	Building footprint area (sqft)
hail_size_in	float	Maximum MESH-derived hail diameter at building location (inches)
max_hail_mm	float	Maximum hail size (mm)
max_damage_indexloss	float	Damage index at peak hail size, 0.0–1.0. Multiply by val_struct for ground-up loss estimate.
duration_periods	int	Number of storm periods in which the building received hail ≥ threshold

Hurricane — buildings_damage_{event_id}.geojson

One feature per building within the wind-field footprint. Wind field modelled using a parametric RVM; damage via HAZUS-MH fragility curves with SI modifiers.

geometry: Point

Field	Type	Description
overture_id	string	Overture Maps building UUID
lat / lon	float	Building centroid coordinates
val_struct	float	Structure replacement value (USD)
bldgtype	string	Building type classification
construction_class	string	Structural class (W1, C3, MH, etc.)
wind_speed_mph	float	Modelled peak 3-second gust at building location (mph)
wind_speed_ms	float	Same, in metres per second
footprint_area_sqm	float	Building footprint area (m²)
med_yr_blt	float	Median year built for the census block group
roof_shape	string	Roof shape (hip, gable, flat, etc.)
damage_ratio	float	Mean damage ratio, 0.0–1.0
damage_state	string	Categorical damage state (none / slight / moderate / extensive / complete)
damage_index	float	Composite damage severity index
rcv_structure	float	Structural replacement cost value (USD)
economic_loss_usd	float	Ground-up economic loss (USD)
loss_usdloss	float	Estimated insured loss (USD)

Wildfire — buildings.geojsonl

GeoJSON Lines format (one JSON object per line). One feature per building within the fire perimeter. Damage modelled using zone-based ignition probability combined with structural vulnerability.

geometry: Point

Field	Type	Description
zone	string	Fire exposure zone: Z1 (within perimeter), Z2 (ember exposure), Z3 (buffer)
BDI	float	Building Damage Index — probability-weighted damage severity, 0.0–1.0
DR_mid	float	Damage ratio, midpoint estimate, 0.0–1.0
val_struct	float	Structure replacement value (USD)
ground_up_lossloss	float	Ground-up economic loss (DR_mid × val_struct) (USD)
occtype	string	Occupancy type (RES1, COM1, etc.)
roof_material	string	Roof covering material — key vulnerability driver for wildfire

Flood — viewer snapshot GeoJSON

One feature per building. Inundation depth from HAND/NWM model, uncertainty bounds from ensemble runs. PMTiles format available for large events.

geometry: Polygon

Field	Type	Description
fd_id	string	FEMA / HAZUS building identifier
depth_ft	float	Inundation depth, central estimate (ft)
depth_lo_ft	float	Lower-bound inundation depth (ft)
depth_hi_ft	float	Upper-bound inundation depth (ft)
p_flooded	float	Probability that inundation depth > 0, 0.0–1.0
data_source	string	Depth model source: HAND (Height Above Nearest Drainage), NWM (National Water Model), SLOSH (storm surge), or pluvial
occtype	string	Occupancy type (RES1, COM1, etc.)
damage_index	float	Structural damage index from depth-damage curve, 0.0–1.0
p_insured	float	Estimated flood insurance take-up rate for the census tract
est_claim_kloss	float	Estimated insured claim (thousands USD)

Straight-line Wind — structures_loss_{event_id}.geojson

One feature per building within the storm footprint. Wind field interpolated from WSR-88D peak gust observations; damage via HAZUS straight-line wind fragility curves.

geometry: Point

Field	Type	Description
structure_id	string	Structure identifier
peak_gust_mph	float	Peak 3-second gust at building location (mph)
eff_gust_mph	float	Effective gust after terrain roughness and exposure adjustments (mph)
const_class	string	Construction class (W1, C3, MH, etc.)
occtype	string	Occupancy type (RES1, COM1, etc.)
footprint_sqft	float	Building footprint area (sqft)
replacement_value	float	Structure replacement value (USD)
p_slight	float	Probability of slight damage state
p_moderate	float	Probability of moderate damage state
p_extensive	float	Probability of extensive damage state
p_complete	float	Probability of complete damage / total loss
mdr	float	Mean damage ratio (probability-weighted across all states), 0.0–1.0
damage_index	float	Composite damage index
wind_multiplier	float	Site-specific wind amplification factor (topography, surface roughness)
calibration_factor	float	Model calibration adjustment applied to this event
event_type	string	Storm type: derecho / convective / tropical_remnant
ground_up_lossloss	float	Ground-up economic loss (mdr × replacement_value) (USD)

PMTiles — {event_id}.pmtiles

Cloud-optimised vector tile archive. Contains the same building-level features as the corresponding GeoJSON, optimised for map rendering at zoom levels 8–16. Use with PMTiles JS or MapLibre GL. Hosted on Cloudflare R2 at https://media.hummingbirdsoftware.net/{peril}/event/{event_id}/{event_id}.pmtiles.

Field	Type	Description
layer name	—	Same as peril: `tornado`, `hail`, `hurricane`, `slw`. Use this as the source-layer in MapLibre style expressions.
damage_index	float	Primary field for choropleth styling (0.0–1.0). Available at all zoom levels.
damage_state	string	Categorical label for tooltip display.
rcv_insured / loss_usd / ground_up_loss	float	Loss field — name varies by peril (see GeoJSON schemas above). Available at zoom ≥ 12.
occtype	string	Occupancy type filter field.

Behavioral Claim Risk Model — behavioral_scores.geoparquet

Census tract–level behavioral risk scores for 23 US metro areas, 21,834 tracts. GeoParquet format (Apache Parquet + embedded WKB geometry). Read with GeoPandas, DuckDB, or any Parquet-compatible tool. Hosted at gs://hummingbird-si/behavioral/behavioral_scores.geoparquet.

Architecture: Raw signals from ACS, CBP (Overture Places), CFPB, HMDA, Zillow, Eviction Lab, HUD, court records, and police dispatch are assembled into four composite indices. Those indices plus all raw signals feed a K-means cluster model that assigns each tract to one of four behavioral cohorts. All index fields are normalized 0–100 within city.

geometry: Polygon (Census tract boundary)

Field	Type	Description
geoid_tractkey	string	11-digit FIPS Census tract code (state 2 + county 3 + tract 6). Join key to any Census dataset.
geometry	geometry	Census tract polygon boundary (WGS 84 / EPSG:4326)
city	string	Metro area name (e.g. "Chicago", "Austin", "Miami")
cohortprimary	string	K-means behavioral cohort assignment — apply as a loss multiplier to peril model outputs: Careful Keepers — low stress, low dispute ecosystem, high maintenance investment. Below-expected claiming. Squeaky Wheels — moderate stress + dispute culture. Baseline claiming. Duct Tape Dynasty — elevated financial stress + deferred maintenance. Elevated claiming. Claim Cavalry — high dispute culture, active legal ecosystem, high stress. Highest claiming propensity.
tipping_zone	string	Claim Cavalry neighbourhood concentration: high (>25% of nearby tracts are CC) / medium (15–25%) / normal (<15%). Contagion signal — CC clusters amplify each other.
cc_pct	float	Share of census tracts within 1 km radius that are Claim Cavalry. Used to compute tipping_zone.
cc_migration_risk	float	0–100 score estimating probability the tract shifts toward Claim Cavalry within 3 years. Driven by: trajectory_direction, tipping_zone proximity, behavioral_trajectory momentum, and ecosystem_growth_rate (Overture Places delta 2024–2026).
COMPOSITE INDICES (0–100, higher = more of that characteristic)
financial_stress_index	float	Weighted composite of financial pressure signals. Inputs (approximate weights): • income_pct_county (28%, ACS — inverted: lower income = higher stress) • renter_share (18%, ACS) • snap_rate (14%, ACS — SNAP recipients) • vacancy_rate (12%, ACS) • rent_burden (10%, ACS — gross rent >30% income) • poverty_rate (10%, ACS) • unemployment_rate (10%, ACS) • single_parent_share (8%, ACS) • eviction_filing_rate (8%, Eviction Lab 2015–18) • income_trend_pct (5%, ACS 2019→2023 delta — inverted) • pawn_shop_index (5%, Overture Places CBP) • linguistic_isolation_rate (5%, ACS) • disability_rate (5%, ACS) • mobile_home_share (4%, ACS) • overcrowded_rate (4%, ACS — >1 occupant/room) • noncitizen_share (4%, ACS — institutional distrust proxy) • no_internet_rate (3%, ACS) • no_vehicle_rate (3%, ACS)
dispute_culture_index	float	Weighted composite of behavioural willingness to dispute (activity and outcome signals — not supply-side infrastructure, which is in dispute_ecosystem_index). Inputs: • cfpb_complaint_index (40%, CFPB Consumer Complaint Database — normalized complaint rate) • claiming_culture_index (30%, derived from CFPB + legal ecosystem co-occurrence) • atty_cfpb_interaction (10%, attorney density × CFPB complaint rate — areas with both legal infrastructure AND active disputing) • dispatch_dispute_rate (10%, police dispatch — property/noise/neighbor calls; Socrata cities only) • court_filing_rate (10%, civil court filing rate; FL only)
dispute_ecosystem_index	float	Weighted composite of professional infrastructure that enables and amplifies claims (supply-side). Inputs: • atty_density_index (30%, Overture Places CBP — law offices per capita) • public_adjuster_index (30%, Overture Places CBP) • chiropractor_index (15%, Overture Places CBP — post-loss billing proxy) • diagnostic_imaging_index (15%, Overture Places CBP — urgent care / imaging centers) • process_server_index (10%, Overture Places CBP)
maintenance_index	float	Weighted composite of active investment and upkeep behaviour. Higher = better maintained, lower claiming propensity. Inputs: • owner_occupancy_rate (25%, ACS) • hmda_home_impr_rate (25%, HMDA — home improvement loan originations per unit) • long_tenure_share (20%, ACS — owner-occupied 13+ years; community embeddedness) • maintenance_response_index (15%, derived — HMDA/age-adjusted investment choice) • permit_rate (15%, county permit records — pulling permits = active investment) • pct_veteran (15%, ACS) • median_housing_age (10%, ACS — inverted: older housing = more deferred maintenance) • discretionary_services_index (10%, Overture Places — pet grooming/vet/fitness/specialty food: investment mindset proxy) • in_migration_rate (5%, ACS — inverted: high in-migration = weaker maintenance culture) • college_rate (5%, ACS — financial sophistication proxy) • mobile_home_share (5%, ACS — inverted) • elderly_share (5%, ACS — inverted: fixed income + physical limitation)
TRAJECTORY DIMENSIONS (0–100, 25% each in neighborhood_trajectory_score)
housing_trajectory	float	Housing market momentum. Inputs: • zhvi_delta_36m (25%, Zillow — 3-year home value change) • zhvi_acceleration (20%, Zillow — ZHVI second derivative; price change rate-of-change) • vacancy_rate_usps (13%, USPS — inverted: lower vacancy = improving) • zri_delta_12m (18%, Zillow — 1-year rent index change) • vacancy_trend_yoy (12%, USPS — inverted: falling vacancy = improving) • zhvi_delta_12m (12%, Zillow — 1-year home value change) • owner_occ_change (8%, ACS 2019→2023 delta) • vacancy_change (6%, ACS 2019→2023 delta)
investment_trajectory	float	Property investment activity momentum. Inputs: • permit_trend_yoy (22%, county permit records — year-over-year permit growth rate) • permit_rate (18%, county permit records — permits per housing unit) • hmda_home_impr_rate (12%, HMDA) • discretionary_services_index (8%, Overture Places — investment confidence proxy) • home_improvement_index (9%, Overture Places CBP) • roofing_contractor_index (7%, Overture Places CBP) • maintenance_response_index (7%, derived) • owner_occupancy_rate (7%, ACS) • long_tenure_share (6%, ACS) • lihtc_unit_density (6%, HUD LIHTC — inverted: high LIHTC = low private investment) • pct_veteran (3%, ACS) • college_rate (2%, ACS) • mobile_home_share (1%, ACS — inverted)
economic_trajectory	float	Economic vitality and trend. Inputs include: • income_trend_pct (17%, ACS 2019→2023 delta) • hmda_purchase_trend (13%, HMDA — purchase mortgage origination trend) • qct_flag (7%, HUD — inverted: Qualified Census Tract = economic weakness) • income_pct_county (8%, ACS — relative income) • unemployment_rate (7%, ACS — inverted) • poverty_rate (7%, ACS — inverted) • lodes_prof+finance combined (6%, LODES — professional/finance employment density) • severe_burden_rate (6%, HUD — inverted: >50% income on rent) • subsidized_share (5%, HUD — inverted) • snap_rate (5%, ACS — inverted) • poverty_change (5%, ACS 2019→2023 — inverted) • discretionary_services_index (5%, Overture Places) • rent_burden (5%, ACS — inverted) • unemployment_change (4%, ACS delta) • lfpr_change (4%, ACS — labor force participation delta) • eviction_filing_rate (4%, Eviction Lab — inverted) • eviction_trend_yoy (4%, Eviction Lab — inverted) • lihtc_expiring_pct (4%, HUD — expiring subsidized units = worsening) • assisted_income_ratio (4%, HUD) • labor_force_participation_rate (3%, ACS) • extreme_poverty_rate (3%, ACS — inverted) • restaurant_ratio_index (3%, Overture Places — full-service vs fast-food: economic confidence) • lodes_health_index (2%), dollar_store_index (2%, inv), rent_to_own_index (2%, inv)
behavioral_trajectory	float	Dispute ecosystem and claiming pressure momentum. Higher = more pressure building. Inputs grouped by category: Ecosystem (~37%): dispute_ecosystem_index (12%), atty_density_index (7%), public_adjuster_index (7%), chiropractor_index (3%), restoration_company_index (3%), ecosystem_growth_rate (5%, Overture Places delta 2024–2026) Claiming (~39%): dispute_culture_index (12%), claiming_culture_index (8%), cfpb_complaint_index (8%), cfpb_trend_yoy (6%), dispatch_trend_yoy (5%) Financial stress proxies (~20%): pawn_shop_index (7%), bail_bond_index (7%), payday_lender_index (6%) Community (~10%): civic_engagement_index (5%), cc_migration_risk (5%)
neighborhood_trajectory_score	float	Equal-weighted average of all four trajectory dimensions (housing 25% + investment 25% + economic 25% + behavioral 25%). Scores 0–100; higher = stronger neighbourhood momentum. Note: behavioral_trajectory is scored positively here but signals claiming pressure momentum — a high overall score reflects gentrification with active claim infrastructure, not just health.
trajectory_direction	string	Qualitative label derived from trajectory dimension scores and ZHVI acceleration: Improving / Stable / Declining / Reversing Down (was improving, now decelerating)

Structure Inventory — {state}_structures.parquet

National building-level inventory. One row per structure. State-level Parquet files on GCS at gs://hummingbird-si/si/. Source: FEMA NSI 2022, enriched with Overture Maps footprints, 3DEP LiDAR roof slope, USGS seismic hazard, FEMA FIRM flood zones, and WUI classification. Underlies all peril damage models.

geometry: Point (lon/lat columns — no embedded geometry)

Field	Type	Description
fd_idkey	int64	FEMA / HAZUS building identifier. Primary join key across all peril GeoJSON outputs.
lon	float	Structure centroid longitude (WGS 84)
lat	float	Structure centroid latitude (WGS 84)
county_fips	string	5-digit FIPS county code
cbfips	string	Census block FIPS code (12 digits)
occtype	string	HAZUS occupancy type: RES1 (single-family), RES1-1SNB/1SWB (subtype), RES2 (manufactured home), RES3A–F (multi-family), COM1–COM10 (commercial), IND1–IND6 (industrial), AGR1 (agricultural), REL1 (religious), GOV1–GOV2 (government), EDU1–EDU2 (education)
bldgtype	string	HAZUS building type / construction class: W (wood), S (steel), C (concrete), PC (precast concrete), RM (reinforced masonry), URM (unreinforced masonry), MH (manufactured housing)
sqft	float	Floor area (sqft)
num_story	int	Number of stories
val_struct	float	Structure replacement cost value (USD) — NSI 2022 estimate
val_cont	float	Contents replacement cost value (USD)
med_yr_blt	int	Median year built for the census block group
found_type	string	Foundation type: S (slab), B (basement), C (crawlspace), P (pier/post)
found_ht	float	Foundation height above grade (ft) — used in flood depth-damage calculations
first_floor_elv	float	First floor elevation above sea level (ft NAVD88)
ground_elv	float	Ground elevation at structure location (ft NAVD88) — from 3DEP DEM
source	string	NSI data source code: P (parcel), A (address), C (census), I (imputed)
parcel_id	string	County assessor parcel identifier (where available)
overture_id	string	Overture Maps building UUID matched to this structure (UUID v4). Null if no Overture footprint within 20m.
footprint_area_sqm	float	Building footprint area from Overture Maps (m²)
height	float	Building height from Overture Maps (m) — null where not available
num_floors	float	Number of floors from Overture Maps — null where not available
roof_shape	string	Roof shape from Overture Maps: gable, hip, flat, gambrel, mansard, etc. Null where not available.
roof_material	string	Roof covering material from Overture Maps: asphalt_shingle, metal, tile, concrete, wood_shingle, etc. Key vulnerability driver for hail and wildfire models.
roof_slope_deg	float	Roof slope derived from 3DEP LiDAR DEM (degrees)
roof_slope_class	string	Roof slope category: flat (<5°), low (5–15°), moderate (15–25°), steep (>25°)
roof_slope_source	string	Source of slope estimate: lidar_1m / lidar_3m / lidar_10m / heuristic (from building type when LiDAR unavailable)
asce7_wind_speed_mph	int	ASCE 7 design wind speed at location (mph) — used in SLW and hurricane vulnerability
firmzone	string	FEMA FIRM flood zone designation: A, AE, X, X500, VE, etc.
ss / s1 / sds / sd1 / pgam	float	USGS seismic hazard parameters (spectral acceleration values at 0.2s and 1.0s periods)
sdc	string	ASCE 7 Seismic Design Category (A–F)
nehrp_site_class	string	NEHRP soil site class (A–F) based on vs30
vs30_ms	float	Time-averaged shear wave velocity in top 30m of soil (m/s)
wui_class	string	USFS Wildland-Urban Interface class: Interface, Intermix, Very_Low_Dens_NoVeg, Non-WUI. Key vulnerability modifier for wildfire model.
enclosure_class	string	Opening protection class for wind/hurricane: enclosed, partially_enclosed, open
enclosure_source	string	Source of enclosure classification: model / assessor / default