U.S. footprint inventory (CHEER)

This recipe creates a footprint-level building inventory for hurricane damage and exposure modeling in the U.S.

It first resolves building footprints from multiple polygon sources (OpenBuildingMap, Microsoft, and local datasets), then enriches each footprint with data from linked parcels, the National Structure Inventory (NSI), Overture dwelling addresses, and FEMA’s USA Structures.

It also integrates deep-learning-based recognition of roof shapes and story counts from Google Satellite and Street View imagery (BRAILS++).

The pipeline is currently being tested in Florida, North Carolina, and Texas.

• Curation recipe: src/openplaces/recipes/US/_all/footprint/cheer/2026/US_footprint-cheer-2026.yaml

• Companion notebook: notebooks/examples/US_curate_footprints.ipynb

This work is supported by NSF’s Coastal Hazards, Economic Prosperity & Resilience hub (CHEER).

Output columns

This recipe produces one .parquet file per county.

Each row represents a deduplicated footprint (or parcel fallback).

Canonical attributes

footprint_id

A unique identifier for each footprint: its 11-digit openlocationcode (duplicate codes receive a numeric suffix, e.g., -1, -2, to guarantee uniqueness).

geometry

The spatial polygon outlining the footprint, or the parcel polygon for fallback records.

geometry_source

Source of the geometry: obm (OpenBuildingMap), microsoft, local state sources (e.g., nconemap), or parcel.<source> for parcel-shaped fallbacks representing unlocated structures.

address

Reconciled street address. Prioritizes the assessor parcel address (where available), falling back to Overture dwelling address components, completed with state names derived from administrative units.

occupancy_type

Canonical occupancy class.

Multi-Family structures are split into HAZUS height bands (Low-Rise: 1-3 stories, Mid-Rise: 4-7 stories, High-Rise: 8+ stories) based on n_stories. Multi-Family rows with no story count keep the plain Multi-Family label.

value

Reconciled structure value in USD. Prioritizes the parcel’s improvement value (apportioned across its primary footprints by floor-area share), falling back to the NSI structure replacement value.

year_built

Reconciled construction year. Prioritizes assessor parcel records over NSI block-median fallbacks.

n_stories

Reconciled number of stories. Prioritizes street-level imagery predictions (BRAILS++), falling back to NSI modeled counts.

n_dwellings

Reconciled count of dwelling units. Prioritizes Overture geocoded address counts over NSI structure counts, falling back to occupancy-class imputation.

height

Reconciled building height in meters (where available from the footprint source, e.g. OpenBuildingMap).

roof_shape

Reconciled roof structure classification (e.g., Gable, Hip, Flat) from visual models.

m2

Calculated footprint area in square meters.

priority_on_parcel

Structural role: primary (main structure), secondary (accessory structure), or unknown (unlinked to parcel).

Provenance sidecars

Indicates which input dataset or curation rule set decided the final canonical value (later steps override earlier ones). The values contain short method/source strings (e.g., parcel, nsi, overture, imputed, geometry, dwellings, secondary, park, keyword, manufactured_home, brails-2026):

• address_source

• occupancy_type_source

• value_source

• year_built_source

• n_stories_source

• n_dwellings_source

• roof_shape_source

Parcel-derived evidence (assessor)

Attributes inherited from the primary parcel linked to the footprint.

parcel_id

Globally unique ID of the primary linked parcel, used to join parcels back onto footprints.

parcel_id_local / parcel_id_local_all

Locally cross-comparable assessor ID of the primary linked parcel, and combined IDs of all intersected parcels (populated only if more than one distinct ID exists). Not unique across administrative units, so it is not used as a join key.

use_group_combined_parcel / use_group_combined_parcel_all

Primary and combined land-use description strings (populated only if more than one distinct use exists).

land_use_class_parcel

Curated parcel land-use classification mapped into 11 potential classes (e.g., Vacant, Manufactured Home, Retail, Office).

occupancy_type_parcel

Assessor-proposed occupancy based on property keywords/groups.

occupancy_type_footprint_fema

Each parcel’s dominant FEMA USA-Structures occupancy class by overlap area (remapped via the US_footprint-fema-2023_occupancy-type-remap crosswalk). Used as second-ranked evidence in the base occupancy cascade.

land_value_parcel

Assessor land valuation. Kept in full on primary footprints and set to missing (null) on secondary footprints.

improvement_value_parcel

Assessor improvement valuation. Apportioned across the parcel’s primary footprints by floor-area share.

year_built_parcel

Construction year recorded in assessor tax records.

address_parcel

Property street address from assessor records (where the parcel geometry source provides it).

overlap_fraction_parcel

Fraction of footprint area intersecting this parcel.

area_intersection_m2_parcel

The overlap area in square meters between the footprint and the dominant parcel.

n_dwellings_parcel

Area-distributed parcel dwelling count from the harmonized spine (may be empty if assessor rolls only exist in the curated parcel lane).

National Structure Inventory evidence (NSI)

Structure-level point attributes matched to the footprint.

n_buildings_nsi / building_id_nsi

Count and ID of matched NSI structure records.

n_dwellings_nsi

Summed NSI unit counts per footprint (excluding records flagged to skip upward correction). Serves as a fallback source for n_dwellings.

occupancy_type_building_nsi / occupancy_type_building_nsi_all

NSI occupancy class (single matches, and combined list of matches if more than one distinct class exists).

group_building_nsi / group_building_nsi_all

NSI occupancy group mapped to openplaces vocabulary (single matches, and combined list of matches if more than one distinct group exists).

structure_value_building_nsi

NSI-calculated replacement value of the structure.

year_built_block_median_building_nsi

Median year built for the associated census block.

source_building_nsi

Underlying source database of the NSI record (e.g., Model).

Overture dwelling and address evidence

Residential unit address points matched from Overture.

n_dwellings_overture

Total count of geocoded residential unit points linked, representing total dwelling units summed across linked address points (address deduplication and multipoint aggregation enabled). Unmatched footprints or those on vacant-use parcels are filled with 0.

address_street_dwelling_overture / address_number_dwelling_overture

Geocoded street name and house number components.

postal_code_dwelling_overture / city_dwelling_overture

Postal code and city components.

Diagnostics and special metrics

Flags and intermediate calculations used for curation and quality control.

n_parcels_per_footprint

Number of parcels intersected by the footprint.

n_footprints_per_parcel

Number of footprints on the associated parcel (includes synthetic fallbacks).

occupancy_type_conflict

Summary of conflicting occupancy labels across inputs (formatted "nsi: X | fema: Y | parcel: Z | overture: W"). This column is only populated where two or more evidence sources disagree. Non-residential classes are collapsed to Non-Residential before comparison, so only residential (or residential-vs-non-residential) disagreements are surfaced.

address_conflict

Summary of conflicting street address values across input sources, only populated where two or more address sources disagree.

occupancy_type_review

Flag (True) for low improvement-value parcel shares indicating potential manufactured homes needing inspection.

value_per_area / improvement_value_parcel_per_area / structure_value_building_nsi_per_area

Calculated values divided by footprint area (USD/m²).

manufactured_home_community

Boolean flag indicating whether the parcel is part of a manufactured home community (having more than 3 final Manufactured Home footprints).

n_manufactured_homes_per_parcel

The total number of curated Manufactured Home footprints residing on the same parcel.

Processing pipeline

The creation of the CHEER dataset proceeds through the core stages of the openplaces pipeline: Ingest (precursors, then entities, then images), Harmonize, Enrich, and Curate.

Stage 1: ingest

This stage downloads and extracts raw boundary, tile, geometry, and reference point datasets.

Precursor datasets

Downloads US Census administrative boundaries (US_admin-census-2021_admin2, US_admin-census-2021_admin3, US_admin-census-2021_admin4) and resolving tile structures (tile-obm-2025). This step generates the spatial indices required to segment raw footprint assets.

Core assets

1. Footprint polygon datasets: Downloads and unzips raw footprints from OpenBuildingMap (OBM), Microsoft, and state/local layers.

2. Parcel datasets: Gathers property tax assessor geometry and tax rolls from local/state agencies.

3. Reference layers: Downloads structure point databases (National Structure Inventory; NSI), geocoded residential address points (Overture), and FEMA USA Structures footprints (used for parcel-level occupancy evidence).

Stage 2: harmonize

This stage merges geometries and links datasets to build the core footprint spine (US_footprint-spine-2026) and parcel spine (US_parcel-spine-2026).

Footprint spine harmonization

Recipe: US_footprint-spine-2026

Creates the spatial exposure units by merging and deduplicating footprint boundaries, linking them to parcels, and attaching point reference data:

• Geometry deduplication: Merges footprints in priority order (OBM first, then Microsoft, then state-specific sources). Footprints smaller than 10 m² are dropped. Candidate shapes are added to the spine only if their Intersection-over-Union (IoU) with existing shapes is < 0.02 and they do not fall within a size-scaled exclusion buffer around large existing spine footprints (area >= 250 m²; buffer distance scales with the square root of the area: keep-out distance = 2 m + 0.5 * sqrt(area)).

• Elongated-footprint filter: Aspect ratios >= 2.5 with aligned axes (within 15°), longitudinal overlap >= 50%, and lateral spacing < 2x width are deduplicated to prevent parallel-shifted footprint representations (common for manufactured homes) from appearing twice.

• Parcel spatial overlay: Intersects footprints with parcels via a polygon identity overlay. Intersections under 10 m² or minor slivers (< 1/6 of the footprint’s largest parcel intersection) are filtered. Systematic spatial displacements between footprint and parcel layers are corrected by snapping chain-displaced footprints to their dominant parcel when minor overlaps are small and land on neighbors with their own buildings.

• Synthetic fallbacks: For parcels where assessor records indicate a structure exists but no footprint is detected, synthetic “footprint” rows are created using the parcel boundary geometry (labeled parcel.<source>). Overlaps against detected footprints are spatial-trimmed. These fallback polygons are included in the parcel footprint count, are backfilled with the ID (parcel_id) and local ID (parcel_id_local) of their source parcel, and are excluded from morphology metrics.

• Reference point linking: Associates structure-level point evidence with the footprint spine using a tiered proximity join (Lochhead et al. 2026).

  Containment: Point is within the footprint.

  Inner Proximity: Points within 10 meters of an edge are assigned to the closest footprint.

  Outer Proximity: Points within 100 meters (NSI) or 50 meters (Overture) are assigned to the closest footprint on the same parcel, ensuring points aren’t misaligned across street boundaries.

  Duplicate Resolution: Colocated duplicate points from low-rank sources (such as outdated ESRI or HAZUS/NSI-2015 records sharing a building ID or location cell with a higher-ranking source) are flagged during the link step and excluded when NSI evidence is merged onto the footprint spine.

• Role classification: Footprints on multi-structure parcels are classified to identify primary vs. accessory structures.

  Sole footprints on a parcel are always classified as primary.

  Footprints with Overture dwelling points are primary, otherwise footprints with NSI building points are primary. If no point evidence exists, all footprints on the parcel default to secondary.

  Footprints not linked to any parcel are classified as unknown, unless they carry dwelling-point evidence, in which case they are promoted to primary.

  Synthetic, parcel-derived fallback geometries representing unlocated structures are always classified as primary.

• Raw attribution: Compiles all raw joined variables into source-suffixed evidence columns in the intermediate spine.

Parcel spine harmonization

Recipe: US_parcel-spine-2026

Creates the parcel-level matching baseline by compiling land boundaries, merging tax assessor records, and linking point evidence:

• Boundary baseline: Merges discovered statewide and local parcel geometry layers into a unified spatial spine.

• Assessor records merge: Joins county and local assessment tables by matching local ID keys and standardizing property use codes.

• Reference data enrichment: Joins building structures (NSI), dominant building groups, and FEMA occupancy, and counts footprint morphology features (such as primary and small elongated footprints) on each parcel. It resolves colocated duplicate NSI points using the same rules (flagging low-rank twins) to exclude them from parcel aggregates.

Stage 3: ingest images

This stage fetches external imagery required for deep-learning visual classification, querying the Google API using footprint geometries from the harmonized spine:

1. Satellite imagery: Scrapes per-building Google Satellite images at Zoom level 20 (image-googlesatellite-z20) via the BRAILS++ scraper.

2. Street View imagery: Downloads street-level Google Street View photos (image-googlestreetview-2026).

Stage 4: enrich

This stage runs deep learning models (BRAILS++) on the ingested imagery to predict visual building attributes (Cetiner et al. 2025):

1. Roof shape prediction: Runs neural network classifiers on the scraped satellite imagery to predict roof_shape.

2. Story count detection: Runs detectors on the Street View photos to infer n_stories.

Stage 5: curate

This stage curates the spines into clean, canonical datasets.

Parcel curation

Recipe: US_parcel-openplaces-2026

Curates the parcel spine to produce clean assessor and land-use attributes:

1. Impute parcel occupancy group: Computes modal building groups per assessor use code using NSI structure counts to fill missing property occupancy groups.

2. Score relative footprint area: Calculates log-space z-scores of the largest footprint area relative to other parcels sharing the same use code, assisting vacant land and townhome classifications.

3. Classify parcel land use: Runs multi-indicator voting rules to identify specific land-use classes (such as Manufactured Home Park, RV Park, Townhome, Vacant, Retail, and Office). Townhome classification leverages morphology indicators—specifically the maximum parcels spanned and dwellings contained by a single footprint—combined with assessor keywords to identify shared-footprint row houses.

4. Reconcile default land use: Determines default land-use class for parcels not claimed by specific rules via consensus voting across three group-vocabulary columns (NSI, FEMA, and parcel use groups).

5. Derive story count from height: Estimates dominant story counts from LiDAR-derived FEMA footprint heights (approximated as height / 3.05, rounded and floored at 1).

6. Standardize data categories: Converts string columns to pandas Categorical types.

7. Format curated parcel schema: Enforces a standard column order on the final curated parcel schema.

Footprint curation

Recipe: US_footprint-cheer-2026

Integrates curated parcels, corrected address counts, reconciled attribute priorities, base occupancy imputation, visual model prediction, and weighted voting to resolve final footprint occupancy classes. The process is organized into seven functional sub-stages:

1. Input integration and value apportionment

   1. Integrate clean assessor data: Matches each footprint in the spine to its corresponding curated parcel, joining attributes like land-use class and FEMA metrics.

   2. Apportion parcel values: Joins property valuation and construction year columns from the curated parcel lane, then splits improvement_value_parcel across primary footprints by floor-area share and keeps land_value_parcel whole on the principal footprint only.

   3. Collect overlapping parcel IDs: Retains the full n:m footprint-parcel membership as a canonical column (parcel_id_all) from the overlay link sidecar.

2. Dwelling and address evidence prep

   1. Correct address evidence: Suppresses Overture dwelling unit counts on vacant-use parcels to filter out platted or pre-construction addresses.

   2. Determine implied Overture occupancy: Assigns a temporary occupancy class to target (occupancy_type_dwelling_overture) based on the corrected Overture count (Single-Family for <= 1 dwelling, Multi-Family for >= 2).

3. Core value and metric reconciliation

   1. Select canonical values: Resolves conflicts between competing source attributes by selecting canonical values for dwelling counts, construction year, and financial valuation.

   2. Reconcile street addresses: Reconciles and harmonizes street addresses from parcel assessor and Overture dwelling address components, completing missing components (like state names).

   3. Zero-fill address counts: Fills missing or suppressed n_dwellings_overture counts with 0.

   4. Compute footprint metrics: Calculates structural area in square meters and value-per-area metrics.

   5. Impute missing residential units: Imputes residential unit counts when no matched source evidence exists based on the occupancy base class.

4. Occupancy consensus and correction

   1. Establish baseline occupancy class: Resolves a weighted consensus vote across present evidence columns (NSI, FEMA, parcel, Overture) where the heaviest class wins. It applies a geometry-based manufactured home fallback and fills residential gaps.

   2. Apply property-use keyword corrections: Refines baseline occupancy using county property-use keyword rules. It also writes the occupancy_type_parcel column, sets review flags, and creates conflicts reports.

5. Imagery enrichment integration

   1. Merge visual model predictions: Merges predicted roof shape and story count attributes from visual model recipes.

   2. Reconcile story counts: Resolves conflicts between competing story count sources (visual predictions n_stories_brails and NSI block-median counts n_stories_building_nsi) by selecting the canonical value.

6. Final occupancy voting and refinement

   1. Score manufactured home probability: Computes a probability score for manufactured home classification based on assessor keywords and footprint morphology.

   2. Resolve occupancy by weighted vote: Resolves final occupancy class (specifically Manufactured Home vs. Multi-Family conflicts) to the target column (occupancy_type).

   3. Split height bands: Splits Multi-Family occupancy into Low-Rise, Mid-Rise, and High-Rise height bands based on the reconciled number of stories.

   4. Flag manufactured home communities: Flags parcels containing more than 3 final Manufactured Home footprints.

7. Schema standardization and formatting

   1. Standardize data categories: Converts string columns to pandas Categorical types.

   2. Cast year built to integer: Rounds and casts the Construction Year column to a nullable integer.

   3. Clean up and order columns: Enforces standard column order and drops transient helper columns.

Technical Annex

• Technical annex
  • Stage 1: ingest
    • Precursor ingestion
    • Core dataset ingestion
  • Stage 2: harmonize
    • Footprint spine harmonization
    • Parcel spine harmonization
  • Stage 3: ingest images
  • Stage 4: enrich
  • Stage 5: curate
    • Parcel curation
    • Footprint curation

Key references

Lochhead M, Zsarnoczay A, Deierlein G (2026) Exposure matters: a synthesis framework for high-resolution building inventory development. International Journal of Disaster Risk Reduction 139 (2026): 106148. doi:10.1016/j.ijdrr.2026.106148

Cetiner, B., McKenna, F., Yi, S.-ri, Wang, B., & Manousakis, I. V. (2025). BRAILS++ (v4.2.0). Zenodo. doi:10.5281/zenodo.17797364