Data Schema

OCR produces two primary types of output data: raster (tensor) datasets and vector (point) datasets. This page documents the structure, variables, and schema of both output types.

Overview

graph TD
    A[OCR Pipeline] --> B[Raster Datasets]
    A --> C[Vector Datasets]
    B --> D[Wind-Adjusted Risk Layers]
    B --> E[Wind Direction Distribution]
    C --> F[Building-Level Risk Samples]
    D --> G[Stored in Icechunk]
    E --> G
    F --> H[Stored as GeoParquet]

Hold "Alt" / "Option" to enable pan & zoom

Raster (Tensor) Datasets

Raster datasets are gridded geospatial layers stored at 30m resolution in EPSG:4326 (WGS84) projection. These datasets are organized by region and stored in the Icechunk format.

Spatial Characteristics

Property	Value
Resolution	30m (~0.0002778 degrees)
Projection	EPSG:4326 (WGS84)
Extent	CONUS
Chunking	Regional chunks (configurable)
Storage Format	Icechunk (Zarr-based)

Wind-Adjusted Fire Risk Variables

The primary output dataset contains the following variables:

Core Risk Variables

Variable	Type	Units	Description
wind_risk_2011	float32	dimensionless	Wind-informed relative risk score (RPS) for 2011 climate conditions. Calculated as burn_probability_2011 × CRPS
wind_risk_2047	float32	dimensionless	Wind-informed relative risk score (RPS) for 2047 climate conditions. Calculated as burn_probability_2047 × CRPS
burn_probability_2011	float32	dimensionless	Wind-adjusted burn probability for 2011 climate conditions, incorporating directional fire spread
burn_probability_2047	float32	dimensionless	Wind-adjusted burn probability for 2047 climate conditions, incorporating directional fire spread

Reference Variables (USFS Data)

Variable	Type	Units	Description
USFS_RPS	float32	dimensionless	Original USFS Relative Risk Score from Scott (2024) baseline dataset (RDS-2020-0016-2)
conditional_risk_usfs	float32	dimensionless	USFS Conditional Risk to Potential Structures (CRPS) from Scott (2024)
burn_probability_usfs_2011	float32	dimensionless	Original USFS burn probability for 2011 from Riley et al. (2025) (RDS-2025-0006)
burn_probability_usfs_2047	float32	dimensionless	Original USFS burn probability for 2047 from Riley et al. (2025) (RDS-2025-0006)

Coordinate Variables

Variable	Type	Description
latitude	float64	Latitude in decimal degrees (WGS84)
longitude	float64	Longitude in decimal degrees (WGS84)

Wind Direction Distribution Dataset

A separate dataset provides the statistical distribution of wind directions during fire-weather conditions:

Variable	Type	Dimensions	Description
wind_direction_distribution	float32	(latitude, longitude, wind_direction)	Fraction of fire-weather hours in each of 8 cardinal/ordinal directions

Wind Direction Dimension: The wind_direction coordinate contains 8 direction labels: ['N', 'NE', 'E', 'SE', 'S', 'SW', 'W', 'NW']

Properties:

• Values sum to 1.0 for pixels with fire-weather hours
• Values are 0 for pixels with no fire-weather hours
• Derived from CONUS404 data using 99^th percentile Fosberg Fire Weather Index (FFWI) as threshold

Data Processing Flow

flowchart LR
    A[USFS Base BP] --> B[Directional Convolution]
    C[Wind Distribution] --> D[Weighted Composite]
    B --> D
    D --> E[Wind-Adjusted BP]
    E --> F[Multiply by CRPS]
    F --> G[Wind Risk Output]

    style A fill:#e1f5ff
    style C fill:#e1f5ff
    style G fill:#ffe1e1

Hold "Alt" / "Option" to enable pan & zoom

Vector (Point) Datasets

Vector datasets contain building-level risk samples stored as a consolidated GeoParquet file covering all buildings across CONUS.

Schema

Geometry Column

Column	Type	Description
geometry	WKB (POINT)	Building point location in EPSG:4326

Risk Attribute Columns

Vector datasets contain the same risk variables as raster datasets, sampled at each building location:

Column	Type	Description
wind_risk_2011	float32	Wind-informed RPS for 2011 at building location
wind_risk_2047	float32	Wind-informed RPS for 2047 at building location
burn_probability_2011	float32	Wind-adjusted burn probability for 2011 at building location
burn_probability_2047	float32	Wind-adjusted burn probability for 2047 at building location
USFS_RPS	float32	Original USFS RPS at building location
conditional_risk_usfs	float32	USFS CRPS at building location
burn_probability_usfs_2011	float32	Original USFS BP (2011) at building location
burn_probability_usfs_2047	float32	Original USFS BP (2047) at building location

Storage Characteristics

Property	Value
Format	GeoParquet (schema version 1.1.0)
Compression	zstd
Geometry Encoding	WKB
Spatial Index	Covering bounding box (bbox)
Coverage	CONUS-wide, single consolidated file
Aggregation	Consolidated from regional processing via DuckDB

Data Quality

• Buildings with NaN values (outside CONUS or unburnable areas) are excluded
• Risk values < 0.01 are trimmed to 0 to match frontend binning
• Building locations sourced from Overture Maps dataset

File Location

The consolidated building dataset is available at:

{building_geoparquet_uri}

This single-file format enables:

• Efficient CONUS-wide spatial queries
• Direct access for analysis tools and workflows
• Simplified data distribution and versioning

Data Validation

Expected Value Ranges

Variable	Expected Range	Notes
Risk Scores (RPS)	[0, ∞)	Product of BP and CRPS
Wind Distribution	[0, 1]	Sums to 1.0 per pixel (if fire-weather hours exist)

Quality Checks

1. Spatial Consistency: All raster layers share identical coordinate systems and extents

2. Missing Data: NaN values appear only in:

   • Unburnable areas (water, urban, etc.)
   • Regions with no fire-weather hours (for wind distributions)

3. Normalization: Wind direction distributions sum to 1.0 (within tolerance of 1e-5) where valid

Metadata Attributes

All datasets include descriptive metadata attributes:

• description: Human-readable description of the variable
• long_name: Extended variable name
• units: Physical units (if applicable)
• composition: Method used for compositing (e.g., "weighted")
• direction_labels: Cardinal/ordinal direction labels for wind data
• weights_source: Source of weights used in calculations

Access Patterns

Raster Data

• By Region: Query specific regional chunks using latitude/longitude slices
• Full CONUS: Access complete dataset via Icechunk storage

Vector Data

• Full Dataset: Query the consolidated CONUS-wide building dataset
• Spatial Query: Use bounding box attributes for efficient spatial filtering
• Attribute Query: Filter by risk threshold using Parquet predicate pushdown with DuckDB or similar tools
• Regional Subset: Extract specific areas using spatial predicates on latitude/longitude

Related Documentation

• Data Downloads: Information on accessing and downloading datasets
• Deployment: Details on data storage infrastructure