Timmahw/conus-h3-land-cover

CONUS H3 Land Cover Density (NLCD 2024)

A continental United States land cover dataset aggregated to Uber's H3 hexagonal grid at resolution 10 (~120m cell edge length). Each cell contains continuous per-channel coverage fractions derived from the USGS National Land Cover Database (NLCD) 2024. Unlike NLCD's native discrete classification, this dataset expresses land cover as continuous multi-channel density values — a cell at a forest-wetland boundary is represented as 0.6 forest and 0.35 wetland rather than being forced into a single class.

Dataset Summary

Property	Value
Geographic coverage	Conterminous United States (CONUS)
Source	NLCD 2024 (USGS/MRLC), 30m resolution
H3 resolution	10 (~15,000 m² per cell, ~120m edge)
Approximate cell count	~100–120M cells
File format	GeoParquet (per state)
CRS	EPSG:4326 (WGS84)
License	CC BY 4.0

Supported Tasks

• Geospatial feature engineering: H3-indexed land cover fractions as ready-to-join features for any H3-indexed dataset (demographic, economic, ecological, health)
• Ecological analysis: Habitat connectivity, transition zone mapping, wetland-agriculture adjacency
• Urban analysis: Development intensity gradients, green space distribution, urban heat island modeling inputs
• Policy research: Healthcare/food/transit desert identification when joined with OSM or other point-of-interest data
• Machine learning: Pre-computed spatial context features for location-based prediction tasks

Dataset Structure

Files

Data is organized as one GeoParquet file per state, named {state_name}_Res10.parquet.

data/
  Alabama_Res10.parquet
  Arizona_Res10.parquet
  ...
  Wyoming_Res10.parquet

Schema

Each row represents one H3 resolution-10 cell.

Column	Type
h3_index	string
centroid_lat	float64
centroid_lon	float64
'raw_11_open_water'	float64
'raw_12_ice_snow'	float64
'raw_21_developed_open'	float64
'raw_22_developed_low'	float64
'raw_23_developed_medium'	float64
'raw_24_developed_high'	float64
'raw_31_barren'	float64
'raw_41_deciduous_forest'	float64
'raw_42_evergreen_forest'	float64
'raw_43_mixed_forest'	float64
'raw_51_dwarf_scrub'	float64
'raw_52_shrub_scrub'	float64
'raw_71_grassland'	float64
'raw_72_sedge'	float64
'raw_73_lichens'	float64
'raw_74_moss'	float64
'raw_81_pasture_hay'	float64
'raw_82_cultivated_crops'	float64
'raw_90_woody_wetlands'	float64
'raw_95_emergent_wetlands'	float64
'feature_count'	int32

Loading the Data

Python (GeoPandas / Pandas)

import pandas as pd
import geopandas as gpd

# Load a single state
wi = pd.read_parquet(
    "hf://datasets/timmahw/conus-h3-land-cover/data/Wisconsin_Res10.parquet"
)

# Filter to mixed forest cells only
forest_cells = wi[wi['raw_43_mixed_forest'] > 0.5]
print(f"{len(forest_cells):,} cells with >50% forest coverage")

# Load as GeoDataFrame with Point geometry
from shapely.geometry import Point
gdf = gpd.GeoDataFrame(
    wi,
    geometry=gpd.points_from_xy(wi.centroid_lon, wi.centroid_lat),
    crs="EPSG:4326",
)

H3 Operations

import h3

# Get a cell's H3 index from a location
lat, lng = 43.073, -89.401  # Madison, WI
cell = h3.latlng_to_cell(lat, lng, 10)

# Look up that cell in the dataset
row = wi[wi['h3_index'] == cell].iloc[0]
print(f"Land cover at Madison, WI:")
for col in [c for c in wi.columns if c.startswith('raw_')]:
    if row[col] > 0.01:
        print(f"  {col}: {row[col]:.2f}")

# Find neighboring cells (1-ring)
neighbors = h3.grid_disk(cell, 1)

# Aggregate to Res-8 neighborhood scale
wi['res8_parent'] = wi['h3_index'].apply(lambda c: h3.cell_to_parent(c, 8))
neighborhood = wi.groupby('res8_parent')[[c for c in wi.columns
                                          if c.startswith('raw_')]].mean()

Dataset Creation

Source Data

The NLCD 2024 Legacy product was obtained from the USGS Multi-Resolution Land Characteristics (MRLC) Consortium. NLCD 2024 is a 30m resolution raster classification of land cover for the conterminous United States, derived from Landsat satellite imagery and ancillary geospatial data.

NLCD is in the public domain (U.S. Government work). The derived dataset is released under CC BY 4.0 with attribution to USGS/MRLC and to this dataset.

Processing Pipeline

1. H3 cell enumeration: All H3 resolution-10 cells within each state boundary were enumerated via hierarchical expansion from a coarser resolution polyfill.
2. Raster sampling (dual formulation): For each raster pixel, its center coordinate was projected from Albers Equal Area (NLCD native CRS) to WGS84, then the containing H3 cell was identified via h3.latlng_to_cell. Pixels were grouped by (cell, NLCD class) and counted.
3. Fraction computation: Raw class counts per cell were summed into channels and divided by the total valid pixel count per cell to produce coverage fractions.
4. Batching: Processing was performed in spatially-coherent batches of ~200,000 cells using a 50km geographic grid for locality. This avoids tile-vertex artifacts that arise from H3-hierarchy-aligned batching.

Cell boundary note

The hierarchical enumeration approach includes a small fringe of cells near each state boundary whose centroid may fall slightly outside the state polygon. These cells are included in the per-state files and will have low feature_count values where the NLCD raster has NoData at their location. They can be identified and filtered by feature_count == 0.

Limitations and Caveats

Temporal vintage: This dataset reflects land cover as of 2024. Development, deforestation, and agricultural changes since 2024 are not captured.

Cell boundary fringe: Cells near state boundaries may include pixels from neighboring states. The per-state files are spatially complete for the H3 cells whose centroids fall within each state, not strictly clipped to state polygon boundaries.

CONUS only: This dataset covers only the 48 conterminous US states. Alaska, Hawaii, and US territories are not included. NLCD 2024 does provide limited coverage for some non-CONUS areas, but those were not processed.

Pixel resolution vs. cell size: At NLCD's 30m resolution, a Res-10 H3 cell (~15,000 m²) contains approximately 16 pixels on average. In transition zones and small features, the per-cell fractions reflect the majority of the cell area rather than precise geometric intersection.

Suggested uses with OSM or other H3 data

Because this dataset uses standard H3 indices, it joins directly with any other H3-indexed dataset by h3_index. Example use cases:

• Join with population density data (Meta High Resolution Settlement Layer) to identify populated areas with low forest coverage
• Join with OSM amenity data (hospitals, grocery stores) to identify developed cells lacking nearby services
• Join with census tract H3 crosswalks for socioeconomic covariates
• Use as feature input for location-based ML models alongside other H3-indexed signals

Citation

If you use this dataset in your research, please cite both this dataset and the underlying NLCD source:

This dataset:

Tim Watson (2026). CONUS H3 Land Cover Density (NLCD 2024).
Hugging Face Datasets. https://huggingface.co/datasets/timmahw/conus-h3-land-cover

NLCD 2024 source:

U.S. Geological Survey (USGS), 2024, Annual NLCD Collection 1 Science Products: U.S. Geological Survey data release, https://doi.org/10.5066/P94UXNTS

H3 geospatial indexing system:

Brodsky, I. (2018). H3: Uber's Hexagonal Hierarchical Spatial Index.
Uber Engineering Blog. https://www.uber.com/blog/h3/

Uber Technologies (2018). H3: Hexagonal hierarchical geospatial indexing system
[Software]. GitHub. https://github.com/uber/h3
Apache 2.0 License.