ARPA-I INSIGHTS Dataset

A large-scale, high-density (~85 pts/m²) geiger-mode aerial LiDAR dataset covering over 1,600 km² across the Salt Lake City, UT and Denver, CO metropolitan regions.

Collected in June 2025 for a U.S. DOT ARPA-I–funded project executed by MIT Lincoln Laboratory, the dataset is designed to support transportation digital twins: virtual inspection, measurement, and inventory of infrastructure assets at corridor and regional scales.

Everything is public and requires no AWS credentials — read it anonymously over HTTPS or with unsigned S3 requests. Dataset page: Registry of Open Data on AWS.

• Coverage: 1,600+ km² across Salt Lake City and Denver
• Point Count: ~140 billion points in 82,429 LAS tiles
• Resolution: ~30 cm vertical/horizontal spacing
• Collection Date: June 2025
• Format: LAS/COPC LAZ point clouds with STAC catalog (GeoParquet index)
• License: CC-BY-4.0 (data), MIT (code)

The release contains three products. They differ in coverage, label provenance, and how much you should trust the labels.

Start from a GeoParquet index rather than listing the bucket — it carries one row per tile with footprints, CRS, point counts, and asset URLs, so you can filter spatially or by content before downloading anything.

Every path below is also reachable as an S3 URI: replace https://arpa-i-insights.s3.us-west-2.amazonaws.com/ with s3://arpa-i-insights/.

Machine-readable class definitions. Map raw Classification values through these before interpreting or combining label products.

One schema covers both products. labels/schemas/class-map.schema.json always serves the current format, with identical mirrors beside each class map.

Class maps carry two version numbers. schema_version is the document format — whether your parser can read it. content_version is the taxonomy — whether the meaning of your labels changed. Only a MAJOR content change alters what a code means, and that forces a new data product version rather than an update in place. Every revision is archived immutably at class-map/<content_version>/, chained through supersedes, and described in a CHANGELOG.md beside each class map. Validate against the URL in the document's own $schema field, which is version-pinned, rather than the unversioned schema above.

s3://arpa-i-insights/
├── lidar/v1/
│   ├── data/<sortie>/...
│   │   ├── tile-level COPC LAZ files        # canonical INSIGHTS-LiDAR access path
│   │   ├── tile-level LAS files             # provenance and legacy-tool compatibility
│   │   └── L3_unified_copc/<sortie>.copc.laz
│   └── stac/
│       ├── collection.json
│       ├── items/<sortie>/...
│       └── index/items.parquet              # STAC GeoParquet index
├── labels/
│   ├── schemas/
│   │   ├── class-map.schema.json                      # current format, validates both class maps
│   │   └── class-map/<schema_version>/...             # every format version, kept for old documents
│   ├── gis-surface/v1/
│   │   ├── index/gis-surface-labels-index.geoparquet
│   │   ├── data/<sortie>/...                          # per-tile COPC LAZ
│   │   └── metadata/
│   │       ├── class-map.metadata.json                # current class definitions
│   │       ├── class-map/<content_version>/...        # immutable copy of each revision
│   │       ├── class-map.schema.json                  # mirror of labels/schemas/
│   │       └── CHANGELOG.md                           # what changed at each revision
│   └── manual-semantic/v1/
│       ├── index/manual-semantic-labels-index.geoparquet   # per-tile index, QC + class counts
│       ├── data/
│       │   ├── qc-splits-final/tier_{0,1,2}/...            # per-tile COPC LAZ
│       │   └── qc-splits-final.zip                         # same 63 tiles, bulk download
│       └── metadata/
│           ├── class-map.metadata.json                     # current class definitions
│           ├── class-map/<content_version>/...             # immutable copy of each revision
│           ├── class-map.schema.json                       # mirror of labels/schemas/
│           ├── CHANGELOG.md                                # what changed at each revision
│           └── manual_semantic_labels_qc.jsonl             # QC record for all 106 reviewed tiles

Load the LiDAR index and plot the dataset's spatial extent (requires geopandas and pyarrow; contextily is optional, for the basemap only).

from io import BytesIO

import boto3
# Optional contextily for basemap visualization
import contextily as ctx
import geopandas as gpd
from botocore import UNSIGNED
from botocore.config import Config

bucket = "arpa-i-insights"
s3 = boto3.client('s3', config=Config(signature_version=UNSIGNED))
STAC_INDEX_URL = "lidar/v1/stac/index/items.parquet"

# Load STAC catalog
stac_gdf = gpd.read_parquet(BytesIO(s3.get_object(Bucket=bucket, Key=STAC_INDEX_URL)["Body"].read()))
print(f"Available tiles: {len(stac_gdf):,}")
print(stac_gdf.head())

# Plot spatial extent of dataset
ax = stac_gdf.plot(figsize=(10,10))
# optionally add basemap
ctx.add_basemap(ax, crs=stac_gdf.crs)
ax.set_title("ARPA-I INSIGHTS Spatial Coverage")

For a guided tour of all three products, see examples/get-to-know-ARPA-I-INSIGHTS.ipynb.

Each sortie is also published as a single unified COPC, best suited to interactive visualization rather than scripted analysis.

Two label products are published alongside the LiDAR. Their Classification codes are product-specific and are not interchangeable — code 64 is a sidewalk in GIS-Surface-Labels but a traffic signal in Manual-Semantic-Labels. Always map codes through the relevant product's class map before combining them.

Surface labels derived by fusing 2024 DRCOG planimetric polygons with the point cloud, available for the DRCOG and I25S2 sorties (37,495 tiles). Tiles are georeferenced COPC LAZ (LAS 1.4, point data record format 7).

Class 0 is not used — every point carries a class, so 1 is this product's background value. Authoritative definitions: class-map.metadata.json.

Class semantics come from the source GIS layers, not from an annotation guideline, so a class here can be scoped differently from the same-named class in Manual-Semantic-Labels. Driveway is the clearest case: DRCOG models driveways in a principally residential sense and does not represent commercial or structured-parking access, whereas the manual class (code 72) covers access to buildings and parking areas generally. DRCOG driveway coverage is also per-municipality — member jurisdictions commissioned the layer individually — so an unlabeled ground point is not evidence that no driveway is there. Compare both class maps before combining the products.

Human-annotated semantic segmentation labels covering transportation surfaces and roadside infrastructure. 63 tiles are released (1.44 km², 16.3 M labeled points of 119.9 M total).

Class codes are ASPRS-compatible: classes with an ASPRS equivalent reuse the standard code, and the rest occupy the LAS user-defined range (64–255). class-map.metadata.json is the authoritative definition.

Per-tile counts for every class are columns in the label index (count:<class name>), so tiles can be selected by class content without downloading any point clouds.

Labels are released QC-stratified rather than as a train/validation/test split, because cuboid-based annotation produces spatially structured errors. Report primary metrics on Tier 1 alone; use Tier 0 to measure false positives; keep Tier 2 results separate.

manual_semantic_labels_qc.jsonl records all 106 reviewed tiles, including the withheld ones, with structured per-class issue fields. Those fields are non-exhaustive: a blank field means "not noted during QC", not "verified absent".

These tiles are normalized, not georeferenced. Each is centered on its own footprint and scaled so its longest axis spans 100 units, so the files carry no CRS, and intensity, GPS time, and return numbers are zeroed. The transform inverts exactly from the norm:* columns of the label index; use examples/scripts/georeference_manual_labels.py to restore projected coordinates, and the index's source_lidar_https_path to recover the zeroed dimensions.

• Notebook: See examples/get-to-know-ARPA-I-INSIGHTS.ipynb for a guided tour of all three products
• Analysis Script: See examples/scripts/analyze_las.py for ground classification and DEM generation
• Georeferencing Script: See examples/scripts/georeference_manual_labels.py to restore projected coordinates to manual-semantic label tiles
• Quality-control analyses: See validation/ for runnable reference implementations of the quality-control measurements reported in the data descriptor — per-sortie noise rates, coverage gaps, footprint area and density, and inter-sortie vertical offsets. Start with validation/README.md, which also covers what to check before committing to a study area.

• Point Cloud: laspy, PDAL, CloudCompare
• Spatial Data: geopandas, pyarrow
• Cloud Access: boto3

• Sponsor: U.S. Department of Transportation ARPA-I
• Maintainer: MIT Lincoln Laboratory

ARPA-I INSIGHTS LiDAR Dataset (2026). MIT Lincoln Laboratory.
U.S. Department of Transportation ARPA-I.
Available at: https://registry.opendata.aws/arpa-i-insights/

DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited.

This material is based upon work supported by the Department of Transportation under Air Force Contract No. FA8702-15-D-0001 or FA8702-25-D-B002. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Department of Transportation.