Havasu Catalog.

The Havasu Catalog is an open data catalog of geospatial datasets, RasterFlow models, solution notebooks, and demo apps from Wherobots. Each dataset lists its schema and sample SQL, so a team can query it in Wherobots right away. Learn more about the Havasu Catalog.

Geospatial data records where things are on Earth. Vector data covers features such as buildings, roads, and boundaries, and raster data covers grids such as satellite and aerial imagery. The Havasu Catalog holds both. Read the introduction to spatial data.

The Havasu Catalog holds 62 entries, including 41 geospatial datasets. Sources include the Overture Maps Foundation, the US Census Bureau, OpenStreetMap, ESA Copernicus Sentinel-2, USDA NAIP aerial imagery, and USGS Landsat.

No. Every Open Data table in Havasu is queryable in place with spatial SQL in WherobotsDB, so a first analysis starts without a download.

Yes. Your own spatial tables and the open geospatial datasets share one catalog, connected through Unity Catalog, AWS Glue, or any Apache Iceberg catalog.

Apache Iceberg is an open table format for large analytic tables in cloud object storage. Havasu began in 2023 as the Wherobots spatial table format that extended Iceberg with geometry and raster types, and Iceberg v3 now includes native geometry and geography types. Read about Iceberg geospatial types.

Wherobots was founded by the original creators of Apache Sedona, which began as GeoSpark research at Arizona State University. The Havasu Catalog continues that lineage for data, models, and apps.

Create a Wherobots account at login.cloud.wherobots.com and query any Open Data table in Havasu. The solution notebooks in the catalog and the Wherobots developer docs give worked starting points.

Overture Maps is open map data published by the Overture Maps Foundation, a Linux Foundation project founded by Amazon Web Services, Meta, Microsoft, and TomTom. It covers buildings, places, transportation, addresses, administrative divisions, and base layers such as land and water. The Havasu Catalog lists each Overture theme as its own table, from buildings and places to the road network graph.

NAIP is the National Agriculture Imagery Program, run by the USDA Farm Service Agency. It captures aerial imagery of the continental United States during the agricultural growing season, typically at 60 cm to 1 m resolution with red, green, blue, and near-infrared bands. In the Havasu Catalog, NAIP is a RasterFlow managed dataset that models such as Fields of the World run on.

Sentinel-2 is the European Space Agency Copernicus mission for multispectral Earth imagery. It records 13 spectral bands at 10 m, 20 m, and 60 m resolution and revisits each location about every five days. The Havasu Catalog includes Sentinel-2 seasonal mosaics and an index of Level-2A surface reflectance scenes.

OpenStreetMap is a free, editable map of the world built by volunteer contributors and published under the Open Database License. The Havasu Catalog includes OpenStreetMap nodes, ways, and postal codes as tables.

A census tract is a small statistical subdivision of a US county, defined by the Census Bureau and usually home to 1,200 to 8,000 people. The Havasu Catalog includes TIGER boundaries for states, counties, census tracts, block groups, and ZIP Code Tabulation Areas.

The American Community Survey (ACS) is an ongoing US Census Bureau survey of population, housing, income, and commuting. Its 5-year estimates cover every census tract and block group. The Havasu Catalog includes ACS 5-year estimates with a variables table that explains each column.

A digital elevation model is a raster grid in which each cell stores the height of the Earth surface. The Havasu Catalog includes two global DEMs at 30 m resolution: Copernicus DEM GLO-30 and NASA and METI ASTER Global DEM v3.

Google Open Buildings and Microsoft Global ML Building Footprints are building outlines detected from satellite imagery with machine learning. The Havasu Catalog combines both into one building footprint table.

Landsat is the joint NASA and USGS Earth observation program, which has imaged the planet since 1972. The Havasu Catalog includes Landsat surface temperature, derived from the thermal infrared bands.

RasterFlow is the Wherobots image preparation and inference engine for large-scale raster processing and geospatial machine learning. It builds mosaics from satellite and aerial imagery, runs computer vision models over them for semantic segmentation, regression, and change detection, and vectorizes the results. RasterFlow is in Public Preview for Professional, Innovation, and Enterprise organizations. Get started with RasterFlow.

RasterFlow includes five pre-trained, open-source models: SAM3 for text-prompted object detection, Fields of the World for agricultural field boundaries from Sentinel-2, Tile2Net for sidewalks and crosswalks, Meta CHM v1 for tree canopy height, and ChesapeakeRSC for rural roads. Each model in the Havasu Catalog links to a sample output table. See the RasterFlow models.

SAM3 is Meta’s Segment Anything Model 3, which detects and segments objects from a text prompt such as “roofs”. RasterFlow runs SAM3 over aerial imagery such as 30 cm NAIP and returns georeferenced polygons or bounding boxes with confidence scores, with no separate vectorization step. Read how SAM3 maps building footprints.

Yes. Export the model with torch.export to a PT2 file, register it, and RasterFlow runs it as an inference task over a mosaic at scale, alongside the built-in models. Follow the bring-your-own-model notebook.

A PT2 file is a PyTorch 2 archive created with torch.export. One file holds the exported model, and it can also carry configuration, hyperparameters, and preprocessing transforms. RasterFlow loads PT2 models from Amazon S3 or Hugging Face, and the built-in models are published as PT2 files in the Wherobots Hugging Face collection.

A PT2 archive can be exported on a CPU and loaded onto a GPU at runtime. It can be compiled for faster execution on CUDA, AMD, or Intel GPUs, and it stores the model metadata and transforms in a standard place, so one file moves from training to large-scale inference.

Yes. RasterFlow builds mosaics from your own Cloud Optimized GeoTIFFs through a GDAL Raster Tile Index, including imagery found through a STAC catalog, in addition to the built-in NAIP and Sentinel-2 datasets. Follow the bring-your-own-rasters notebook.

RasterFlow writes mosaics and model predictions as Zarr stores and writes vectorized results as GeoParquet, which joins with vector data in WherobotsDB. Wherobots-GL shows mosaics, prediction stores, and vectorized results on one map in Wherobots Cloud or a notebook.

RasterFlow bills four tasks, mosaicking, inference, building multiscales, and vectorization, in RasterFlow Spatial Units. Mosaicking and inference apply a complexity factor that varies by dataset and by model. See Wherobots pricing.