Skip to content

2. Bringing In Cloud-Hosted Data

9:00 - 10:00 a.m.

Lecture 15 min · Hands-on 35 min · Break 9:50 - 10:00

Recording of section 2: Bringing In Cloud-Hosted Data

Learning goals

By the end of this section you will be able to:

  • Find any data source in the Add data menu and say which of its five groups (Files, Web services, Cloud formats, 3D layers, Databases) it belongs to.
  • Load vector data by URL: GeoJSON, CSV with coordinate columns, GeoParquet (streamed in place), FlatGeobuf, and PMTiles.
  • Load a Cloud-Optimized GeoTIFF and read its real georeferencing from the file header.
  • Search a STAC catalog for Sentinel-2 imagery and add scenes to the map.
  • Add an OGC API - Features layer and a WMS layer, and explain the difference between a map service and a feature service.
  • Organize a growing layer stack with groups, and know what a saved project keeps for each kind of source.

Lecture (15 min)

One menu, five kinds of source

Almost all data enters GeoLibre through the Add data menu in the top toolbar. It is grouped by where the data lives, not by file extension:

Group What it holds Examples in this section
Files Files from your disk or a URL, read in the browser Vector layer, Delimited text layer, Raster layer
Web services Live services that answer requests WMS, OGC API - Features, ArcGIS, STAC layer
Cloud formats Single files built for partial reads over HTTP GeoParquet, FlatGeobuf, PMTiles, Zarr
3D layers Point clouds, 3D Tiles, glTF models, splats Section 6
Databases DuckDB, PostgreSQL/PostGIS, SQL Server, Apache Iceberg Section 4

Three other routes reach the same loaders: dragging a file onto the map, the Browser panel on the left rail (saved services and databases as a tree), and the catalog browsers under Plugins → Web services (STAC, Source Cooperative, USGS National Map, NASA Earthdata, and about 30 more). Whatever the route, the result is an ordinary layer in the Layers panel that you can restyle, reorder, and save.

Coming from QGIS or ArcGIS Pro?

There is no driver list and no "choose a provider" step, and you do not save a connection before you can browse a service: paste a URL and the capabilities request happens inline. The Browser panel plays the role of the QGIS Browser or the ArcGIS Pro Catalog pane when you do want saved connections. If an Add data item mentioned here is missing, Settings → Interface may be set to a beginner experience level that hides it; switch to Advanced.

Who reads your file

GeoLibre detects the format from the file itself, then hands it to the reader that does that format best. Knowing which reader handled a file explains most odd behavior:

Format Read by Worth knowing
GeoJSON The browser's own JSON parser Human-readable and universal, but read whole and several times larger in memory than on disk
CSV / TSV, Excel Delimited text reader Coordinate columns, a WKT column, or addresses geocoded on import
GeoParquet DuckDB-WASM Columnar and compressed; remote files can be streamed by range request
FlatGeobuf DuckDB-WASM Carries a spatial index, so features stream in order
GeoPackage sql.js (SQLite in WebAssembly) One file, many layers; a picker asks which to load
Shapefile (zipped) shpjs, DuckDB fallback The .prj sets the CRS; remote zips load in the browser too
KML / KMZ GeoLibre's own parser Keeps Google Earth styling; folders become layers in a group
COG / GeoTIFF cog-tiler-wasm (WebAssembly) Tiles and overviews are fetched for the current view only
PMTiles PMTiles reader One tile per request from a single archive
Zarr Zarr reader Chunked N-dimensional arrays (time × lat × lon)

When a format has no reader in the browser, DuckDB-WASM Spatial converts it to GeoJSON inside your tab (INSTALL spatial; LOAD spatial; SELECT * FROM ST_Read(...)). Nothing is uploaded.

Every vector layer is reprojected to WGS84 (EPSG:4326) on load, using the CRS stored in the file: a shapefile's .prj, a GeoParquet geo block, a GeoPackage's metadata. A State Plane shapefile, a UTM GeoPackage, and a WGS84 GeoJSON therefore line up with nothing to configure. CSV files and CAD drawings carry no CRS, so for those you tell GeoLibre which one to use.

Cloud-native formats: a small index, read first

Section 1 introduced HTTP range requests. Every cloud-native format uses them the same way: it keeps an index a client reads with one small request, and it chunks the data so each later request returns a useful piece.

Format Data Index Chunk Use it to
COG Raster TIFF header with tile offsets, plus overviews 256 or 512 px tile Display imagery, elevation, land cover
GeoParquet Vector table Parquet footer: schema, row groups, column statistics, bounding boxes Row group × column Filter, aggregate, and join large tables
FlatGeobuf Vector features Packed Hilbert R-tree Feature Stream features by bounding box
PMTiles Vector or raster tiles Directory of tile offsets One z/x/y tile Pan a big layer quickly at every zoom
Zarr N-dimensional arrays zarr.json or .zmetadata Array chunk Slice climate, weather, and ocean model output
COPC / EPT Point clouds Octree hierarchy Octree node Stream LiDAR (Section 6)

Rule of thumb: COG to display a raster, GeoParquet to analyze a table, PMTiles to show a big vector layer fast (its geometry is simplified per zoom, so it is for display, not analysis), Zarr to slice through time or depth.

Two requirements that catch everyone

In a browser, the server must send CORS headers that allow the app's origin, and single-file formats (COG, GeoParquet, FlatGeobuf, PMTiles, COPC) need HTTP Range support. A URL that works with curl but fails in GeoLibre is almost always missing one of these. Object stores such as Amazon S3 and Source Cooperative support both, which is why the workshop data lives on data.source.coop.

Large data: what happens automatically

GeoLibre picks a strategy by size so the tab keeps responding:

Threshold What changes
More than 50,000 features The layer is tiled in the browser (vector tiles generated in a worker) instead of drawn as one GeoJSON source
More than 100,000 features GeoLibre asks before converting every feature to GeoJSON in memory
Files of 100 MB or more The file is streamed through DuckDB off the main thread

When a file is genuinely too big, convert instead of fighting it: Processing → GeoLibre toolbox → Conversion has Vector to GeoParquet, Vector to PMTiles, and Raster to COG, all running in the browser.

Services and catalogs: maps versus features

Service Returns You get
WMS Rendered images A picture of the data; attributes only through GetFeatureInfo
WMTS / XYZ Pre-rendered image tiles Fast display
WFS Features as GeoJSON or GML Geometry and attributes
OGC API - Features GeoJSON over a REST API, paged Geometry and attributes
WCS Raw raster values Numbers you can analyze
ArcGIS FeatureServer / MapServer / ImageServer Features, map images, or imagery A real vector layer from a FeatureServer; raster layers from the others
STAC API JSON describing items and their asset files (usually COGs) A search engine that leads straight to cloud-native files

Use a map service when you need a reference picture, and a feature or coverage service when you need data to analyze. A STAC catalog groups data into collections (sentinel-2-l2a), items (one scene: one footprint at one time, with properties such as eo:cloud_cover), and assets (the files of an item: a true-color COG, single bands, masks).

What a saved project keeps

A project stores where each layer comes from, not a copy of the data, so a URL-based layer reopens from its URL. Two exceptions matter today: points built from a CSV are embedded in the project (so the file grows), and a feature service result (OGC API, WFS) is saved as the features you fetched. A Zarr layer is saved as its store URL plus the variable, time selector, and color settings, and reopens from the store. Credentials (API keys, S3 keys, ArcGIS sign-ins) are stripped from shared links and exports; a local save asks before keeping any.

Hands-on (35 min)

Start a new project: Project → New…, set Project name to NGIS 2026 - 2 Bringing in data, click the Positron basemap, and click Create.

All workshop files are read in place from https://data.source.coop/opengeos/data/workshops/ngis-2026/.

Exercise 2.1: GeoJSON by URL (3 min)

  1. Choose Add data → Vector layer. The Add Vector Layer panel opens over the map.
  2. Paste this URL into the URL box and click Load:

    https://data.source.coop/opengeos/data/workshops/ngis-2026/nv_counties.geojson
    
  3. Paste a second URL and click Load again:

    https://data.source.coop/opengeos/data/workshops/ngis-2026/reno_sites.geojson
    
  4. Expected result: nv_counties (17 polygons) and reno_sites (8 points) appear in the Layers panel, and the map zooms to the last layer added, the Reno-Tahoe sites (zoom about 10).

  5. Click Close panel (×) on the Add Vector Layer panel.

Exercise 2.2: Earthquakes from a CSV (4 min)

The USGS earthquake catalog exports CSV with latitude and longitude columns. This extract holds 8,031 M2.5+ earthquakes from 2016 through 2025.

  1. Choose Add data → Delimited text layer.
  2. Set Layer name to Nevada earthquakes. Keep Source type at Delimited text URL and paste:

    https://data.source.coop/opengeos/data/workshops/ngis-2026/nv_earthquakes.csv
    
  3. Click Retrieve columns. Expected result: the dialog reports Retrieved 8 columns. and sets Longitude field to longitude and Latitude field to latitude.

  4. Leave Import as at Coordinates and Coordinate system blank (the coordinates are already WGS84), then click Add layer.
  5. Expected result: a Nevada earthquakes point layer, and the map zooms out to the whole extract (zoom about 6).

Points in the Atlantic?

If a CSV holds projected eastings and northings instead of longitude and latitude, enter its EPSG code in Coordinate system (or pick one from Common coordinate systems…). Skip that step and the points land far from where they belong. Import as → Addresses (geocode) builds points from an address column instead, sending each address to a geocoding service.

Exercise 2.3: Stream a GeoParquet file in place (4 min)

  1. Choose Add data → GeoParquet layer. The same Add Vector Layer panel opens.
  2. Tick Stream GeoParquet (no copy), paste the 3,235 US counties, and click Load:

    https://data.source.coop/opengeos/data/workshops/ngis-2026/us_counties.parquet
    
  3. Expected result: the panel's layer list shows us_counties tagged geoparquet · 3.2k ft · geojson · stream, and the map zooms out to the file's full extent. US counties run from Guam to the Aleutians, so that is most of the globe.

  4. Close the panel, hide us_counties with its eye icon, and click Zoom to layer (magnifier) on the nv_counties row to come back to Nevada.

Streaming or not?

Streaming queries the file where it sits instead of copying it into memory. It suits large, spatially sorted files: this one was written with SORT_BY_BBOX in row groups of up to 256 rows (14 in all), so nearby counties share row groups and a reader can skip the rest. For small files, leave streaming off: the layer is handed to GeoLibre and gets the full Style panel and attribute table.

Exercise 2.4: The same counties as PMTiles (4 min)

  1. Choose Add data → PMTiles layer. The PMTiles Layer panel opens.
  2. Paste the URL into PMTiles URL:

    https://data.source.coop/opengeos/data/workshops/ngis-2026/us_counties.pmtiles
    
  3. Click Fetch. Expected result: Found 1 source layer(s), with the source layer counties ticked.

  4. Tick Enable feature picking (click to inspect), set Layer Name to US counties (PMTiles), and click Add Layer. The panel confirms PMTiles layer added (vector).
  5. The map jumps to the archive's own extent, a world view. Close the panel (×) and click Zoom to layer (magnifier) on the nv_counties row to come back to Nevada.
  6. Click any Nevada county. Expected result: a Counties popup. For Nye County it reads GEOID 32023, NAME Nye, ALAND 47090939214.
  7. Zoom in and out. Each zoom level fetches a handful of tiles from one file, with no tile server involved. Hide the PMTiles layer when you are done.

The US counties PMTiles archive with feature picking, showing the attributes of Nye County

Discussion

You now have the same 3,235 counties as streamed GeoParquet and as PMTiles. Which would you use to compute each county's area in SQL, and which to show counties on a public web map? Why can't PMTiles answer the first question reliably?

Exercise 2.5: A Cloud-Optimized GeoTIFF and its metadata (5 min)

  1. Choose Add data → Raster layer. Leave Rendering engine at cog-tiler-wasm (WASM).
  2. Paste the Reno-Tahoe elevation model and click Load:

    https://data.source.coop/opengeos/data/workshops/ngis-2026/reno_tahoe_dem.tif
    
  3. Expected result: the map zooms to the Reno-Tahoe area (zoom about 9) and a grayscale elevation image appears, bright on the Carson Range and dark in the valleys. Close the Add Raster Layer panel.

  4. Click the Metadata button (the ⓘ icon) on the reno_tahoe_dem.tif row. Under Source details, find what GeoLibre read from the GeoTIFF header: EPSG:4326, width 2520 × height 3240, float32, nodata -9999, Deflate compression, 512 × 512 tiles, and 3 overviews.
  5. Close the dialog. The file is 22 MB, yet it appeared almost at once: the browser fetched the header and only the overview tiles this view needs.

Styling comes next

The gray ramp is the default. Section 3 gives the DEM a terrain color ramp and a stretch. Inspect pixel values on the layer row reads elevations now if you are curious.

Exercise 2.6: Search Sentinel-2 with STAC (7 min)

  1. Type Lake Tahoe in the search box at the bottom left and pick Lake Tahoe, Placer County, California, United States. The map frames the whole lake. To bring Reno into view as well, choose View → Set view…, set Zoom to 9, and click Go.
  2. Choose Plugins → Web services → STAC catalogs. The panel docks on the right.
  3. In STAC catalog or API URL, enter Element 84's Earth Search and click Connect:

    https://earth-search.aws.element84.com/v1
    
  4. Expected result: the panel reads Earth Search by Element 84 · STAC API and lists nine collections. Click Sentinel-2 Level-2A.

  5. Keep Limit search to the current map extent ticked. Set Start date to 07/01/2025 and End date to 08/31/2025. Tick Max cloud and set it to 5 %. Set Sort by date to Newest first.
  6. Click Search items. Expected result: Showing 11 of 189 items, newest first, starting with S2B_10SFH_20250831_0_L2A, and a STAC search footprints layer on the map.
  7. Lake Tahoe straddles two Sentinel-2 tiles, 10SGJ and 11SKD. For S2B_10SGJ_20250831_0_L2A and then S2B_11SKD_20250831_0_L2A, open the Assets menu and pick True color image — COG. The panel confirms Added True color image to the map. each time, and the map recenters on the new scene.
  8. Hide STAC search footprints and reno_tahoe_dem.tif, and close the STAC panel. Clear the search box (×) to remove the pin.

Two Sentinel-2 true-color scenes from August 31, 2025 over Lake Tahoe, found and added through the STAC catalogs panel

Assets, bands, and indices

The Assets menu lists every file of an item; ones GeoLibre cannot draw are greyed out as not addable instead of failing. The Bands menu next to it builds composites and indices from the bands, such as Color infrared (vegetation) and NDVI (vegetation). Raster rendering options at the bottom of the search form set bands, colormap, min/max, and nodata for assets you add next.

Discussion

Raise Max cloud to 100 % and search again. How many more items match, and why does a summer date range over Lake Tahoe still return mostly clear scenes?

Exercise 2.7: A feature service and a map service (5 min)

  1. Choose Add data → OGC API - Features.
  2. Set Layer name to USGS monitoring locations (OGC API) and paste the USGS Water Data API into Service URL:

    https://api.waterdata.usgs.gov/ogcapi/v0
    
  3. Click Retrieve collections. Expected result: Select from 39 collections. Click Monitoring locations (monitoring-locations); the Collection box fills in monitoring-locations.

  4. Leave Max features at 1000, click Use view to fill the Bounding box from the map, and click Add layer.
  5. Expected result: stream gages, wells, and other sites appear as points around the lake. Open Layer actions (⋯) → Open attribute table on the new layer: the footer reads 1000 features, with columns such as agency_code, agency_name, and the site name (for example TRUCKEE RV ABV D...). Close the table.
  6. Now a map service from the Browser panel. Click Browser on the left rail, expand Services → WMS (three built-in services), and double-click NLCD 2021 Land Cover (MRLC). Click Layers on the left rail to return.
  7. Expected result: a NLCD 2021 Land Cover (MRLC) layer tagged WMS covers the map with land-cover colors: dark green forest, tan shrubland, red developed land around Reno. Drag its opacity to about 0.5, compare it with the imagery, then hide it.

USGS monitoring locations from OGC API - Features over the Sentinel-2 scenes, with the attribute table reporting 1000 features

Discussion

The attribute table shows exactly 1,000 features. Is that how many monitoring locations exist in the view, or something else? What would you change to get them all? And why can you sort and filter the monitoring locations, but not the NLCD land-cover classes?

Exercise 2.8: Organize the stack with a group (3 min)

Eleven layers is already a long list. Groups are folders in the layer stack.

  1. On the S2B_10SGJ_20250831_0_L2A — True color image row, open Layer actions (⋯) → New group from layer. A Group 1 folder appears with the layer inside.
  2. Double-click Group 1, type Sentinel-2 (STAC), and press Enter.
  3. On the S2B_11SKD… row, open Layer actions → Move to group → Sentinel-2 (STAC). Do the same for STAC search footprints.
  4. Expected result: one collapsible Sentinel-2 (STAC) group holding three layers, with its own eye icon and opacity slider. Hide the group, then show it again.
  5. Save the project with Project → Save (Ctrl+S, Cmd+S on macOS).

Why this project file is larger

The saved project is about 3.8 MB: the CSV earthquakes and the 1,000 monitoring locations are embedded as features, while every file-based layer (GeoJSON, GeoParquet, PMTiles, COG, Sentinel-2) is stored as a URL of a few hundred bytes.

Bonus: if you finish early

  • FlatGeobuf. Add data → FlatGeobuf layer opens a panel with Format already set to FlatGeobuf. Paste https://data.source.coop/opengeos/data/workshops/ngis-2026/us_counties.fgb into Vector URL, set Layer Name to US counties (FlatGeobuf), and click Add Layer. The panel reports Added 3235 features (flatgeobuf)., the layer arrives under your name tagged FLATGEOBUF, and the map zooms out to a world view. Compare the load time with the streamed GeoParquet copy, then remove it (trash icon, then Remove).
  • Zarr. Add data → Zarr layer, then Load sample data… → Sea surface temperature, monthly (NOAA). The form fills in the store URL, the variable sst, and a display range of -2 to 32 °C. Tick Pickable (click to show pixel value), name the layer, and click Add Layer (Zarr layer added successfully.). Set View → Set view… zoom to 2 to see the globe. The temperatures cover the oceans only: land cells are stored as no data, so the continents show the basemap through them. Change Selector (JSON) to {"time": 12} and add a second time step. The Diagnostics counter in the status bar goes to 1: the reader first checks for a Zarr v3 zarr.json, gets a harmless 404 from this v2 store, and falls back to .zmetadata. Zarr layers save with the project and reopen from the store, so you can keep them.
  • A deep link. Open this URL in a new tab. GeoLibre starts with the earthquake GeoParquet already loaded and zooms to it, no menus involved:

    https://web.geolibre.app/?data=https://data.source.coop/opengeos/data/workshops/ngis-2026/nv_earthquakes.parquet
    
  • Search your own data. Type Nevada in the search box. Before the place results, GeoLibre lists matches from the attributes of loaded layers (for example the place field of the earthquakes), grouped by layer.

Check your work

Open the finished project

The finished project (02-bringing-in-data.geolibre.json) opens over Lake Tahoe with the Sentinel-2 group, the monitoring locations, the earthquakes, and the Nevada layers visible. The DEM, the two US county layers, and the NLCD WMS are saved hidden; toggle them on in the Layers panel.

Going further

  • More catalogs. Plugins → Web services has ready-made browsers for Sentinel-2 explorer (every Sentinel-2 scene, read from a STAC-GeoParquet index with no API), NASA Earthdata, Earthdata GIS, JAXA Earth, Source Cooperative, US state GIS (including Nevada's portal), USGS National Map, and OSM downloader. Processing → Planetary Computer searches Microsoft's catalog.
  • ArcGIS services. Add data → ArcGIS layer reads FeatureServer layers as real, editable vector layers, and MapServer, ImageServer, and VectorTileServer URLs as map layers.
  • Live layers. A GeoJSON feed such as the USGS "all earthquakes, past day" feed (https://earthquake.usgs.gov/earthquakes/feed/v1.0/summary/all_day.geojson) can refresh on a schedule: Layer actions → Auto refresh works on WFS and GeoJSON URL layers, and the project remembers the cadence.
  • Private buckets. Settings → Cloud storage adds S3 connections (and S3-compatible stores such as MinIO and R2). Layers keep the s3:// URI and GeoLibre signs each request, so a saved project never contains a key. Plugins → Web services → S3 browser walks a bucket.
  • Reuse your work. Layer actions → Save to My Data stores a fully configured layer (source, style, labels, filters) in the Browser panel's My Data section for any later project, and exports as JSON for a team.
  • Convert first. Processing → GeoLibre toolbox → Conversion turns files into GeoParquet, FlatGeobuf, PMTiles, or COG in the browser.
  • Share a dataset as a link. ?data= takes GeoJSON, GeoParquet, PMTiles, a COG, or a LiDAR file, and ?stac= opens the STAC panel on a catalog, for example https://web.geolibre.app/?stac=https://earth-search.aws.element84.com/v1/collections/naip.
  • Read Adding Data, Web Services, Cloud-Native Data, and Supported Data Formats.

Break: 9:50 - 10:00. Next: 3. Visualization and Styling