Skip to content

5. Analysis and Processing

1:00 - 2:00 p.m.

Lecture 12 min · Hands-on 38 min · Break 1:50 - 2:00

Recording of section 5: Analysis and Processing

Learning goals

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

  • Explain where each kind of analysis runs in GeoLibre (your browser, a WebAssembly toolbox, a Python sidecar, or a remote service) and pick an engine on purpose.
  • Find any of the 1,000+ processing tools from the Processing menu search.
  • Buffer features, attach attributes with a spatial join, and summarize the result without writing SQL.
  • Bin points into H3 hexagons and test whether the clusters you see are statistically significant hot spots.
  • Derive hillshade and slope from an elevation model, and run a viewshed and a drive-time isochrone from a right-click.
  • Chain several tools into a reusable workflow with Model builder.

Lecture (12 min)

Where analysis runs

Section 4 used SQL for analysis. This section uses the point-and-click tools, and the first thing to know about them is where they compute. Most run on your own machine, inside the browser tab:

Engine Where it runs Used for
Client (Turf.js) Browser The default engine of the GeoLibre toolbox vector tools
Python (Pyodide) Browser The same GeoPandas and Shapely code as the sidecar, compiled to WebAssembly
Client raster tools Browser Hillshade, Slope, Aspect, Clip by extent, Raster calculator, Spectral index, Reclassify, Focal statistics
DuckDB-WASM extensions Browser DGGS tools (H3, A5, DGGRID)
Spatial statistics Browser Moran's I, Getis-Ord Gi*, kernel density, Emerging Hot Spot, composite scores
WhiteboxTools and GeoLibre tools (WebAssembly) Browser The 1,000+ tool Whitebox toolbox and Model builder
Python sidecar (rasterio, GeoPandas) Desktop app, or a self-hosted server Reproject, Resample, Contour, Zonal statistics, Polygonize, and batch runs over folders
Valhalla routing server Remote service Isochrones, OD cost matrices, routes

Everything in this section runs in the web app with no installation. The one exception to "your data stays on your machine" is routing: a drive-time isochrone sends the clicked coordinate to a public Valhalla server, and GeoLibre asks before it does.

Two toolboxes, one menu

The Processing menu holds two toolboxes, and category names such as Vector and Raster appear in both:

  • Whitebox toolbox and the category submenus under it (Hydrology, LiDAR, Terrain, Remote sensing, and others) are the WebAssembly catalog: more than 1,000 tools from WhiteboxTools and GeoLibre's own Rust tools. Hydrology and LiDAR live only here.
  • GeoLibre toolbox holds guided dialogs: Vector, Raster, Network, Spatial statistics, DGGS, and a few more. These are the tools with an engine picker.

You do not need to remember which is which. The Search processing tools... box at the top of the Processing menu searches both, and pressing Enter opens the highlighted tool. Search for Hillshade and you get the GeoLibre dialog first, then the Whitebox Hillshade, Lidar Hillshade, Hypsometrically Tinted Hillshade, and Multidirectional Hillshade tools.

Coming from QGIS or ArcGIS Pro

The tool names are the familiar ones: Buffer, Clip, Dissolve, Spatial join, Hillshade, Slope. GeoLibre toolbox plays the role of the QGIS Processing Toolbox's native algorithms, the Whitebox toolbox is the WhiteboxTools plugin, Model builder is ModelBuilder or the Graphical Modeler, and Processing → History is the Geoprocessing History pane.

Choosing a vector engine

Every GeoLibre toolbox vector tool offers three engines in an Engine menu:

Client (Turf.js) Python (Pyodide) Sidecar (GeoPandas)
Where it runs Browser Browser Local Python process
Works on the web build Yes Yes No
How a buffer gets to meters Per feature, in its own local projection One UTM zone for the whole layer One UTM zone for the whole layer
Layer spanning more than 180° of longitude Fine Refused Refused
First-run cost None Downloads the Python runtime once (tens of MB) Needs the vector extra
Matches a GeoPandas script exactly Approximately Yes Yes

For a study area like ours (a county, a state) all three agree to well within anything you could see on a map. For a continental or global layer, the Client engine is the safer choice, because a single UTM zone stretches badly far from its central meridian. We use the Client engine today because it has no download and the venue network is shared.

Since v3.7, lengths and areas computed from geometry (field calculator $area, movement tools, distance thresholds) are measured on the WGS 84 ellipsoid with GeographicLib, the same algorithms as PROJ and pyproj. Distance parameters are typed in meters, kilometers, or miles and converted for you; you never need to reproject first just to buffer.

Grids and spatial statistics

Eight thousand earthquake points are hard to read; 500 hexagons with counts are easy. A discrete global grid system (DGGS) bins the world into cells of nearly equal area, so counts are comparable, and every cell has a stable ID that does not change when a boundary is redrawn. H3, Uber's hexagonal grid, is the most widely used:

H3 resolution Approximate edge length Good for
5 about 8 km Regional patterns (today's exercise)
7 about 1.2 km City-wide patterns
9 about 170 m Neighborhoods
11 about 25 m Buildings

Binning shows a pattern. Spatial statistics ask whether the pattern is distinguishable from chance. Getis-Ord Gi* compares each cell and its neighbors with the whole layer and labels statistically significant hot and cold spots at 90, 95, and 99% confidence.

Three ways to fool yourself with a hot-spot map

  1. Significance is not importance. With many cells, small differences reach significance.
  2. The cell size is a choice. Rerun at a coarser and a finer resolution as a sensitivity check. A pattern that survives all of them is robust to cell size, not proven real: a bias in the data, such as the one in point 3, survives every resolution too (this is the modifiable areal unit problem, and hexagons do not make it go away).
  3. Counts are not rates. A hot spot of incident counts is often a hot spot of population. For earthquakes there is no population denominator, but there is a detection one: the seismic network is denser in some places than others.

Terrain from an elevation model

The client raster tools compute on the layer already on the map, so a hillshade of a 2520 × 3240-pixel DEM takes seconds and needs no server. Two notes:

  • The Z factor converts vertical units to horizontal units. The browser engine handles our geographic (EPSG:4326) DEM correctly with a Z factor of 1; if you take the same DEM to a desktop GIS, reproject it to a metric CRS first or the hillshade looks like crumpled paper.
  • The right-click Quick analysis menu runs a viewshed or line of sight on the map's own global terrain tiles, so it needs no DEM at all. It is reconnaissance: the quick viewshed ignores Earth curvature and refraction (immaterial over a few kilometers), and the terrain tiles are a generalized global model that knows nothing about buildings or trees.

From clicks to a reproducible workflow

Every toolbox and quick-analysis run is recorded in Processing → History with its parameters, engine, and duration, and can be re-run, edited, or copied as JSON or Python. Model builder goes one step further: wire tools into a graph, and the whole graph is validated before anything runs, so a missing input or a mismatched data type is a red message on the canvas rather than a failure forty minutes in. A saved model re-runs on other layers, exports to a .model.json file, or becomes a Python script for the geolibre package (Section 8).

Hands-on (38 min)

Start a new project: Project → New..., name it NGIS 2026 - 5 Analysis and processing, and click Create. Use Add data → Vector layer to load these three URLs (paste each and click Load), then close the panel:

https://data.source.coop/opengeos/data/workshops/ngis-2026/nv_counties.geojson
https://data.source.coop/opengeos/data/workshops/ngis-2026/nv_earthquakes.parquet
https://data.source.coop/opengeos/data/workshops/ngis-2026/reno_sites.geojson

Then use Add data → Raster layer to load the elevation model, and close that panel too:

https://data.source.coop/opengeos/data/workshops/ngis-2026/reno_tahoe_dem.tif

Hide nv_counties and nv_earthquakes (eye icon) so the map is not cluttered. The earthquakes still take part in every analysis while hidden.

Exercise 5.1: Buffer the landmarks (4 min)

  1. Open the Processing menu and type Buffer into Search processing tools.... Press Enter on the first result. (The long way is Processing → GeoLibre toolbox → Vector → Buffer.)
  2. The Vector tools dialog opens with Buffer selected. Notice the list on the left: Geometry, Overlay, Join, Select, DGGS, Movement & time, Data management, and Data quality tools all live in this one dialog.
  3. Set Input layer to reno_sites, Distance to 10, Units to Kilometers, and Buffer side to Outside (grow). Leave Dissolve result unticked and Engine on Client (Turf.js).
  4. Click Run. The output reads Buffered 8 feature(s) by 10 kilometers (outside), and a Buffer layer appears with one circle per landmark.

Three buffer sides

Inside (shrink) erodes a polygon (a setback, a riparian interior); Both sides keeps a band on either side of a boundary (a shoreline zone). Direction comes from the side setting, never from a negative distance.

Exercise 5.2: Which earthquakes fall in each buffer? (6 min)

A spatial join attaches the attributes of one layer to the features of another based on where they are. Keep the dialog open.

  1. Click Spatial join in the Join group.
  2. Input layer: nv_earthquakes. Join layer: Buffer. Spatial relationship: Within. Join type: Inner (keep only matched features).
  3. Click Run. The output reads Spatial join: produced 29 feature(s), and a Spatial join layer appears: the earthquakes within 10 km of a landmark, each now carrying that landmark's name and kind.
  4. Close the dialog. On the Spatial join row, open Layer actions (⋯) → Open attribute table. The footer reads 29 features.
  5. Click Explore all fields at a glance in the table toolbar. The Column explorer shows the distribution of every field. Under name:

    Landmark Earthquakes within 10 km
    Mount Rose summit 13
    Nevada State Capitol, Carson City 7
    Tahoe City 5
    Incline Village 2
    Reno-Tahoe International Airport 1
    Virginia City 1

    The same counts come out of a SQL ST_Intersects join in the SQL workspace. The workshop venue and Downtown Reno are missing: an inner join drops landmarks with no matches.

  6. Close the Column explorer and the attribute table.

Discussion

The name counts add up to 29, but the time card in the Column explorer reports only 27 unique values. Two earthquakes were counted twice. Look at the map: why? (Hint: the Mount Rose and Incline Village buffers overlap.) When would you want Dissolve result ticked on the buffer instead?

Exercise 5.3: Bin earthquakes into H3 hexagons (6 min)

  1. In Search processing tools..., type DGGS Binning and press Enter.
  2. DGGS type: H3. Input point layer: nv_earthquakes. Aggregate: Count. Resolution (0-15): 5.
  3. Click Run. The output reads Binned points into 511 H3 cell(s) at resolution 5, and the map zooms out to an H3 bins (res 5) layer covering Nevada and eastern California.
  4. Close the dialog and open the new layer's attribute table. It has two fields, h3 (the cell ID) and count. Click the count header twice to sort from high to low. The top three cells hold 802, 765, and 719 earthquakes.
  5. Copy the top cell's ID, 85298eb3fffffff, and paste it into the search box at the bottom left. The match in your H3 bins (res 5) layer is listed first; under Places, pick Go to H3 cell 85298eb3fffffff (res 5). The map frames a hexagon in the Monte Cristo Range of west-central Nevada: the M6.5 earthquake of May 15, 2020 and its aftershocks.
  6. Close the attribute table and clear the search box.

Cells are the join key

The search box accepts any H3 index, as hex or as a 64-bit integer, so a cell ID in a spreadsheet, a log file, or a colleague's message becomes a place you can look at. Plugins → DGGS → H3 draws the live grid and reports a clicked cell's parent, children, and neighbors.

Exercise 5.4: Are the clusters real? Getis-Ord Gi* (5 min)

  1. In Search processing tools..., type Getis and press Enter. The Spatial statistics dialog opens on Getis-Ord Gi* hotspots. The list on the left also has Global and Local Moran's I, Average nearest neighbor, Kernel density, Emerging Hot Spot, and Composite score.
  2. Layer: H3 bins (res 5). Value field: count. Keep Spatial weights on K nearest neighbors and Neighbors (k) at 8.
  3. Click Run. The output reads Getis-Ord Gi* for "count": 34 hot-spot, 0 cold-spot features (n=511). A new layer, H3 bins (res 5) — Gi* (count), appears. Its attribute table adds count_gi_z, count_gi_p, count_gi_bin, and count_gi_class (for example Hot spot 99%).
  4. Close the dialog. Hide H3 bins (res 5), then click the palette icon on the Gi* layer. Under Style suggestions, click Categorize by count_gi_bin.
  5. The significant cells (bin 3, 99% confidence) turn yellow in two clusters: the Monte Cristo Range in west-central Nevada, and the Ridgecrest and Coso area of eastern California, home of the M7.1 Ridgecrest earthquake of July 2019.

Getis-Ord Gi* hot spots of earthquake counts in H3 resolution 5 cells: two 99% clusters, in west-central Nevada and around Ridgecrest, California

Discussion

The tool found 0 cold spots. Is the rest of Nevada really "normal"? Think about which cells are in the input layer at all. Binning creates a cell only where at least one earthquake fell, so the quiet places that would make cold spots are missing. How would you fix that? (Hint: DGGS Generator fills an area with empty cells.)

Exercise 5.5: Hillshade and slope in the browser (6 min)

  1. Hide the Gi* layer and click Zoom to layer on reno_tahoe_dem.tif.
  2. Search for Hillshade in the Processing menu and press Enter. The Raster tools dialog opens.
  3. Engine: Client (browser) (the sidecar option is greyed out on the web). Use a layer on the map: reno_tahoe_dem.tif. Keep Azimuth (degrees) 315, Altitude (degrees) 45, and Z factor 1.
  4. Click Run. The log reports Loaded raster: 2520×3240, 1 band(s)., a note that the raster is large (8.2 M pixels), then Computed "Hillshade" in the browser. and Added "hillshade.tif" to the map.
  5. Click Slope in the same dialog. Choose the DEM again, keep Units on Degrees, and click Run. A slope layer appears.
  6. Close the dialog. In the Style panel for slope, open Color ramp and pick magma. Steep canyon walls and the Carson Range escarpment glow; Lake Tahoe and the valley floors go dark. Click Inspect pixel values on the slope row and click a few places to read slopes in degrees.
  7. Hide slope. Drag the hillshade opacity slider to about 0.5, and give the DEM the terrain color ramp in its Style panel. Color now shows elevation and the hillshade shows relief.

Computed rasters are not saved

The hillshade and slope live only in browser memory. When you save, GeoLibre warns that local raster layers were not saved with the project. Click Download GeoTIFF in the tool dialog to keep one, then host it and add it by URL. The finished project points at a hosted copy, reno_tahoe_hillshade.tif. Vector results (buffers, joins, bins) are embedded in the project file instead.

Exercise 5.6: Viewshed and drive time from a right-click (5 min)

  1. The reno_sites points now sit under the DEM and the buffers, so go to Mount Rose by coordinates: View → Set view..., Longitude -119.9179, Latitude 39.3438, Zoom 13, and press Enter. The summit is now at the center of the map. Right-click the center and choose Quick analysis → Viewshed from here (15 km).
  2. A Viewshed (15 km) image layer appears within a few seconds and the map zooms out to its extent. Green is ground visible from an observer standing on that spot. No DEM was needed: it read the same global terrain tiles as 3D terrain.
  3. Hide the viewshed and the DEM. Go to the workshop venue the same way (View → Set view..., -119.816, 39.5444, zoom 13), right-click the center of the map, and choose Quick analysis → Drive time from here (5, 10, 15 min).
  4. The first time you run a network tool, a dialog, Network analysis uses an external routing server, explains that the coordinate will be sent to the public Valhalla server at valhalla1.openstreetmap.de. Click Continue. The browser remembers your answer, so later runs skip the dialog.
  5. A Drive time layer appears: three nested areas reachable by car in 5, 10, and 15 minutes, stretching along I-80 and US-395 and thinning into the hills. In its Style panel, click the Categorize by contour suggestion to color the three bands.

Discussion

Compare the 15-minute drive-time area with a 10 km circle around the venue. Which places are inside the circle but not reachable in 15 minutes, and why? Which question does each answer?

Exercise 5.7: A stream network with Model builder (6 min)

  1. Choose Processing → Model builder. A floating panel opens: the tool palette on the left, the canvas in the middle, settings on the right, and a message log at the bottom.
  2. Click Samples and pick DEM stream network extraction. The log reads Added sample data "Sample DEM" to the map. and Loaded sample "DEM stream network extraction". The canvas shows an Input node wired to a chain of Whitebox tools.
  3. Point the model at our DEM instead of the sample: click the Input node, and in the settings on the right set Source layer to reno_tahoe_dem.tif.
  4. Click Run in the Model builder toolbar. The log steps through Fill Depressions Wang And Liu, D8 Pointer, D8 Flow Accum, Extract Streams, and Raster Streams To Vector, converting each raster output to a Cloud-Optimized GeoTIFF, and ends with Run finished — 1 output(s) added. This takes a minute or two on a laptop.
  5. Close Model builder. An Extracted Stream Network layer appears, 3,133 line features. The map is still on the Oregon sample, so remove the Sample DEM layer (trash icon, then Remove) and click Zoom to layer on Extracted Stream Network: streams drain the Carson Range into the Truckee Meadows and Washoe Valley.

The extracted stream network and the 15-minute drive-time area from the workshop venue, over the hillshaded Reno-Tahoe DEM

Why a model instead of five tool runs?

The graph is validated before it runs, intermediate rasters are discarded unless you press Keep this result on a node, and the same model re-runs on another DEM by changing one input. Save stores it in the project, Export writes a .model.json file, and after a successful run Copy Python script copies a m.run_model_builder(...) call for the geolibre package.

Save the project (Project → Save).

The buffer and spatial-join exercise: 10 km circles around the Reno-Tahoe landmarks, the 29 matched earthquakes in red, over the hillshaded DEM

Bonus: if you finish early

5.8 Processing history (2 min). Choose Processing → History. The Processing history dialog lists every toolbox and network run in this project, newest first (seven so far), each tagged with its engine (client) and duration. Every entry has Edit & re-run and Copy JSON; the vector, statistics, and network tools also have Re-run, and the vector and statistics tools have Copy Python. Click Edit & re-run on the DGGS Binning entry: the Vector tools dialog reopens with the same settings. Change Resolution to 6 and click Run: Binned points into 996 H3 cell(s) at resolution 6. Does the Monte Cristo Range still stand out?

5.9 Line of sight (3 min). Go back to Mount Rose with View → Set view... (-119.9179, 39.3438, zoom 10), right-click the center of the map, and choose Quick analysis → Line of sight from here, then click a target, such as Carson City or a point across Lake Tahoe. The Line of sight card reports Target visible or Target not visible, the Distance (about 23.7 km to the Capitol), the First obstruction and Highest point along the path, Ground in view, and the Terrain resolution, with a profile chart underneath. The Observer height (m) defaults to 1.7. If the first obstruction is only a few meters away, your observer point landed just off the summit and the peak itself blocks the view; Clear and right-click closer to the top. Untick Earth curvature and refraction and watch whether the answer changes; at tens of kilometers it can. Set Target height (m) to 30 to ask whether a mast, rather than its base, is visible.

5.10 Historical imagery (3 min). Search for the venue and zoom in to about zoom 15. Choose Plugins → Historical imagery. An Esri Wayback layer named for the newest release date (for example Esri Wayback 2026-10-29) appears, and the panel's Release timeline spans every Esri World Imagery release since 2014-02-20, about 200 of them. Drag it back through the years and watch the campus change; tick Only versions with local changes to skip releases where nothing changed here. Add persistent layer keeps a release as an ordinary layer.

Check your work

Open the finished project

The finished project (05-analysis-processing.geolibre.json) opens on the buffers, the joined earthquakes, and the drive-time area over the DEM and the hosted hillshade. The H3 bins, the Gi* hot spots, and the stream network are in the Layers panel, hidden; toggle them on. It is about 3 MB because the computed vector layers are embedded in it.

Going further

  • Overlay tools. In Vector tools, compare Clip with Intersection: both cut to the overlap, but only Intersection brings the overlay's attributes along. Union keeps no attributes on either engine.
  • Engines side by side. Run the 10 km buffer again with Engine: Python (Pyodide) (it downloads the Python runtime once) and compare the circles with the Client result at a close zoom.
  • More statistics. Run Global Moran's I on the H3 counts before mapping hot spots, or Emerging Hot Spot (space-time cube) on the raw earthquakes with a weekly time step to see where activity is new, intensifying, or fading.
  • Hydrology by hand. Search the Processing menu for Fill Depressions, D8 Flow Accum, and Extract Streams and run the chain one tool at a time; then load the Strahler stream order or Height above nearest drainage sample in Model builder.
  • Loops and branches. In Model builder, Iterate features with Collect results runs a tool once per feature (a clip per county), and If/Else switch sends features down different branches.
  • Routing. Processing → GeoLibre toolbox → Network has Isochrone / service area, OD cost matrix, and Sequential route (directions). For sensitive locations, point the tool at your own Valhalla server.
  • Share a tool run. A Copy link in a Whitebox tool dialog builds a ?tool= URL that reopens the tool with its form filled in, the most precise bug report there is.
  • Read Processing Tools, Model Builder, Vector Analysis, and Terrain Analysis.

Break: 1:50 - 2:00. Next: 6. Rendering Engines, 3D, and Time