Sentinel-2 NDVI, NDMI and NBR pre-post delta screening for ACR IFM verification, as a browser-based GeoLibre plugin.
This is a canopy change detection tool implemented as a panel. It produces the same classified rasters and histograms as the QGIS and geemap workflow it replaces, with no QGIS install, no Python environment and no plugin setup.
It needs no account. Imagery is read straight from the Copernicus Sentinel-2 L2A archive published as cloud-optimised GeoTIFFs on AWS Open Data, found through Element 84's public Earth Search catalogue. Both answer anonymous requests, so there is no sign-in, no Cloud project, no OAuth client, no test-user list and no billing. Composites, indices, deltas, histograms and classification are all computed in the browser tab.
Outputs complement, but do not replace, ground plots and developer monitoring reports.
One page, condensed from the SOP of 3 September 2026 and covering the
interactive tool. Click it for the PDF at docs/cheat-sheet.pdf.
The documentation library is the same set of guides that ships inside the app, reachable from the Disturbance Check menu in the toolbar and from the Help footer at the bottom of the tool panel.
Start with Using the tool if you are running a check, Data and access if you want to know where the imagery comes from.
Eight sections, in order:
- Imagery, where the data comes from and which cloud mask is running.
- Area of interest, as typed bounds, pasted GeoJSON, or an uploaded project boundary.
- Reporting periods, the pre and post windows, one or many, plus the cloud ceiling, with the daily climate at the site drawn under them so the windows can be placed at matching points of the season.
- Severity thresholds, the Low, Moderate and High cut points for each of the three differenced indices, editable before the first run.
- Site data, the project boundary, streamside management zones, and plot points, uploaded as zipped shapefile, GeoJSON or KML.
- Results, with overpass counts, histograms with draggable breaks, and class areas.
- Visual check, the run's own before-and-after true colour on the 20 m working grid, optionally sharpened to 2.5 m, beside Esri's dated high-resolution archive.
- Corroboration, the same ground seen through four independent records, namely mapped fire, LCMS change, LANDFIRE disturbance agent, and the FACTS management record.
The operator sets an area of interest and one or more reporting periods. The panel then, per period:
- Asks Earth Search which Sentinel-2 scenes cover the area in each window, and folds them into overpasses so a scene published in two UTM zones is not counted twice.
- Reads only the pixels it needs, by HTTP range request, out of the cloud-optimised GeoTIFFs, correcting the +1000 DN baseline offset on any scene the catalogue reports still carrying it.
- Masks cloud, shadow and snow, then reduces each window to a per-pixel median. The mask is the scene classification layer by default, or OmniCloudMask, a 57 MB model run in the tab, when a run asks for it.
- Derives NDVI, NDMI and NBR, and the three deltas, masking water at the delta stage so the RGB layers keep water for context.
- Accumulates a fixed histogram over every pixel and reads its shape.
- Classifies each delta into Low, Moderate and High.
- Counts per-class hectares on the native Sentinel-2 UTM grid, clipped to the boundary polygon rather than its bounding box.
- Paints five layers per period, the pre and post true-colour composites and the three classified rasters, and adds them to GeoLibre with the classified rasters on top.
Every constant in the analysis traces to a section of the SOP in
src/defaults.ts.
Each differenced index is divided into four classes, undisturbed, Low, Moderate and High. Undisturbed pixels are masked server-side, so the classified rasters arrive with transparency already in them and only disturbed cells are drawn over the site.
The thresholds ship as the SOP Step 6 defaults and are editable in the opening panel, before the first run, so a colleague can set them for their own site without waiting for a result. After a run they can also be dragged directly on each histogram, against the distribution they are cutting. In either case, a value moved off its default marks that index as adjusted and asks for a written justification, which is saved with the project. The breaks are kept in order, so Low always stays below Moderate.
Project boundary, streamside management zones and plot points load from a zipped shapefile, a GeoJSON file, or a KML. Files are parsed in the browser and are not uploaded.
Plot points are labelled on the map with their identifier, so a screenshot of a
disturbance polygon can be tied to a plot without a separate legend. The
identifier column is detected automatically, preferring plot-specific names like
Plot ID or PLOT_NO over generic ones like OBJECTID, and the detected field
is shown and can be changed. Loading a project boundary also sets it
as the area of interest, rather than making the operator supply the same extent
twice.
Sentinel-2 at 10 metres shows that an index changed but not what changed, because a road, a landing, a cutblock edge and a blowdown patch occupy the same handful of pixels. Section 7 addresses that question in two ways on one screen.
The tool's own before-and-after true colour is built from the same masked observations the indices were built from, on the 20 metre working grid, so it shows the composite that actually produced the number. Either view can be sharpened to 2.5 metres by SEN2SRLite, a 236 kB model run in the tab on WebGPU where the browser offers it and WebAssembly where it does not. Sharpening is a separate pass over the archive and does not feed the analysis. Hectare counts continue to read the 20 metre grid, so a generated pixel can affect what a verifier sees but not the figures the tool reports.
Beside it sits Esri's World Imagery Wayback, the dated archive of that basemap, frequently sub-metre and served anonymously. It replaces the Google Earth historical timeline the SOP relies on, and it needs no account.
Section 8 puts the same ground against four records built by other people from other data, none of which is an input to the calculation.
Mapped fire comes from three registries with different roles. MTBS assesses severity from imagery a year or more after the event, the interagency perimeter feed is same-season and covers the years MTBS has not reached, and the Canadian National Burned Area Composite which runs from 1972 and carries projects north of the border. LCMS, the Forest Service Landscape Change Monitoring System, classifies change annually across the conterminous United States from the full Landsat and Sentinel-2 record by a method unlike this tool's. LANDFIRE names the cause where LCMS only names the change, with Fire, Clearcut, Harvest, Thinning, Mastication, Insects, Disease and Weather as distinct classes. FACTS is different in kind from the other three, because it is not a measurement of the canopy but the record of what was done, entered by the people who did it, with a date and an acreage.
Agreement between an unrelated method and a threshold set in this tool carries more weight than either on its own, and where the two disagree, the disagreement is the matter worth investigating.
The practical lessons recorded in the SOP are built in as diagnostics, so they do not depend on recall.
- Dormant-season windows. A window outside July to September raises a warning, because senescence drives SWIR1 reflectance up before leaf-fall and produces a uniform false moisture-stress signal in dNDMI.
- Mismatched pre and post windows. Phenology drift between periods is the most common source of spurious inter-period change.
- Histogram shape. A unimodal distribution with a long right tail and no gap is flagged as composite contamination rather than disturbance. A bimodal distribution with a clear gap suggests where the Low break belongs.
- Thin composites. Fewer than four scenes per window is flagged, because the median normaliser is unstable below that.
- Reversed coordinates. Bounds are normalised, so a swapped east and west cannot silently produce an empty geometry.
Moving a class break off its default marks the delta as adjusted and asks for a written justification, which is saved with the project.
Earlier versions served map tiles against an access token that lapsed after an hour, and the layers went with it. The layers are now images the tab painted, so they last as long as the tab does.
Only parameters are saved into a project, not results, which means a saved project is a description of a check rather than a snapshot of one. Re-running it costs seconds and needs no credentials.
The published site is a GeoLibre web build with this plugin baked in, served from GitHub Pages at https://seamusrobertmurphy.github.io/disturbance-checker/. GeoLibre is not vendored here; the deploy workflow checks it out at a pinned tag, drops in the built plugin, builds, and publishes.
No secrets are required. Set Pages to build from GitHub Actions and push, as
described in docs/first-run.md.
Bump GEOLIBRE_REF in .github/workflows/deploy.yml
to move to a newer GeoLibre.
npm install
npm run build # typecheck, then bundle to dist/
npm test # load the bundle and assert the plugin contract and SOP defaults
npm run package # produce the drop-in layout under build/To try it against a local GeoLibre checkout, run npm run package, then copy
build/disturbance-check/ into
apps/geolibre-desktop/public/plugins/ and restart the GeoLibre dev server.
Discovery happens at build and dev-server start, so a restart is required after
adding or updating the folder.
The plugin is one self-contained ES module because GeoLibre's external-plugin
loader executes the entry through a blob import and does not resolve relative
imports inside the bundle. Its three runtime dependencies are geotiff for
reading cloud-optimised GeoTIFFs, proj4 for the UTM transforms and shpjs
for boundary imports, all MIT.
MIT. See LICENSE.
