What's physically on the ground at a GPS point, or all along a path. Road, trail, forest, lawn, building, water, wetland, rail, crops, bare ground or snow: Underfoot fuses ten free open-data sources, today's weather and the newest satellite pass among them, and returns a probability for each, with its working shown.
Open the app · Download underfoot.html · How it works · Roadmap · Changelog
A 900 m line across Yosemite Valley: from the Ansel Adams Gallery over two roads, through Cook's Meadow, across the Merced River and into the forest to the Valley Loop Trail. Each point along it is a sounding.
No API keys, no account, no server. It runs in your browser, on a phone, or from a single HTML file on disk.
Draw a line, paste coordinates, or drop in a GPX, CSV or GeoJSON track. Stations are spaced along it, and every mapped road, trail, rail line and stream the line crosses gets a station of its own, named from OpenStreetMap. Where the line follows a mapped trail or road, its stations move onto it and read as what it is, and a GPS track's weave across the trail it's on doesn't count as crossings. The transect underneath shows the posterior along the whole line over its elevation profile, and a route card above it sums the line up: how much of it is forest, grass, path or road, what it crosses, how much it climbs, and which stations it's unsure about. A station worth checking on the ground, because the call is close or the sources disagree, gets an amber halo on the map and a mark on the transect's doubt band, and the answer says why: grass 49% or forest 48% · the map says grass, land cover says forest. The five most worth a look, 40 m or more apart, make a numbered check list for your next walk, which exports as GPX waypoints.
Right or wrong. Under any answer, say whether it's right, wrong or you're not sure, what's really there and how you know. A mark keeps what each source said at that moment and stays in your browser; the Marks menu lists, edits and exports them. With ten or more, Refit fits the source weights to your marks, shows how many each set gets right on marks it never saw, and lets you choose; Reset puts the defaults back exactly.
On the demo line, the gallery comes out building 97%, Northside and Southside Drives paved 98%, Cook's Meadow grass at up to 98%, the Merced River water 98%, and the Valley Loop Trail, under 55% tree canopy in an evergreen-forest pixel, path 91%. Smoothing along the line cleans up one-station flicker inside a run of forest or meadow but never blurs a road, building or crossing.
Outline a lot and Underfoot reads all of it. Field maps of 2 m cells tile across the outline, each cell taking the map, the imagery and the land cover, canopy and satellite pixels under it. The Area card gives acres per class, the map shows each cell's call as a mosaic, and GeoJSON (the outline and each class's cells) and CSV take it away. A 5-acre lot on Cook's Meadow reads in about 5 s: forest 2.5 acres, grass 1.9, the trail across it 0.16. The classes always sum to the outline's own area. One outline takes up to 64 tiles, about 180 acres.
Every station is also rated for crossing it on foot, on an ATV and in a truck: a speed and a band, go, slow or no-go. The rating reads what the ground probably is, its grade and steepness, its roughness, today's soil water on soil that drains poorly (USDA's soil survey in the US), the trees overhead, the type of mapped path, and what the line crosses. The transect gets a band per preset, the route card sums the line (Foot 2.6 km/h, blocked at 654 m: the river, no bridge), and each station's panel shows every factor. It's a model of the ground, not of access or the vehicle. How it works, and what it doesn't know.
Under the answer, the time machine reads every distinct picture Esri's archive of past imagery holds of the spot (World Imagery Wayback, releases since 2014 with captures back to about 2010). Most of its 197 releases reuse the tile underneath, so asking each release where its tile came from walks back through only the dozen or so that changed it. Each picture is scored by the imagery classifier, the captures are smoothed across the years, and a change is flagged between the two captures either side of it: at Giga Texas, changed between Nov 2020 and Jan 2022: bare ground → building. Tap a picture to see that capture on the map. The past is the photo alone, with no historical map, and the panel says so. On six places that changed and five that didn't, it flags 5 of the 10 changes and none of the unchanged places, with two false flags at a gravel lot (every place).
Drop in a CSV, GeoJSON or GPX of separate points (survey sites, plots, sensors) and say they're points, not a line. Up to 1,000 are read as a batch: points near each other are read together, 40 at a time, so they share map tiles, terrain tiles, one weather request, one satellite search and one land-cover request. The table lists the least sure first; open a row for that point's full answer, place name included. CSV and GeoJSON keep the file's own columns. Every finished group is kept in your browser, so a tab closed partway offers the batch back and Resume reads the rest. 500 points in 25 places, half of them abroad, read in 43 s on the live network with no request refused (details).
Phone fixes wander 3–10 m. Choose ±3, ±5 or ±10 m and the answer becomes the 2 m field map averaged under that disc, which is what a fix actually tells you.
At the Ansel Adams Gallery the exact coordinate says building (97%). At ±3 m it's 94%, at ±5 m 83%, and at ±10 m it spreads to 56% building, with the forest and grass around it taking most of the rest.
The ledger shows what each source said, how many bits it moved the answer, and its reasoning in plain words. At the trail crossing: the line crosses Valley Loop Trail here → P(on)=93%, NLCD 2021 · 42 — Evergreen forest, tree canopy 55%, no surveyed structure within 45 m.
Every weight has a slider, and so does the correlation discount, so you can disagree with a source and watch the answer move.
When a call looks wrong, Export → Report a wrong call opens a GitHub issue with the sounding linked. Each report becomes a test case, and later training data.
On a phone the map fills the screen and the answer sits in a sheet over it. Drag the sheet up or down by its head, or by its content until it's full, to see more of the map or more of the answer. A line heads the sheet with its strip: drag along it, or use ‹ ›, to move between stations. Under the call, Right, Wrong and Not sure mark it without leaving the map. The tabs below hold the rest: the answer, the route, getting across, the years, and the evidence.
Tap to probe, tap to draw a line with Undo and Done, drag a vertex to reshape it, long-press to delete one, pinch to zoom. Lock freezes a finished sounding so stray taps only pan, zoom, pick stations and peek at the field map.
Here follows your phone and reads the ground under you as you walk. The newest fix is read together with the last 200 m of fixes behind it, so on a trail it names the trail, even when the fix itself is 15 m off in the trees. It re-reads only when you've moved farther than the fix's own accuracy. Rec keeps the walk; Stop reads it as a line you can export or share. A mark started while you walk holds Here until it's kept, so the question doesn't move on before you've answered.
Share card (Export, or the button under the answer on a phone) makes one picture of the call: the map, the answer and its runners-up or a line's share of length, the place and date, and the link that reopens it. 1200 × 630, about 150 KB. A phone hands it to the share sheet, and Export's GeoJSON and CSV can go there too.
| Place | Underfoot says |
|---|---|
| The demo line, Yosemite Valley | building → roads → meadow → river → forest, with every crossing named |
| The Ansel Adams Gallery, Yosemite | building 97% |
| Half Dome summit | bare ground 93% |
| Grand Prismatic Spring, Yellowstone | water 97% |
| Paradise snowfield, Mount Rainier | snow / ice 70%, bare next |
| Angels Landing summit, Zion | bare ground 55%, scrub next |
| The Everglades | wetland 64%, water next |
| Lake Louise, Banff | water 98% |
Calls as of October 2026; they shift a little as OpenStreetMap and the imagery are updated. Ten park trails to load as GPX are in tests/fixtures/trails.
Each source turns what it sees into a log-likelihood over the twelve classes. They're added in log-odds space to a prior and normalised.
| Source | What it reads | Coverage |
|---|---|---|
| OSM polygons | buildings, water, woods, farmland, parks and land use enclosing the point (OpenFreeMap vector tiles) | global |
| OSM lines | distance to every road, path, rail and stream centreline against a modelled half-width | global |
| Building footprints | FEMA USA Structures: footprint, occupancy, height | US |
| Imagery pixels | colour and texture of Esri World Imagery under the point, scored by a classifier fitted to 2,494 labelled patches | global |
| Land cover | NLCD 2021 30 m class, read as a mixture | CONUS |
| Canopy and impervious | NLCD tree-canopy and impervious fractions, plus the descriptor that tells road from roof | CONUS |
| Terrain | slope, roughness and relief from USGS 3DEP in the US (1 m lidar where it has been flown, a 3 m or 10 m DEM otherwise), or AWS's Terrain Tiles elsewhere (mostly SRTM at 30 m), with Open-Meteo's ~90 m DEM where a tile won't load | global |
| Gazetteer | Nominatim reverse geocode, counted only when its polygon contains the point | global |
| Today | the Open-Meteo weather model at the point: snow depth, soil moisture, recent rain and snowfall. Fresh snow lies on top of the ground; wet soil leans to wetland | global |
| Newest pass | the Sentinel-2 scene class at the 20 m pixel from the newest pass with a clear view, found through Element 84's Earth Search and read in the browser. Fades with age; a cloudy pixel abstains | global |
| World cover | Impact Observatory's 10 m land cover (with Microsoft and Esri), one map a year, read as a mixture. It answers wherever NLCD has no class, so in the US nothing changes | outside NLCD |
What keeps it honest:
- Correlation discount. OSM, the NLCD rasters and the photo partly see the same trees, so the summed evidence is divided by τ = (active weight) / N_eff, with N_eff ≈ 3.5. Abroad, where the footprints and NLCD rasters go dark, τ falls on its own.
- Sub-pixel floor. A 2 m trail inside a forest polygon, under canopy, inside a forest pixel would be voted out by sources that can't resolve it. Area-scale sources may support a narrow class but not refute it, and only where something has seen that class nearby.
- Crossings are facts. Where a line crosses a mapped road, the road's probability comes from its existence odds directly, outside the discount.
- Nothing reads 100%. A 2% uniform mixture stands for "any source can be wrong".
The full method is in docs/method.md.
On 36 places outside the lower 48, six per continent, labelled by eye before the engine ran, 27 are called correctly and 32 have the truth in the top two (npm run spots). On 18 engine fixtures, 17 are exact and all 18 are in the top two. The first release's 25 hand-labelled spots (22 right, all 19 in the US) can't be rerun: their coordinates weren't kept.
Along ten trails mapped by the National Park Service, from Angels Landing to the Anhinga boardwalk, it calls path at 479 of 480 stations, and at 87–94% along three hikers' GPS tracks of the Mist Trail, which wander 4–8 m off it at the median. A line that runs along a mapped trail is matched to it (how). Before that, in 1.0.0, it was 8% and 2–7%: it named the forest, scrub or wetland around the trail. A transect drawn beside a trail, or across one, isn't matched; the demo line crosses paths nine times and roads three times, and follows none.
The details, and every miss, are in docs/validation.md.
- Outside the contiguous US, the FEMA footprints and both NLCD layers have nothing to say. World cover stands in for land cover; nothing stands in for canopy, imperviousness or footprints, and accuracy drops with them.
- Today's weather is a model on a grid a few kilometres wide, not a measurement. It can carry snow that isn't there: at the Aletsch glacier in October it held 43 cm with none fallen in a week. So snow only lies on top of the ground after recent snowfall, and not at all when a clear satellite pass in the last ten days saw none and none has fallen since.
- Satellite photos can be years old or leaf-off. Seasonal snow over mapped bare rock fooled it at the Aletsch glacier, which it called bare at 97%. It still says bare. The photo says snow, and the doubt map marked the spot as worth a look until World cover sided with the map; now it doesn't.
- A single point on a trail still reads as the land around the tread; only a line that follows the trail is matched to it. A GPS track 10–15 m beside a trail gets matched too; switch follow off on the transect for that.
- OpenStreetMap completeness varies; where the map is thin, absence counts for less.
- The imagery classifier sees one 30 m patch at a time. Abroad it reads most crops as grass, it still reads some large white roofs as bare ground, and the time machine has only it to go on.
Read the percentages as calibrated opinion, not measurement.
Offline: download underfoot.html from the latest release and open it. It needs a network connection for the data, nothing else.
From source:
npm install
npm run dev # http://localhost:5173
npm run check # format, types, unit tests, both builds
npm run e2e # browser tests: soundings, mouse, touch, lock
npm run trails # the real-trail validation (docs/validation.md)
npm run spots # 36 places abroad, labelled from imagery (docs/validation.md)
npm run years # the time machine on places whose history is known
npm run batch # 500 points read as a batch (add -- --live for the live network)npm run build writes the site to dist/ (deployed to GitHub Pages from main) and the offline file to dist-single/underfoot.html (attached to each release). In the browser console, underfoot exposes the state, the engine and the parsers.
src/
core/ classes, priors, math, geometry, DOM helpers, shared types
data/ one module per external source, plus the vector-tile decoder
engine/ geometry, evidence, fusion, field map, smoothing, path following, doubt, narration (pure)
app/ state, stations and crossings, the sounding run, user actions
map/ canvas map, drawing, interaction, lock
ui/ console, transect, menus, dialogs
io/ coordinates, files, search, links, history, export, GPX, marks
model/ imagery classifier weights
calib/ how the imagery classifier was fitted (Python)
tests/unit engine fixtures and parsers (Vitest)
tests/e2e browser tests (Playwright)
tests/fixtures ten National Park Service trails as GPX, 36 labelled places abroad
scripts/ README images, trail fixtures, the trail, places-abroad and field-map checks, the community refit
supabase/ the community store's table and rules (not switched on yet)
docs/ method, validation, go / slow / no-go, architecture, community weights, roadmap
The code is moving to strict TypeScript file by file; core/, the fusion engine and path following are done. See docs/architecture.md.
Six milestones are out:
- M1, the phone-first release: following trails, live GPS, a route surface report and share cards.
- M2, the learning loop: mark calls right or wrong and refit the weights in your browser.
- M3, community weights: built and switched off until there are walkers to share marks.
- M4, Now: today's snow and soil, and the newest satellite pass.
- M5, Read the ground: land cover outside the US, go / slow / no-go on foot, ATV and truck, and area mode.
- M6, History and scale: how a spot changed over the years, and up to 1,000 points at once.
That completes the roadmap as planned. Each item, what "done" meant and what fell short is in docs/roadmap.md.
The most useful thing you can do is report a wrong call: it's one click from the Export menu. Code, data-source ideas and fixes are welcome too; see CONTRIBUTING.md. Every data source has to be free, keyless and readable from a browser.
Marks are kept in your browser (IndexedDB) and nowhere else, and so is a batch of points with its answers: nothing about it leaves except the requests each point needs. Marks ▸ Share is how they leave it, a batch at a time and only when you choose. The community store isn't switched on yet; until it is, Share saves the batch as a file, exactly as it would be sent. How the store works and how it gets switched on: docs/community.md.
| A shared mark carries | It never carries |
|---|---|
| what each source said there (11 sources × 12 classes, log-likelihoods to 4 places) | the coordinates, unless you tick Also share the exact points (then to about a metre) |
| whether you said right or wrong, what was really there, and Underfoot's call with its probability | the place name |
| how you know (standing here, a photo, the imagery, local knowledge), if you said | the link to the sounding |
| the 1° cell it's in (N37W120 is 37–38°N, 119–120°W) | the day and time |
| the month, the app version and the prior in use | not-sure marks |
| the mark's random id, and a random id for your browser so the fit can count one person as one |
The code is MIT-licensed. The data each source returns stays under its provider's terms, and the app credits them on the map.
- Map data © OpenStreetMap contributors (ODbL), served by OpenFreeMap
- Imagery © Esri, Maxar, Earthstar Geographics and the GIS user community, under Esri's terms of use
- USA Structures: FEMA · NLCD: MRLC consortium / USGS · 3DEP: USGS
- Today's weather and the elevation fallback: Open-Meteo (CC BY 4.0; the free tier is for non-commercial use)
- Elevation outside the US: Mapzen Terrain Tiles on AWS Open Data, from SRTM and GMTED2010 (courtesy of the U.S. Geological Survey), ETOPO1 (NOAA), EU-DEM (Copernicus) and national datasets; full attribution
- The newest pass: contains modified Copernicus Sentinel data, found through Element 84's Earth Search and read from the Sentinel-2 cloud-optimised GeoTIFFs on AWS Open Data
- World cover: Sentinel-2 10m Land Use/Land Cover by Impact Observatory, Microsoft and Esri, read through Esri's image service under Esri's terms of use
- Soils: USDA NRCS Soil Survey Geographic Database (SSURGO), through Soil Data Access
- Geocoding: Nominatim (ODbL data; the app keeps to its one-request-a-second policy)
- Past imagery: Esri World Imagery Wayback (Esri, Maxar, Earthstar Geographics and the GIS user community), under Esri's terms of use
- Basemaps: Esri Dark Gray Canvas (Esri, HERE, Garmin, © OpenStreetMap contributors), © OpenTopoMap (CC BY-SA)
If you deploy Underfoot commercially, check the Esri and Open-Meteo terms first.
Before its first public release, Underfoot was a single HTML file called SOUNDING. A single probe is still called a sounding.
MIT © 2026 dogum