Palimpsest — a land-cover time machine on real Annual NLCD
The site you walk onto is one frame of a landscape mid-transformation. Pick a spot and scrub 1985–2025 of real 30 m land cover — the ground’s pixels flip class year by year — and read how much “normal” actually moved. Try the predict-first version ↗
1 · Choose a site
Start on one of these starter teaching sites — real places where 30 m land cover changed dramatically since 1985 — or drop your own pin / search an address (conterminous US only).
You can also click the map to drop a pin.
Sources & methods
Land cover: Annual NLCD Collection 1.2 (CONUS, 30 m, 1985–2025) — the keyless USGS/MRLC WMS served from the USGS EROS GeoServer (dmsdata.cr.usgs.gov). Years are a single layer’s TIME dimension; the map reveal is WMS GetMap raster tiles and the grade is WMS GetFeatureInfo point queries. Both send Access-Control-Allow-Origin: *, so the browser calls them directly — no proxy.
Trajectory: the land cover is sampled on a 3×3 point grid at 5-year steps (1985, 1990, … 2020, 2025), ~72 queries, each degrading to null on failure, then folded into six-bucket shares per year. In the opt-in ?predict=1 variant, your painted bucket, two direction bets and three chips are graded against those shares into a confidence-weighted surprise score, 0–100.
Honest limits. NLCD is 30 m — it cannot see a single tree, a rain garden or a narrow creek, so pixel-scale claims are coarse by construction. The canopy bet is graded from the forest-class share (labelled “canopy ≈ forest-class share at 30 m”), not a per-year tree-canopy service — no “22%→6%” precision we can’t source. Sites below ~0.5 ha abstain (below the record). Grading floors at 1985(NLCD’s first year); there are no pre-1985 aerials and no AI historian in this version. Conterminous US only.