Taiwan tree cover 1984–2026 (tw-tree)

Tree fraction (0–100 %) per 30 m Landsat pixel for each year, with per-pixel provenance back to the single raw Landsat scene each value came from. Data for the 3D viewer 臺灣樹冠時光機 (https://chenhunghan.github.io/tw-tree/).

Releases:

• taiwan/ — island-wide v3, normalised (2026-09-29). 761 tiles in UTM zones 51 and 50 covering the main island, Penghu, Kinmen and Matsu (36,319 km² of land; 3.57 M ha of land pixels after the water mask), with a per tile-year reflectance normalisation that removes most of the drift over time (see Method). Code: https://github.com/chenhunghan/tw-tree.
• taiwan_raw/ — the same release without the normalisation (the first complete v3 publish). Provenance tiles are byte-identical to taiwan/; only f<year>, the overview and the totals differ.
• pilot/ — Taipei pilot v2 (20 tiles in zone 51, sensor-calibrated). Kept unchanged.

Layout (taiwan/ adds own, land and overview.arrow.gz)

taiwan/index.json                 tiles, years, scene counts per zone, model, island totals (land pixels only)
taiwan/overview.arrow.gz          480 m block means per tile: zone, i, j, x_utm, y_utm, z_m, own, land, f<year>
taiwan/summary.parquet            per tile x year: owned/land/valid pixels, observed and gap-filled tree hectares
taiwan/<zone>/<i>_<j>.frac.arrow.gz   x_utm, y_utm, z_m, own, land, built, bh, bs, f<year>
taiwan/<zone>/<i>_<j>.prov.arrow.gz   x_utm, y_utm, s<year>, n<year>

• own: 1 = canonical pixel for this location. Where a zone-50 tile overlaps a zone-51 tile along the zone boundary, the zone-50 pixels are own = 0 (use the zone-51 tile). Every location is counted once.
• built, bh, bs (display context, resampled onto the grid, not measured by this project): first year the pixel was impervious minus 1900 from GISA 1972–2021 (72 = by 1972, 78 = 1978–84, 0 = never); building height in metres (JRC GHSL GHS-BUILT-H 2018, 100 m); built-up surface share % (GHS-BUILT-S 2018, 10 m averaged). The viewer draws illustrative buildings from them. The overview adds bs (mean), built (median first-built year) and built_share.
• land: 1 = land (ESA WorldCover 2021 permanent water < 50 % of the pixel). f is still stored for water pixels (it is the model's raw output), but totals and the viewer use land pixels only.

Pilot layout:

pilot/index.json                  tiles, years, per-year scene counts, model metrics
pilot/summary.parquet             per tile x year: valid pixels, tree hectares, scenes
pilot/51/<i>_<j>.frac.arrow.gz    x_utm, y_utm (int32), z_m (int16), f<year> (uint8)
pilot/51/<i>_<j>.prov.arrow.gz    x_utm, y_utm, s<year> (uint16), n<year> (uint8)

Each file is an Arrow IPC file, gzipped as a whole (so the browser can decompress it with DecompressionStream). Schema metadata key tpetree holds a JSON description.

• x_utm, y_utm: exact pixel centre in EPSG:32651 (zone 51) or EPSG:32650 (zone 50). Values are multiples of 30 m on the native Landsat Collection 2 grid; no resampling.
• f<year>: tree fraction percent, 255 = no clear observation that year. In taiwan/ the frac metadata's normalisation.params[year] holds that tile-year's per-band gain ×6 and offset ×6 (blue, green, red, nir, swir1, swir2), applied as sr · gain + offset to the calibrated reflectance before the model.
• s<year>: index into scenes[year] in the provenance metadata (Landsat product IDs), 65535 = none. The raw pixel is col = (x_utm − 15 − scene_ulx) / 30, row = (scene_uly − y_utm − 15) / 30.
• n<year>: number of distinct clear acquisition dates in that year.
• z_m: Copernicus DEM GLO-30, bilinear to the pixel centre (display only).

Tiles are 256 × 256 pixels. Tile (i, j) covers x ∈ [15 + 7680·i, 15 + 7680·(i+1)), y ∈ [15 + 7680·j, 15 + 7680·(j+1)).

import gzip, pyarrow.ipc as ipc, polars as pl
t = ipc.open_file(gzip.open("pilot/51/46_360.frac.arrow.gz").read()).read_all()
df = pl.from_arrow(t)            # one row per pixel, one column per year

Verification

taiwan/ (v3): Earth Engine returns the raw Collection 2 DN of each pixel's medoid observation; calibration, features and the model run locally. For the densest tile-years (2022–2023 around 23.9° N, up to ~150 scenes), Earth Engine ran out of memory, so 173 tile-years were fetched as reflectance and inverted to the identical DN, and 34 had their medoid picked locally from the per-date raw DN (checked against Earth Engine on 40- and 122-scene tile-years: scene index, DN and date count identical for 100 % of pixels). dn_source in the build cache records the path.

Normalised release (taiwan/, 2026-09-29), recomputing from the Planetary Computer DN with the calibration and the normalisation recorded in the tile metadata: 60/60 integer raw pixels and clear, DN identical in 59/59 same-processing cases, tree fraction 59/59 exact in those cases (the 60th, a 2023 USGS reprocessing with different DN on Planetary Computer, is off by 1 point).

Complete island release before normalisation (now taiwan_raw/, 2026-09-29): 60 random stored pixels gave an integer raw pixel in 60/60, clear in QA_PIXEL in 60/60, raw DN identical to Planetary Computer in 59/59 same-processing cases (the 60th is a 2023 USGS reprocessing), and the recomputed tree fraction matched 60/60 exactly. A further 20 pixels drawn only from the locally picked tile-years: 20/20 integer positions and clear, DN identical in 19/19 same-processing cases, tree fraction 20/20 exact (5 products not mirrored on Planetary Computer were skipped).

First 8 tiles (2026-09-27): on 60 random stored pixels, the position formula gave an integer raw pixel in 60/60 and the pixel was clear in QA_PIXEL in 60/60. The six raw DN read from the independent Planetary Computer copy were identical to the DN Earth Engine returned in 59/59 cases with the same processing version (the 60th is a Landsat 9 scene USGS reprocessed in 2023; Planetary Computer serves the 2022 processing). Recomputed tree fraction: 59/60 exact, 1 off by 1 point (that reprocessed scene). Before the switch, a pilot tile was exported both ways: medoid scene index identical for 100 % of pixels, local tree fraction within 1 point of Earth Engine's (99.8–100 % exact).

pilot/ (v2):

verify_provenance.py (v2) checked 60 random stored pixels (25 TM, 25 ETM+, 6 OLI, 3 OLI-2, 1 Landsat 4) against the raw USGS files on Microsoft Planetary Computer, an independent copy of the data. For 60/60 the position formula gives an integer raw pixel, and that pixel is clear in QA_PIXEL. Recomputing locally from the raw DN with the recorded calibration gave the same tree fraction in 59/60 cases. The remaining case differed by 2 points, which fits float rounding at a single split of the 60-tree model.

Method

• Landsat 4/5/7/8/9 Collection 2 Level-2, Tier 1 only, processed in Google Earth Engine.
• Per year: clouds, shadows, snow and saturated pixels are masked (QA_PIXEL, QA_RADSAT), and same-date overlapping scenes are mosaicked. Then a medoid composite is built: for each pixel, the one real observation closest to the per-band median. That is why every value traces back to one scene.
• Cross-sensor calibration: LT04/LT05 use a TM→ETM+ reduced-major-axis fit (same-year per-sensor medians, northern Taiwan, 2000–2009 overlap years) followed by Roy et al. (2016) ETM+→OLI. LE07 uses Roy et al. (2016). LC08 is the reference. LC09 is unchanged (no measured bias). Coefficients: pipeline/model/sensor_calibration.json; each tile's metadata records them.
• taiwan/ model rf_tw_2021: random forest (scikit-learn, exported as Earth Engine tree strings; the rebuild matches to 2e-16) trained on 2021 all-sensor composites at 22,048 points stratified by WorldCover tree-fraction bin × elevation band in every tile. Held-out (3 km blocks) agreement with WorldCover 2021: R² 0.46, MAE 21 points, bias −0.3 points. This is lower than the pilot's northern figure because the island sample includes many mountain grassland/bamboo pixels. A multi-year (2014–2025) model was tested and was no more stable over time, so the single-year model is used.
• Normalisation (taiwan/): stable closed forest (Hansen tree cover 2000 ≥ 80 %, no loss 2001–2025, WorldCover 2021 tree ≥ 95 %, never built) should read the same every year, but its composite reflectance darkens over time in red and SWIR (most in the north after 2013) and the model read that as more trees (80 % in 1987–98 → 91 % in 2022–26). For each tile-year and band, a line sr' = gain · sr + offset maps the median reflectance of two anchor sets, stable forest and stable open land (Hansen 2000 and WorldCover 2021 tree < 10 %, never built), onto the same tile's 2021 medians (the model's training year). Medians are smoothed within each UTM zone (Gaussian σ ≈ 23 km); tiles with too few anchors borrow from a wider neighbourhood (up to ~90 km); Penghu, Kinmen, Matsu and a few islet tiles have no anchors nearby and are not normalised. A forest-only correction was tested and rejected: it flattened forest but pushed stable open land and partial forest down by ~5–10 points. Parameters: pipeline/model/normalisation_taiwan.json, and each frac tile's metadata.
• pilot/ model: trained on Landsat 8 only (2021), the reference sensor.
• A random-forest regression on 6 bands plus NDVI, NDMI, NBR and NDWI, trained on 2021 composites against ESA WorldCover 2021 (tree class averaged to 30 m). On held-out 3 km blocks it agrees with WorldCover at R² ≈ 0.71 and MAE ≈ 13 points. That measures agreement with the reference; it is not an independent accuracy assessment.
• Code: the pipeline/ scripts (train_model.py, export_tiles.py, verify_provenance.py).

Limitations

Drift over time (normalised in taiwan/). On a 4,000-pixel sample per tile, from 1987–98 to 2022–26: all land 62.4 → 69.6 % raw vs 63.8 → 66.5 % normalised; stable forest 80 → 91 vs 84 → 88; partial forest (not used in the fit) 57 → 57 vs 57 → 54; long-built pixels 18 → 11 vs 17 → 11. Mean year-to-year change of the all-land mean: 1.7 → 0.8 points. Against Dynamic World (Sentinel-2, 2016–2025 tile means), the island slope is +0.04 points/yr; raw +0.39, normalised +0.08. Trade-off: raw's island-wide year-to-year wiggles follow Dynamic World's in direction (r 0.83, ~5× larger), and the normalisation removes them, so single-year changes of a few points over a region are not meaningful in either release. Stable forest still reads ~4 points lower in 1987–98 than after 2014.

The post-2013 rise (investigated before the normalisation; describes taiwan_raw/ and pilot/). The pilot total rose 54 % → ~64 % between 2013 and 2021. It is mostly a measurement effect, not new canopy: a Landsat-8-only rebuild keeps the trend (so it is not a sensor switch), but stable closed forest (Hansen tree cover 2000 ≥ 80 %, no loss, WorldCover ≥ 0.95) rises just as much (77 % → 99 %), tracking NDVI and inversely the L8 aerosol QA level, while Sentinel-2 Dynamic World stays flat on the same pixels. Island-wide, only northern Taiwan's forest drifts this way (+8 points 2014–2021); central/southern forest is flat. Treat northern trends smaller than ~8 points after 2013 as unreliable. Island-wide, stable forest also reads ~4–5 points lower in the TM/ETM+ era (2000–2010) than in the OLI era, and steps up ~3–4 points in 2022 and 2024–25.

Sensor transitions (v2): v1 used Roy et al. (2016) coefficients alone, and yearly totals stepped at each sensor change. Measuring the same pixels in the same year showed that Landsat 5 TM read about 4.9 points lower than Landsat 7 ETM+. v2 adds a TM→ETM+ correction fitted on northern Taiwan overlap years, which brings that gap to −0.4 points. Pilot-area means for 1987–98 / 2000–11 / 2013–21 / 2022–26: v1 59,947 / 66,417 / 73,096 / 76,185 ha; v2 62,345 / 65,009 / 70,402 / 74,594 ha. The step around 1999 shrank from +6,470 ha to +2,660 ha. The rise after 2013 happens gradually within the Landsat 8 era, so it is not a sensor switch. It may be real regrowth or another effect (for example, Landsat 7's shrinking share in the composites) and has not been independently verified.

• Island totals (gap-filled land tree area, share of 3.57 M ha): taiwan/ (normalised) 66.7 % for 1987–98, 68.3 % for 2000–12, 69.1 % for 2014–21 and 69.3 % for 2022–26 (2021 70.3 %, 2022 70.5 %). taiwan_raw/ 65.2 %, 65.7 %, 70.0 %, 72.4 %, with a single-year high of 75.3 % in 2022; that post-2013 increase is largely the drift described above. 1984 and 1986 have few scenes and rely on gap filling.
• No Tier-1 scenes over Taipei in 1982, 1983 or 1985. 1984 and 1986 have only 3 scenes each (low confidence).
• Raw geolocation is about 12 m RMSE, so there can be sub-pixel shifts between years.
• Same-year residual differences between sensors after calibration are within about ±3 points in a given year. Landsat 7 has SLC-off gaps from 2003.
• The model is trained on one year and one reference. Applying it to the 1980s–1990s assumes the spectral relationships held then.
• A full-year composite mixes seasons, so there is some phenology noise. Most street trees are smaller than a 30 m pixel.

Attribution

• Landsat Collection 2 Level-2 courtesy of the U.S. Geological Survey.
• ESA WorldCover 2021 v200 © ESA WorldCover project, CC BY 4.0.
• Copernicus DEM GLO-30 © DLR e.V. 2010–2014 and © Airbus Defence and Space GmbH 2014–2018, provided under COPERNICUS by the European Union and ESA.
• Hansen, M. C. et al. Global Forest Change 2000–2025 v1.13 (University of Maryland), CC BY 4.0 (stable-forest checks).
• GISA 1972–2021: Ren, H., Huang, X., Yang, J., Zhou, G. (2025), ISPRS J. Photogramm. Remote Sens. 220, 354–376. CC BY 4.0.
• JRC GHSL P2023A GHS-BUILT-H and GHS-BUILT-S (2018), European Commission, Joint Research Centre.

This dataset (derived tree fraction, provenance and summaries) is released under CC BY 4.0.