# Data-access probe for steps 2–3 (Colombia)

Probed on 2026-10-03 from the project server, with no credentials unless
stated. "Remote window" means a Colombia subset can be read over HTTP range
requests (GDAL `/vsicurl/`, `/vsizip//vsicurl/`) without downloading the global
file. "Login: yes, available" means the source needs an account that Anne
already has (NASA Earthdata or CDSE). Access was tested with netrc-based curl
range requests; no credential contents were read or printed. Licences marked
"(not re-verified)" are taken from the provider page or the task brief and were
not checked against a licence file during this probe.

Step 1 inputs (the authors' Zenodo record 10.5281/zenodo.7428804 and GADM 4.1)
are downloaded by `notebooks/01_data_download.py` and are not repeated here.

| # | Input (role) | Source / URL | Size | Licence | Login? | Remote window OK? | Notes |
|---|---|---|---|---|---|---|---|
| 1 | Fagan et al. 2022 natural-regrowth / plantation polygons (labels) | `https://gfw2-data.s3.amazonaws.com/plantations/pantropical_tree_plantation_expansion/pantropical_tree_plantation_expansion_2000_2012.gpkg` | 4.29 GB | CC BY-NC 4.0 (from brief; not re-verified) | no | **yes**: GeoPackage with R-tree per layer; Colombia read in 77 s over `/vsicurl/` | 30 layers `PredEnsRF1_Finp4v3_mergeArea_traindata_bal3passBbDisp_1…30` split by longitude, EPSG:4326, polygons. Fields include `pred3class` (regrowth / plantation / open), `conf_rank` (High / Intermediate / Low), `fin_conf`, `prob.RF`, `area4326`. Polygons intersecting the GADM Colombia polygon: **regrowth 188,921** (High 177,491; Intermediate 10,176; Low 1,254); plantation 15,514 (High 7,696; Intermediate 3,930; Low 3,888); open 6,110 (Low). Σ`area4326` of regrowth ≈ 2.80 × 10⁹ (≈ 0.28 Mha if m²; the unit is not documented in the file). |
| 2 | Hansen GFC tree cover 2000, gain, loss year | `https://storage.googleapis.com/earthenginepartners-hansen/GFC-2025-v1.13/Hansen_GFC-2025-v1.13_<layer>_<lat>_<lon>.tif` (latest found: v1.13; v1.14 → 404) | treecover2000: 7 tiles for Colombia (20N_090W, 20N_080W, 10N_090W, 10N_080W, 10N_070W, 00N_080W, 00N_070W), 67–558 MB each, ≈ 2.1 GB; gain and lossyear tiles are 20–85 MB each | CC BY 4.0 (not re-verified) | no | technically yes (1 s for a 0.1° window), but tiles are **strip-organised, not tiled/COG** (block 1 × 40000), so whole-tile download is more efficient for a country | Same 0.00025° lattice as the authors' map (tile edges on 10° multiples; authors' tiles at 5° offsets), so no resampling is needed between GFC and the authors' raster. The paper needs tree cover 2000 (training) and a 2018 tree-cover layer (prediction) whose construction is not specified (`00_paper_summary.md` gap 2). |
| 3 | ESA CCI Land Cover, 300 m, annual 1992–2015, v2.0.7 | CEDA: `https://dap.ceda.ac.uk/neodc/esacci/land_cover/data/land_cover_maps/v2.0.7/ESACCI-LC-L4-LCCS-Map-300m-P1Y-<year>-v2.0.7.tif` (also `…v2.0.7b.nc` and a 1992–2015 multiband `.tif`) | ~300 MB per year (GeoTIFF) | ESA CCI data policy, free with attribution (not re-verified) | **no** | **yes**: internally tiled (256 × 256, LZW); full Colombia window for 2000 read in 5 s | **A Copernicus Climate Data Store (CDS) account is not needed** for v2.0.7 1992–2015; CEDA serves it anonymously. CDS hosts the C3S continuation (2016+ / v2.1.1); CDSE does not host ESA CCI LC. Years ≤ 1999 needed for the step-3 "pre-2000 land cover" variant are available. |
| 4a | SoilGrids250m **v2017** (names used in the paper: OCDENS, PHIHOX) | `https://files.isric.org/soilgrids/former/2017-03-10/data/{OCDENS,PHIHOX}_M_sl{1..7}_250m_ll.tif` | ~1.6–5.0 GB per layer (OCDENS_M_sl1: 5.04 GB; PHIHOX_M_sl4: 1.60 GB) | CC BY 4.0 (not re-verified) | no | **yes, but slow**: EPSG:4326, 0.00208°, deflate, **strip-organised** (1 × 172800) with overviews; a 1° × 1° window read in 11.5 s, so a Colombia window reads ~14 % of each file | Top 30 cm needs depths sl1–sl4 (0, 5, 15, 30 cm) and trapezoidal depth weighting. That is 8 layers, ≈ 5 GB of range reads. Matches the paper's variable names exactly. |
| 4b | SoilGrids **v2.0** (`ocd`, `phh2o`, depth intervals 0–5, 5–15, 15–30 cm) | `https://files.isric.org/soilgrids/latest/data/<var>/<var>_<depth>_mean.vrt` | VRT over COG tiles | CC BY 4.0 (not re-verified) | no | **yes**: VRT of tiled COGs (128 × 128, overviews); 1° window in 3.5–4.7 s | Interrupted Goode Homolosine (needs warping). v2.0 is a different model from v2017 (deviation if used). Prefer 4a for fidelity. |
| 5 | WorldClim v2.1, 30 s, bio1–bio19 (1970–2000) | `https://geodata.ucdavis.edu/climate/worldclim/2_1/base/wc2.1_30s_bio.zip` | 10.4 GB zip, 19 GeoTIFFs | WorldClim terms (free for research and non-commercial use; not re-verified) | no | **yes**: `/vsizip//vsicurl/…/wc2.1_30s_bio_1.tif` works; Colombia window (2436 × 1824) in 53 s per variable | Each tif is strip-organised LZW. **PCA is not reproducible from the paper**: Supp. Table 1 (MOESM3) gives only SD / proportion / cumulative proportion of PC1–PC5 (SD 780.4, 362.0, 146.8, 103.1, 59.7; cumulative 0.994). No loadings, centring or scaling. Methods: R `prcomp` on 1 million random **global land** points, then `predict.prcomp`. PC1 SD of 780 implies unscaled (covariance) PCA, i.e. `prcomp` defaults (centred, `scale.=FALSE`). Re-derivation needs global sampling, so downloading the 10.4 GB zip beats remote reads. Check the re-derived SDs against Supp. Table 1. |
| 6 | Biome: RESOLVE Ecoregions 2017 (Dinerstein et al. 2017) | `https://storage.googleapis.com/teow2016/Ecoregions2017.zip` | 149 MB (shapefile) | CC BY 4.0 (not re-verified) | no | download (small) | Biomes from `BIOME_NAME` / `BIOME_NUM`: 1 = tropical and subtropical moist broadleaf, 2 = dry broadleaf, 3 = coniferous. |
| 7a | NPP, MOD17A3 annual 2000–2015 (Zhao et al. 2005, Supp. Table 2 ref. 8) | NTSG (Univ. Montana): `http://files.ntsg.umt.edu/data/NTSG_Products/MOD17/GeoTIFF/MOD17A3/GeoTIFF_30arcsec/MOD17A3_Science_NPP_<year>.tif` and `…_NPP_mean_00_15.tif` | ~385 MB per file | NTSG / NASA open (not re-verified) | no | download (30 arcsec global GeoTIFFs) | Collection 055, annual 2000–2015, 30 arcsec. A ready-made **2000–2015 mean** exists and matches the paper's description. Alternative: MOD17A3HGF v061 (LP DAAC, 500 m), Earthdata **login: yes, available** (HTTP 206 on a test granule). |
| 7b | Burned area (Artés et al. 2019, GlobFire / GWIS), monthly 2001–2017 | PANGAEA 10.1594/PANGAEA.895835; shapefiles 10.1594/PANGAEA.895898 → `https://hs.pangaea.de/Maps/MCD64A1_burnt-areas/MODIS_BA_GLOBAL_1_<m>_<yyyy>.zip` (monthly, 1.6–21 MB each) | ~200 monthly zips (≈ 2–3 GB) | CC BY-SA 4.0 (PANGAEA metadata) | no | n/a (shapefiles) | **Probe (2026-10-03): HTTP 503 on three attempts. Full run (2026-10-03, 01b): all 204 months downloaded from GlobFire, no fallback, but 161 files needed retries (433 retries, worst file 7 of 8 attempts); the step took 2 h 27 min.** Availability is real but unreliable: a rerun may take many hours or end on the MCD64A1 fallback. The derived raster the paper used (monthly average, resolution unreported) is not published. Fallbacks: MCD64A1 v061 (the MODIS product GlobFire is built on), LP DAAC, Earthdata **login: yes, available** (granules found via CMR); or GEE `JRC/GWIS/GlobFire` (needs an Earth Engine account). |
| 7c | Road density (GRIP4, Meijer et al. 2018) | `https://dataportaal.pbl.nl/data/GRIP4/GRIP4_density_total.zip` (density raster); vectors `GRIP4_Region2_vector_fgdb.zip` (South & Central America) | 3.6 MB (density); 282 MB (vectors) | open with attribution (globio.info; not re-verified) | no | download | Density raster is the 5 arcmin (~8 km) m per km² product. The paper gives no reference year (dataset compiled ~2018). |
| 7d | Protected areas (WDPA) | Current release: `https://d1gam3xoknrgr2.cloudfront.net/current/WDPA_Oct2026_Public.zip` | 2.0 GB | WDPA terms of use (non-commercial; not re-verified) | no (current) | download | The paper used the **Aug 2020** release plus a June 2017 China supplement. Historical releases are not on the public CDN and must be requested from UNEP-WCMC. Not in the final 10-variable model. |
| 7e | Distance to water (Kummu et al. 2011) | No distributed raster: the PLoS ONE paper describes its derivation from VMAP0, GLWD and WDB II, but no download is given | — | — | — | — | **Not obtainable as published.** It would have to be re-derived; HydroRIVERS (`https://data.hydrosheds.org/file/HydroRIVERS/HydroRIVERS_v10_sa_shp.zip`, 95 MB, no login) plus GLWD would be a deviation. Not in the final model. |
| 7f | Slope from SRTM 1 arc-second | SRTMGL1 v003 (LP DAAC), e.g. `…/lp-prod-protected/SRTMGL1.003/N04W075.SRTMGL1.hgt/N04W075.SRTMGL1.hgt.zip`; NASADEM_HGT v001 also on LP DAAC | on the order of 100–200 1° land tiles for Colombia (not counted), ~10–25 MB each | NASA open (not re-verified) | Earthdata **login: yes, available** (HTTP 206 on test granule) | per-tile download | No-login alternatives: NASADEM GeoTIFFs on OpenTopography S3 (`https://opentopography.s3.sdsc.edu/raster/NASADEM/NASADEM_be/NASADEM_HGT_n04w075.tif`, 19 MB, HTTP 200), or Copernicus DEM GLO-30 COGs on AWS (`https://copernicus-dem-30m.s3.amazonaws.com/…`, no login; also on CDSE). Both are deviations from SRTM. Not in the final model. |
| 7g | (socioeconomic, model comparison only) GDP / HDI, Kummu et al. 2020 | Dryad 10.5061/dryad.dk1j0 (`GDP_PPP_30arcsec_v3.nc` 316 MB, `HDI_1990_2015_v2.nc` 11 MB, …) | 482 MB total | CC0 | **Dryad download API returned 401/403 to anonymous scripted requests** | download | Downloads probably need a browser session or Dryad API token; not tested further. |
| 7h | (socioeconomic) GHS-POP R2019A | `https://jeodpp.jrc.ec.europa.eu/ftp/jrc-opendata/GHSL/GHS_POP_MT_GLOBE_R2019A/GHS_POP_E2000_GLOBE_R2019A_4326_30ss/V1-0/GHS_POP_E2000_GLOBE_R2019A_4326_30ss_V1_0.zip` | 263 MB (epoch 2000, 30 ss) | CC BY 4.0 (not re-verified) | no | download | Epochs 1975/1990/2000/2015 available. |

## Summary

- **Everything needed for the ten-variable final model is reachable without new
  accounts:** Fagan labels (remote GPKG), Hansen GFC, ESA CCI LC (CEDA,
  no CDS account), SoilGrids v2017 OCDENS/PHIHOX, WorldClim 2.1, and RESOLVE
  biomes.
- **Step-3 extras:** NPP (NTSG, no login) is fine. Burned area: GlobFire downloaded in full
  in the full run after many PANGAEA 503 retries (slow, unreliable host; MCD64A1 via Earthdata as fallback). Road density (GRIP4)
  is fine. Slope works via Earthdata, or with no login via NASADEM or Copernicus
  DEM. WDPA 2020 needs a request to UNEP-WCMC. Distance to water (Kummu 2011)
  is not distributed.
- **Re-derivation needed:** bioclim PCA loadings (only variance is published);
  tree cover 2018 construction; land-cover simplification (31 → 11 classes).