Link-Dev/canopy-height-data

Canopy height data: NEON and five international sites

Training and evaluation data from our canopy height mapping study, packaged for testing. All source data are public (see Sources and licences).

Archive	Content
Canopy_height_data_NEON.tar.gz	neon_test/
Canopy_height_data_MRF.tar.gz	intl_test/MRF_stack/, intl_train/MRF_train_stack/
Canopy_height_data_EBR.tar.gz	intl_test/EBR_stack/, intl_train/EBR_train_cover400_stack/
Canopy_height_data_SER.tar.gz	intl_test/SER_stack/, intl_train/SER_train_cover400_stack/
Canopy_height_data_SPC.tar.gz	intl_test/SPC_stack/, intl_train/SPC_train_cover400_stack/
Canopy_height_data_MUR.tar.gz	intl_test/MUR_stack/, intl_train/MUR_train_cover400_stack/

All archives unpack into Canopy_height_data/ and include stats/ (per-channel mean and standard deviation of the 76-channel input). Every array is a stack of 256 × 256 chips at 10 m resolution, stored as a NumPy .npy file with the chips on the first axis. Stacks of more than 400 chips are split into part001 and part002.

International sites: MRF = Mount Richmond Forest (New Zealand), EBR = Entlebuch Biosphere Reserve (Switzerland), SER = Sepilok and Danum Valley (Malaysia), SPC = São Paulo (Brazil), MUR = Middle Usumacinta (Mexico).

Weights

weights/: trained models for chm-tool.

Folder	Files
source/	UNet-ALS.pth (input AE), UNet-A-ALS.pth, UNet-E-ALS.pth, UNet-T-ALS.pth, UNet-TE-ALS.pth
NEON/	UNet-SLS.pth, RF-SLS.joblib
MRF/, EBR/, SER/, SPC/, MUR/	UNet-SLS.pth, RF-SLS.joblib, KG-UNet1.pth, KG-UNet2.pth

Files

neon_test/: chips are ordered by site: ABBY (18), BART (30), HARV (61), LENO (25), MLBS (36), NIWO (28), SCBI (25), SJER (25), SRER (37), TALL (30), TEAK (46), WREF (43). The site-year is 2019, except NIWO (2020) and MLBS (2021).

File	Content
X_part001.npy	annual input, 76 channels
X_part001_TS.npy	seasonal input, 44 channels
y_part001.npy, chm_p95.npy, chm_p98.npy	ALS canopy height, p90 / p95 / p98
y_GEDI_part001.npy	GEDI RH95
y_eth_part001.npy, y_umd_part001.npy, tolan_p90.npy	HRCH, GFCH, GMTCH
chip_index.csv	site and biome of each chip

intl_test/<SITE>_stack/:

File	Content
<SITE>_partNNN.npy	annual input, 76 channels
<SITE>_chm_partNNN.npy, <SITE>_chm_p95_partNNN.npy, <SITE>_chm_p98_partNNN.npy	ALS canopy height, p90 / p95 / p98
<SITE>_GEDI_part001.npy (SER, SPC, MUR)	GEDI RH95
<SITE>_eth_partNNN.npy, <SITE>_umd_partNNN.npy, <SITE>_gmtch_partNNN.npy	HRCH, GFCH, GMTCH
<SITE>_index.csv (<SITE>_label_index.csv for EBR and MRF)	chip locations

intl_train/:

File	Content
<SITE>_train_cover400_stack/<SITE>_train_cover400_partNNN.npy (EBR, SER, SPC, MUR)	annual input, 76 channels
<SITE>_train_cover400_stack/<SITE>_train_cover400_GEDI_partNNN.npy	GEDI RH95 training labels
<SITE>_train_cover400_stack/<SITE>_train_cover400_index.csv	chip locations
MRF_train_stack/MRF_GEDI_partNNN.npy	GEDI RH95 training labels for MRF; MRF is trained on its evaluation chips, whose input is intl_test/MRF_stack/MRF_partNNN.npy
MRF_train_stack/MRF_train_index.csv	chip locations

Data

Annual input (uint16; 0 = no data):

Channels	Layer	Decode
0–63	Satellite Embedding, annual	value / 10000 − 1
64	SRTM elevation	m
65–66	Sentinel-1 VV, VH, annual median	value / 100 − 50 (dB)
67–75	Sentinel-2 L2A B2, B3, B4, B5, B6, B7, B8, B11, B12, annual median	value / 10000 (reflectance)

Seasonal input (X_part001_TS.npy, same encoding): channels 0–7 are Sentinel-1 VV and VH for seasons 1–4 (VV1, VH1, VV2, VH2, …); channels 8–43 are the nine Sentinel-2 bands for seasons 1–4. The seasons are December (of the previous year) to February, March–May, June–August and September–November.

ALS canopy height (float32, m; −999 = no data): the 90th, 95th and 98th percentile of the 1 m ALS canopy height within each 10 m cell.

GEDI RH95 (float32, m; −999 = no footprint): median GEDI L2A RH95 of 2019–2021 on the 10 m grid.

Global products (float32, m): umd = GFCH, eth = HRCH, gmtch / tolan_p90 = GMTCH (90th percentile of the 1 m map within each 10 m cell).

International index files: one row per chip in stack order (part, row_in_part), with the CRS (epsg or crs) and the upper-left corner of the chip (tile_x0, tile_y1 or x_ul, y_ul).

Biome codes in chip_index.csv: TRF = temperate rain forest, TSF = temperate seasonal forest, BF = boreal forest, WS = woodland/shrubland, SD = subtropical desert.

Sources and licences

Data	Source	Licence
NEON ALS	NEON, Ecosystem structure (DP3.30015.001): RELEASE-2025 https://doi.org/10.48443/jqqd-1n30; RELEASE-2024 https://doi.org/10.48443/zzz8-pr54 (SCBI); provisional data (MLBS 2021)	CC0 1.0 when obtained
SPC ALS	São Paulo City Hall (2024), OpenTopography, https://doi.org/10.5069/G9NV9GD1	GPL-3.0
MUR ALS	Inomata (2022), OpenTopography, https://doi.org/10.5069/G95B00NF	CC BY 4.0
SER ALS	Coomes & Jackson (2022), NERC EDS CEDA, https://doi.org/10.5285/dd4d20c8626f4b9d99bc14358b1b50fe	OGL v3.0
EBR ALS	swisstopo swissSURFACE3D Raster and swissALTI3D, https://www.swisstopo.admin.ch	swisstopo OGD
MRF ALS	LINZ / Marlborough District Council (2020–2022), OpenTopography, https://doi.org/10.5069/G97D2SB0	CC BY 4.0
Satellite Embedding	Google and Google DeepMind (2025), https://developers.google.com/earth-engine/datasets/catalog/GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL	CC BY 4.0
SRTM	NASA JPL (2013), https://doi.org/10.5067/MEaSUREs/SRTM/SRTMGL1.003	public domain
Sentinel-1, Sentinel-2	Copernicus Sentinel data 2018–2021	Copernicus Sentinel data terms
GEDI L2A	Dubayah et al. (2021), https://doi.org/10.5067/GEDI/GEDI02_A.002	public domain
GFCH	Potapov et al. (2021), https://doi.org/10.1016/j.rse.2020.112165	CC BY
HRCH	Lang et al. (2023), https://doi.org/10.1038/s41559-023-02206-6	CC BY 4.0
GMTCH	Tolan et al. (2024), https://doi.org/10.1016/j.rse.2023.113888	CC BY 4.0

Licence

CC BY 4.0, except the three intl_test/SPC_stack/SPC_chm_* files, which are derived from the GPL-3.0 São Paulo lidar survey and are released under GPL-3.0-only. See LICENSE and LICENSE-GPL-3.0.txt. The weights in weights/ are released under the MIT licence.