TreeSatAI-CHM

Extension of the TreeSatAI benchmark with canopy height model (CHM) labels derived from German LiDAR airborne surveys (LGLN). Supports joint prediction of tree species (15 genera) and canopy height from Sentinel-1 SAR. Study area: Lower Saxony, Germany.

Code and documentation: GitHub

Files

treesatai-chm/
├── training-data-60m/      # NumPy arrays, ready to use (33 MB)
│   ├── train_x.npy         # (40292, 4, 6, 6) float32
│   ├── train_y.npy         # (40292, 15) float32 — multi-hot species labels
│   ├── val_x.npy / val_y.npy
│   ├── test_x.npy / test_y.npy
│   ├── *_filenames.npy     # original patch filenames
│   └── classes.npy         # 15 class names
├── chm/                    # Raw 1 m CHM GeoTIFFs (13 GB, 6,810 tiles)
├── chm_60m/                # CHM resampled to 60 m (13 GB, 57,191 tiles)
└── lgln-opengeodata/       # LGLN tile index GeoJSONs (49 MB)

Input channels (*_x.npy)

Ch	Description	Unit	Range
0	Sentinel-1 VV	dB	−21.7 to +8.7
1	Sentinel-1 VH	dB	−32.2 to +0.7
2	VV/VH ratio	dB	−9.1 to +7.1
3	CHM	metres	0–44.7 (nodata = −9999)

Splits: Train 40,292 / Val 5,038 / Test 5,043 patches.

Classes: Abies Acer Alnus Betula Cleared Fagus Fraxinus Larix Picea Pinus Populus Prunus Pseudotsuga Quercus Tilia

Data sources

Component	Source	Resolution	Vintage
SAR patches	TreeSatAI	10 m/px, 60 m patch	2017–2019
Species labels	Lower Saxony forest records	patch-level	2017–2019
DSM (DOM1)	LGLN	1 m	2020
DTM (DGM1)	LGLN	1 m	2016
CHM	DSM − DTM	1 m → 60 m	—

Usage

import numpy as np

X_train = np.load("training-data-60m/train_x.npy")  # (40292, 4, 6, 6)
y_train = np.load("training-data-60m/train_y.npy")  # (40292, 15)
classes = np.load("training-data-60m/classes.npy", allow_pickle=True)

chm = np.where(X_train[:, 3] == -9999, np.nan, X_train[:, 3])  # metres

from huggingface_hub import hf_hub_download
import numpy as np

X_train = np.load(hf_hub_download(
    repo_id="siyux1927/treesatai-chm",
    filename="training-data-60m/train_x.npy",
    repo_type="dataset"
))

Citation

@article{ahlswede2023treesatai,
  title   = {TreeSatAI Benchmark Archive: A Multi-Source Dataset for Tree Species Classification in Remote Sensing},
  author  = {Ahlswede, Steve and others},
  journal = {Earth System Science Data},
  year    = {2023}
}

LGLN data: Landesamt für Geoinformation und Landesvermessung Niedersachsen, https://opengeodata.lgln.niedersachsen.de/

License

• Training arrays & CHM tiles: CC BY 4.0
• LGLN GeoJSONs: Datenlizenz Deutschland – Namensnennung 2.0