longdpkrsub/PureForest_UDAsubset

PureForest — UDA target cache

A reprocessing of a 4-species subset of PureForest (IGN Lidar HD, France) as the unlabelled target domain of a sim→real domain-adaptation pair with Central European Virtual ALS (Czechia). Geometry only, 50 × 50 m patches, voxel 0.25 m, cap 40,000 points.

This is a derivative work. All data, annotations and acquisition credit belong to IGN.

patches	35,456 (train 21,061 / val 4,683 / test 9,712)
size	12.4 GiB, 13 shards
splits	PureForest's own polygon-disjoint official split, preserved verbatim

Patches per class

split	Fagus sylvatica	Picea abies	Pinus sylvestris	Quercus robur
train	7,008	2,579	11,330	144
val	1,626	627	2,429	1
test	4,036	868	4,506	302

⚠ Quercus robur is 0.33 % of PureForest and comes from only 6 forest polygons (4 train / 1 val / 1 test). Its 302 test patches are all one forest, so any oak number is a single-site result. Prefer a 3-class run (label != 3) as the headline, or report oak's polygon support explicitly. group_id is the polygon — count polygons, not patches, when reporting diversity.

Labels use the exact species, not PureForest's 13-class grouping: class 0 ("Deciduous oak") also contains Q. petraea, Q. pubescens and Q. rubra, which do not exist in the source domain.

Licence and attribution

The source dataset is published by IGN under the Licence Ouverte / Open Licence 2.0 (etalab-2.0), which permits redistribution and derivative works provided the source is attributed. This reprocessing is distributed under the same licence.

• Source data: IGNF/PureForest — IGN
• LiDAR: Lidar HD programme (2020–2025), IGN
• Annotations: derived from BD Forêt and the French National Forest Inventory, curated by IGN photointerpreters

Cite the data paper, not this repackaging:

@misc{gaydon2024pureforest,
      title={PureForest: A Large-Scale Aerial Lidar and Aerial Imagery Dataset for Tree Species Classification in Monospecific Forests},
      author={Charles Gaydon and Floryne Roche},
      year={2024}, eprint={2404.12064}, archivePrefix={arXiv},
      url={https://arxiv.org/abs/2404.12064}, primaryClass={cs.CV}
}

Row schema

column	type	meaning
patch_id	string	unique within the domain
dataset	string	pureforest
domain	string	target
split	string	train / val / test
label	int8	0 Fagus_sylvatica · 1 Picea_abies · 2 Pinus_sylvestris · 3 Quercus_robur
species	string	exact Latin binomial — the cross-dataset join key
group_id	string	the annotation polygon (one forest); patches within it are correlated
num_points	int32	≤ 40,000, unpadded
xyz	binary	float32[num_points, 3] C-order = x, y, height-above-ground, metres

import numpy as np
from datasets import load_dataset

ds  = load_dataset("longdpkrsub/PureForest_UDAsubset", split="train")
row = ds[0]
pts = np.frombuffer(row["xyz"], dtype="<f4").reshape(row["num_points"], 3)   # [N, 3]

x/y are centred on the patch centroid; z is height above an interpolated ground surface (2 m cells, ASPRS class 2, per-cell 5th percentile, 3×3 median, nearest fill, bilinear resample), so this half shares one vertical datum with its counterpart.

Preprocessing

Ground normalization → xy centring → voxel 0.25 m (barycentre/mean reduction, matching PureForest's reference pipeline) → cap 40,000 points, uniform without replacement, seeded per patch_id so a rebuild is byte-identical. Rows are variable-length; nothing is padded.

Geometry only. The virtual-ALS simulator writes a constant 0 for intensity and number_of_returns, so intensity, per-point RGB/NIR colorization and aerial imagery are all absent from the shared feature space — a UDA method cannot use a feature the source lacks.

Point order inside a blob is spatially sorted (ascending voxel key), not random — it compresses ~13 % better that way. Any further subsampling must draw indices, never slice: p[:N] takes a spatial slab. Pointcept/LitePT is unaffected (it serializes the cloud itself).

Feeding the models

LitePT-S (Pointcept convention — variable length, cat + cumulative offset):

coord  = torch.cat(blocks)                        # [ΣN, 3] float32 metres
data   = {"coord": coord, "feat": coord, "grid_size": 0.25,
          "offset": torch.cumsum(torch.tensor([len(b) for b in blocks]), 0)}

grid_size must be 0.25 — the voxel the points were reduced on. in_channels = 3.

RandLA-Net (fixed N — pad/resample at collate, not in the cache):

N = 40,000
idx = rng.choice(len(p), N, replace=len(p) < N)    # never p[:N]
batch = np.stack([p[idx] for p in blocks])         # [B, N, 3]

KNN / pooling / upsampling indices are deliberately not precomputed: at k=16 over 4 layers they are ~10 MB per patch and they depend on the drawn sample, so they belong in the loader.

Built by tools/uda_ceva_pureforest/. Full provenance — every shard, the exact settings, per-class and per-group counts — is in cache_manifest.json and _build_config.json.