lidar_rendu/docs/GROUND_CLASSIFICATION.md

Ground classification — options, benchmark and references

Status (up to date): the pipeline default is --ground-classification ign with --ign-classes sol (IGN vendor ground class 2), extracted directly with laspy (_extract_ign_ground in dtm.py), with PDAL as fallback. auto, smrf and csf remain selectable via --ground-classification. Map-triggered generation always uses ign (no classification/reconciliation setting is exposed in the map UI).

Synthesis note for choosing/improving the ground-detection algorithm. Context: IGN LiDAR HD tiles (Lambert 93), areas of strong relief / rock outcrops / dense forest where the ground is under-classified and the DTM shows large holes.

Reference tile: LHD_FXX_0999_6778_PTS_LAMB93_IGN69

• 65,047,919 points, ~1 km², resolutions 0.5 m and 0.2 m.
• Class breakdown (vendor pre-classification): class 2 (ground) 25.79%, class 5 (high vegetation) 63.2%, class 3 (low vegetation) 7.85%, class 4 (medium vegetation) 2.66%, class 1 0.47%, class 6 0.01%. No points in class 0.
• Holes in the existing DTM (before correction): 43.5% at 0.5 m, 44.1% at 0.2 m (tile extent, header bounds).

Measured benchmark (PDAL, 1 km²)

Method	Time	Ground points	Ground surface*	Holes*
IGN (vendor pre-classification)	9.4 s	25.8%	84.6%	15.4%
SMRF	326.3 s	36.1%	90.4%	9.6%
CSF	355.2 s	17.6%	46.9%	53.1%

* "Ground surface" computed over the point extent (point cloud bounding box), at 0.5 m. Final DTM hole percentages (header bounds, larger) are higher: see the reference tile above.

A. Geometric filters (current PDAL stack)

• IGN (vendor pre-classification, class 2): the fastest (~9 s). Reliable where the vendor has confidence; holes under dense forest / steep relief. No parameter to tune.
• SMRF — Pingel, Clarke & McBride 2013, ISPRS J. Photogramm. Remote Sens. 77:21-30. Raster-based filter (operates on a DSM, not on points), hence faster than point-based filters; minimizes type I errors (ground omission) → well suited when ground is scarce (forest). Best coverage of the three here (90.4%) but ~5.4 min/tile.
• CSF — Zhang et al. 2016, Remote Sensing 8(6):501. Inverted cloth draped over the point cloud; simple, accurate, but the cloth no longer touches the ground on steep/hilly terrain → poor classification. Slowest here and worst on this tile. Best reserved for urban areas.
• PTD/PTIN (Progressive TIN Densification) — Axelsson 2000, ISPRS Congress. Literature's winner: most robust on complex terrain + forest (Moudrý et al. 2020, Measurement 150:107047; Cai et al. 2019, Remote Sensing 11(9):1037) and the fastest (lidR benchmark: PTD ~20 s vs CSF ~156 s vs PMF ~1800 s). NOT available in the PDAL version bundled in this image (filters.ground / TIN missing) — would need to be added to use it (or via lidR / a custom implementation).

B. Fast hybrid (chosen for implementation)

PTD / Wack & Wimmer principle (Wack & Wimmer 2002, ISPRS Archives XXXIV/3A:293-296: DTM from lowest return, excluding the lowest 1% per cell to discard outliers):

1. Base = IGN pre-classification (class 2, ~9 s, reliable and official).
2. Measured gap filling: for each cell with no ground point, take the robust lowest return (min of the 99% of points in the cell) → adds measured ground where the vendor failed (rock outcrops, clearings, forest floor).
3. Topographic inpainting of the remaining gaps (terrain-aware interpolation already implemented in dtm.py:_interpolate_holes).

Expected: continuous DTM (0% holes), robust in forest/relief, ~10-15 s/tile instead of 326-355 s. No GPU dependency, no training.

C. AI / ML models (supervised — require labels)

Caveat (Qin et al. 2023, ISPRS J. Photogramm. Remote Sens. 202:246-261): everything is supervised; the main risk is generalization — a model trained on one region degrades elsewhere. No fully unsupervised DL filter published to date.

Point-based (3D):

Model	Year	Architecture	Accuracy	Speed (~/km², GPU)
KPConv / RandLA-Net (Qin, OpenGF)	2021	KPConv / RandLA-Net	97.8% OA, DTM RMSE 0.20 m, ground IoU 95%	0.5-2.5 min
PFCN (Jin, IEEE JSTARS 13:3958)	2020	point-FCN	Te 1.73%, Kappa 93.9%	~1/3 the cost of PointNet++
Terrain-Net (Li, Remote Sensing 14(22):5798)	2022	KPConv + self-attention	OA 98%, mIoU 0.933	parameter-free at transfer
MSVC (Štroner, Remote Sensing 17(4):615)	2025	9x9x9 voxel DNN	beats CSF on F-score	—

Rasterized (directly output the DTM — closest to our need):

Model	Year	Architecture	Result
Precursor (Rizaldy, ISPRS Annals IV-2:231)	2018	2D FCN	Te 5.22%, 78x faster
DeepTerRa / ALS2DTM (Lê, IEEE JSTARS 15:2778)	2022	GAN pix2pix (U-Net)	DTM RMSE < 1 m, filter + interpolation in one pass
DSM2DTM (Bittner, ISPRS Annals X-1/W1-2023:925)	2023	U-Net (EfficientNet)	non-ground mask + per-pixel ground height

Training datasets: OpenGF (Qin et al., CVPRW 2021, arXiv:2101.09641 — 47.7 km², 542 M pts); ALS2DTM (Lê et al. 2022, arXiv:2206.03778 — 52 km², 1.66 billion pts, urban/forest/mountain).

Costs / obstacles for our case: (1) labels → to be generated as pseudo-labels (high-quality SMRF/PTD output on a representative sample of our tiles) or pre-training on OpenGF/ALS2DTM; (2) generalization across the diverse terrain of LiDAR HD (plain/forest/mountain/urban); (3) infrastructure: checkpoint + GPU inference path in the Docker image.

Best fit if going the AI route: a rasterized U-Net (DSM2DTM-style) — rasterize the point cloud into multi-channel grids (elevation, slope, curvature, density, return statistics), output a ground mask + ground height. 2D = very fast and trivial to deploy on GPU, merges filtering + interpolation. KPConv/RandLA-Net is more accurate in pure 3D but heavier to deploy.

Synthesis / decision

• "Fast" hard constraint + low maintenance → fast hybrid (B) (~10-40 s/tile, zero training, zero GPU). ← chosen approach, IMPLEMENTED
  • Base = IGN pre-classification (fast, ~10 s). auto prefers it as soon as ≥ 20% of points are classified as ground (threshold lowered from 30% to 20%, since the DTM is subsequently completed — see below).
  • This gap-filling note is superseded: gap filling in the DTM is no longer a distance-based fillnodata pass over small holes. It is now a morphological closing bounded to the point envelope (_fill_small_gaps in dtm.py): the closing radius follows the local point spacing (measured over 5 m, staged at 1/1.5/2/3 m), nothing is extended beyond measured pixels, and islands under 1 m² are removed. Large holes (dense forest, steep relief where ground is under-classified) still remain as nodata (black in the renders). Deliberately no floor at the lowest return: under dense canopy that return is vegetation, which would print trees into the DTM.
• Maximum quality in hard cases (steep + dense), ~1-2 min/tile + GPU + training accepted → rasterized U-Net (C). (not implemented)
• Best geometric filter available in PDAL → SMRF (A) (best coverage 90.4% but 5.4 min/tile). Selectable via --ground-classification smrf.
• Literature's absolute "winner" (fast + robust) → PTD/PTIN (A): to be integrated (not in the current PDAL stack).

References

• Axelsson (2000), PTIN/PTD, ISPRS Congress.
• Pingel, Clarke & McBride (2013), SMRF, ISPRS J. P&RS 77:21-30.
• Zhang et al. (2016), CSF, Remote Sensing 8(6):501.
• Wack & Wimmer (2002), lowest-return DTM, ISPRS Archives XXXIV/3A.
• Moudrý et al. (2020), CSF/PTIN/PMF/SMRF comparison, Measurement 150:107047.
• Cai et al. (2019), CS+PTD, Remote Sensing 11(9):1037.
• Qin et al. (2021), OpenGF, CVPR Workshops (arXiv:2101.09641).
• Qin et al. (2023), dataset + evaluation + survey, ISPRS J. P&RS 202:246-261.
• Lê et al. (2022), DeepTerRa/ALS2DTM, IEEE JSTARS 15:2778 (arXiv:2206.03778).
• Bittner et al. (2023), DSM2DTM, ISPRS Annals X-1/W1-2023:925.
• lidR book (PTD/CSF/PMF comparison): https://r-lidar.github.io/lidRbook/gnd.html