English | Français · Providers · LiDAR CLI reference · Visualization guide · Back to the main documentation

Standing structures with LAZ, DFM, and CSF

National bare-earth digital terrain models are excellent for earthworks, but they can erase the very feature being searched for when it still stands above the ground. lidar2map's point-cloud mode rebuilds a surface from the full classified LAZ/LAS source so that low standing structures can remain available to LRM, SVF, openness, VAT, and every other relief visualization.

This is a candidate-generation tool, not a wall or archaeology classifier. Vegetation, rocks, modern objects, and classification errors can return with the walls. Keep the area small, compare independent layers, and validate interpretations on the ground where access and heritage rules allow.

Why a bare-earth DTM can lose a ruin

National producers deliberately classify a terrain model as bare earth. For IGN LiDAR HD, ground/water/virtual-ground returns form the terrain base, while a wall still standing above roughly 1 m is commonly classified as vegetation or unclassified. The official DTM interpolates through the resulting gap. Once the wall has been removed from that raster, no hillshade or other visualization computed from the DTM can restore it.

This creates a useful distinction:

low banks, ditches, terraces, and heavily collapsed walls often remain in the official DTM and are best surveyed with the normal lidar2map workflow;
roofless walls and other low standing structures may require the full point cloud and a reconstructed Digital Feature Model (DFM).

The DFM concept—terrain plus archaeological standing features in one model—comes from Štular et al. (2021). lidar2map's automatic class/height selection is its own first-pass heuristic, calibrated on two sites in the Var, France. The published workflow uses more deliberate, often semi-manual reclassification. The implementation must therefore be read as prospection assistance, not as the published method reproduced exactly.

Enable point-cloud mode

In the GUI, select a provider that exposes a point-cloud mode and enable DFM mode next to it. In the CLI, select the normal parent provider and add --laz:

python lidar2map.py \
  --lidar --provider fr-ign --laz --download \
  --zone-gps <lat>,<lon> --zone-width 1 --zone-name site \
  --shadings lrm svf --file-formats mbtiles

All requested visualizations are then computed from the reconstructed point-cloud surface instead of the official DTM. DFM changes the input surface; LRM, SVF, VAT, and the other outputs still have their normal meaning. See the visualization guide when choosing how to render the result.

Start with a small zone around a known or suspected structure. A department- or region-scale point-cloud run is rarely a sensible first experiment.

The GUI exposes the class-based and CSF ground-base methods separately while keeping the same LiDAR outputs.

Two reconstruction methods

lidar2map exposes two ground-base methods under the same --laz mode. Output names distinguish them as laz_dfm (class re-injection) and laz_csf (cloth filter).

Class-based re-injection

Select --laz-ground classes. On IGN, this is the default.

The pipeline:

bins the provider's selected terrain classes to form a ground base;
fills that base temporarily to estimate each point's height above ground;
selects low non-ground points inside the hmin–hmax band;
re-injects those points only into cells where the terrain base has a hole;
performs a final interpolation bounded to 200 m and writes the DFM GeoTIFF.

For fr-ign, the default selected classes are 1,2,3,4,9,66. Classes 2,9,66 form the terrain base and 1,3,4 are candidates for re-injection between 0.4 and 2.5 m above it.

This mode is fast and preserves producer knowledge, but its result depends on the source's class semantics. One ground return inside a wall cell keeps the ground and prevents the wall from being re-injected there. Dense scrub can also return as speckle. If a wall looks incomplete, first check the provider's class scheme; on IGN, class 5 can be tested explicitly when walls were classified as high vegetation.

Cloth Simulation Filter

Select --laz-ground csf. Most non-IGN point-cloud providers use it by default because their producer classes are heterogeneous or insufficiently documented.

The Cloth Simulation Filter inverts the cloud and drapes a virtual cloth over it. lidar2map deliberately uses a soft cloth so low continuous structures can be absorbed into the reconstructed surface while canopy points are rejected. A class-independent canopy pre-filter first keeps points within 3.5 m of a 5 m local minimum; on the calibration clouds this retained about 57% of the points.

On the two Var validation sites, CSF retained an equivalent wall signal with a cleaner background than class re-injection. It is much slower, remains sensitive to terrain and density, and is not a wall classifier.

hmin, hmax, and classes are ignored in CSF mode. Conversely, csf-* settings are ignored in class mode; lidar2map reports either mismatch.

Choose	Best starting point	Main cost or risk
`classes`	Well-documented producer classes; rapid retuning; IGN	Speckle and missed walls where a ground return already occupies the cell
`csf`	Heterogeneous classes; cleaner background; most international providers	Several minutes per tile, ~3 GB RAM per conversion, parameter sensitivity

Parameters

Defaults can vary by provider and are shown by the GUI. The values below are the common defaults; fr-ign defaults to classes while most other point-cloud providers default to csf.

CLI option	Common default	Applies to	Effect
`--laz`	off	both	Switch from the official DTM to the provider's classified point cloud
`--laz-ground classes\|csf`	provider-specific	both	Select class re-injection or cloth reconstruction
`--laz-hmin M`	0.4 m	classes	Minimum candidate height above the reference ground
`--laz-hmax M`	2.5 m	classes	Maximum candidate height above the reference ground
`--laz-classes LIST`	provider-specific	classes	Comma-separated terrain and candidate LAS classes
`--laz-csf-threshold M`	0.5 m	CSF	Maximum point-to-cloth distance absorbed as ground; increasing it can retain more degraded walls and more scrub
`--laz-csf-resolution M`	0.5 m	CSF	Cloth grid size; valid range 0.1–3.0 m
`--laz-csf-rigidness N`	1	CSF	`1` steep/soft, `2` intermediate, `3` flat/rigid; `3` approaches bare earth and can erase standing walls
`--laz-parallel N`	1	conversion	Concurrent LAZ conversions; budget roughly 3 GB RAM per conversion

Threshold and cloth resolution accept 0.1–3.0 m. The solver's time step, iteration count, and canopy pre-filter are intentionally fixed: they are implementation controls rather than interpretable site parameters.

Advanced class behaviour:

omitting class 2 produces a slice-like output containing band-selected objects on a transparent background; heights are still referenced to class-2 ground;
selecting no re-injected class produces approximately a rebuilt DTM.

These are diagnostic configurations, not recommended first runs.

Time, storage, RAM, and cache

Costs depend on point density, archive format, server access, CPU, and the selected method. Measurements below are orders of magnitude, not guarantees:

Source/example	Download or density	Classes	CSF	Memory note
IGN LiDAR HD, France	~205 MB/km²; dense COPC	~20–25 s/tile	~3–4 min/tile	~2.9–3 GB peak on a ~45 M-point tile
swissSURFACE3D	~125 MB/km²	provider-dependent	around 6 min/tile	Keep the area targeted
Denmark DHM/Punktsky	~82 MB and ~12 M points/km²	~19 s/tile	~6.4 min/tile	Denser cloth simulation is slower
NRCan COPC sample	~40 points/m²	provider-dependent	provider-dependent	1 km² can mean ~1 GB temporary LAS and ~3 GB RAM

Times were measured on specific machines and datasets. A 4-core machine showed no useful gain from running several conversions concurrently because one CSF conversion already uses multiple cores. Increase --laz-parallel only on a larger VM with both spare cores and roughly N × 3 GB available RAM.

Cache behaviour

The downloaded LAZ/LAS cloud is kept in the tile cache.
The derived DFM/CSF GeoTIFF lives in production output.
The same cloud is shared across class, height-band, and CSF experiments.
Non-default settings are encoded in derived tile and project names, preventing DTM, class-DFM, and CSF outputs from being silently mixed.
Retuning rebuilds from the cached cloud without another network download.
Use --download-overwrite only when the source itself must be fetched again.

Conversions are written atomically through a temporary file, so an interrupted conversion is not accepted later as a complete GeoTIFF. Dependencies (laspy, lazrs, and CSF when requested) are checked before a heavy download and installed on demand in the managed Python environment; standalone bundles already include them.

Multi-provider scope

DFM mode requires the full, dense, classified point cloud. A bare-earth raster or ground-only cloud has already lost the standing structures and cannot support it. The current point-cloud implementations are summarized below; the provider catalogue remains authoritative for live availability, credentials, coverage, and source sizes.

Parent provider	Point-cloud notes	Default	Validation status or caution
`fr-ign`	IGN COPC, ~205 MB/km²	classes	Class and CSF methods field-checked on two Var sites
`ch-swisstopo`	swissSURFACE3D `.las.zip`, ~125 MB/km²	CSF	End-to-end conversion; field validation recommended
`pl-gugik`	~28 points/m²; PL-2000 CRS varies by zone	CSF	End-to-end validated
`ee-maaamet`	~4 points/m² in the tested standard cloud	CSF	Technically valid but marginal for a 0.5 m grid; field validation pending
`be-flanders`	~11 points/m²; OpenLidar classified cloud	CSF	End-to-end validated
`ca-nrcan`	Windowed remote COPC, up to ~40 points/m² in the test	CSF	Only the requested bbox is read; dense windows remain RAM-heavy
`ca-quebec`	Direct LAZ, ~10 points/m² in the test; multiple MTM zones	CSF	End-to-end validated; project-based coverage
`us-3dep`	Windowed Planetary Computer COPC; ~5 points/m² in the tested older survey, often denser in newer projects	CSF	No account for LAZ; project-based coverage
`dk-datafordeler`	~12 points/m²; API key required	CSF	End-to-end validated; CSF measured around 6.4 min/tile
`fr-craig`	Very dense regional campaigns, up to ~60 points/m² in the test	CSF	Named regional campaigns, not wall-to-wall France

“End-to-end validated” means discovery, download, CRS handling, conversion, and GeoTIFF output were exercised. It does not mean that archaeological recall and false-positive rates were field-validated for that country.

Visual comparison

This roofless house ruin in the Var, France has walls around 1.5 m high under scrub. The aerial image barely hints at them. The classic LRM from the official DTM shows surrounding terraces but not the ruin. Both point-cloud methods bring the rectangular footprint back; CSF produces the cleaner background.

Aerial ortho	Classic LRM from the DTM

Walls lost under vegetation	Terraces show; the ruin does not
DFM-LRM: class re-injection	DFM-LRM: CSF cloth base

Rectangular building reappears, with speckle	Same walls, cleaner background

Interpretation and field validation

Never interpret the DFM layer alone. A practical review stack is:

official-DTM LRM or another familiar bare-earth visualization;
class-DFM and/or CSF LRM using the same scale;
DFM minus DTM delta where available;
current and historical orthophotos;
SVF, openness, slope, or another independent visualization;
field observation where lawful and safe.

Read continuous lines, corners, repeated geometry, and agreement between independent views more seriously than isolated bright/dark dots. Scrub commonly appears as speckle; rocks, scarps, terraces, forestry traces, buildings, and modern debris can mimic archaeological forms.

Known limitations:

class re-injection can miss a wall when a single ground return occupies its cell;
CSF can retain vegetation or erase walls, especially with an unsuitable threshold or rigidness=3;
low-density clouds may not sample narrow masonry well enough for a 0.5 m output;
the integrated output does not yet provide measured/interpolated, density, confidence, or re-injection masks;
interpolation can bridge data gaps by up to 200 m, so a smooth shape is not proof of direct measurement;
minimum-height binning remains sensitive to an abnormally low return that was not marked as noise/withheld; a robust low-quantile alternative still needs field calibration;
cloth filtering has no cross-tile halo yet; inspect suspicious features at tile boundaries;
CSF uses OpenMP and is not guaranteed bit-identical between runs. Validate the final raster and interpretation, not raw CSF point-index lists;
acquisition year, classification scheme, and survey density vary between and sometimes within providers.

Avoid launching separate processes against the same point-cloud cache and area at the same time: derived files are atomic, but a cross-process LAZ conversion lock is not implemented yet.

The pipeline automatically rejects incompatible CRS/unit combinations where they can be resolved, filters ASPRS noise classes 7/18 and withheld points before minimum-height binning, and uses nominal tile bounds where available to reduce seams. If a ZIP unexpectedly contains several point clouds, it warns and currently keeps only the largest one.

Field validation remains the decisive test. The current heuristic needs adversarial checks on known ruins and negative areas containing rocks, cliffs, terraces, and scrub, including structures that cross several source tiles. Do not publish precise coordinates of sensitive remains, dig, or use the output to bypass archaeological and land-access rules.

Standalone QGIS comparison tool

For a targeted IGN France comparison outside the main pipeline, tools/dfm_ruines.py downloads the COPC LAZ, performs class-based reconstruction, and writes three georeferenced EPSG:2154 GeoTIFFs:

<prefix>_lrm_mnt.tif — LRM of the reconstructed ground-only DTM;
<prefix>_lrm_dfm.tif — LRM of the DFM, with standing candidates and scrub;
<prefix>_delta.tif — DFM minus DTM in metres.

python tools/dfm_ruines.py \
  --center <lon>,<lat> --rayon 150 --out site

python tools/dfm_ruines.py \
  --bbox <west>,<south>,<east>,<north> --out zone --cache laz_cache

Standalone defaults are 0.5 m output, height band 0.4–2.5 m, candidate classes 1,3,4, LRM sigma 7.5 m, and ./laz_cache. Class 6 can be included for a specific built-feature test. This script does not run CSF.

In QGIS, a useful first comparison is to display *_lrm_dfm.tif in grayscale around −0.5 to +0.5 m over the orthophoto, then threshold *_delta.tif around 0.4–1 m to inspect candidates. Adjust those display limits to the site and always compare with *_lrm_mnt.tif. Lines and rectangles are candidates; speckle is commonly scrub.

One IGN source tile is roughly 205 MB/km², so this remains a few-km² inspection tool. It uses laspy/lazrs, rasterio, SciPy, NumPy, and pyproj; the managed lidar2map environment already provides these dependencies.

Navigation Menu

Standing structures with LAZ, DFM, and CSF

Why a bare-earth DTM can lose a ruin

Enable point-cloud mode