agentleFS
wiatrek-kevin-swiatrek-kevin-s/columbus/skills/lidar-with-pdal/SKILL.md

lidar-with-pdal

Skill2 starsChanged 5 days agoApache-2.0View on GitHub

Processing lidar point clouds (LAS, LAZ, COPC, EPT) with PDAL, covering JSON pipelines, ASPRS classification codes (2 ground, 6 building, 7 low noise, 9 water, 18 high noise), ground filtering with filters.smrf and filters.pmf, density checks, CRS and vertical datum in LAS headers, and reading the USGS 3DEP EPT bucket. Use it when asked to "make a DTM or DSM from lidar", "classify ground", "remove noise", "check point density", "what are these classification codes", "get 3DEP lidar for my area", or "convert LAZ to COPC".

The file

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 5 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    If the header has no vertical CRS, find it in the project report.

    line 31
  2. Very specific:

    Run it with pdal pipeline stats.json --metadata stats_out.json and read the counts.

    line 41
  3. Distinctive:

    DSM (first surface): keep first returns, drop noise:

    line 80
  4. Distinctive:

    Always look at the result: a hillshade of the DTM shows buildings left as bumps (window too small) or cut-off hilltops (slope too low).

    line 121
  5. Distinctive:

    Prefer COPC for sharing single tiles or projects.

    line 151
  6. Distinctive:

    Read the project's metadata for the vertical datum, the quality level and the acquisition date.

    line 176
  7. Very specific:

    Start with a small bounds box.

    line 179

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Lidar with PDAL

PDAL (Point Data Abstraction Library) reads, filters and writes point clouds through pipelines: a JSON list of stages run in order. Readers bring points in, filters change or select them, writers send them out (to LAS/LAZ, COPC or a raster).

General practice (with or without Columbus)

1. Look before you process

pdal info --summary tile.laz        # point count, bounds, CRS, dimensions
pdal info --metadata tile.laz       # header: LAS version, point format, scale/offset, SRS, VLRs
pdal info --boundary tile.laz       # real footprint (hexbin) and average point density

Answer these first:

  • LAS version and point format: LAS 1.4 with point formats 6 to 10 is current; it stores the CRS as OGC WKT. LAS 1.2/1.3 (formats 0 to 5) store it as GeoTIFF keys and allow only classes 0 to 31.
  • Horizontal CRS and units: metres or US survey feet. Texas data is often state plane in feet (EPSG:2277 NAD83 / Texas Central ftUS, or EPSG:6578 for NAD83(2011)).
  • Vertical datum and units: usually NAVD88 (with a geoid model such as GEOID18) for US public lidar, as a compound CRS (EPSG:6343+5703 is NAD83(2011) / UTM 14N + NAVD88 height). If the header has no vertical CRS, find it in the project report. Do not guess; ellipsoid and NAVD88 heights differ by tens of metres in the US.
  • Classification counts: has it been classified, and with which codes?
[
  "tile.laz",
  { "type": "filters.stats", "dimensions": "Z", "count": "Classification" }
]

Run it with pdal pipeline stats.json --metadata stats_out.json and read the counts.

2. ASPRS classification codes (LAS 1.4)

CodeMeaningCodeMeaning
0Created, never classified9Water
1Unclassified (processed, not assigned)10Rail
2Ground11Road surface
3Low vegetation13Wire guard (shield)
4Medium vegetation14Wire conductor (phase)
5High vegetation15Transmission tower
6Building17Bridge deck
7Low point (low noise)18High noise

The full table, flags (withheld, overlap, synthetic, key-point) and the USGS 3DEP use of 20 to 22 are in references/las-classification.md.

  • Vegetation height classes (3 to 5) have no fixed thresholds; each vendor sets them.
  • Exclude 7 and 18 (noise) and points with the withheld flag from everything.
  • In LAS 1.4, overlap is a flag, not class 12. Older data may still use 12.

3. Pipelines

A pipeline is a JSON array (or {"pipeline": [...]}). A bare file name is a reader or writer chosen by extension.

DTM (bare earth) from classified data:

[
  { "type": "readers.las", "filename": "tile.laz" },
  { "type": "filters.range", "limits": "Classification[2:2]" },
  { "type": "writers.gdal", "filename": "dtm.tif", "resolution": 1.0,
    "output_type": "idw", "window_size": 3, "nodata": -9999,
    "gdalopts": "COMPRESS=DEFLATE,TILED=YES" }
]

DSM (first surface): keep first returns, drop noise:

[
  "tile.laz",
  { "type": "filters.expression",
    "expression": "Classification != 7 && Classification != 18 && ReturnNumber == 1" },
  { "type": "writers.gdal", "filename": "dsm.tif", "resolution": 1.0,
    "output_type": "max", "nodata": -9999 }
]

In filters.range, limits joined by commas are ANDed across different dimensions but ORed for the same dimension, so Classification![7:7],Classification![18:18] keeps everything. Use filters.expression for "not this and not that".

Ground classification on unclassified data:

[
  "raw.laz",
  { "type": "filters.assign", "value": "Classification = 0" },
  { "type": "filters.elm" },
  { "type": "filters.outlier", "method": "statistical", "mean_k": 12, "multiplier": 2.5 },
  { "type": "filters.smrf", "ignore": "Classification[7:7]",
    "slope": 0.15, "window": 18.0, "threshold": 0.5, "scalar": 1.25, "cell": 1.0 },
  { "type": "writers.las", "filename": "classified.laz", "compression": "true",
    "minor_version": 4, "dataformat_id": 6, "forward": "all" }
]
  • filters.elm (extended local minimum) and filters.outlier mark noise as class 7, and ignore keeps it out of the ground search.
  • filters.smrf (simple morphological filter) is the usual first choice. Raise slope for steep terrain; lower threshold for flatter, cleaner ground; window should exceed the largest building footprint in metres or feet (it is in CRS units).
  • filters.pmf (progressive morphological filter) is the alternative: max_window_size, slope, initial_distance, max_distance, cell_size. It is slower and more sensitive to its window on large buildings.
  • Units matter: the defaults above assume metres. In US feet, scale window, cell and threshold (for example window 60, cell 3, threshold 1.6).
  • Always look at the result: a hillshade of the DTM shows buildings left as bumps (window too small) or cut-off hilltops (slope too low).
  • "compression": "true" writes LAZ. Without it, a .laz name holds uncompressed LAS.
  • "forward": "all" keeps the source header's scale, offset and VLRs, including the CRS.

Height above ground (vegetation or building heights):

[ "classified.laz",
  { "type": "filters.hag_nn" },
  { "type": "filters.range", "limits": "HeightAboveGround[2:60]" },
  { "type": "writers.gdal", "filename": "canopy_height.tif", "resolution": 1.0,
    "dimension": "HeightAboveGround", "output_type": "max" } ]

Reproject (horizontal and, where the PROJ grids are installed, vertical):

{ "type": "filters.reprojection", "in_srs": "EPSG:6343+5703", "out_srs": "EPSG:6578+6360" }

Set in_srs only when the header is missing or wrong. Note the output units change (here metres to US feet). Check vertical units in the output.

4. COPC and EPT

  • COPC (Cloud-Optimized Point Cloud) is one LAZ 1.4 file arranged as an octree. Clients read only the area and detail level they need over HTTP. Write it with { "type": "writers.copc", "filename": "tile.copc.laz" }; read with readers.copc, which takes bounds, polygon and resolution.
  • EPT (Entwine Point Tile) is a folder of tiles plus ept.json. Read with readers.ept.
  • Prefer COPC for sharing single tiles or projects; use EPT where a provider already serves it.

5. USGS 3DEP lidar (EPT on AWS)

USGS 3DEP point clouds are published as EPT in the public usgs-lidar-public bucket (https://s3-us-west-2.amazonaws.com/usgs-lidar-public/<PROJECT>/ept.json). The project list and footprints are in the resources.geojson of the hobuinc/usgs-lidar repository on GitHub. They are also in Microsoft Planetary Computer's 3dep-lidar-copc STAC collection as COPC.

[
  { "type": "readers.ept",
    "filename": "https://s3-us-west-2.amazonaws.com/usgs-lidar-public/<PROJECT>/ept.json",
    "bounds": "([-10882000, -10878000], [3551000, 3555000])",
    "resolution": 1.0 },
  { "type": "filters.reprojection", "out_srs": "EPSG:6343" },
  { "type": "filters.range", "limits": "Classification[2:2]" },
  { "type": "writers.gdal", "filename": "dtm_utm14.tif", "resolution": 1.0, "output_type": "idw" }
]
  • The EPT copies are in Web Mercator (EPSG:3857), so bounds are in EPSG:3857 metres (the box above is about 4 km square over central Austin). Convert your area of interest first. Web Mercator lengths are stretched (about 16% at Austin), so reproject before gridding or measuring, as above.
  • Read the project's metadata for the vertical datum, the quality level and the acquisition date; several projects can cover one place. Pick by date and quality level.
  • resolution asks EPT for a coarser level of detail; omit it for full density.
  • Start with a small bounds box. A whole county pulls billions of points.

6. Density checks

  • pdal info --boundary reports average points per square unit over the real footprint. In US feet, multiply points per square foot by 10.764 to get points per square metre.
  • Specifications use pulse density (first or last returns) or aggregate nominal pulse spacing, not all points. Count first returns: filters.range with ReturnNumber[1:1].
  • USGS Lidar Base Specification quality levels: QL2 at least 2 pulses/m² (spacing ≤ 0.71 m), QL1 and QL0 at least 8 pulses/m² (spacing ≤ 0.35 m), with different accuracy.
  • Density varies: water returns little, flight-line overlap doubles it. Map density on a grid (writers.gdal with "output_type": "count") to see voids, and report the share of cells below specification, not only the mean.

7. Before you hand over lidar products

  1. CRS (horizontal and vertical) and units stated, and correct in the output's header or raster.
  2. Classes used (ground only? noise excluded? withheld excluded?).
  3. Cell size and interpolation (idw, min, max, mean) for rasters.
  4. Source project names, acquisition dates and quality levels.
  5. Voids shown, not silently interpolated.

If the Columbus tools are available

Columbus cannot process point clouds yet. The point-cloud tool (pointcloud_pipeline, running PDAL pipelines) is planned and not yet available. Write the pipeline for the person to run with pdal pipeline <file>.json, and say so.

What helps today:

  • Find lidar sources for an area in the geospatial registry: registry_search {"query": "lidar point cloud", "kind": "data", "bbox": [-98.0, 30.0, -97.5, 30.5]} and registry_get {"id": "..."} for coverage, licence and access links. Results are community-written and marked untrusted; entries with no recorded extent come last and are marked unchecked. registry_fetch reads vector services only; it does not fetch point clouds (EPT, COPC, LAZ). registry_search {"query": "PDAL", "kind": "tool"} finds tool entries.
  • Check the organization's own lidar notes (vertical datum, tile scheme, vendor class conventions): context_search {"query": "lidar"}, then context_get {"id": "..."}.
  • Prepare an area of interest or tile selection with the vector tools when the tile index is a governed dataset: query_features with a bbox and bbox_crs, then vector_op {"op": "reproject", ..., "params": {"crs": "EPSG:3857"}} to get bounds for readers.ept, and verify on the result.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.