Point Clouds for Science, Gaussian Splats for Advocacy

Every capture we produce generates two primary outputs: a point cloud and a Gaussian splat. They are representations of the same forest, derived from the same measurements, published at the same time. They look superficially similar — both render a three-dimensional forest environment — and they are used in entirely different ways.

The distinction matters because choosing the wrong representation for the wrong audience is a real error. A Gaussian splat in a scientific appendix is not just aesthetically wrong; it is epistemically wrong. A raw point cloud in a public-facing embed is not just technically inconvenient; it is communicatively inert.

The Point Cloud

A point cloud is a collection of three-dimensional coordinate measurements, each one a discrete sample of a surface. In our LiDAR captures, each point has an XYZ coordinate, a return intensity value, and — when we fuse colour from the camera sensor — an RGB colour value. The point cloud for a 1,000-acre capture might contain 2 billion of these measurements.

The point cloud is what we analyze. TreeLearn runs on it. Stem identification runs on it. Canopy height models are derived from it. DBH calibration validation is done against it. When we say "this block contains 847 large-diameter stems," that claim is traceable to operations on a point cloud.

Point clouds are also what we archive. The .laz format (compressed LAS) is a well-established geospatial standard supported by every major GIS platform and scientific software environment. A .laz file from a 2023 capture can be opened in 2040 by anyone with free software. It can be cited in a paper. It can be attached to a legal filing. It can be handed to a university research group and used as primary data.

For browser visualization, we use Potree — an open-source point cloud renderer that handles billion-point clouds through a hierarchical octree structure, loading only the points visible at the current view distance and detail level. A Potree viewer is not beautiful in the way a Gaussian splat is. It is legible in a different way: you can see the gap between the measured ground and the first-return points. You can identify measurement artefacts. You can trust what you're looking at as data rather than interpretation.

The Gaussian Splat

A Gaussian splat represents the forest differently. Rather than individual measurements, it represents the surface as a field of oriented ellipsoids, each one encoding colour and opacity derived from the photographic source data. The rendering is continuous — no visible point density, no scan-line structure, no artefact gaps in the geometry. The splat of a redcedar bark surface looks like bark, not a statistical representation of bark.

That continuous, photographic quality is what makes splats effective for advocacy. A person who has never seen a LiDAR scan in their life can walk through a Gaussian splat in a web browser and understand they are inside an old-growth forest. The spatial scale is legible. The age is visible in the trunk diameter. The complexity is overwhelming in exactly the way the living forest is.

This is not decoration. The response we observe in hearing rooms where we deploy immersive splat renders is qualitatively different from the response to slides, photographs, or maps. People stop treating the forest as an abstraction. The combination of point-measured geometry and photographic texture produces something that functions like presence — you are inside the space, not looking at a document about it.

The limitation of the splat is that it is not data. You cannot query individual trees in it. You cannot measure a DBH from a rendered Gaussian. You cannot run statistical analysis on it. It is optimized for communication, not computation.

When to Use Each

For scientific publications, regulatory filings, and technical appendices: point cloud. The .laz archive, a Potree viewer link, and a methods section describing the capture parameters.

For public campaigns, donor presentations, hearing rooms, and website embeds: Gaussian splat. The .spz file, a browser-accessible viewer, and a clear statement that the visualization is derived from LiDAR capture data.

For the most effective advocacy work: both, clearly labelled, with an explicit note that the viewer is a visualization and the underlying data is published and downloadable. The splat earns the emotional response. The point cloud earns the credibility.

The forest is the same. The representation depends on who needs to understand it, and what understanding needs to produce.