November 29, 2022

Tips for Building and Vegetation Classification

Written by: Amanda Lind

 

Welcome back to Part 2 of the Tips for Troubleshooting Point Cloud Classification series, focusing on helping users better understand Global Mapper’s automatic classification tools and how they can be adjusted for variations in point clouds. In the first blog, we covered noise and ground classification. Next, we move to the Non-Ground Classification Tool, used for building and vegetation classification. 

First, be sure noise and ground points have been classified in the dataset. Remember from the previous blog that the classification tools are loosely made to be run in a specific order: noise, then ground, before buildings and vegetation. The order is important because part of the building and vegetation classification process includes looking for points a specific height above the ground. A poor ground classification lays poor groundwork, excuse the pun, for future classifications. 

Points classified as buildings are orange, while vegetation is green. 

There are two different classification methods in the Non-Ground Classification Tool: Segmentation and Gridding. These options assess the data through different methods and are each suited for different data structures based on how they were collected. As with any automatic classification tool, before adjusting the settings, it’s always good to run the data, or part of the data, with the default settings to see what structures of points are being missed. 

Gridded Method

The Gridded Method is the original legacy method. Lidar used to be predominantly gathered by fixed-wing aircraft, and that’s the type of data that this method is built to use. This method has been largely superseded by the segmentation method. For data collected by means other than fixed-wing aircraft, use the segmentation method. 

The Gridded Method works by looking for planar or flat building features. Basically, it carves up the data into grid cells (hence the method name) and fits the points in each cell to a plane. These planes are flat building features. Points that are far from the plane are marked as vegetation, meaning that points that aren’t roughly on the same plane, or flat surface, as their neighbors are assumed to be vegetation. 

If building points aren’t included, try increasing your Maximum Clo-Planar Distance setting. This determines how far points can be from the plane/wall/flat surface and still be counted as part of the surface. Increasing this setting may also help to keep points at the tops of buildings from being misclassified as vegetation. 

The Minimum Height Above Ground and the Minimum Vegetation Distance settings are the more important parameters for high vegetation classification. The minimum Height Above Ground sets a threshold for the start of the non-ground classification. This setting helps to weed out other ground features like low vegetation, rocks, and cars. The Minimum Vegetation Distance sets a minimum distance between points for an area to be considered vegetation. This is useful as vegetation areas tend to have more spread-out points compared to a more solid surface like a building.

If vegetation is being erroneously classified at the edges of buildings, try increasing the Minimum Vegetation Distance. This is how far a point has to be from a plane before being classified as vegetation. 

If the tool isn’t classifying dense vegetation, make sure you have enough true ground points in vegetated areas to differentiate the ground level from the vegetation. Since some of the parameters in the non-ground classification and extraction tools refer to height, you need to have classified ground points that can be used to determine the relative height of other points. 

See how these settings can affect different structures in the data in this diagram from the Knowledge Base

Segmentation Method

In addition to lidar collected via fixed wing, we now also see lidar collected from UAVs, helicopters, terrestrial scanners, photogrammetric point clouds, and more. The segmentation option was developed to better classify and work with those other point cloud sources. 

Similar to the newer stand-alone Segmentation Tool, the Segmentation Method groups point together with their neighbors based on how similar they are to each other. By looking at the points in a 3D space, as opposed to a gridded 2D space, it does a better job identifying building points near or under vegetation. 

The Neighborhood Size is the resolution. A 3.5-point spacing is typically good for most aerial lidar. For more dense terrestrial data, it’s recommended to use a foot or so instead. Keep in mind if the local neighborhood being evaluated is too small, lots of things look flat. For example, the side of a tree looks flat if the area being evaluated is only a couple of inches. 

The Standard Deviation parameter controls how aggressively points are clustered into segments. As it’s the segments themselves that are evaluated for classification, usually, 2 or 3 standard deviations are enough. Once the points have been grouped into clusters, the Minimum Cluster Size is used to toss out segments that are too small to be able to tell what they are. 

These settings adjustments should get you started and enable you to get better results from the automatic classification of buildings and vegetation. Stay tuned for future blogs covering the other classification tools: powerline and powerpole, and segmentation.  

 

Contact geohelp@bluemarblegeo.com for more specific help in classifying your data. More information about Global Mapper lidar classification tools and settings can be found here in the Knowledge Base

For full lessons on how to manage your data in Global Mapper, consider enrolling in an online, hands-on training class

For blog topic requests, please email marketing@bluemarblegeo.com

Companies using Blue Marble’s geospatial technology