For many customers who do engineering surveying and are new to LIDAR as a raw data source, the concept of data classification is novel. I thought I would use this week’s Bulletin to provide an overview of this important concept.
As you know, LIDAR systems send out a pulse of laser light and measure the time until a return “echo” is detected by the sensor (similar to sonar). Since the speed of light is nearly a constant in air, the range to the point where the reflection occurred can be computed.
Separately, ancillary data such as position (X,Y,Z) and orientation (Pitch, Roll, Yaw) system (POS) information is collected. This POS information is combined in a set of (usually) post-processing steps to generate a pseudo-random point cloud in the spatial reference system (SRS) chosen by the user.
Each point represents a location in “object” space (e.g. a point on a building, the ground, vegetation, etc.) or, on occasion, a “noise” point. The information inherent to each point is simply the X, Y, Z location; the LIDAR has no idea the type object from which the echo bounced.
If we want to derive secondary information from the point cloud (e.g. use the data for something other than visualization), we need to segregate at least some of the points that will feed our analysis. One of the most common products is a topographic map represented as a random collection of echo points that impacted the ground (point cloud model), a raster elevation model (Digital Elevation Model or DEM) or a set of topographic contours. To achieve this result, we need to identify enough of the points that reflected from the ground to support our model.
The process of segregating LIDAR points according to the type object from which they reflected is termed “Classification” since we are labeling each point according to the class of reflecting object. This class information is stored on an individual point basis within the point storage file. We use a common file format called LAS (as in LASer) which contains, in addition to X, Y,Z attributes, slots for other information such as Classification. An example of the attributes stored on a per-point basis in a True View workflow is shown in Figure 1. I draw a red box around the attribution we are discussing here, Classification. You can see for my selected point, the Classification is Ground.
A list of standard American Society for Photogrammetry and Remote Sensing (APSRS) classes is depicted in Figure 2 (this is a Class tab in our Live View control from True View EVO, the software included with every GeoCue True View sensor).
When we first create a point cloud from the raw sensor input data, all points are set to the class 0 – “Created, Never Classified.” Note that class 1, “Unclassified” looks to mean the same thing and it does. When the LAS format was first defined, both 0 and 1 were being used for the same representation. We (the ASPRS LAS format committee) made a distinction to use class 1 to mean a point that was classified to something other than unclassified and then returned to the unclassified state but this never really received wide-spread use. Do be careful, however, since agencies such as USGS demand that only class 0 be used as the unclassified marker.
Nearly all LIDAR viewing tools can display points colorized by class and, furthermore, allow you to set the display color on a class by class basis. You can see this in the dialog snippet of Figure 2 where we have Unclassified as grey, Ground as orange, vegetation as shades of green, Building as red and so forth. Figure 3 shows a region of a point cloud colorized by class using the color scheme of Figure 2.
Of course, the 64,000 dollar question is how did these classes get assigned since all points came in from the “geocoding” step in the Unclassified state? Obviously we could provide some interactive “painting” tools that allow you to manually change the class attributes on points (and we do provide a rich set of these tools in EVO) but manually classifying enough ground points to allow the generation of a set of topographic contours would be daunting task!
An area of current and intense research in LIDAR data processing is the development of semi-automatic and fully automatic tools for performing classification. EVO contains a number of these tools to automatically classify:
There are also “geometry” classification tools such as above/below a polygon, within distance of a polyline and so forth.
The general approach to classification is to first clean up any noise in the data set (a subject for a separate Bulletin article), run an automated classification algorithm and then clean up the classification using the many manual tools in EVO. However, before embarking on the classification task portions of a workflow, it is critical to pause and ask what products need to be created.
One of the primary motivations for our creation of the concept of a 3D Imaging Sensor (3DIS®) is to eliminate the need to classify data for visualization purposes. In the Colorize LIDAR point step of our post-processing workflow, each LIDAR point is ray-tranced to the “best” RGB image (images are simultaneously collected by GeoCue 3DIS) and RGB tags on the point are populated with these values. This provides a natural color visualization point cloud that is much more informative (and better looking, I might add!) than the colors rendered from colorizing multiple classes. The best part is that this colorization is fully automated and achieved within just a few minutes post-flight. Figure 4 depicts a colorized point cloud of a USDA dam site with (clockwise from upper left) a plan view, 3D view and profile view. This natural, 3D view of a site is really terrific eye candy for your customers, especially considering we provide a free 3D viewer you can hand out for use in visualization!
So with viewing taken care of in a rather spectacular fashion, what remains for classification? Usually the answer to this question is driven by the derivative products that you need to create for your customers. Some examples include:
Figure 5 illustrates one of the beautiful products created by True View EVO. This is our test site (the GeoCue campus) where I have classified Low Noise and Ground. The rendering is Ground as a solid model (a Triangulated Irregular Network, TIN) and all non-Ground points shown in their RGB color. Overlaid on this is a set of 1m topographic contours. I think this makes for quite an impressive product!
Whatever you do, don’t get into the habit of using the same workflow regardless of the output product. Many times I have seen folks doing an arduous ground classification as part of the workflow for doing stockpile volumes; a completely unnecessary step. This “one size fits all” approach can add a lot of labor to a project without increasing project revenue. As a final note, we offer a number of free and paid training opportunities where we discuss best practices for a variety of common products; see you in Class!