In the previous installment, we examined the process of rigorous geometric correction of LIDAR data. It is advisable to fully correct the data and perform a project "readiness review" prior to moving on to data classification. While it is possible to perform classification prior to final geometric correction, it is generally not a good idea for the following reasons. The first is simply a matter of wasted time. If you invest hours in classification and then discover an unrecoverable geometry problem (meaning you will have to perform a reflight), the classification effort will be completely wasted time. The second reason is that incorrect geometry can cause errors in automated classification algorithms. This is particularly true in flight line overlap areas. Here an elevation difference can cause an automated algorithm to classify the lower points as either low points or ground but perhaps treat the higher side as building rooftops.
Classifying LIDAR data is the process of assigning a numerical value to the LAS Classification field that corresponds to the object type from which the point reflected. Classification can range from a minimum of determining a bare earth surface to a complex classification involving dozens of categories such as a transmission corridor mapping project. Obviously the project budget determines the degree of classification completeness.
LIDAR point cloud data are emitted from the proprietary sensor geocoding software classified as either Created, Never Classified (class == 0) or Unclassified (class == 1). Many ground classification algorithms start with the lowest last return data and work up in forming the bare earth surface layer. For this reason, the first step in classification is to find points lower than the ground surface and tag these as "low" points. Low points occur in LIDAR data as a result of several factors:
Multipath GPS signal, making a return appear farther away than it is in reality
Abnormally low reflectance intensity
Unknown system "glitches"
A standard automated algorithm for detecting low points looks at individual LIDAR returns and measures their vertical position with respect to the immediate neighbors. An example of a low point is illustrated in Figure 1.
Following the reclassification of low points, the ground points are classified. There are a wide variety of automated ground classification algorithms in use in commercial LIDAR processing software. A common, general purpose algorithm is one that places a grid over the data and selects the lowest point in each grid cell as a "seed" ground point. The algorithm then constructs a Triangulated Irregular Network (TIN) of this surface and assumes that this is the first rough approximation of the ground. You can see that a low point would distort this initial approximation and hence the initial low point removal pass. The algorithm next attempts to add additional points to the TIN by examining the properties of the facet that would be created as a result of adding a candidate point. The usual inclusion criteria include the distance the point lies in a vertical direction from the existing TIN and the angle the newly added facet will make with respect to adjacent facets that are already in the ground class. This process is iterative, gradually building up a surface model. The algorithm generally includes refinements such as edge climbing tests to allow sharp surface changes (such as berms) to be included in the ground surface.
Ground classification is generally run in a fairly conservative manner because most algorithms only add points to the ground class. This means that if non-ground points migrate into the ground class, they will never be removed. As a result, a typically classified LIDAR project will still have many points that rightly belong in the ground class still in the "unclassified" state. An example of this is shown in Figure 2. For this reason, it is very important for a density check to be included in the Quality Check procedures of the data.
Ground classification can often include "noise" in the ground classified points. An example of this is very low vegetation being classified as ground. One approach to cleaning data that exhibit this problem is to run a statistical noise algorithm against the ground class. This algorithm inspects local regions of the ground class and rejects points that exceed a user specified multiple of the standard deviation of the ground class point heights. While this algorithm can assist in smoothing noisy data, it does have the deleterious side affect of thinning the ground class.
Following ground classification and de-noising, either interactive classification is performed or further automatic classification. We will examine these scenarios in the next edition.