LIDAR Best Practices: Part V Intermediate Surface Classification

In the previous installment, we examined the process of the first step of automatic LIDAR data classification. Recall that LIDAR point data in LAS format contains an attribute called "Class." This attribute is used to designate the object from which the point (a laser return or "echo") reflected. All points are assigned a value of "Created, Never Classified" (class == 0) or "Unclassified" (class == 1) by the sensor manufacturer’s post-processing software. In the previous edition, we examined the steps necessary to move from a state of unclassified to removing low noise and automatically classifying some of the candidate ground (or "bare earth") points.

Much research and development (R&D) work has been conducted on extracting "bare earth" LIDAR points (this was discussed in some detail in the previous edition of this article). There are two drivers for this. First of all, for topographical mapping the bare earth surface (as opposed to the "first surface" or "canopy" model) is the most useful and hence most requested extraction. Secondly, the bare earth surface is often used as a foundation "known" in other automated extraction algorithms. For example, a building roof extraction algorithm will typically use the distance above a previously extracted ground surface to define a search range. This prevents the algorithm from identifying areas of the ground (a nice, planar surface) as roof planes.

The number of classification categories that will be extracted in a LIDAR processing project is a function of the client needs as well as budgetary considerations. Obviously the more classes desired, the more costly the processing. An additional important consideration is the nature of the classes that are desired. For example, a bare earth classification that includes at-grade roads (roads at the same approximate elevation as the surrounding ground) in the ground class is considerably easier to extract than a specification that requires roads (including at-grade) in a separate "road" class. For example, in Figure 1 a ground surface of points is depicted (without intensity shading). Note the "holes" in the data where building footprints and other non-ground objects occur.

Fig 1- A classified ground surface

Figure 2 depicts this same area, rendered as a shaded relief Triangulated Irregular Network (TIN) using the intensity return of the LIDAR data (the very valuable "I" in LIDAR). Note that in this rendering, the roads are clearly visible. However, they are not contained within the LIDAR data class attribute.

Fig 2 – A TIN rendering of the surface of Figure 1, clearly showing roads

Adding classifications to data where there is little to no distinguishing 3D attribute (e.g. relative height, slope, texture and so forth) often reverts to image processing techniques. For example, a simple (and admittedly not very effective) technique of distinguishing roads is to classify by Intensity using source points that have already been prequalified in some way. For example, in Figure 3 are depicted the results of classifying roads by using intensity bands as the discriminator. The LIDAR return pulses from the road surface exhibit lower return intensity than the surrounding areas (due to the propensity of the road surface to more readily absorb infra-red light).

Fig 3 – Roads classified (purple areas) using LIDAR intensity discrimination

Surprisingly, these sorts of techniques (that is, image processing combined with 3D analysis) are in their infancy with few functional algorithms appearing in commercial LIDAR data processing software. There are two primary reasons for this lack. The first is that normalization of LIDAR intensity is fairly difficult and LIDAR manufactures are just now introducing correction algorithms into their post-processing software. By normalization I mean correcting the return intensity for effects caused simply by variations in range (i.e. a pulse impacting a road from an altitude of 1,000 m will have a higher energy return than a pulse impacting that exact same area of road but from a sensor at an altitude of 2,000 m). The second reason for the slow emergence of these algorithms is the configuration of the projects. It is only fairly recently that cameras have become de rigueur with corridor LIDAR missions and are still often not part of a areal LIDAR mapping campaign. Thus there has not been a high level of motivation among software providers to develop algorithms that rely on what may be non-existent supplemental image data (there are notable exceptions – The National Geodetic Survey routinely uses 4 band supplemental imagery fused with LIDAR to delineate shoreline).

Easier to discriminate are objects that exhibit elevation or elevation derivative differences relative to the ground surface. In the next installment, I will review some of the more common of these.