A 1.215Mb PDF of this article as it appeared in the magazine complete with images is available by clicking HERE
We have an enterprise systems division of our company that primarily develops technology for organizing and processing point cloud and raster data. Within this body of work and projects we have commercial product development such as our LIDAR Server technology as well as custom development such as a very large system hosted in Amazon Web Services that will be used, among other data, for managing a near real time space-based hyperspectral imager.
One of the primary choices one faces when designing such systems is the format that will be used for storage, presentation and delivery (which are not necessarily all the same) of various types of data. We are engaged in a large LIDAR data management and dissemination project for one of our key US government customers. We recently became involved in a discussion of raster versus point cloud for LIDAR data storage and delivery. I thought this debate had died long ago but it still occasionally comes up (especially from companies who support only one format!). I think it continues to surface because point clouds are a real problem to work with. Rasters are so nice and simplecan’t we just make everything a raster?
Rasters are single layer, quantized representations of object space. Due to the quantization, they are very easy to represent. Their X, Y locations need not be stored because they are on a uniform grid. I need only know the spacing between posts, the origin and the orientation (e.g. is it rotated?). Raster elevation models typically represent only one height value (though they could easily represent more, albeit with poor storage efficiency). Rasters cannot represent point features, period. This is their biggest downfall (for you electrical engineers out there, it is like trying to sample a randomly located delta function with a uniform sampler).
Consider modeling a "vee" shaped culvert. When a raster is superimposed over the culvert, a raster post will very seldom fall directly over the bottom of the vee. This means that this critical (critical as in calculus, meaning the point at which the gradient is zero) simply cannot be modeled. This is the reason that GIS folks adopted the Triangulated Irregular Network (TIN). A good way to think of a raster model is as a lossy compressor that loses information in the X, Y plane due to quantization and loses all information in the vertical plane except the particular attribute chosen for storage (for example, is it bare earth or canopy?).
Point clouds have several advantages over raster when modeling the earth. The first advantage is that we never move the sampling points (though it is correct that they could have been moved at the front end of the sampling process if the sample is something other than a traditional scanning LIDAR). A second advantage is that they need not be uniform in terms of the attributes being represented in object space. For example, one point can represent ground, the adjacent point a building, a third point low vegetation and so forth. Most importantly, they can represent both random and duplicate points in X, Y space. In a standard raster, you cannot simultaneously represent the bare earth location in elevation at a given planimetric point as well as a tree branch directly above this same point. An example of a multi-dimensional point cloud is depicted in Figure 1. Here the ground is rendered as a wireframe TIN (orange). Superimposed over this TIN are building points (red) and vegetation (green). The data are also multi-dimensional in terms of point attributes, showing echo (return), classification, acquisition time and a number of other attributes. This type of information is lost when a point cloud is compressed into a raster.
Of course, this is not meant at all to say that the ubiquitous raster is a bad format! It is just that we have variety of transport formatsit is important to choose the correct one for the particular type of data being represented. It would be a poor choice indeed to represent evenly spaced, single height elevation data in a point cloud.
I hope we can dismiss the idea of compressing point clouds down into rasters unless the application truly is more amenable to a raster model. Point clouds should, in general, never be spatially quantized into a raster. It is a terrible waste of money to very carefully process point cloud data and then corrupt it at the end by squeezing it in to a raster!
Lewis Graham is the President and CTO of GeoCue Corporation. GeoCue is North America’s largest supplier of LIDAR production and workflow tools and consulting services for airborne and mobile laser scanning.
A 1.215Mb PDF of this article as it appeared in the magazine complete with images is available by clicking HERE