LIDAR acquires a three-dimensional cloud of points with irregular sampling.
The measurements hit on objects such as buildings, vehicles, vegetation, and bald terrain. To generate a DTM, the core task involves the separation of non-terrain points from LIDAR datasets. This task is referred to as filtering, classification, non-ground measurements removal. The filtering task dominates the topic of DTM generation. A number of algorithms have been reported in the literature, but there are a number of conditions that make filtering a very difficult problem. DTM generation from LIDAR data has proven to be a challenging task.
Kraus and Pfeifer (1998, 2001) utilized linear least squares interpolation iteratively to remove tree measurements and generate DTMs in forest areas. The least squares interpolation is based mainly on calculating the residuals, i.e., the distance from the surface to the measurement points. It is assumed that terrain points are likely to have negative residuals, whereas non-terrain points are more likely to have positive residuals. These residuals are used to compute weights. Points with large negative residuals have maximal weights and they attract the surface. Similar methods have been adopted by Lohmann et al. (2000). Pfeifer et al. (2001) extended this method by a hierarchical strategy to assemble data from coarse to fine. Lee and Younan (2003) used a modified linear prediction technique, followed by adaptive processing and refinement. However, this method fails to model terrain with steep slopes and large variability. In addition, extreme low points can be easily misclassified as terrain points as a result of the negative errors.
Kilian et al. (1996) used morphological filters to eliminate non-terrain points.
Typically these filters need to predefine a search window size. These filters may have problems with dense forest canopy or large buildings. If a small window size is used, large-sized buildings cannot be removed. On the other hand, larger window size causes the filter to over-remove terrain points or chop off hills. Kilian et al.
(1996) suggested using a series of windows to progressively filter the terrain.
Petzold et al. (1999) proposed a filtering algorithm. A rough terrain model is calculated by the lowest points found in a moving window of a rather large size. All points with a height difference exceeding a given threshold are filtered out, and a more precise DTM is calculated. This step is repeated several times by reducing the window size. The final window size and the final threshold have great influence on the results: small window size leads to points on large buildings remaining, while a high threshold in the final step leads to non-terrain points being classified as terrain points. Therefore, the parameters depend on the terrain variation and have to be adjusted for flat, hilly and mountainous regions.
The concept of resizing window size was adopted by Zhang et al. (2003, 2005);
they proposed a progressive morphological filter. By gradually increasing the window size of the filter and using elevation difference thresholds, the non-terrain points are removed, while terrain points are preserved.
Lohmann et al. (2000) compared two algorithms, namely the use of linear prediction and the use of dual rank filters. The use of linear prediction showed satisfactory results in forest areas, whereas areas with steep terrain showed problems.
The linear prediction filter needs to be improved by being locally adapted to the shape of the terrain (Lee and Younan, 2003). The dual rank filter is a mathematical morphology filter which is applied to a grayscale image. Dual rank filters need to be improved through interactive control and some pre-knowledge to properly set the
necessary parameters.
Vosselman (2000) proposed a slope-based filter. This method was commonly applied for DTM generation from LIDAR data. A measurement is classified as a terrain point if the height differences of this measurement point and any other point within a given circle are smaller than a predefined threshold. Choosing a threshold of too small height results in removing some detail of the terrain or cutting hill peaks;
too large a threshold results in preservation of non-terrain points. Furthermore, the fixed threshold procedure limits the use of slope-based filter to terrain with gentle slopes. This technique will give satisfactory results only when non-ground objects (trees, buildings, etc.) and background terrain slopes are distinct and are uniform throughout the full coverage.
The next parameter of the filter is the radius of search region. The region size should be large enough to enclose the larger non-ground objects, such as buildings, but not so large as to cover different terrain aspects (i.e., ridges should not be crossed).
Small search regions tend to cover uniform background slopes. Large regions tend to have larger topographical undulations than smaller regions. Adaptive techniques are required to overcome the effects of nonuniformities in background slopes.
Slope-based filter has been shown to be equivalent to the erosion operator in morphology. Sithole (2001) has attempted to improve this filter.
Axelsson (1999, 2000) described an adaptive TIN (Triangulated Irregular Network) model to process ground points in dense urban areas where discontinuities may occur. A course TIN is formed based on seed points selected from low points.
The course TIN iteratively adds more points if their parameters are below threshold values. The problem with adaptive TIN is that it is difficult to detect various sizes of non-ground objects by using a fixed parameter. Raber et al. (2002) showed an adaptive vegetation removal process. The different thresholds were adaptively given
based on a vegetation map for various land cover types.