For each return, the three-dimensional (3D) position is determined and its intensity
recorded. Hence the raw Lidar data consist of a dense cloud of 3D points with
associated return intensity. Each point provides the location where the laser hit Earth’s
surface during the scanning process, whereas the intensity is a digital representation of
the fraction of pulse energy reflected at that location.
The z coordinate of points corresponds to terrain elevation with respect to a
horizontal datum, typically the mean sea level, plus the height of nonterrain features
in some instances. In order to produce a DTM by interpolation of ground points,
a discrimination of ground from nonground points must be carried out first. This
discrimination process is referred to as ground filtering and is generally considered
a preprocessing for generating not only the DTM but also the height of nonground
components (Axelsson, 1999). Methods to ground filtering can be grouped into two
major categories, point-based and raster-based methods. Methods in the first category classify directly the point cloud whereas methods in the second category first
interpolate the point cloud onto a regular grid surface, namely, the digital surface
model (DSM). Each approach has advantages and disadvantages. In general, rasterizing the data first allows to take advantage of digital image processing algorithms
which run much faster than point-based operations, whereas point-based processing
tends to be more accurate (Axelsson, 1999). Point-based methods tend to include
techniques such as clustering analysis (Roggero, 2001), local surface fitting, discrimination by slope/terrain difference/surface curvature (Sithole, 2001), active
contours, and adaptive triangulated irregular networks (Axelsson, 2000), whereas
in raster-based methods approaches such image segmentation, edge detection
(Brovelli, 2002), mathematical morphology (Zhang et al., 2003), repetitive interpolations, and analysis–synthesis framework are more common. The method presented
in this chapter falls in the last category.
Filtering methods vary in complexity, accuracy, and sensitivity to changes in
parameters. A recent review of methods and critical issues of the ground filtering
problem have been revised by Meng et al. (2010). The International Society for
Photogrammetry and Remote Sensing (ISPRS) Working Group III/3 has conducted
a test to determine the performance of several filters developed over the past years to
extract bare-Earth points from point clouds. Among other things, the study concluded
that, in general, filters that estimate local surfaces are found to perform best (Sithole and
Vosselman, 2004). Other comparative studies have found that using several window
sizes in a progressive filtering tends to be more accurate as it removes features of
different sizes and tends to be less sensitive to parameter selection (Zhang et al., 2003;
Zhang and Whitman, 2005). Interestingly, the need for processing at multiple scales/
resolutions has been increasingly recognized by several studies. For instance, the socalled progressive morphological filters (Chen et al., 2007), which apply morphological
operators with structuring elements of increasing sizes, have been shown to outperform
implementations with a single-size operator. Other methods have taken advantage of the
progressive smoothing in a scale-space representation for detecting high surface
curvatures that occur at ground/nonground transitions (Evans and Hudak, 2007). These
notions are more formally articulated in multiresolution image decompositions, which
other studies have also explored in the context of ground filtering; these include the
268
MULTISCALE APPROACH FOR GROUND FILTERING FROM LIDAR
recorded. Hence the raw Lidar data consist of a dense cloud of 3D points with
associated return intensity. Each point provides the location where the laser hit Earth’s
surface during the scanning process, whereas the intensity is a digital representation of
the fraction of pulse energy reflected at that location.
The z coordinate of points corresponds to terrain elevation with respect to a
horizontal datum, typically the mean sea level, plus the height of nonterrain features
in some instances. In order to produce a DTM by interpolation of ground points,
a discrimination of ground from nonground points must be carried out first. This
discrimination process is referred to as ground filtering and is generally considered
a preprocessing for generating not only the DTM but also the height of nonground
components (Axelsson, 1999). Methods to ground filtering can be grouped into two
major categories, point-based and raster-based methods. Methods in the first category classify directly the point cloud whereas methods in the second category first
interpolate the point cloud onto a regular grid surface, namely, the digital surface
model (DSM). Each approach has advantages and disadvantages. In general, rasterizing the data first allows to take advantage of digital image processing algorithms
which run much faster than point-based operations, whereas point-based processing
tends to be more accurate (Axelsson, 1999). Point-based methods tend to include
techniques such as clustering analysis (Roggero, 2001), local surface fitting, discrimination by slope/terrain difference/surface curvature (Sithole, 2001), active
contours, and adaptive triangulated irregular networks (Axelsson, 2000), whereas
in raster-based methods approaches such image segmentation, edge detection
(Brovelli, 2002), mathematical morphology (Zhang et al., 2003), repetitive interpolations, and analysis–synthesis framework are more common. The method presented
in this chapter falls in the last category.
Filtering methods vary in complexity, accuracy, and sensitivity to changes in
parameters. A recent review of methods and critical issues of the ground filtering
problem have been revised by Meng et al. (2010). The International Society for
Photogrammetry and Remote Sensing (ISPRS) Working Group III/3 has conducted
a test to determine the performance of several filters developed over the past years to
extract bare-Earth points from point clouds. Among other things, the study concluded
that, in general, filters that estimate local surfaces are found to perform best (Sithole and
Vosselman, 2004). Other comparative studies have found that using several window
sizes in a progressive filtering tends to be more accurate as it removes features of
different sizes and tends to be less sensitive to parameter selection (Zhang et al., 2003;
Zhang and Whitman, 2005). Interestingly, the need for processing at multiple scales/
resolutions has been increasingly recognized by several studies. For instance, the socalled progressive morphological filters (Chen et al., 2007), which apply morphological
operators with structuring elements of increasing sizes, have been shown to outperform
implementations with a single-size operator. Other methods have taken advantage of the
progressive smoothing in a scale-space representation for detecting high surface
curvatures that occur at ground/nonground transitions (Evans and Hudak, 2007). These
notions are more formally articulated in multiresolution image decompositions, which
other studies have also explored in the context of ground filtering; these include the
268
MULTISCALE APPROACH FOR GROUND FILTERING FROM LIDAR