try, three sets of reference measurements are used. Two of these datasets are tachymetric measurements
Ž .
Fig. 1 carried out in March 1996. The third set is a DTM of the area and the coordinates of the roof
Ž .
corners of buildings Fig. 1 . This DSM has been generated by means of photogrammetric measure-
ments on true colour aerial photographs at a scale of 1:4000 also acquired in March 1996. The expected
planimetric and height accuracies of the measured features are 6 and 7 cm, respectively.
4. Information content of the laser data
In this section, it is assessed which of the infor- mation in Table 1 can be extracted from laser mea-
surements. For this purpose, the laser-DTM and the laser-DSM mentioned above are used. The informa-
tion that cannot be extracted from the laser data has to be collected from existing sources of information,
such as databases. In this context, some considera- tions are also given to the combination of laser data
with Dutch 2D and 3D databases.
4.1. Extraction of information from laser data alone Extraction of information comprises detection,
identification, modelling, measuring, and labelling. The last two processes provide the geometric and
semantic information of the features, while the first three permit one to recover this information. The
modelling procedure will have a great impact on the
Ž .
final accuracy Section 5 . Some of the procedures may be considered together like detection and identi-
fication, and modelling and measuring. Thus, the terms ‘detection’ and ‘modelling’ are used hereafter
in this broader sense.
Extraction of information is considered both man- ually and automatically. To extract manually geo-
metric and semantic information, visual inspection of the laser data is required. This can be done with the
help of so-called laser images created by assigning a grey value to, for example, the height information.
The laser images can also be displayed as a perspec- tive view or in stereo, which facilitates interpreta-
Ž tion. The display of an image in stereo can easily be
achieved by creating two perspective views from two different viewpoints and by using a screen with
. stereo capability.
By using the laser images, the operator is only able to extract 2D geometric information. The Z
coordinate has to be estimated afterwards by means of automatic modelling. The information that can be
extracted in this way from the laser-DSM is tabu- lated in Table 2. From this table, it can be seen that
it is possible to extract all the features that describe the terrain relief. Concerning objects on the terrain,
the semantic information of 60 of the object types Ž
. i.e., 4 out of the 7 features of category 2 in Table 1
cannot be recovered. The automatic extraction of information is only
partly possible because its labelling is still unfeasi- ble. Automatic extraction of the geometric informa-
tion in Table 1 was only implemented for those features that correspond to abrupt changes in terrain
Ž .
slope hereafter referred to as breaklines . This means that only the features that describe the terrain relief
were extracted by using the laser-DTM. Objects on the terrain cannot be extracted automatically solely
from laser range data. The research on this subject
Ž .
combines laser with images Haala, 1994, 1995 ,
Ž .
existing databases Haala and Brenner, 1997 , or the intensity of the laser beam recorded simultaneously
Ž .
with the range data Hug and Wehr, 1997 . To extract the breaklines automatically, first a
laser-DTM image was created. The Laplacian opera- tor is applied to this laser image to detect edges. Fig.
2a illustrates automatically detected breaklines. The corresponding manually extracted breaklines are de-
picted in Fig. 2b. As it can be seen, there are more breaklines detected manually than automatically.
Furthermore, the breaklines detected automatically need to be edited. As for the manually extracted
breaklines, the automatically extracted breaklines have to be converted from 2D to 3D features. This
procedure signifies modelling, which may be done in several ways. If only linear interpolation is used, the
Ž quality of the modified laser-DTM with the break-
. Ž
. lines will not be improved see Section 5.2 . This is
because, although knowledge was increased, in the sense that a breakline is known to exist, no adequate
modelling was carried out. Consequently, in order to increase the quality of the laser-DTM, adequate
modelling is required. If it is assumed that breaklines are straight lines, their modelling can be done by
L.M. Gomes
Pereira, L.L.F.
Janssen r
ISPRS Journal
of Photogrammetry
Remote Sensing
54 1999
244 –
253 248
Table 2 Geographic information that can be extracted from laser measurements alone or in combination with existing databases
DSM-category DSM features
Laser Laser and 2D databases
Laser and 3D databases Ž
. Ž
. GBKN and Top10vector
DTB-roads and DTM-roads Geometricrsemantic
Geometric and semantic Geometric and semantic
Ž .
Terrain relief Breaklines
yesryes yes
yes DTB-roads Ž
. GBKN or Top10vector
Ž .
Lines describing flat surfaces yesryes
no yes DTM-roads
Ž .
Lines describing highly curved yesryes
no yes DTM-roads
surfaces Ž
. Individual points
yesryes no
yes DTM-roads Ž
. Ž
. Objects on terrain
Separation line between two yes if breakline rno
yes GBKN yes
Ž .
paved surfaces in DTB-roads after some processing
Ž .
Separation line between paved yes if breakline rno
no yes
Ž .
and unpaved surfaces in DTB-roads after some processing
Ž .
Lines painted on roads norno
no yes DTB-roads
Ž .
Ž .
Protection walls yesrno
yes GBKN yes DTB-roads
Ž Water boundaries
yes if no laser pulse return yes
yes .
Ž Ž
. for measurements on water ryes
GBKN, and Top10vector if DTB-roads after some processing
width of ditches is bigger .
than 2.5 m Ditches
yesryes yes
yes Ž
Ž .
Top10vector if ditch DTB-roads after some processing
. width 2.5 m
Ž .
Roof gutters and ridges yesryes
yes GBKN yes
Ž .
DTB-roads, but Z is inaccurate
Ž . Ž .
Fig. 2. a Automatically detected breaklines from a laser-DTM image; b the corresponding manually measured breaklines.
surface approximation methods. A simple model is to reconstruct a breakline as the intersection of two
planes, computed by using the least squares method. Before modelling, some processing may be needed
to close broken lines, and delete small lines as well as those lines that result from wrongly detected
edges. This editing process is totally or partly man- ual.
4.2. Extraction of information from laser data in combination with existing databases
In Section 4.1, it was clearly stated that laser data alone does not allow the extraction of all the infor-
mation needed to generate a DSM for road planning and design. Furthermore, the quality of information
may need to be improved. Thus, combination of laser with existing information is a prerequisite. In
this section, only the combination of laser data with 2D and 3D Dutch databases is considered. No testing
was carried out for the feasibility of such combina- tion.
The databases considered are the 2D databases GBKN and Top10vector, and the 3D databases
DTB-roads and DTM-roads. GBKN is a large-scale topographic base map, and Top10vector is a large-
scale topologically structured vector database in scale 1:10000, generated by photogrammetric means.
DTB-roads is a large scale database for road mainte- nance, produced by photogrammetric and terrestrial
means, while DTM-roads is a database for road construction, generated by terrestrial means.
Information stored in such databases can be used not only to label some of the features in the DSM,
but also to locate them, e.g., boundaries of buildings or lines painted on roads. If the database is not
object-oriented, as is the case of GBKN, before feature extraction, e.g., of buildings, from the points
and lines, some processing has to be carried out to
Ž group them into meaningful objects Lemmens et al.,
. 1997 . Their location in the laser data furnishes the
height values for the objects in the database. This process, i.e., the projection of the objects or features
in the database onto the laser data, requires some care. Their related semantic information should be
used for a more adequate estimation of the heights. For objects on the terrain, like buildings, this estima-
tion must be carried out using the raw measurements Ž
. laser-DSM . Obviously, if the database is 3D and
the heights are accurate, modelling may not be re- quired.
Table 2 lists which of the information in Table 1 can be collected by combining laser data with the
databases mentioned above. Since laser allows the extraction of the relief information, combination with
databases is in principle needed solely to extract objects on the terrain.
5. Accuracy assessment of the laser data