The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
317
Point clouds in each window are automatically classified as
trees, for certain combinations of point density in the four
subregions. These combinations generally request that at least
one upper subregion has high density, while at least one lower
subregion should have lower density. In addition, we request
that the difference Zmax - Zmin in each window exceeds a
threshold. For the upper subregions a threshold T for the point
density is used, where T is defined as:
(1)
UP and up are the mean value and standard deviation of the
number of points in each of the upper subregions in all the
windows.
Figure 10. Extracted trees (black points) overlaid on CIR
orthoimage.
In the case of Zurich Airport data, the density threshold T is 6
points / 2.5m by 2.5m. The extracted tree class has been
compared with the tree class extracted from NDVI and nDSM.
73% of tree points were correctly classified, while 8% were not
detected. The density of point cloud directly affects the quality
of the result. As it can be seen in Figure 10, visible errors in the
results are small objects as vehicles and aircrafts. In addition,
some tree areas could not be extracted because of the low point
density in the whole dataset. As mentioned above, the points in
the raw DSM not present in DTM-AV describe buildings,
vehicles and high or dense vegetation. After extracting the trees
using point density analysis, buildings are obtained by
subtraction of the tree layer from the DSM points,
corresponding to voids and low density, and filtering of small
objects (Figure 11). The accuracy analysis shows that 92% of
building pixels are correctly classified, while 17% of buildings
could not be detected.
Figure 11. Extracted building points after elimination of tree
points.
5. ANALYSIS OF RESULTS
The accuracy results of the four methods described in Section 4
are summarized in Table 2. Method 4, based only on Lidar data,
performs best in terms of correctness, but is the worst in terms
of completeness. It does not detect all buildings, but those
detected are correct. On the other hand, Method 1, again based
on Lidar data, but with NDVI contribution, can extract the
largest part of buildings but other objects are included, resulting
in the worst correctness value. The other two methods have
basically the same performance, lying in the middle between
Method 1 and 4 results. It should be noted that results using
Lidar data strongly depend on average point density, but also
number of echoes that are registered per pulse and whether
Lidar data acquisition was with leaves on or off.
Due to time restrictions, only a first simple fusion of the results
has been attempted. By union of the four building detection
results, the omission rate decreases (8%) but also correctness
too (81%), while the intersection of all results gives the best
correctness (96%). The correctness of each method could be
improved by developing an automatic detection of objects like
aircrafts, which are classified as buildings in all methods.
Taking into account the advantages and limitations of each
method, currently we can not recommend a single solution for
building extraction. By intersecting the results from method 4,
based on Lidar data analysis, and method 2, based on
supervised classification, the best correctness rate is achieved,
but the completeness is poor. The other building layer
combinations in Table 2 led to worse results.
Method 1
Method 2
Method 3
Method 4
Iu2u3u4
In2n3n4
lu4
ln4
2u4
2n4
nDSM+NDVI
Class.+nDSM
Voids+
NDVI
Lidar
Correctness (%)
76
86
87
92
81
96
83
96
83
95
Omission error (%)
10
13
13
17
8
29
10
29
9
27
Table 2. Results of building extraction using four methods and combinations thereof. The best results are shown in bold.
6. CONCLUSIONS
In this paper, different methods for object extraction (mainly
buildings) in Lidar data and aerial images have been presented.
In each method, the basic idea was to get first preliminary
results and improve them later using the other available data.
The methods have been tested on a dataset located at Zurich
Airport, Switzerland, containing aerial colour and infrared
images, Lidar DTM and DSM point clouds and regular grids
and vector data for accuracy assessment. The results showed
that correctness values up to 92% can be achieved using Lidar
data only, while the highest completeness is obtained by the
combination of image and Lidar data.
Future work will include the improvement of building
extraction from aerial images and Lidar data. The algorithms
will be tested also on other airport locations. However, the
main focus will be on a better fusion of the individual results,
use of image edges for better building delineation and more
detailed building modelling.