ul 2004 
or the 
(4) 
on the 
flat). 
'esult. 
AN of 
n the 
insfer 
steps: 
Eq.5 
mage 
using 
wing 
the 
ages. 
bject 
ts of 
on IS 
one 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
The difference between the two contours is caused by different 
photographic conditions and inaccuracy in the model. There are 
three basic approaches for merging two polygons: polynomial 
transformation, triangulation transformation and polyline 
projection (Filin and Doytsher, 1999). The third part consists of 
generalization operations such as simplification and smoothing, 
which can be implemented by using a knowledge base, in order 
to facilitate the final mapping of the building. Application of 
this part is neither presented nor detailed in this article. 
Sm(Z Zn xix, 7 x )rhrs X(yyr aln) 
XsSx(nx(X,-x)trix(y,-)-nxf)* X, 
Vz SKK NT X VEN, 
(5) 
‚MX -Xo)+r2(Y-Yo)+rz3i(Z-—Zo) 
r13(X — X0)+r23(Y — Yo)+r3(Z — Zo) (6) 
Pr2(X — X 0) +r2a(Y — Yo)+r32(Z — Zo) 
ri( X — X 0) +r23(Y — Yo) + r33(Z — Zo) 
  
x= xo- f 
  
y = yo -— f 
3. IMPLEMENTATION AND EXPERIMENTS 
In the course of research, a semi-automatic system for building 
mapping from a medium image scale (~1:40,000) was 
developed in order to examine the algorithm efficiency. The 
system enables opening a pair of aerial images, in order to 
perform a manual vote on the wanted building roof in the left 
image, mapping the 3D building contour and transferring it to a 
Geographic Information System in a local coordinate system. 
Since the aim was to develop a semi-automatic approach for 
constructing and updating the buildings layer of the GIS, the 
Israeli national GIS was chosen as the pilot environment and 
the same conditions used for its construction and updating were 
retained. The Israeli national GIS has been characterized by 
Peled and Raizman (1997) as follows: (a) mapping is based on 
photogrammetric mapping of 1:40,000-scale air photographs by 
I and 2™ class photogrammetric stereoplotters; (b) the 
planimetric and altimetry accuracies of the mapping are + 2 
meters, suitable to the 1:5,000-scale traditional mapping; (c) the 
level of mapped details is according to regional mapping at 
1:10,000 scale; (d) the DEM is measured at 50 meter resolution. 
The experiments were conducted on two residential building 
areas in Tel-Aviv using two medium scale (~1:40,000) aerial 
images scanned at a pixel size of 14 Lm . The first test area was 
in north Tel-Aviv and included 80 residential buildings, while 
the second was in central Tel-Aviv and included 97 residential 
buildings. The chosen test areas were large enough and 
represented buildings in a flat crowded urban area. The 
buildings had 4-24 corners and most had few floors and flat 
roofs. Figures 2, 3 present the 2D building extraction in the left 
and right images of both areas. Figure 4 presents the semi- 
automatic building mapping (dark) upon the manual mapping 
(bright) in both areas. 
4. ANALYSIS AND DISCUSSION OF THE RESULTS 
Results are analyzed separately, as qualitative results and 
quantitative results. 
4.1 Qualitative Results 
In the qualitative analysis, the aim is to evaluate whether the 
approach is practical, i.e., what percentage of buildings can be 
TH 
mapped using this approach. For this evaluation Eq. 7 was 
employed, where BSM is the number of buildings successfully 
mapped, BPM is the number of Buildings partially mapped, 
BNM is the number of buildings not mapped and K is a weight 
for evaluating success in the BPM category (k=0.5). Table 1 
presents the success percentage in each test area. These results 
show that a significant percentage (76%) of the buildings was 
mapped. However, the major innovation is that the operator can 
see at a glance all buildings that can be mapped using this 
approach. Therefore, even if the success rate was smaller, it 
would still be efficient to use this approach initially and 
complete the mapping by using the traditional method. 
Buildings mapping rate = 
BSM * k- BPM (7) 
  
  
  
  
  
_— x 100 
BSM + BPM + BNM 
Test Area North Center 
BSM 62 66 
BPM 
BNM 12 26 
Mapping Rate (%) 81% 71% 76% | 
  
  
  
  
  
Table 1: Success percentage in each test area 
4.2 Quantitative Results 
The quantitative analysis was based on comparison between the 
3D buildings contours extracted using the semi-automatic 
approach and 3D manual mapping of these buildings made by a 
professional operator using the same images and same solution 
model. The deviation vector d = [dx dy dz] of each 
building corner from the manual mapping and the closest point 
on the semi-automatic building contour were measured. 
Altogether 1444 deviation vectors belonging to 139 buildings in 
the test areas were measured. 
  
  
  
  
  
  
Testarea | dX (m) | dY (m) | dZ (m) 
Mean North 0.22 -0.12 0.44 
728 
RMS 0.72 0.59 1.06 
vertexes 
Mean Center 0.27 -0.19 0.12 
716 
RMS 0.70 0.67 1.35 
vertexes 
Relative 1444 
: 0.71 0.63 1.21 
Accuracy vertexes 
  
  
  
  
  
Table 2: mean and RMS of the deviation vectors. 
Based on 1444 deviation vectors between the manual and the 
semi-automatic mapping, the relative accuracy (semi-automatic 
to manual) was calculated. In Table 2, the mean and the RMS 
of the deviation vectors in each area are presented.