inbul 2004 
olygon by 
lygon If it 
following 
polygon (5) 
: of P^ n, 
int density 
parameter 
old. 
lygon if it 
following 
1-76) 
ir the total 
hold is py, 
umbers of 
ing to P". 
m 
   
    
filtering 
ping 
T 
over the 
ing map) 
success of 
he ground 
Ordnance 
| of detail 
F the OS 
| quality 
buildings 
nos image 
ucted at a 
mage and 
'enerated. 
sterMap” 
of Ikonos 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
image. As a result, very small features cannot be clearly 
recognized in the Ikonos image. Since this research has an 
interest to evaluate the quality of the UCL building map, 
rather thane the OS MasterMap®, those inherent faults of 
the OS MasterMap” was removed from the UCL data, and 
small polygons whose member points are less than 100 
points were also excluded before the quality evaluation. 
  
    
hi Jared " A 
(b) OS MasterMap” ground plan 
Figure 11. Building extraction result and OS MasterMap” 
    
A number of objective evaluation metrics suggested by 
Shufelt (1999) was adopted in order to provide a 
quantitative assessment of the developed building extraction 
algorithm. These metrics can be defined as follows: 
building detection percentage (96) 2 100 x TP /(TP - TN) 
branching factor = FP/TP (8) 
quality percentage (%)=100XTP/(TP + FP+ FN) 
where 7P (True Positive) is a building classified by both 
datasets, 7N (True Negative) is a non-building object 
classified by both datasets, FP (False Positive) is a building 
classified only by the UCL building map, and FN (False 
Negative) is a building classified only by the OS 
MasterMap". Table 1 shows the pixel classification results, 
and the evaluation on the UCL building map computed by 
Eq. 8 is presented in table 2. 
Table 1. Pixel classification results 
Pixel classification Pixels 
True Positive (TP) 67085 
True Negative (TN) 255794 
False Positive (FP) 4344 
False Negative (FN) 14639 
Table 2. Building extraction metric result 
  
Building extraction metric Evaluation result 
Building detection percentage — 93.92 (94) 
Branching factor 0.22 
Quality percentage 77.94 (%) 
6. DISCUSSION 
As can be seen in table 2, the proposed building extraction 
technique detected building objects with 94 % detection rate 
(building detection percentage), and showed 0.2 delineation 
performance (branching factor). Finally, the overall success 
of the technique was evaluated as 78 % extraction quality 
(quality percentage). These results suggest that the 
developed system can successfully acquire accurate 
detection and description of building objects using Ikonos 
images and lidar data with a moderate point density. 
  
However, the UCL building map contains certain amount of 
building extraction errors (FP and FN), which should be 
reduced for achieving a more accurate extraction of building 
objects. The errors apparent in the result generated by the 
developed system can generally be divided into three 
categories: 
Building detection error: most of FN pixels in Eq. 8 were 
generated by under-detection of the terraced houses (see 
blue coloured polygons in figure 12). This problem is mainly 
caused by the fact that the NDVI classification described in 
83.3 tends to over-remove "building" points over those 
building with long and narrow structures such as a row of 
terraced houses and results in a very small “blob”, whose 
member points are fewer than 30 points. This problem can 
be resolved by modifying the NDVI classification from 
point-wise to region-wise approach. That is, in order to 
ensure larger numbers of member points are obtained, 
“high-rise” points populated in $3.2 are clustered in a 
number of single objects, and then a cluster-by-cluster tree 
detection is made by the NDVI classification. This 
modification may make terraced houses detectable since 
more member points are retained. 
Building delineation error: these errors are caused when 
boundaries of building objects are not properly extracted by 
the building description process (see red coloured pixels in 
figure 12). Those errors are related to the inherent 
planimetric accuracy of input data (i.e., Ikonos image, lidar 
data, and OS MasterMap ), and the point density of lidar 
data. Most of boundary delineation errors are deviated from 
the OS reference data with one or two pixels if lidar