'. Istanbul 2004 
f the successful 
, 57 additional 
d from a range 
' included the 
d photographs 
otographs of a 
hotomosaic of 
scale of these 
rements of the 
to 256 x 256 
al test data for 
. Although no 
fier identified 
ending on the 
errors occurred 
ality compared 
  
were able to 
iple data and 
om training is 
ther work that 
mate the true 
eter set that is 
general trends 
ial kernel and 
ith the linear 
all produced 
between these 
mage content. 
the second in 
alization was 
resolution (16 
jarsest image 
of these tests 
| many more 
mputationally 
in terms of 
> cost of more 
s) and a very 
n results are 
thod of pre- 
influence the 
| not apparent 
The training 
Mbytes up to 
ion level, the 
e coefficients 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Accuracy estimates Accuracy for Number 
after training Out-of-sample testing Misclassified 
Test No. ofkernel | Error |Recall| Precision | Accuracy | Recall | Precision an Non- T 
Number | evaluations | (<=%) |(<=%)| (<=%) (%) (%) (%) Buildings buildings Total 
2-3a 13621468 14.7.1 87.2 86.7 82.9 81.5 90.8 S770 258 837 
2-7a 1326497 24.4| 76.7 79.1 81.7 84.8 86.2 476 424 900 
2-7b 8279892 25.31 815 75.4 87.4 88.6 91:5 358 258 616 
3b 9431103 24.1]. 78.3 74.6 84.7 90.7 86.1 290 459 749 
3-2a 275218851 3391 661 66.1 84.8 89.1 87.3 341 406 747 
3-4b 29534759 2601 76.1 77.2. 84.7 82.4 92.9 551 198 749 
3-7a 1132296 24.11 776 79.0 82.0 86.9 85.2 410 474 884 
3-7b 8575675 29.9} 772 717 85.7 92.5 86.1 234 466 700 
4-3a 14108050 14.7] 87.1 86.7 83.0 81.5 90.8 578 258 836 
4-7a 1326937 24.4| 76.7 79:1 81.7 84.8 86.2 476 424 900 
4-7b 7690877 25.3]. 81.5 75.4 87.4 88.6 91.5 358 258 616 
6-1b 11131518 2394 77.8 79.3 83.5 85.3 88.4 460 350 810 
6-3b 18612448 21.51 800 81.3 87.4 89.0 91.0 345 274 619 
6-4b 3002719641 28.4| 74.1 75.0 87.1 85.9 93.3 442 193 635 
6-7a 21478593 30.9] 74.9 71.4 84.0 89.7 85.9 324 462 786 
6-7b 36570891 25.5. | 78.5 70.3 84.2 90.3 85.7 305 471 776 
6-8a 117950052 18.9 | 83.4 82.9 85.0 92.7 85.1 230 508 738 
  
  
  
  
  
  
  
Table 1. Results of classification and testing using large training sample 
A general set of optimal parameters for the pre-processing of 
image data and the training of the SVM is difficult to 
determine. It is likely that while some general principles can be 
established, fine tuning of the classification approach is data 
dependent and must be reviewed on a case-by-case basis. 
Based on the tests in this research, over-sampled wavelet 
coefficients at a resolution of 16 x 16 appear to offer the best 
trade off between classification accuracy and computational 
efficiency. Combined with normalisation in the image domain 
(test 3-7b), this set of parameters produced fewer errors in 
classifying the buildings but at the expense of a higher false 
positive rate. The classifier produced by this test also achieved 
the highest recognition rates with the additional test data. 
7. CONCLUSION 
Machine learning methods have been used successfully in 
several image processing and machine vision domains. The 
research presented here extends this to building recognition for 
photogrammetric applications. 
An important aspect of machine learning in vision applications 
is to extract a representative set of characteristics from the 
image. The multi-resolution approach of wavelets achieves this 
effectively and leads to a solution that is computationally 
feasible. One potential limitation of the wavelet approach is that 
for large training sets, the coefficient files can become very 
large and unwieldy. 
With sufficient training data, an effective classification model 
can be obtained using a polynomial kernel with the support 
vector machine. This classification model performs well in out- 
of-sample testing and has a success rate of more than 80% in 
correctly recognizing building image patches. 
While these techniques cannot satisfy the metric requirements 
of photogrammetry, they can provide useful starting points and 
heuristic filters in the area of automated object extraction. With 
some refinement, this method could be incorporated into a 
building extraction system as a heuristic filter and be used to 
ensure that only image patches with a high probability of 
containing a building were passed to the algorithms that 
performed the extraction. 
REFERENCES 
Agouris, P., Gyftakis, S. & Stefanidis, A. 1998. Using A Fuzzy 
Supervisor for Object Extraction within an Integrated 
Geospatial Environment. In: International Archives of 
Photogrammetry and Remote Sensing. Ohio, USA, 
XXXII(HI/1), pp. 191-195. 
Baltsavias, E. P., Gruen, A. & Van Gool, L., Eds. 2001. 
Automatic Extraction of Man-Made Objects from Aerial and 
Space Images (III). Zurich, A. A. Balkema. 
Bellman, C. J. & Shortis, M. R., 2002. A Machine Learning 
Approach to Building Recognition in Aerial Photographs. In: 
International Archives of the Photogrammetry, Remote Sensing 
and Spatial Information Sciences, Graz, Austria. Vol. XXXIV 
(3A), pp 50-54. 
Canny, J. F., 1986. A Computational Approach to Edge 
Detection. /EEE Transactions on Pattern Analysis and Machine 
Intelligence 8(6), pp. 679-686.