International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
7 0.06 
0.05 
0.04 
0.03 
5 0.02 f 
0 T F T T 1 ¥ T 
ON AR d oV o o S dE P m o gu e ur A LS a a o o 
S V a a a D a SE FD SS a SU I X SS STD TFS e 
BEN FT TR SS SS us FESS e. ELF FF 
Q S qe SS NV > S ES A 7 > > v v v o 8 M e / ? [C] 
s Ff FLEETS ELSES L AT eS e at o 
& wx f N RT CS al f f A : S 
N SP OP as AO a qu a OY o p SP o? 
N.N a. OU SD ON eT Or Or 2° CC 
Qu oS a M CT DS SS SG dq ov € 
T & SS N P EL & 
ES N LL 
NC 
Figure 3. Random Forests-based feature importance 
REFERENCES 
Arefi, H., Hahn, M., and Lindenberger, J., 2003. Lidar data 
Iterative Feature Selection classification with remote sensing tools, Proceedings of the 
0.08 ISPRS Commission IV Joint Workshop: Challenges in 
> : Geospatial Analysis, Integration and Visualization II, 08-09 
& 0.07 September, Stuttgart, Germany, pp. 131-136. 
> 
2 0.06 Breiman, L., Friedman, H., Olshen, A. and Stone, J., 1984. 
o 0.05 Classification and Regression Trees. Monterey, CA: Wadsworth 
= & Brooks/Cole Advanced Books & Software. 
9 0.04 
9 0.03 Breiman, L., 1996. Bagging predictors. Machine Learning, 24 
a 
= (2), 123-140. 
a 0.02 
z 0.01 Breiman, L., 2001. Random Forests. Machine Learning, 45, 5— 
32. 
48 .39 | 36... 30 26. 221.15 
Feature Numbers Chehata, N., Guo, L. and Mallet, C., 2009. Airborne lidar 
  
  
  
Figure 4. Iterative feature selection 
4. CONCLUSIONS 
In this study Random Forests is successfully applied to the 
feature selection for land-use classification. There are 48 
features extracted from lidar data and imagery. Making use of 
the Random Forests, an assembling classification tree, that 
provides feature importance index, we iteratively eliminate 
features with less important index until the mean decrease 
accuracy is stable. The extensive experiments are conducted to 
describe the Random Forests' characteristics and prove its 
performance. Classification results suggest that much more 
feature cannot guarantee the improvement of classification 
accuracy, and confirms that the selected features can obtain the 
satisfied classification results. Overall, the classification results 
indicate that the selected features agree the existing 
physiological knowledge. 
    
feature selection for urban classification using random forests. 
In: F. Bretar, M. Pierror-Deseilligny & G. Vosselman, eds. Laser 
Scanning, IAPRS, 38(3/W8). s.l.:s.n. 
Congalton, G., 1991. A review of assessing the accuracy of 
classifications of remotely sensed data, Remote Sensing of 
Environment, 37(1): 35-46. 
Congalton, G. and Green, K., 2009. Assessing the Accuracy of 
Remotely Sensed Data: Principles and Practices, Taylor & 
Francis, Boca Raton, FL. pp. 85-89. 
Diaz-Uriarte, R., and Alvarez de Andres, S. 2006. Gene 
selection and classification of microarray data using random 
forest, BMC Bioinformatics, 7(3). 
Gislason, P. O., Benediktsson, J. A. and Sveinsson, J. R., 2006. 
Random Forests for land cover classification, Pattern 
Recognition letters, 27:294-300. 
Guo, L. Chehata, N., Mallet, C. and Boukir, S., 2011.