separate non-slum 
jects. Hence the 
he relative border 
the NB2 and NBI 
) image objects. It 
ngth of an image 
igned to a defined 
ative border of an 
ass is 1, the image 
ative border is 0.5 
of its border. The 
ed in pixels. The 
; used to separate 
S. 
  
shadow in built-up 
1m areas 
object level, it can 
1owledge to obtain 
ene or to aid in 
1s are tends to be 
se to river and 
efore Distance to 
vere also used to 
  
Id photos 
  
ining built-up was 
1 formal was first 
ts can be better 
understood. Since the informal settlements tend to cluster 
themselves, it can be differentiated easily from other formal 
areas based on their area. Therefore ‘area’ feature was used to 
identify the slums in eCognition environment. The false 
positives 1.e., non-slum areas looking like slums were 
eliminated by using adequate rules sequentially. The false 
positives are happening because of non-visible slums 
exemplified by planned but deteriorating inner cities (Turkstra, 
2008). The false positives were reclassified into non-slum 
areas, by using geometry of the object such as area, asymmetry, 
shape index and also contextual parameters such as distance to 
the object. Similarly some slum areas which were surrounded 
by non-slum areas also included in non-slum areas. To rectify 
this, relation to the border of the object and distance to 
parameters were used. So that the slums which included in 
non-slum areas were classified correctly Thus finally the image 
was classified (figure 8) into slums and non-slums along with 
other non- built-up areas. 
  
  
N 
0 250500 1.000 1,500 2.000 
UNE am Meters 
Figure 8. Classified Image of central part of Pune city 
3.4 Accuracy assessment 
The evaluation of a classification is a complex concept that 
includes the reference to several criteria. The main idea is to 
determine the accuracy of this classification by comparing the 
results with data provided from the reality in the field. These 
realities come from the slum survey carried out by city based 
NGO’s and Pune Municipal Corporation’s environmental 
status report(2010). 
  
  
  
User Class \ Serncde | Formal HO | sum 
Canfunion Mairie 
Formal & 36 657 
Informal B 40 46 
uralssetied 8 48 5 
Sum Es 124 
Accuracy 
PEN (3767302 03225308 
ser 9635552 
SIRE (346 (9635552 
Overall Accuracy 0.9714296 
KIA 0.5123278 
  
de 5 opm [Eee] 
Figure 9. Accuracy assessment 
  
Through field survey and primary data collected from slum 
dwellers, the slum map was created. The classification result 
was compared with slum map prepared based on Slum Survey. 
The overall accuracy is 87 % (Figure 9). 
4. CONCLUSION 
The issue of slums is very complex. Detecting slums might be 
one of the most challenging tasks within urban remote sensing. 
Though the present study demonstrated the advantage of VHR 
data and OOA approach in detecting the slums, it required 
local knowledge of existing slums and their characteristics. 
Using thematic layers such as roads and water bodies saved the 
time and reduced the complexity of rule set on extracting roads 
and water bodies from the image and thus helped to 
concentrate on detecting and discriminating slums from non- 
slums. One major issue in this analysis is false positives. 
Thorough understanding of study area is essential to develop 
the rule set to diminish the false positives. Detailed field visit 
is also essential to develop the rule set and can help to achieve 
reasonable accuracy. But complete removal of false positives is 
not possible in the inner/old city area because of its 
complexity. The present study area is mainly covering the 
indigenous city of Pune. So, complete clean-up was not 
possible and not done in the present study. 
Applying the same rule set to other scenes of quick bird data 
was also tried and the results are promising. But the threshold 
values for various rules such as brightness values, GLCM 
values etc. have to be modified as per the scene characteristics. 
Thus the present work may provide a basis for more advanced 
research to generalise a rule set which can be applied to 
various scenes of the same city and to various cities. 
5. ACKNOWLEDGEMENTS 
The author would like to thank Dr. Richard Sliuzas and 
Dr. Norman Kerle of ITC-Faculty of Geoinformation Science 
& Earth Observation, University of Twente, The Netherlands 
for their continuous support and valuable inputs during this 
work. Special thanks to Divyani Kohli and Deepti Durgi, (Ph.D 
scholars, ITC) for assisting in learning the eCognition software 
and sharing the Quickbird data. Sincere thanks to European 
Commission for providing Post Doctoral Research Fellowship 
to complete this work. 
References 
Angeles, G., Lance, P., Barden-OFallon, J., Islam, N., 
Mahbub, A. Q. M., & Nazem, N. I. (2009). The 2005 census 
and mapping of slums in Bangladesh: design, select results and 
application. [Article]. International Journal of Health 
Geographics, 8, 19. 
Baud, I., Kuffer, M., Pfeffer, K., Sliuzas, R., & Karuppannan, 
S. (2010). Understanding heterogeneity in metropolitan India: 
The added value of remote sensing data for analyzing sub- 
standard residential areas. International Journal of Applied 
Earth Observation and Geoinformation, 12(5), 359-374. 
Baud, I, Pfeffer, K., Sridharan, N., & Nainan, N. (2009). 
Matching deprivation mapping to urban governance in three 
Indian mega-cities. Habitat International, 33(4), 365-377.