Figure 1: 
  
   
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B8, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
The state of 
Maharashtra as obtained from two DMSP-OLS images of 2001. (a) Maharashtra shown using the stable lights dataset. (b) Maharashtra 
shown using the radiance calibrated dataset (showing brightness values) 
percentage of households with cars, jeeps and vans; percentage 
of households with television; percentage of permanent census 
houses and percentage of households using electricity as power 
source. Maps are produced for the villages in the districts of 
Pune in the state of Maharashtra. The use of multi-scale data led 
to the consideration of issues arising from MAUP and 
ecological fallacy which are also described in this paper. 
2. METHOD 
The study uses two types of DMSP-OLS data products: the 
stable light data set and the brightness data. The stable light 
data was part of the latest average DN data series and was 
obtained from the National Geophysical Data Centre (NGDC) 
website (National Geophysical Data Centre 2006). In this 
image, the data values range from 1-63. Background noise in 
the data is represented using zero while areas with no cloud-free 
observations are denoted by the value of 255. The second 
DMSP-OLS image used in the study is the global composite of 
brightness data for 2000 — 2001. It was prepared from fixed 
gain images taken from satellites F12 to FI5 by NGDC. 
However, this data contained brightness values ranging from 0 
to 653 and was not calibrated to radiance (Tuttle 2008). 
The mean and standard deviation of stable lights and brightness 
were calculated for 32 districts and all the taluks in the state of 
Maharashtra. There are 35 districts in the state of Maharashtra. 
Of these, the districts of Mumbai, Greater Mumbai and Thane 
were not included for sample selection as they had very high 
values of both mean and standard deviation of brightness and 
stable lights compared to others. From the remaining 32 
districts, 24 were randomly selected and 8 districts were 
withheld for model validation. Although the census accounts for 
354 taluks, data was available for only 286 taluks. As a result 
the analyses were conducted on the available taluks. 196 taluks 
were randomly sampled for model development and the 
remaining 90 taluks were withheld for model validation. 
Five demographic metrics and four socio-economic metrics 
were chosen from the census. Ten metrics were shortlisted after 
a number of statistical tests and are listed in table 1. 
From the ten census metrics selected for this study, only three 
variables were available from the Indian census at the scale of 
a village. They are number of households per square kilometre, 
total population per square kilometre and total workers per 
square kilometre. The models proposed at the district and taluks 
were used to predict and map the metrics for the villages that 
are unavailable from traditional census statistics. These metrics 
include: number of female literates per square kilometre; 
percentage of households with cars, jeeps and vans; percentage 
of households with television; percentage of permanent census 
houses and percentage of households using electricity as power 
source. Maps were produced for the villages in the districts of 
Pune in the state of Maharashtra. The district of Pune has the 
million plus city of Pune as its district headquarter along with 
some very rural areas appearing dark in the satellite image. 
3. RESULTS AND DISCUSSION 
3.1 Models proposed at district and taluks 
Linear regression models and multiple regression models were 
proposed. The selected census metrics were chosen as the 
dependent variables and mean and standard deviation of 
brightness and stable lights obtained from the images were used 
as the independent variables. The models were validated using 
the withheld districts and taluks. The models which best 
predicted the census metrics (€ 2596 error margin) for the 
highest number of districts and taluks were identified as the 
most appropriate models (Roychowdhury et al. 2011b; 
Roychowdhury et al. 2010). 
The selected census metrics showed positive correlations with 
both the mean and standard deviation of brightness and stable 
lights. The adjusted r^ of these models ranged from 0.8 to 0.97 
at 95% confidence interval at the district level. The correlation 
coefficients (r)) achieved at 95% confidence interval for all the 
census metrics ranged from 0.2 to 0.8 for the taluks. The 
adjusted r? values of the models are presented in details in 
previous works by the authors (Roychowdhury et al. 2011b; 
Roychowdhury et al. 2010). 
  
   
Internati