km by 10 km to 80 km by 80 km. The case study area was 
divided into square grids as shown in Figure 3. The 
correlation was calculated between the number of bright 
pixels (rather then the digital number of pixels) in a grid and 
population in the same grid. In this manner, it was hoped 
that the problems associated with "saturated pixel" could be 
avoided. 
  
Figure 3: Case Study Area as divided into 40 km 
by 40 km grid squares 
The results of the analysis are shown in Table 1. These 
models (for square grids) apparently give better results 
(with R = 0.59 to 0.97), as compared with the above 
mentioned pixel-by-pixel correlation (between the 
digital number and population per pixel). It 1s thus 
suggested the number of "bright" pixels in a grid (in this 
case between 100 and 6400 sq. km.) obtained bv 
DMSP/OLS may serve as a better indicator of 
  
  
population. 
Gnd Size Coefficient N of Samples 
10km X 10km 0.59 1050 
20km X 20km 0.77 321 
40km X 40km 0.86 104 
80km X 80km 0.97 34 
  
  
  
Table 1: Accuracy of Estimated Population for 
Square Grids in Hokkaido 
CASE STUDY IN EAST COST OF U.S.A. 
Another case study was also conducted on the east coast of 
U.S.A., whichis composed of (part of) six states as shown 
in Figure 4. The area includes such large cities as 
Washington D.C., Baltimore and Philadelphia. For this 
study, "synthesized" city lights data, as mentioned in 
(Elvidge, 19972) was employed rather than the "one shot" 
imagery used for the above mentioned analysis in Hokkaido. 
Population data as developed by the U.S. Department of the 
Interior were obtained through the ftp cite of CIESIN. Only 
the number of bnght pixels in a grid was used to estimate 
and population in the same gnd. 
  
Figure 4: Nighttime Imagery of East Coast of 
U.S.A. 
The outcome of the study is as shown inTable 2. Thus case 
study showed a better results in terms of accuracy in 
estimating population, as compared with the study 
conducted for Hokkaido. The reason why two case studies 
showed some difference in accuracy of estimation is subject 
to further study. 
  
Gnd Size Coeflicient N of Samples 
10km X 10km 0.81 1608 
20km X 20km 0.86 440 
40km X 40km 0.92 126 
80km X 80km 0.93 37 
  
  
  
Table 2: Accuracy of Estimated Population for 
Square Grids in the east coast of U.S.A. 
A possible explanation is that the "synthesized" city lights 
data used for the U.S.A. case study of may represent the 
human activity of a region more accurately than the one 
time imagery applied for the Hokkaido case study, but it 
should be only proved by follow-up studies. 
458 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
RE 
Cro 
spac 
Elvi 
Nig 
Sys 
Sen: 
Hvi 
sett] 
cont 
Biol