The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
288 
ADS40 Forest straight edge 
Threshold 
Quantize Window green=3x3, blue=5x5, red=7x7 
Figure 6. Threshold vs. CBO of "straight forest edge" 
qw=7x7 th=0.65 CBO=11.2 
¿2» 0 
ET 
/X * ¿¡W JX 
J 
qw=7x7 th=0.95 CBO=9.8 
Figure 7. CBO "straight forest edge" 
Figure 6 reveals a non-linear characteristic of the error measure 
CBO=f(th,qw) as in Figure 3, but the regression line for all 
quantization window sizes shows that higher threshold values 
yield lower errors for any window size. Nevertheless 7x7 
window size seems to be generally closer to the ground truth 
than smaller sizes. 
qw=7x7 th=0.25 CB028.5 
qw=7x7 th=0.60 CBO= 14.8 
5. DISCUSSION AND CONCLUSION 
Parameter evaluation with the city block optimization (CBO) 
has its main strengths due to its robustness and when only 
limited ground truth is available. By using different ground 
truth examples, CBO allows to optimize segmentation results 
with a robust and simple metric. The evaluation in the image 
coordinate space avoids also quantization problems like sliver 
polygons, which are often encountered as side effect in the 
result of vectorization algorithms. 
The minimum of the CBO values represents the best 
segmentation compared to the corresponding ground truth. But 
varying illumination within the NFI image database limits the 
ground truth comparison to images with a similar lightness 
distribution. The color quantization of the used JSEG 
segmentation is very sensitive to strong lightness variations. 
The distribution of the parameter values (Figure 3 and 6) can be 
used to estimate optimized threshold values and to narrow down 
the initially wide parameter range of values, but it is also 
evident, that the non-linear function characteristic does not lead 
to a closed solution. At least a narrowed parameter range of 
values allows in a further process to search the parameter space 
with randomized variations (what we initially tried to avoid 
mainly due to computational time constraints), which should 
yield an even better optimization result.