International Archives of the Photogrammetry, Remote S ensing and Spatial Information Sciences, Vol XXXV , Part B2. Istanbul 2004 
  
Interna 
as well, and not only on the ground pixel size (which is often hours c 
used wrongly as synonymous to image resolution and image running 
quality). | configu 
| process 
Table 2 also shows that all quality measures are gradually | large u 
deteriorating with decreasing pixel size. One cause for less | km ofr 
completeness is that paths in fields are only partially extracted | 
because the path surface is blurred and road edges are very 4.2 Gi 
weak (Fig. 6), while in small villages performance was also 
worse (Fig. 7). However, the quality deterioration is much less Our sy 
than the pixel size reduction. E.g. for 60-cm vs. 20-cm pixel (Swissi 
size, in the first case we have 9 times less data, but Howev 
   
  
   
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
completeness, correctness and accuracy deteriorate only by HRS d 
i : 7.5%, 4.5% and 40%. On the other hand, this slight quality | image, 
Figure 3. Examples of road extraction and junction generation decrease may still mean expensive additional manual editing, so narrow, 
in scenes with small settlements. the question of pixel size choice should be carefully considered. i Sys 
e inc 
Quality measures | Aerial | Aerial | Aerial | ADS40 visible 
20cm | 50cm 60cm 30cm edges 2 
Completeness (%) |95.44 [90.49 |8828 95.27 compar 
Correctness (%) [94.65 |92.53 190.68 19424 comple 
Length of 42.08 [42.08 [42.08 42.08 rem 
reference (km) Qu: 
Length of 4243 |41.15 140.97 42.54 | meas 
extraction (km) [Compl 
RMSemor | x [045 [0.60 [0.72 1.00 onan 
(m) y (0.43 0.61 0.75 0.97 Length 
Mean error x [0.04 0.17 0.29 0.57 peterenr 
(m) y [008 [0.22 1019 0.44 Length 
Processing time (s) | 1872 |774 |620 1184 exten 
RMS 
Table 2. Quality evaluation of the results in Thun site. error (mn 
3 id] Mean 
Figure 4. Examples of road extraction and junction generation error (n 
in scenes with small settlements.  Proces: 
Tab 
It is ap 
depend 
Figure 6. A path extracted in 20-cm pixel size image (a), but too, an 
only partially extracted in 60-cm pixel size image (b). within 
The black lines are the extracted results and the white influenc 
lines are the reference data. conditio 
much le 
to airbo 
| object d 
is simil: 
| than In 
caused 
coordin: 
ic | 8 system. 
Figure 5. Results in a scene with a first-class road. | is simil: 
Figure 7. A village road extracted in 20-cm pixel size image (a), | HRS oi 
Our system delivers the best results with the 20-cm orthoimage. but not in 60-cm pixel size image (b). The black lines | DSM/D 
About 95% of the roads are correctly extracted with an are the extracted results and the white lines are the | 8 shows 
accuracy of about 50 cm. The non-extracted or falsely extracted reference data. from the 
roads are mainly in small villages. Taking into account the = = | 
discrepancy between the ADS40 and the 20-cm film Table 2 shows that the processing speed of our system is high In the ( 
orthoimage (as indicated by the large mean values), our system (e.g. ca. 30 minutes for more than 40-km roads in 20-cm pixel inside t! 
performed equally well on ADS40 data. Indeed, visual check size images), and that processing time decreases almost 1:1 are class 
over the whole test site shows that the results from ADS40 data With orthoimage pixel size. The processing time also depends those in 
ADS40 are sharper and have a better radiometric quality scene, increasing with them. Extensive tests at swisstopo with fields is 
compared to scanned film. Image quality is important in general 50-cm orthoimages show that roads In an average road density false exi 
and for object extraction, and depends on such image properties [:25,000 map sheet covering 210 km? can be extracted in 3-4 from th 
744 
| 
2 % - > ro: densi accer extent o0 ov iv he | us C 
are actually at the road centers. Furthermore, images from on the road density and to a lesser extent complexity of the | Fig. 90