iture 
d as 
it is 
ures 
ons. 
" 
2" or 
hich 
but 
the 
y per 
Iracy 
f the 
s the 
s the 
issed 
ucers 
jages 
re no 
nly a 
e not 
to be 
  
  
  
Such data (like the zero for bridges on this half of the test area) 
had to be eliminated from the analysis. By doing so the sample 
size for this feature in this test was reduced by approximately 
50%, which makes the data less significant. 
Fortunately most of these blunders concerned only small sam- 
ples, so that the OPA for the feature groups are practically not 
changed by the elimination of those blunders. 
The overall producers accuracy (OPA) shows a clear increase 
with a reduction of the pixel size (as can be expected). When 
looking to the pixel size on the ground we find the following: 
the résults for the 1.8 m ground pixels are practically the same 
for the two image scales, whereas for the 0.9 m ground pixels 
the smaller scale photography gives the same result only for line 
features. It is slightly inferior for point features and clearly 
worse for area features. 
This indicates, that the step in ground pixel size from 0.9 m to 
1.8 m has a much bigger influence, than the step in photoscale 
from 1:30,000 to 1:60,000. Only for the area features the change 
of photoscale has a significant influence on the interpretability 
of images scanned with a ground pixel size of 0.9 m. 
The missed rate (M) is a good measure for the amount of mea- 
surements required in the field completion. In our tests it is in 
most cases 100% - P. This shows, that there was not much con- 
fusion between the considered features, as the number of detec- 
ted but misclassified features is 100% - P - M (See "fault rate" 
in chapter 2.7). Misclassified features would only require field 
identification, not measurement. 
The users accuracy (U) is very low, indicating that many of the 
digitized features were wrongly included. This is no big problem 
for the field completion, as it only requires changing the classi- 
fication or deleting the feature. 
point features are strongly dominated by the isolated small 
houses. The missed rate, which varies between 7696 (1:60,000, 
60 um) and 32% (1:30,000, 15 um) shows the considerable 
amount of additional field measurements needed. The (very low) 
user accuracy is probably unrealistic. It can be caused by the 
limitation of the test to houses, which are small and single. 
Many of the features, which were wrongly digitized as "small 
single house" may still be houses, but the fact, that they do not 
qualify for "small" or "single" might have been undetectable. In 
the reference data houses larger than 120 m? were digitized, and 
the corresponding houses in the test data omitted, but this did 
not catch all of these "errors". 
line features show very diverse data. The single track railway 
was always identifiable and the result on single lane roads was 
also rather good (between 5796 and 8596). On the other hand is 
the interpretability of foot paths and ditches strongly depending 
on the ground pixel size. The producers accuracy varies between 
zero for the 3.6 m pixels and 8596 for the 0.45 m pixels. 
area features: vineyards, which are the dominating feature in 
this group are usually identified by a pattern of tiny parallel 
lines, less than 1 m wide on the ground (the lines of wines and 
their shadow) spaced less than 3 m. In this context the success 
rate of more than 30 % for the 1.8 m ground pixels is more sur- 
prising than the zero for the 3.6 m ground pixels. 
3.2 mono interpretations 
The second part of the study, which compares mono interpreta- 
tion of 1:50,000 orthophotos with unaided eyes to stereo inter- 
pretation of the analogue aerial photographs with considerable 
optical magnification is summarized in table 5. 
Point features were not detectable in these tests, this feature 
group is not shown in the table. From the context can be seen, 
where single houses and bridges should be found, but they could 
not be positively identified. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
photoscale 1:60,000 1:30,000 
image pixel size 60um (3.6m) 30um (1.8m) 15pm (0.9m) 60pm (1.8m) 30um (0.9m) 
ortho pixel size 100um (5m) 100um (Sm) 50um (2.5m) 20um (1m) 50um (2.5m) 
Feature P U M P U M P U M p U M P U |M 
footpaths 27.1. 44 ] 73,.].44 1. 52 1-51 | 85 .1759.] 40.1.53 ] 45.1 47 1.43 |]. 36 | 48 
single lane road | 32 | 47 | 68 | 53 | 66 | 44 | 65 | 64 | 33 | 67 | 51 | 33 | 40 | 48 | 59 
ditches 33 | 43 |.67.|.52 1.66.1 48 | 55 | 67 | 45 | 52 771 a8 | 52 | 63 | 4% 
OPA-line 31 51 62 63 41 
vineyards 0 - 1001 26 | 61] 63 | 37 {31 63-|40 [#55 1748 9 39 189 
orchards 48.1043 1552.1:56.| 431 44 | 59 155 | 31 | 52. 1.43 :| 39.7 33 |. 39 | 59 
OPA-area 14 38 43 47 17 
  
  
  
  
  
  
  
Table 5: Global results of the mono interpretations 
For line- and area features the results are surprisingly close to 
the stereo interpretations. In the orthophotos prepared from the 
photography 1:60,000 line- and area-features show, that for the 
Output ground pixel of 5 m reducing the input pixel size from 
36m to 1.8m gives a considerable improvement. This 
contradicts the assumptions made by Leberl (1992) and by 
Schiewe and Siebe (1994), that the pixel size on the ground can 
be the same for the input and the output images, and supports 
the final suggestion of Schiewe and Siebe to use a smaller input 
pixel size. 
Reducing the output pixel size from 100 um (5 m) to 50 um 
309 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996