anbul 2004 
zorithm, a 
' noted that 
integrity of 
lesigned to 
cale colour 
model and 
| the height 
| of 54 and 
650 meters 
're scanned 
- The arca 
agricultural 
meters grid 
or had the 
efined grid 
TIN rather 
otograph of 
ly collected 
of different 
s have been 
"asurements 
heoretically 
equal to the 
a base of 90 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
mm (ISM, 1998). The precision of the measurements is less 
than half of the pixel size in ground units, hence better than 
the theoretical accuracy expected. 
4. TESTS AND VALIDATION OF THE PROPOSED 
METHOD. 
4.1 Application over synthetically distorted data. 
The first attempt to check the algorithm was based on an 
artificially distorted DEM. An inclined plane (fig. 3) has been 
calculated and added to the manually collected DEM, ranging 
from +4 to —5 meters. From this distorted DEM two 
orthophotographs were created, one from the left and one 
from the right photographs of the pair used for the manual 
collection of the DEM points. The orthophotographs were 
covering the aforementioned area (fig. 2) with a ground pixel 
size of 0.25 meters. 
The proposed algorithm has been applied on the two 
orthophotographs. À combination of a feature and area based 
matching algorithm was used incorporating adaptive template 
and elliptical areas instead of the standard square templates 
(Skarlatos, Georgopoulos, 2004). It is essential to note that 
the algorithm provides matched points in an almost 
predefined grid, but the feature extraction slightly distorts the 
grid (Forstner, 1986) so that to match in interest points. In 
addition it doesn’t interpolate, nor fixes the grid in the 
predefined positions, returning a TIN. Therefore it is quite 
often to have small gaps particularly in areas where the grey 
tones are smooth. 
The matching has been performed with a 25 pixel spacing 
(equals to 6.25 meters in ground units) and maximum 
template of 1521 pixels (equivalent to 39x39 square 
template), necessary for the quarry area, where there is no 
adequate grey tone variation. The resulting estimated 
corrections, filtered with a rotational filter (Sonka et al, 1993) 
can be compared with the artificial distortions (inclined 
plane) in fig. 3. It is obvious that the algorithm has correetly 
detected the pattern of the inclined plane. There are of course 
some undulations and the covered area is a bit smaller than 
the original, because the matching algorithms cannot match 
pixels on the edge of the image. The reasons of the 
undulations are the aforementioned problem of matching 
points close to the edges and the DEM interpolation, 
necessary for visualization of the differences, especially 
along the edges where missing points create problem in the 
interpolation. 
   
Wo : LR 
89 
Figure 3. The artificial difference applied to the original 
DEM, and the differences detected from the 
algorithm as filtered DEM. The peak is due to 
interpolation, since there were no points measured 
in the close vicinity. 
In order to verify that the algorithm is working properly, it 
Was necessary to apply the proposed corrections by the 
algorithm to the distorted DEM under investigation and 
CA 
cA 
compare it to the manually collected DEM. The algorithm 
returns X, Y, dZ coordinates in random positions. In order to 
compare it with the reference data, which are also in TIN 
format (manually collected) it is necessary to convert one of 
the two in DEM format. It is obvious that during this 
interpolation there is some loss of accuracy (Al-Tahir et al, 
1992). It was decided to convert the manually collected TIN 
in DEM with 5 meters grid spacing. This ensures that the 
deterioration will be hopefully held at a minimum. The 
corrected points from the algorithm are then compared to this 
surface and the differences from it are calculated and 
statistically analyzed. In order to reduce the mismatches of 
the matching algorithm, the mean (expectance zero) and 
standard deviation of the corrections is calculated and a 9596 
two-tail reliability check is performed on the corrections. 
This is more like an internal precision check, without any 
external data, which can be performed easily and does not 
affect the validity of the proposed algorithm. The remaining 
values are also statistically analyzed and presented in fig. 4. 
The corrected DEM (the filtered one) has a mean of —0.11 
meters, Standard Deviation (SD) 0.59 meters, Mean Absolute 
Deviation (MAD) 0.36 meters and Root Mean Square (RMS) 
error 0.60 meters. Two conclusions can be easily deducted 
from these measures: 
The mean of the corrected DEM is well under the precision 
of the manually collected points 
The MAD is equal to the expected accuracy (one pixel in 
ground units, ISM 1998). 
Difference of points Délsrence of poirts (25%) 
  
  
  
  
  
  
  
5 ; Wo 
1 H P. o Jl ! | 
i i ] 5 OR i í 
10 | : hh : Er ! of 
wl ; AE m ml A 7 
A i + à . . y 
IS. M, | - S 1500. 7. i ; 
1 teet 1500 im ^w, d n eode T5800 
» 277 can > c m 
£ - yo 
mn sn Sn _-" an 
a aV 
0 o 0 6 
Dif Distribution & Stat Dif Distribution & Stat (57$) 
1 n . 7 1 . ; . 3 
i 
Gg Ï 09 
cs il 0.0 | 
1 | 
a7} input Values: 36311 | 07} input Vaues: 4720 | | j 
Vslid Values; 36311 br percent 100 Valid Walues: 34738 jog percent 100 
O6} rmean Dis | 08} mean 091172 || 
trimmed maan 2X phases trimmed mean 2% 41097277 
DAP nbd dev 0738857 || 06} utd dev DAEBE | | 
mean absolute sodas 0.40094 mean absolute devidtien 0.36065 
QA rage ALLY | | D.4| range 174756 E 
skewness: -23464 | | skewness -2002 | 
03] kutosis: 316% | | 03! kunosis: 16.9165 | 
|} e 
02 {a 02 B 
i 
i 
01 J 0.1 hd 
/ \ 
l t a 
-0 5 0 5 -10 5 0 0 
Figure 4. Visualization and statistical 
analysis of the 
differences between the manually collected and 
the corrected by the algorithm DEM.