International Archives of the Photogrammetry, Remote S 
in the same manner as for the conventionally orientated DEMs. 
Although additional checkpoints were available — those points 
that were used as GCPs in the conventional orientation — the 
configuration was not changed, meaning that the same 
assessment was made. Z differences were again taken between 
the DEMs and the checkpoints (Table 3), allowing the heighting 
precision graph to be revised (Figure 4). 
  
  
  
  
0.6 ——— = 
0.5 
e» ga | 
S k^ | 
= 03 | 
$ | 
à 02 | 
01 
i 
ed 
270 450 i 
Flying height (m) 
—e— expected —i#— conventional —#— matching 
  
  
  
Figure 4. Effect of flying height on DEM precision 
5. DISCUSSION 
It is interesting and useful to note that employing surface 
matching to the orientation problem had a positive influence on 
the precision of the resulting DEMs, the matching removing 
systematic errors present across the whole of the surface 
models. This suggests that for small area DEMs the method is 
more advantageous and efficient than using conventional 
photocontrol, especially if a reference DEM already exists, as is 
becoming increasingly common in an age of digital terrain 
modelling. Examining the resulting precisions of both the 
conventionally controlled and match-controlled DEMs suggests 
that the expected precisions of the aerial photography are 
optimistic for such low flying heights. However, the fact that a 
less steep line gradient exists for both test datasets means that 
intersection with the linear expected values may occur at some 
greater height; further data and testing would be required to 
confirm this. 
It is apparent from Table 3 that the conventionally controlled 
DEMs have systematic error affecting them (as indicated by the 
higher mean of differences), to a greater extent than the match- 
controlled surfaces. The cause of this is unclear but could be 
related to a host of factors such as the image configuration, poor 
bundle adjustment results, or an imperfect GCP distribution. 
Similar factors may affect the match-controlled surfaces, but 
the effect of surface matching is to minimise the height 
differences, removing bias 
It is noted that the surfaces extracted from the digital SFAP 
were more sensitive to small errors in the aerial triangulation 
stage. An example was seen during conventional processing: 
despite low image and control residuals, after adjustment and 
comparison with ground points a systematic offset of 0.5 m was 
present in the coordinate differences. When the triangulation 
was checked and modified by a slight adjustment to the tie 
point configuration, this error was reduced. The problem 
reinforces an important advantage of the surface matching 
approach — that of having an independent surface model that as 
well as being used in the critical registration procedure, also 
allows validation of the photogrammetric DEM. Additionally, 
34 
   
  
  
  
    
   
    
    
   
   
   
  
  
  
  
   
   
  
  
   
  
  
  
  
   
    
   
   
   
  
   
   
   
  
  
  
   
   
  
    
  
   
    
   
  
   
   
     
     
  
   
ensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
the DEMs created from lower scale imagery contained more 
noise than from the higher 600 m data, having connotations on 
the accuracy of the surfaces created. Consequently, more tests 
with conventional imagery at lower scales are required to 
determine whether the results are true for all scales. If so, the 
technique would be most valuable for photogrammetry, having 
the potential to reduce expenses incurred in collected ground 
control considerably, where a DEM already exists or can be 
easily acquired. 
The surface matching algorithm is not the superlative solution 
to the problem of DEM registration, as issues affecting the 
success of the technique remain. Because the two matched 
DEMs were collected using different terrain modelling 
techniques — GPS and digital photogrammetry — each model is 
in effect a wholly different representation of the same real- 
world surface. The DEMs have different structures, different 
point configurations, differing accuracies. Each of these 
introduces disparities, not only with each other but with the 
‘true’ surface. In addition, each of the measurement techniques 
records the terrain in a different manner. For example, 
kinematic GPS measures to the true ground height, as the detail 
pole is in contact with the terrain surface at all times, while 
photogrammetry images all vegetation and surface objects. 
Because of this, discrepancies are again introduced. Although 
not a problem in this research, the matching of DEMs collected 
at different epochs may again introduce change between the 
surfaces, causing further differences. All of these effects mean 
that in the least squares minimisation of height differences, it is 
not the same true surface that is being compared; rather, it is 
two similar models which therefore leaves the solution open to 
absorbing error into the output parameters, resulting in an 
imperfect solution. This may not be apparent from the output 
statistics, as the absorbed interpolation error will appear to 
result in a good minimisation of differences; however, 
parameter standard deviations may be optimistic (Maas, 2000). 
The most likely outcome of discrepancies influencing the end 
solution is that tränsformation parameters may contain error. In 
a similar way, because of the ill-posed nature of the matching 
problem, multiple solutions may be attainable, with the only 
change between matches being the choice of initial parameter 
estimates or outlier exclusion threshold. Multiple solutions may 
result in multiple parameter sets, and a significant problem is to 
determine how these affect the position in space of the 
transformed DEM. Small (or reasonably large) parameter 
changes may have little or no effect on the final position of the 
matched DEM. However, any change will mean that the 
position of the model will be moved and, though height 
precision may be higher, the planimetric position may suffer. 
Difficulties with assessing the planimetric accuracy of an 
essentially ‘featureless’ terrain representation occur, requiring 
ancillary data such as intensity values with airborne laser 
scanning DEMs (e.g. Maas, 2000), to ensure true conjugate 
surface patches are compared. Despite this pessimism, the 
effect of differing absolute orientations, with alternate 
configurations of GCPs, has not been fully investigated, leaving 
the registration to be, at best, an approximation of reality. 
6. CONCLUSIONS 
This paper has addressed the subject of photogrammetric DEM 
absolute orientation, and the attempt to improve the efficiency 
of this notoriously expensive and manual process. A surface 
matching algorithm was developed and used to minimise height 
    
International Arc! 
differences betwe 
DEM in a model « 
of the procedur 
conventional orie 
using digital SF 
imagery capturec 
extraction and or 
checkpoints was 1 
each method. 
orientated improv 
DEMs, while the 
larger systematic « 
These results are 
photogrammetry à 
systematic error, 
increase in autom 
data in the forr 
versatility, as data 
further research 
robustness of the a 
solutions on the al 
addition, the use « 
due to the scale 
configuration whe 
photography. T 
therefore be of gre 
A 
This work was car 
University of New 
Engineering and F 
grant GR/N23721. 
Surveyors (RICS I 
of the University - 
data collection. 
Bes, P.I and ? 
registration of 3- 
Analysis and Mach 
Buckley, S. and ! 
integrating the woi 
coastal erosion. Su 
Chandler, J.H., Sh 
2001. Measuring fl 
Kodak DSC460. Pj 
Flotron, A. and Kc 
civil engineering. 
pages. 
Hofmann-Wellenhc 
Global Positioning 
Springer, Vienna. 3 
Light, D., 2001. ; 
Photogrammetric E 
1299-1305.