bsolute 
  
  
  
  
  
  
  
  
  
ng (left) 
‘absolute 
umber of 
  
  
grid points used for DEM matching and based on that, we 
further examined the feasibility of automatic DEM matching 
using a large quantity of grid points extracted automatically 
over the whole area. Table 2 shows the accuracy of absolute 
models from DEM matching with different number of grid 
points. The accuracy was assessed using true GCPs, 18 points 
as shown in table 1. 
Firstly, the different number of selected grid points were 
tested by removing outliers though visual checking and it is 
observed that more reliable visually checked grid points lead to 
more accurate DEM matching results. Continuously, the 
potential of automated extraction of grid points was tested for 
DEM matching. We extracted automatically about 6,000 points 
at a regular grid interval. The accuracy of the absolute models 
from DEM matching with the automatic gird points was shown 
in table 2. Slight accuracy degradation was observed. Since 
there are limitations on manual acquisition and visual checking 
of many grid points for DEM matching, grid points generated 
from automatic process are preferred. Moreover, automatic grid 
points selection is a meaningful process in that it leads to 
automatic mapping of satellite image without ground control 
points. 
Considering several DEM matching error sources such as 
resolution difference between two DEMS, respective errors on 
each DEM, in particular the reference DEM, we could expect 
that fully automated grid points including outliers may lead to 
accuracy degradation of DEM matching. Nevertheless, it is 
observed that DEM matching still stably worked to SPOT-5. 
Through automatic DEM matching, we generated a DEM at 
grid spacing of 10m based on absolute orientation using 
automatic grid selection. The DEM is shown in figure 3. We 
carried out accuracy assessment further by comparing the 
reference dataset. The DEM extracted from digital map on a 
scale of 1:5,000 were employed as reference data and it is 
shown in figure 4. 
  
Figure 3. SPOT-5 DEM from automatic DEM Matching 
  
Figure 4. The reference digital map and extracted DEM 
We compared about 3,000 points of the resultant DEM with 
corresponding points of the DEM extracted from the digital 
map. Height differences between the two DEMs were calculated 
by Mean Absolute Errorr«:MAE) and Root Mean Square 
Error(RMSE). Our DEM yielded MAE of approximately 7 m 
and RMSE of approximately 12m. Overall, through the DEM 
matching experiments, we confirmed the feasibility of reliable 
DEM generation without ground control points. 
4.2 Applicability of the Existing 90m Global Elevation Data 
There exist several difficulties coming from differences of 
resolution between the relative DEM and reference DEMs such 
as the errors contained in the reference DEM when applying 
DEM matching. Considering these factors, DEM matching 
looks more challenging when using much lower resolution 
elevation dataset. For testing this feasibility, we here extended 
our DEM matching technique by using the existing 90m global 
elevation data. If it is proved that the global DEMs can be 
sufficiently used for DEM matching, our proposed method can 
be used as an efficient solution to absolute orientation of high 
resolution satellite image without ground control points. 
Using SRTM-derived DEMs(DTEDs) with gird spacing of 
90m, we performed DEM matching with automated grid points 
selection and check the accuracy of absolute models. Table 3 
shows the accuracy with 90m elevation data and the previous 
30m elevation data. We could achieve the accuracy of about 3 
pixel on the image space and of about 9m in horizontal and 2m 
in vertical direction on the object space. Although the issue 
whether this is sufficient remains, the result reflects that our 
DEM matching worked well for the existing elevation data at 
90m spacing. This is very encouraging in terms of attaining 
wide applicability with the existing 90m resolution elevation 
data available worldwide. 
5. SUMMARY AND CONCLUSION 
In this paper, we exploited the DEM Matching technique 
further using existing global elevation dataset for absolute 
orientation of high resolution satellite image without ground 
Table 3. Automatic DEM matching results using 90m existing global elevation data 
Left model errors (pixel) 
  
Right model errors (pixel) Object space errors (m) 
  
  
  
  
Col Row All Col Row All Horizontal Vertical 
30m existing elevation data 0.62 2.12 2.21 1.41 1.66 2.18 5.30 3.98 
90m existing elevation data 2.60 2.44 3.57 2.36 2.10 3.16 9.03 1.99