IAPRS & SIS, Vol.34, Part 7, "Resource and Environmental Monitoring", Hyderabad, India, 2002 
  
been developed for hydrological analysis, 
(http://cres.anu.edu.au/outputs/anudem.html). ^ Hutchinson 
(2002) discusses a locally adaptive approach to the 
interpolation of DEMs and indicates the importance of not 
loosing information when interpolating from an irregular 
network to a grid representation. 
Takagi (1996) has investigated different methods and their 
suitability for different applications of DEM. The methods 
he has investigated are maximum value, minimum value, 
mean value and nearest neighbour methods of resampling, 
which were supported in the software package GLOBE. For 
slope aspect accuracy he found that DEM generated using 
maximum value method provided better accuracy and for all 
other applications such as slope inclination accuracy, 
accuracy of derivation of drainage pattern nearest neighbour 
method was more suitable. 
Schneider, (2002) from the University of Basel, discusses 
the representation of DEMs as surfaces. In his method the 
original data is preserved but surfaces are described by an 
interpolation method, which may be selected to suit a 
particular surface. 
The techniques of item response theory, used to analyse 
questionnaires by statisticians, might be used. This 
technique aims to separate out errors, rather than lump them 
together; it inspects the theoretical knowledge of causal 
links amongst variables, test empirical data against 
computed values. The outputs are submitted to Linear 
Structural Relation Modelling that further segregates and 
breaks down the components. The resulting variances aid in 
quantifying absolute and relative errors. 
4.4 Data fusion 
The opportunities for data fusion are greatly increased as 
more sensors are launched and data becqmes more easily 
available and often less costly. If more than one data set is 
available then solutions have been proposed for exploiting 
any synergy that is present. (Honikel, 2002, Hahn and 
Samadzadegan 1999). Fox and Gooch (2001) have proposed 
generating 2 DEMs from the same data to improve the 
result. 
A useful discussion on data fusion is given by Honikel 
(2002) who recommends three steps: 
1. Data alignment during which all data is transformed 
to the same reference system in the same units. 
2. Data association during which data is grouped and 
edited so that common points are merged and 
erroneous points are removed; 
3. Estimation during which a final DEM is created 
which best fits to the multiple observations. 
Honikel is concerned with fusing ERS IfSAR data with a 
DEM from SPOT and demonstrates how the synergy of 
these two data sets can be exploited to make use of the 
strengths of both sets of data to give a DEM that is better 
than either of the initial sets of data. In his case the SPOT 
DEM is used to improve the phase unwrapping and to 
remove systematic trends; associated data is used to remove 
blunders and get a better estimate for points to which more 
than one observation refer, and by working in the frequency 
domain the strengths of both data sets can be combined. 
From this we can propose some generic techniques: 
e Fusion of data with different spacing (quasi or true 
grid) to get a better estimate of individual points. 
Examples: Hahn and Samadzadegan (1999) use wavelets to 
combine DEMs of different resolution and accuracy. 
e Fusion of data that has different qualities. 
Examples: IfSAR can be very accurate where coherence is 
high and SPOT can be accurate where correlation is high, 
either can perform better on particular types of feature 
depending on aspect, time difference etc. Stereo SAR might 
be used with SPOT for similar reasons. 
LIDAR data might give an indication of where buildings or 
trees occur, which could control matching of optical data. 
e Fusion in a coarse to fine strategy. 
Examples: A coarse DEM can be sufficient to give initial 
values to generate a fine DEM and to indicate areas where 
problems might occur. 
A coarse DEM can assist with phase unwrapping of IfSAR 
data and remove trends due to atmospheric effects or base 
line errors. (Honikel 2002) 
e Fusion of different types of data. 
Examples: Use of rivers, spot heights, lakes, breaklines etc. 
within the matching process. 
Two general questions remain to be answered in respect to data 
fusion: at what stage should fusion take place? And how to exploit 
the full information from an irregular network? 
5.0 FUTURE STRATEGIES 
In COMET it is intended to use 3 main sources of optical 
stereoscopic images from space. At present these are SPOT and 
ASTER, from which data is being collected now and there is also a 
large archive. Both are available and inexpensive for scientific 
use. In 2004 data from the ALOS PRISM sensor will also be used. 
In order to test some of the ideas set out above some experiments 
have been carried out with SPOT and ASTER data. 
The strategies that will be followed in the COMET project are as 
follows: 
e Compare existing packages to determine which are most 
suitable for landscape evolution using the following criteria: 
Accuracy; 
Flexibility; 
Ability to modify the software. 
e Improve stereomatching techniques to allow fine detail to be 
extracted where it is needed, making use of breaklines and 
iterative strategies. 
e Develop new interpolation algorithms, 
stereomatching algorithm. 
e Develop data fusion techniques using existing data sets, such 
as SRTM or GTOPO30 that will allow better matching and 
interpolation from the higher resolution optical data. 
linked to the 
It is hoped that these strategies will produce more efficient and 
more accurate DEMs that can be used in terrain evolution studies 
and other applications. 
REFERENCES 
Ackermann F, 1980. The accuracy of digital height models. 
proceedings of 37th Photogrammetric Week. University of 
Stuttgart. 234 pages:113-143. 
Caner H, 2001. Improving the accuracy of automatically generated 
DEMs. UCL MSc project report. 96 pages. 
620