process with minimum variance. Theoretically, the variance
of the prediction errors decreases with increasing orders p
and q. Hence this críterion alone is not sufficient. There-
fore, Akaike /3 / introduced his information criterion,
which determines the order by minimizing the function
AICrerine,? + 2(p+q) / N
A further criterion to be considered is the significance of
the parameters a and b. The maximum order p and q is
determined by the number of significant parameters.
6. Modelling Digital Terrain Profiles
The statistical properties of ARIMA processes and their
relations to autocorrelation functions and to spectral
analysis justify to consider ARIMA processes as an adequate
model for digital terrain profiles. Here the process
orders, the statistical properties of the prediction errors
and the variations of the process parameters within the
same type of terrain are of main interest. To achieve this
we studied terrain profiles from different regions. The
experiments comprised 52 profiles out of 9 regions. Most
profiles were measured on the occasion of the ISPRS-Test of
Comm..III/3.-(:.Torlegárd /15/ ) and sare described in
Frederiksen /7/ and Jacobi and Kubik /9/. Table 1 gives a
survey of the terrain profiles analysed.
For process identification every profile was modeled by
ARIMA processes of increasing order. Stationarity could be
reached with the first derivatives (i.e. the slopes of the
profile) on flat terrain. The second derivatives (i.e. the
curvatures) are sufficiently stationary in all other
profiles.
With regard to the process order, the empirical investiga-
tions established that the profiles of a local region
(similar type of terrain) can be modeled by the same
process order. Table 2 records process orders established
for the different types of terrain. It is a remarkable
result that the process orders are very low. Consequently a
very limited number only of parameters are required to
describe the profiles. A maximum order of p+d+q = 6 is
sufficient for all cases considered.
Besides, table 2 includes the statistical properties of the
prediction errors: the mean values of the standard devia-
tions, the expectations and the correlation coefficients.
Within a given region the standard deviations of the pre-
diction errors are very homogeneous. The expectations and
the correlation coefficients are not significant in any
case. Therefore the prediction errors may be described by a
white noise process. Consequently terrain profiles satisfy
a condition of ARIMA processes.
- 433 -