Tl T5 T6
5 1.1 5 1.3 10 2.5
62.5 10 1.8 31.25} 10 2.2 15.625] 20 4.3
5 1.6 5 1:7 10 2.8
125 10 2.0 62.5 10 2.6 31.25 120 4.6
Ain T os Ax in T %s Ax nin T s
Table 1: Accuracy of progressive sampling and linear interpolation; all
values are given in meters
T1 T5 T6
5 1.35 5 1.58 10 2.91
62.5 10 2.26 31.251 10 3.023 15.625| 20 4.57
5 1.38 5 1.62 10 3.06
125 10 2.27 62.5 10 2.93 31.25 120 5.24
Ax in T Og AX in T Og AX in T Og
Table 2a: R.M.S.E. estimates from composed transfer function, a = 1/8
Tl T5 T6
5 1.0 5 1.3 10 2.9
62.5 10 1.6 31.25] 10 2.4 15.625| 20 4.0
5 1.4 5 1.4 10 3.1
125 10 1.6 62.5 10 2.4 31.25 120 4.1
AX in T Og AX in T Og AX in T Og
Table 2b: R.M.S.E. estimates from composed transfer function, a = 1/10
At least for the test sites used, the actual r.m.s.e. of the DEMs are
smaller, in general, than the estimates computed with o = 1/8, and
larger than those computed with o - 1/10. The estimates obtained with
the composed transfer function using oa = 1/10 are closer to the actual
r.m.s.e. than those with a = 1/8. Comparing table 2b with table 1, we
find the largest discrepancy for Tl with a minimum grid spacing of 125 m
and T = 10 m. There the power spectrum of the profile is computed from
only 96 values (12000/125); thus it is likely to be affected by alias-
ing, resulting in inferior estimates (see /11/). Whether composed
transfer function and power spectrum of a profile also yield fair DEM
accuracy estimates for real terrain must be verified by further
investigations. If so, a simple procedure would be available for ac-
curacy assessment, not requiring additional measurements. Computing the
power spectrum by the Fast Fourier Transform algorithm relies on an
equispaced sample. A profile extracted from a DEM created by PS, must
- 661 -