International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
small slope, significant correlation score. These points are
sampled and triangulated in order to produce a dense DTM.
Superimposing with this DTM all aboveground objects defined
by external vectors then produces a “clean” DSM. Prior to
superposition, the aboveground object elevations are
preprocessed (noise filtering, interpolation) in order to produce
dense data which exactly matches the external vectors. Finally,
additional "significant" aboveground objects (trees, sheds, etc)
can be detected and added in the DSM.
3.2.4 Self-evaluation
The DTM accuracy is assessed using reference vectors selected
in the beginning of the process. The DSM accuracy is assessed
using the input 3D points associated to each building polygon.
4. PERFORMANCE ANALYSIS
The accuracy of the DTM and the DSM produced with both
manual and semi-automatic approaches has been assessed. In
particular, we analysed the following aspects:
e Influence of the accuracy of input vectors,
e Influence of the detail level of input vectors,
e Influence ofthe number of input breaklines,
e Estimation of the capture cost.
4.1 Description of the test data set
Three test areas are presented, with various scales and
significant height variations. Each area is approximately
2000x2000 pixels, with a content representative of urban areas
(see Figure 8). Detailed characteristics are given in Table 1.
Name Scale Xyre | Zres | Zmin | Zmax (m)
S (m) | (m) | ground | build
(m)
Deauville 1:15000 | 0.21 | 0(36 1] 0 33 46
Kerlaz 1:20000 | 0.28 | 0.53 0 47 54
Le Havre 1:25000 | 0.32 | 0.56 | 4 41 56
Table 1: characteristics of test areas
For each area, a set of 3D vectors was produced by
photogrammetric capture, according to the rules mentioned in
section 2. Various DTM and DSM were computed for each
area, with both manual and semi-automatic approaches (see
examples of semi-automatic DSM in Figures 9a, 9b, 9c).
Raster accuracy is assessed during the self-evaluation stage of
AutoDEM (see section 3), but also using complementary
independent vectors captured for this purpose: ground points
(20m grid) and 3D points located anywhere on building roofs
(one point per roof).
The accuracy is estimated with the average error (Avg), the
error standard deviation (Std), the root mean square error
(RMS), the maximal error (Emax), and finally the percentage of
“reliable” points (%Pts+/-1), characterized by a measured error
within [-1;1]. NbPtsRef is the number of reference points used
for assessment. The production cost is estimated with the total
number of captured points (NbPtsUsed); this estimation gives
an indication about one particular stage of the process, but it
can not be representative of the whole cost.
4.2 DTM Production
The DTM accuracy was assessed by varying the breakline
number, complexity and agguracy:
e Random selection of 0% to 100% of available
breaklines, or exclusive use of the main road network,
e Polyline simplification,
e Random perturbations on Z-values.
4.2.1 Influence of the number of breaklines
Manual approach. The accuracy of the "manual" DTM is
presented in Table 2. By using all available breaklines, the
RMS is always below 0.85cm, with 85% to 95% reliable points.
The evolution of RMS with the number of breaklines shows
that a small reduction does not significantly decrease accuracy
(see Figure 5). However, by using only the road network
(around 35% of the breaklines over test areas), the percentage
of reliable points goes down to 75%, with a RMS between |
and 1.70m (see Table 3).
Semi-automatic approach. |t gives similar results to the manual
approach when using all the breaklines (see Table 2). However,
when using less breaklines, the semi-automatic approach
generally improves accuracy (see Figure 6 for a comparison). It
is particularly true with the road network only (see Table 3): it
is then possible to get a RMS around Im and a proportion of
reliable points between 77 and 89%, for a capture cost of 65%
less than for the traditional manual approach with all the
breaklines.
AREA Deauville Kerlaz Le Havre
Cost 4027 9553 2884
(NptsUsed)
NptsRef 378 684 934
DTM Man | Auto Man Auto Man Auto
Accuracy
Avg 0.39 | 0.42 0.20 0.21 0.48 0.51
Std 0.73 0.72 0.48 0.61 0.69 0.68
RMS 0.83 | 0.84 0.52 0.64 0.85 0.85
Emax 6.02 |. 7.09 4.09 7.60 5.58 5.68
%Pts+/-1 | 85.98 | 86.24 | 96.05 | 96.35 | 84.80 | 83.68
Table 2: Cost and accuracy of DTM produced with all available
breaklines (manual and semi-automatic process)
AREA Deauville Kerlaz Le Havre
Cost 1551 3199 972
(NptsUsed)
NptsRef 378 684 934
DTM Man | Auto Man Auto Man Auto
Accuracy
Avg 0.17 | 0.41 | -0.36 0.08 0.35 0.37
Std 1-00 |} 0,89 1.65 0.86 1.09 0.99
RMS 1.01 | 0,98 1.69 0.86 1.15 1.06
Emax 8.06 | 6.80 |-10.93 7.50 6.13 6.96
%Pts+/-1 | 75.13 | 79.95 | 78.51 | 88.61 | 73.49 | 77.52
Table 3: Cost and accuracy of DTM produced with the main
road network (manual and semi-automatic process)
2 5 EHE SE
2
>
[5]
S
3 p
2 15] —e— Kerlaz
= —æ—Le Havre
= *
e Deauville
WA :
c
©
=
0,5
0 d
0 20 40 60 80 100 120
Percentage breaklines
Figure 5: RMS of manual DTM for the 3 test areas, against the
percentage of used breaklines
International Archi
MERCI TEE
04 E
Figure 6. RMS of
percenta
42.2 Influence o
The input breaklin
displacement along
2m, no segments sh
Kerlaz. For both the
e The poly
affect the
the captur«
e The z-disj
the DTM,
the semi-:
process (lc
Input | Cost
Brea | NbPts | NI
klines | Used F
REF | 7100 | 6
Dz1 - 6
Dz2 - 6
Spl2 | 4428 | 6
| SpIS | 4357 | 6
REF | 7100 | 7
Dzl - 7
Dz2 - 7
Spl2 | 4428 7
Spls | 4357 | 7
Table 4: DTM cost z
varies (K
4.2.3 DTM produ
The best accuracy :
below 85cm), but
process allows a si
time saved) for a sm;
If input vectors are
approa.h is prefere
simplifying the breal