|. Istanbul 2004 
nual” DTM is 
breaklines, the 
reliable points. 
eaklines shows 
crease accuracy 
road network 
the percentage 
‚MS between | 
s to the manual 
le 2). However, 
natic approach 
comparison). It 
see Table 3): it 
a proportion of 
ire cost of 65% 
h with all the 
Le Havre 
2884 
934 
Man Auto 
  
  
0.48 0.51 
0.69 0.68 
0.85 0.85 
5.58 5.68 
  
4.80 | 83.68 
ith all available 
ic process) 
Le Havre 
972 
  
934 
Man Auto 
  
  
).35 0:37 
1.09 0.99 
LAS 1.06 
5.13 6.96 
  
  
3.49 | 77.52 
with the main 
1atic process) 
  
—$— Kerlaz 
&— Le Havre 
Deauville 
as, against the 
   
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
Kerlaz 
2 
1,8 
18 5777177 — Semi-auto DTM = 
he -- Manual DTM i 
  
  
  
  
  
0 20 40 60 80 100 120 
Percentage breaklines 
Figure 6. RMS of manual and semi-automatic DTM, against the 
percentage of used breaklines (Kerlaz) 
42.2 Influence of the breakline accuracy 
The input breaklines have been randomly modified: random 
displacement along z-values with a standart deviation of 1 or 
2m, no segments shorter than 2 or 5m. Table 4 shows results on 
Kerlaz. For both the manual and the semi-automatic process: 
* The polyline simplification does not significantly 
affect the DTM accuracy, but it has a drastic effect on 
the capture cost (35 to 40% reduction over test areas), 
e The z-displacement introduces a significant error on 
the DTM, which is, however, less important within 
the semi-automatic process than within the manual 
process (loss of 0.65 to 0.95m instead of I to 1.2m). 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Input | Cost Accuracy 
Brea | NbPts | NbPts | Avg | Std | Emq | Emax | %Pts 
klines | Used Ref +/-1 
Manual DTM 
REF | 7100 | 684 0.18 | 0.49 | 0.52 | 4.19 | 95.47 
Dz1 - 684 027 | 0,97 | 1,01 | 4.67 | 9491 
Dz2 - 684 | 0.35 1 1.73 | 1.76 5.14 140.79 
Spl2 | 4428 | 684 0185 10541 057 | 3:09 193,57 
| SpIS | 4357 | 684 0.16 10,58 | 0.61 | 4.33 | 92 84 
Semi-automatic DTM 
REF | 7100 | 725 0.25 10.67 | 0.71 | 10:96 | 93.03 
Dz1 - 718 0.30 | 1.01 | 1.06 | 10.37 | 80.22 
Dz2 - 715 0.36 | 1.021 1.66 | 1127 153.57 
Spl2 | 4428 | 722 0.20 | 0.86 | 0.89 | 10.57 | 95.38 
Spls | 4357 | 721 0.19 } 0.77 | 0.79 | 10.36 | 93.20 
  
Table 4: DTM cost and accuracy when input breakline accuracy 
varies (Kerlaz) 
4.2.3 DTM production: conclusion 
The best accuracy is provided by the manual process (RMS 
below 85cm), but at an important cost. The semi-automatic 
process allows a significant productivity gain (65% capture 
time saved) for a small loss in accuracy (final RMS around 1m). 
If input vectors are not very accurate, then the semi-automatic 
approa.h is prefereable. In both cases, it is always worth 
simplifying the breaklines. 
4.3 DEM production 
4.3.1 Manual approach 
Within the manual process, all the buildings are represented 
with a constant height (highest elevation, see Figure 7). The 
DEM accuracy is therefore naturally limited by this model. For 
highest precision, a very detailed vector description with 
frequent building block division is required (separation every 
2m height in our process). 
  
  
(c) (d) 
Figure 7: Extract from Kerlaz: left image (a), captured vectors 
(b), manual DEM (c) and semi-automatic DEM (d) 
4.3.2 Semi-automatic approach 
It provides a realistic representation of roof shapes and 
superstructures (see Figure 7d). In order to precisely assess the 
influence of input vectors, we have modified the nature and the 
accuracy of the input polygons: 
e Ref: reference polygons as described in section 2: 
variable z values and block division every 2m; 
e  Zmed: polygons with constant z + division every 2m; 
e Blocks: polygons with variable z but no division 
e Dz: random displacement along z axis (standard 
deviation 1, 2 or 3m); 
e Dxy: random displacement in the XY plane (standard 
deviation 1m); 
e Spl: contour simplification (no segment shorter than 
3m) 
Results are given in Table 5. 
The roof accuracy using reference polygons is around 1m, and 
the proportion of reliable points is above 75%; at least 9396 of 
the checked roof points are less than 2m away from the 
reference point (column Pct+/-2 of Table 5). These figures 
show that the matching algorithm AutoDEM provides a roof 
description close to ground truth, with an accuracy appropriate 
to many applications. 
The perturbations introduced on input polygons affect accuracy 
as follows: 
e A constant z-value on polygons does not affect 
accuracy, 
e Using whole building blocks instead of independent 
adjacent contours implies a small loss in accuracy (0