International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
Neural Nets 72 GCPs and 10 CPs 
  
  
  
  
  
  
  
  
  
  
  
  
  
Noles RMS. RMS, RMS 
number (cell n.) (cell n.) (cell n.) 
GCPs CPs GCPs CPs GCPs. | CPs 
3 12,32 12,85 16,61 17,01 20,68 (21,32 
4 5,40 7,86 1.1.57 8,09 2,4128 
3 4,24 4,14 11,57 7,64 12.33 18,69 
6 4,00 5,43 8,14 7,26 9,07 | 9,07 
7 1,43 1,89 4,33 3,01 4,56 | 3,55 
S. 1,58 2,14 6,17 4,04 6,37 | 4,57 
9 1,14 2,09 3,82 3,57 3,99 | 4,14 
10 0,77 1,61 2,44 3,67 2,56 | 4,00 
11 0,70 1.32 3,80 4,06 3,86 | 4,27 
12 0,66 1,65 2,94 3,83 3,01.—1-4,17 
13 0,42 1,34 2,26 3.03 2,29 | 3,32 
  
  
  
  
  
mean Res, | mean Res mean Res 
Nodes $ n (n. celle) 
number (n. celle) (n. celle) : 
GCPs C Ps GCPs CPs GCPs | CPs 
3 0,00 -3,64 0,00 -1,75 16,99 |17,64 
  
0,00 -1,79 0,00 0,27 10,97 |10,08 
  
   
7 nodes 13 nodes 
6 nodes 
  
  
  
  
  
Figure 6 — Images obtained from NN test with nodes increases 
and cartographic map overlap results. 
3.7 Series 3 analysis results 
As for the RFM method now it is possible to evaluate the 
sensibility of the NN method while varying the number of 
GCPs for a constant number of nodes (10). 
The results in correspondence of the four different 
configurations of points (the same used for Series 1) are shown 
in table 5. From their analysis it is not possible to deduce a 
direct dependence of the level of attainable accuracy with the 
number of GCPs, all the RMS maintain very similar each other. 
This is also confirmed by the qualitative analyses. The best 
configuration is N=61, as for the RFM.. 
  
0,00 -1,07 0,00 -0,93 | 10,16 | 7,53 
  
  
0,00 -1,39 0,00 -0,30 3:68 | 3.19 
0,00 -0,83 0,00 -0,34 525 | 391 
  
Neural Nets 10 nodes and 10CPs 
  
RMS. RMS RMS 
n 
GCPs 5 ; 
(cell n.) (cell n.) (cell n.) 
  
number 
GCPs CPs GCPs CPs GCPs CPs 
  
4 
5 
6 0,00 -2,51 -0,15 0,56 7,70 | 8,42 
7 
8 
9 
0,00 -1,18 0,00 -0,23 3,41 | 3,76 
10 0,00 -1,07 0,00 -0,51 2.17.1 3,21 
11 0,00 -0,49 0,00 -0,93 3,06:;{ 3,69 
12 0,00 -0,07 0,00 -0,58 236 13.70 
13 0,00 -0,14 0,00 -0,97 1,82. .| 2,96 
  
  
  
  
  
  
  
  
  
  
  
  
  
Table 4 — RMS and mean residuals obtained from Series 2 test. 
The RMS behaviour for the GCPs and CPs, seems to suggest a 
good generalization capability while the number of nodes 
increases. 
The general qualitative analysis confirm that the corrected 
images do not introduce particular anomalies (see Figure 6). 
Bd , a » 
i & d a 2 » 
Pd : 6 E ; 
d 
10 nodi 13 nodi | 
     
  
7 nodi 
Figure 6 — Images obtained from NN test on series 2 while 
increasing the nodes. 
Only in a few cases we have noticed some small discontinuities 
at the upper limits of the image, however these are entirely 
negligible in comparison to those found in the RFM method. 
On the other hand the overlap of the map furnishes satisfactory 
results as well (in each test the accuracy is constant throughout 
the entire image, also along the uneven zones where the GCP 
are not present) and confirms the previous errors, above 7 nodes 
the improvements are limited and located only to some points 
(see Figure 7). 
39 0,16 5,76 2,49 5,37 2,49 7,87 
  
50 0,58 3,44 4,46 4,30 4,49 5,50 
  
61 0,75 1,76 3,02 2,38 3.42 2,96 
  
  
  
  
72 0,77 1,61 2,44 3,67 2,56 4,00 
  
mean Res, mean Res, mean Res 
GCPs : 
(cell n.) (cell n.) (cell n.) 
number 
  
GCPs CPs GCPs CPs GCPs CPs 
  
39 0,00 1,75 0,00 -0,36 2,04 6,81 
  
50 0,00 -0,88 0,00 -0,47 3,83 4,99 
  
61 0,00 -0,97 0,00 -0,91 2,51 2,32 
  
  
  
  
  
  
  
72 0,00 -1,07 0,00 -0,51 2,17 3 21 
  
  
  
Table 5 — RMS and means of residuals obtained from 
the Series 3 tests. 
    
| 39 GCPs 50 GCPs 61 GCPs | | 726cPs 
Figure 7 — Images obtained with the NN correction approach 
on Series 3 while increasing the GCPs. 
876