nbul 2004 
  
)des 
  
increases 
aluate the 
iumber of 
different 
are shown 
deduce a 
y with the 
ach other. 
The best 
  
RMS 
cell n.) 
s | CPs 
) 7,87 
) 5,50 
) 2,96 
) 4,00 
— 
  
  
  
  
  
  
  
ean Res 
cell n.) 
'S CPs 
  
  
i 6,81 
  
  
  
  
  
  
  
72 GCPs 
  
approach 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
3.8 Normality Test on the residuals 
7 . . 
The y^ test on the residuals has been carried out on the 
following tests: 
- RFM with 20 coefficients and 72 GCPs; 
- NN with 10 nodes and 72 GCPs. 
The test has been positive for the residuals obtained with the 
NN approach. Residuals along & direction obtained with the 
RFM method do not pass the test; a systematic effect can be 
hypothesize in this case. 
4. CONCLUSIONS 
The obtained results from the performed tests allow to conclude 
that: 
- RFM method introduces instability problems when using third 
order polynomials, underlined quantitatively by the elevated 
variability values of RMS and qualitatively by the presence of 
strong anomalies in the corrected images; 
- such anomalies are not present if the NN approach is adopted; 
- the RFM method shows the worst behaviour along & (also 
confirmed by the results of the residuals normality test), not 
underlined by the other two methods; it could be due to a 
unknown systematic errors induced on the data by the MIVIS 
whiskbroom scanning system or by the correction method itself; 
- the method based on the NN shows good generalization 
capabilities when observing the decreasing number of errors on 
the GCPs and CPs as the number of parameters grow, not 
observed for the other two methods; 
- the level of attainable accuracy from the two methods can be 
judged as good. Also considering that the qualitative analyses 
have often revealed that it is greater than the RMS numerical 
values suggested. 
5. MIVIS 2 IMAGE GEOMETRIC CORRECTION AND 
MOSAICKING 
MIVIS 2 image has been corrected using NN method and 72 
GCPs. Therefore MIVIS 1 and MIVIS 2 NN orthocorrected 
images have been linked in a mosaic. The final result is shown 
in the final figure showing good coherence. 
  
877 
References 
Bello; M. G. , 1992, Enhanced training algorithms, and 
integrated — training/architecture selection for multilayer 
perceptron networks, /EEE Trans. on Neural Networks, 3,864- 
875. 
Borgogno Mondino E., Giardino M., Perotti L. - Neural 
Network : a method for the application of hyperspectral images 
analysis to the Cassas landslide hazard assestements (Susa 
Valley, NW-Italy) 1° EGU, Nice 2004. 
Breuer M., Albertz J. Geometric Correction of airborne 
whiskbroom scanner imagery using hybrid auxuliary data, in 
IAPRS - International Archives of Photogrammetry and Remote 
Sensing, vol. XXXIII, 2000. 
Demouth H., Beale M., Neural Network Toolbox User's Guide, 
vol. 4, Matlab, 2003. 
Dowman, I., Tao, V., 2002, An update on the use of rational 
functions for photogrammetric restituition, /SPRS, Vol. 7, N°3, 
September 2002, pp. 26-29 
Giardino M., Gomarasca Mario A., Piero Boccardo, Enrico 
Borgogno Mondino, Luigi Perotti Airborne Sensor Mivis 
Images For Landslide Phenomena Analysis In Mountain Areas 
5th ISEMG, Thessaloniki, Greece 2004. 
Howard Demuth, Mark Beale, MATLAB Neural Network 
Toolbox User's Guide, The Mathworks, Version 4, pp.5-8,5-30. 
Tao, C., Hu, Y., 2001, A comprensive study of the Rational 
Function Model for Photogrammetric processing, 
Photogrammetric engineering & Remote Sensing, December 
2001, pp.1347-1357.