rn part 
'egion) 
2d. The 
above 
  
t-2 
several 
ale and 
se of an 
odology 
he data 
alysis is 
hod 
  
The first two steps are the same used for the flooded 
areas assessment (§5); then, the analysis is split into two 
different methods: the first is related to image 
classification, while the second is related to feature 
extraction techniques. 
a) First methodology. In the post-flood image a 
radiometric range that characterises the different 
known landslides has been determined, forcing all the 
other Dns to a null value. In the pre-flood image the 
same area Dns have been then recorded; a difference 
between these two new synthetic images has pointed 
out the existence of landslides caused by the heavy 
November precipitation. In order to reduce the 
radiometric noise generated by the extraction of non- 
landslide features that show a similar reflective 
behaviour, a general consideration has been 
deducted. The majority of the slides in that area 
occurred in N-W dipping, with an angle dip in a range 
comprised between 6°-15° with respect to the 
horizontal plane. Thus, a clivometric model has been 
generated determining the dipping plane as the 
projection on the horizontal plane of the interpolating 
normal one, which was previously calculated for each 
slope angie. The new image locates the dipping 
direction in each cell. An automatic procedures has 
then been implemented in order to verify the three 
restrictions previously determined: 
-— 
the radiometric range; 
2. the slope range; 
3. the N-W dipping. 
b) Second methodology. The aim of the second approach 
is to perform a feature extraction (from the geometric 
and radiometric point of view) based on landslides of 
known morphologic characteristics. The suitability of 
this methodology is based on some basic 
assumptions: 
e the majority of the slides occurred in a N-W 
direction; 
e the slides presents a series of parallel fractures 
perpendicular to the flow direction (SW-NE). 
An automatic feature extraction procedure has thus 
been implemented based on a target (both extracted 
from the real images and synthetically generated) that 
reproduces those morphologic characteristics. The 
algorithm adopted is based on the restitution of the 
correlation coefficient (R) calculated on the target and 
search area; the correlation coefficient is considered 
acceptable when R»0.75. The output image is 
composed by null values when the restriction over R 
is not satisfied, where the original values are restored 
when satisfied. This method allows one to determine 
similar features ever where not directly visible 
(shadowed slopes), because the correlation index is 
independent from the original reflective values, and 
takes only the morphology of the subset (Dns 
geometric arrangement) into account. The different 
synthetic images generated (one for each target) 
using the above mentioned algorithm have then been 
added, and the resulting image has been substituted 
to the radiometric range image (1.) used in the 
previous methodology. In fig. 6 a colour composite 
(converted B/W values) of the resulting image is 
shown. 
69 
7. FINAL REMARKS AND FURTHER DEVELOPMENTS 
The extraction of flooded areas when compared to 
plotted ones, shows a remarkably correspondence; 
further developments concern the usage of ERS-1 Sar 
images, that have not yet been integrated in the model 
because of the extreme difficulty to fit them to the 
absolute georeferenced one (discrepancies grater than 
1.5 the ground resolution). 
With regards to  landslide assessment, the two 
methodologies presented show a good description of the 
phenomena. Further developments will regard the 
integration with multispectral remotely sensed data (such 
as Thematic Mappers sets) in order to evaluate possible 
correlation with humidity (extracted by TM5 band) and 
vegetation index (extracted mainly from TM4 band). 
f 4 : ur 
: D 7 
"4 a > » 
t pans x 
z eM ae 
wow f 
' T x = £ 
2a ; ja z * 
^ ww 
ef ow af 
  
Figure 6 - B/W composite showing the final image; 
circled areas are recognised slides, while the black 
values remaining are noise (less than 20-25%). 
8. ACKNOWLEDGEMENTS 
This research has been conducted in collaboration with 
Prof. John Mc. Mahon Moore and Dr. Philippa Mason of 
the Centre for Remote Sensing at the Imperial College of 
Science and Medicine in London. The majority of the 
analysed images have been provided by C.S.l. Piemonte 
in Turin; without their aid the experimental image 
processing could not have been done. A special thank to 
everybody at the Dipartimento di Georisorse e Territorio 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996