are properly considered as share of snow and green. The 
result of the performed Fuzzy classification is very 
satisfactory, the number of blunders is very low, and the 
approach to mixed pixels is correct. 
The band with the snow membership degrees is the core 
of the identification of blanket of snow. 
As a first result of our analysis, pixels with snow 
membership degree around 1 can be considered as 
completely covered by snow, the other ones proportionally 
covered to their membership degree. 
10. ANCILLARY DATA FOR SNOW LAYER 
LOCALISATION 
Snow layer is not allocated on terrain by chance and 
unexpectedly, especially in late spring when heavy 
snowfalls have to be ruled out in low-altitude areas. 
Presence and melting of snow blanket is close linked to 
local conditions especially to the topographic parameters. 
First of all height, aspect and slope of mountainside rule 
the snow layer presence-absence and so information from 
Digital Elevation Model (DEM) of Alps region can be very 
useful to check and improve the classification result for 
snow layer identification. 
The DEM we used is a regular grid with a 250 meters step 
and it is modelled as a raster image having terrain height 
as a pixel attribute and the step as the pixel size. Single 
information about aspect and slope of terrain can be 
obtained as derived raster images from the DEM itself. All 
this information can be joined to each pixel of NOAA 
image and of the fuzzy classified images. 
First of all NOAA image has to be registered so that each 
terrain area has the same position in the DEM raster 
image and in the NOAA raster image. The NOAA image 
registration is done by a third degree polynomial 
transformation using the nearest neighbour resampling 
method. The polynomial parameters are evaluated by last 
square adjustment having about eighty Ground Control 
Points. The GCP are points of known pixel coordinates 
both in DEM image and in NOAA image. 
T. ^ / 4: " " PPP - y 
  
Ji. A: 
? $ 
  
Figure 14: DEM of central Alps arc and the ground control 
points. 
First of all the terrain information can be used as filters to 
remove the especially blunders of second type that is 
when sure non-snow pixels are classified as snow ones 
(assigned to snow class in traditional classification and 
with snow membership degree about 1 in fuzzy 
approach). 
The mountainside type can be very useful not only to find 
and eliminate classification blunders but also during the 
classification itself. Since snow layer presence is linked 
also to topographic elements considering their 
contributions during classification and not only in the end 
can dramatically improve the quality of results. 
In multisource classification, topographic features become 
an integral part of the classification procedure. Raster 
images describing aspect and slope of mountainsides are 
extracted from DEM and so each pixel has these 
topographic specifications too 
Careful observations of snow layer during the years 
provide experts with practical knowledge to characterise 
the topology of the snow covered area in the different 
seasons. This precious knowledge becomes the “rule” for 
the multisource classification in order to identify areas with 
high / medium / low probability to be snow covered 
season by season. 
These rules take action on the fuzzy membership degrees 
with the aim of making the best use of all the available 
information, pixel by pixel. 
Fitting parameters for the rules are the basis of this new 
part of classification and so they have to be calibrated and 
updated every season. 
Multisource classification tries to reproduce the human 
decision process: it starts from the acquired knowledge, 
here expressed with the rules, and then tries to decide 
each situation comparing it to the rules. If no rule matches 
a situation it can not be solved. Rules observe all the pixel 
values (membership degree to snow class, to green 
class; 7, height, aspect and slope). 
Rules have the IF-THEN structure; they input all the pixel 
information and output a value (high — medium- low) for 
snow-nsk. 
An example of rule can be the following: 
IF snow IS high AND IF green IS low AND IF cloud IS low 
AND IF building IS low AND IF vegetation IS low AND IF 
lake IS low AND IF sea IS low AND IF cultivated-field IS 
low AND IF rice-field IS low AND IF height IS low AND IF 
aspect IS low AND IF slope IS low THEN SNOWRISK is 
high 
After the Fuzzy classification each pixels is described by: 
a) nine membership degrees: 
membership degree to class snow, 
membership degree to class green, 
membership degree to class cloud, 
membership degree to class vegetation, 
membership degree to class lake, 
membership degree to class building, 
membership degree to class cultivated-field, 
membership degree to class rice-field, 
. membership degree to class sea, 
b) three topographic features: 
10. mountainside height, 
11. mountainside aspect, 
12. mountainside slope. 
o 0o NOUO RON: 
In this way each pixel is equipped with 12 features 
contributing to snow layer identification. In order to 
perform this new part of classification all the information 
are collected together in a multiband file. 
334 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
  
Figure 
black 
The fi 
espec 
while 
Then 
fuzzy 
The fi 
on Alf 
areas. 
areas 
estim: 
Acknt 
Speci: 
staff v 
perfor 
UCLA 
made 
Multim 
Milan 
0521).