Fuzzy 
use of it. 
ms since ‘ 
sis in the 
allows to 
iltisource 
ow in an 
ervations 
portant in 
needs 
the time 
s are the 
arth area 
s a day), 
resolution 
it; on the 
! (around 
y contain 
-to snow 
following 
e there 
  
3. CLASSIFICATION MISTAKES OF MIXED PIXELS 
The core of the remote sensing approach for this analysis 
is the image classification. Since snow and non-snow 
covered soils, have quite different radiometric behaviours 
they can be easy identified if they are observed in 
different pixels. 
The classic remote sensing approach performs snow 
layer monitoring regardless of mistakes and lacks of 
precision due to mixed pixels ignoring. In fact the classical 
approach to the classification problem, assigns each pixel 
to an exact class, basing on radiometric considerations, 
allowing the unclassified class for the too much uncertain 
ones. Each pixel of a classified image belongs to a unique 
class and does not belong to all the other classes; the 
different classification methods (maximum likelihood, 
minimum distance, parallelepiped, and so on) are different 
rules to choose the classes. Mixed pixels can be assigned 
only to a single class and so their allocation is very 
delicate; in fact the overall behaviour of mixed pixel (e.g. 
class A plus class B) can be more similar to the behaviour 
of another class C than to the original class A and class B 
behaviours and so the pixel A+B is assigned neither to A 
nor to B but to C class. 
A mixed pixel (A+B) can be assigned in the following 
wrong ways: 
1 Assigned only to class A or only to class B; 
Il. Assigned to unclassified class; 
It. Assigned to a class C. 
The tree above wrong ways have quite different 
meanings. In fact the first and the second are similar to 
information losses, in the first case only a part of the pixel 
nature is ignored and in the second all is ignored; 
otherwise, the third one is a completely wrong assignment 
and it can be very dangerous because it is a kind of 
information flake. 
The following images in figure 2 show the three wrong 
classification ways : 
Real pixel composition 
  
  
real pixel composition: snow in light grey and green in 
dark grey 
| wrong way 
  
mixed pixels are assigned to the more present class 
Il wrong way 
  
unclassified 
| 
  
unclassified : | 
unciassiffed | 
Tn ed | 
| unclassified 
  
  
  
  
  
   
  
mixed pixels are assigned to the unclassified class 
Ill wrong way 
  
  
  
  
  
  
  
mixed pixels are assigned to an extraneous class 
Figure 2: three wrong assignment ways for mixed pixels 
4. CLASSICAL CLASSIFICATION APPROACH 
Mixed pixels are not treated as mixed using a classical 
approach. For instance, snow layer borderline is classified 
as cloudy area since the radiometric behaviour of snow- 
green mixed pixels is often similar to the cloudy areas 
one. 
In order to highlight the mixed pixels variable classification 
different classic classification ways are performed and the 
results compared together. 
The first image is part of a NOAA image of Alps region 
recorded second May 1997 at 12.30. (Figure 3). 
The AVHRR scanner carried on NOAA satellites records 
data at 1.1 Km resolution in the following channels: 
0.50 - 0.68 um 
0.725 — 1.10 um 
3.55 — 3.93 um 
10.30 — 11.30 um 
11.50 — 12.50 um. 
> a * > à aie M 
05 N} à m ^ 
- 3 YA Wet eL a 
+ 
wid 
  
Fig. 3 Part of NOAA image used to perform the analysis. 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 329