Mixed pixels problems and multisource classification 
for snow layer detection 
Alessandra Colombo 
"P.h.D attendent at Politecnico di Milano, Dip. IIAR sez. Rilevamento 
P.za Leonardo da Vinci, 32 -20133 Milano ITALY 
e-mail: alessandra.colombo@polimi.it 
ISPRS Commission VII Symposium, Working Group 4 
KEY WORDS: 
Remote-Sensing, image analysis, snow monitoring, database, DEM, Expert-System, mixed pixel, classification, Fuzzy 
ABSTRACT 
Monitoring the thawing of snow layer in spring is very useful for estimating its water equivalent and making good use of it. 
Remote sensing classic techniques and low-cost images have allowed a fairly efficient approach to this problems since ' 
the eighties. All the main problems of the classical approach are connected with the high number of mixed pixels in the 
images, since classical approach is completely unsuited to analyse mixed pixels. Nowadays the fuzzy approach allows to 
define degrees of membership to classes so that one can treat mixed pixels as they really are. Then a multisource 
classification processes membership degrees and DEM information jointly in order to localise the blanket of snow in an 
enough controlled and exact way. 
1.INTRODUCTION 
An efficient methodology for detecting snow layer 
borderline on large mountain areas is locating it by remote 
sensing images since snow blanket and snow-uncovered 
soil have brightly different spectral signatures. The 
borderline between uncovered-soil and snow is usually 
non continuos and irregular and so, many pixels in a 
remote sensing image are mixed, each of them recording 
the reflected-emitted radiation from both snow and snow- 
free soil. In springtime blanket of snow is surrounded 
principally by green and so we consider mixed pixels 
along borderline made up only by snow and green. 
Traditional classifiers assign each pixel to a single class, 
or at least to the unclassified-pixel class, depending on 
classifier's characteristic rules. 
In this way part of the mixed pixels radiometric value gets 
lost and any mixed pixel after the classification is 
considered as a non-mixed one. 
A fuzzy approach, on the contrary, allows to define a 
membership degree to each class, and in this way all the 
pixels are actually classified as mixed and no information 
is lost. 
Many different elements can be considered later to assign 
the pixels to a precise class if this information is required 
for further analysis. . 
A multisource classifier can take action after the fuzzy 
classification so different elements can be used to 
evaluate the former mixed pixels nature; so they can be 
classed consistently with a non radiometric information 
too. 
In the case of snow layer borderline a multisource 
classification following a fuzzy classification, considers 
topographic elements mountainside height, slope and 
aspect. This is our procedure to decide which class the 
mixed pixels should be assigned to, and so identify the 
snow borderline in an enough controlled and reliable way. 
2. SNOW LAYER MONITORING SYSTEM 
Monitoring systems require many repeated observations 
and so the choice of the image type is very important in 
order to satisfy the  image-management needs 
(radiometric resolution and spectral resolution), the time 
resolution and the costs. Very good compromises are the 
NOAA images because they survey the same earth area 
several times each days (from three to six times a day), 
their radiometric resolution and their spectral resolution 
are quite suitable and, finally, they are low cost, on the 
contrary their geometrical resolution is not high (around 
1.1 km x 1.1 km at nadir position) and so they contain 
many mixed pixels, particularly snow free soil-to snow 
borderline is recorded in mixed pixels as the following 
image shows. 
no-snow 
  
Fig. 1 Along the snow-to snow free soil borderline there 
are many mixed pixels. 
328 Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
  
3. CI 
The € 
is th 
cove 
they 
differ 
The 
layer 
preci 
appre 
to ar 
allow 
ones 
class 
differ 
minir 
rules 
only 
delici 
class 
of an 
beha 
nor t« 
A mi 
wron 
I. 
I. 
The 
mear 
inforr 
natur 
other 
and 
inforr 
The 
class 
real | 
dark