170 
Picture 2: Zoom into the map obtained by a 
photcinterpretaticn. Snow covered areas are white 
coloured. 
Snowpack borderlines are very jagged, there is not a 
clear border between snow and free soil and so it is quite 
difficult to bound snowpack and then to evaluate its 
extension. This way requires a iot of time, the accuracy 
of the snow identification is not well-know and the 
snowpack extension is empirically vaiuated. 
4. SNOW COVERED AREA EXT1MATION BY 
AUTOMATIC HARD CLASSIFICATION 
Snow covered areas can be identified by an automatic 
classification since snow has a typical spectral behaviour 
different from snow free soil. Next picture 3 shows snow 
spectral behaviour. 
Picture 3: Snow typical spectral behaviour [D. K. Hail, J. 
Martinecj 
Since 80’s computer development and multispectra! 
images availability allow to perform automatic 
classifications of remote sensing images. Automatic 
classification techniques allow to train computers to 
recognise automatically snow-covered areas and so to 
produce thematic maps in short time. Human experts 
have only to identify some well-known areas into the 
images, called training set, and after algorithms 
automatically identify thematic classes according to the 
information of selected training set. 
Classification accuracy can be evaluated comparing, for 
a weli-known area call testing set, the true situations to 
the classification results. The comparison is done into a 
square matrix n x n, called confusion matrix, where n 
represents the number of categories. The left-hand side 
is labelled with the categories true, the upper edge is 
labelled with the categories evaluated and the values in 
the matrix represent numbers of pixels, inspection of the 
matrix shows how the classification represents well- 
known area, in the matrix the sum of diagonal entries 
gives total number of correctly classified pixels, column 
marginal sums give the total number of pixels of each 
class as assigned by the classifier anc row marginal 
sums give the total number of pixels of each class in the 
testing set. 
Next picture 4 shows an example of confusion matrix. 
Classified 
snow wood rock 
total 
T 
snow 
r 
wood 
H 
Row 
om/ginais 
u 
e 
rock 
total 
■ 
■ 
Cci'jmn marginate 
Sum at dtagon&f 
elements givps iai&i } 
number of correctly 
classified pixos 
; 
Picture 4: exempla of a confusion matrix 
Automatic classifiers allow to identify snowpack in short 
time and to evaluate the classification accuracy 
compiling the confusion matrix. 
Traditional classification algorithms are called hard 
because they are able to assign each image pixel only to 
one class. In this way they do not allow intermediate 
state due to mixture end the mixed pixels after the 
classification are forced into one class, at the most into 
the unclassified class. NOAA images, (6 hours time 
resolution, 1 km x 1 km pixel size and low cost) are the 
most suitable for a snowpack monitoring during melting 
season, but they contain many mixed pixels. Transition 
between snowpack and snow free soil is recorded in the 
images as a mixture zone and the traditional classifiers 
do not fit the problem. Infect the overall behaviour of 
mixed pixels made up by snow and wood is often more 
similar to cloud then to pure snow and pure wood. These 
cause misciassification errors in transition zones and so 
the accuracy of snowpack identification decreases a lot 
Next picture 5 shows a zoom into a hard classification 
result, many pixels on the snowpack borderlines are 
attributed to cloud class and so the classification 
accuracy is low.