4 METHOD
4.1 MULTISPECTRAL SUPERVISED CLASSIFICATION
One of the methods to classify an image containing several information channels is multispectral supervised
classification. This is a technique widely used in Remote Sensing, and has proved to be very robust and reliable during
the many decades it has been used (Bryan, 1979) (Campbell, 1987) (Chuvieco, 1995).
The classification process provides an array of probabilities to assign each image pixel to each label in the legend. Then
we can look for the highest probability in that array and build a new image with the class label assigned to each pixel in
order to analyse the spatial distribution of the classification. Anyway, in the method we present, the core process will
really manage the probabilities obtained during the classification of each data set combining them to obtain a good
merge.
An estimation of the accuracy of our classification can be obtained through a statistical value usually called kappa index
κ. (Richards, 1993). This statistical value is obtained through the construction of the so called confusion matrix. The
closer to 1 the more accurate is our classification. We have to note that if we perform a random classification using N
classes, we can obtain percentual indexes of right classification for each class that will be 100/N %. For example, if we
have only 2 classes, in a purely random classification we would obtain a 50 % of rightly classified pixels. But if we also
calculate the κ=index, we would obtain a null value. This value is telling us that our classification is completely
inaccurate, and that the results it shows are completely unrelated to the real ground classification.