The recognition of the different types of objects present on the ground (roads, rivers, towns, forests, land under cultivation, deserts, etc.) constitutes one of the most important tasks performed. The process of assigning each pixel of an image to a particular class is called "classification".

A human operator, called a "photo-interpreter" is specialized in "reading" satellite images. For him, the classification operation consists in analysing the image and, from his own experience, in classifying objects "manually" by determining each class of object that it is possible to find in an image of a given region of the globe and its localization. Moreover, he works on a high level: he first distinguishes great sets of objects present in the image (overall vision) and then focuses on a particular area to refine the previous classification.

A computer, on the other hand has neither experience nor overall vision of the image, and if we wish to automate this classification operation, many problems must be solved. The lack of experience can be filled by use of expert system. The machine has at its disposal a knowledge data base about the different classes of objects that it is possible to find in an image (at least the objects that we are interested in). This leads to the elaboration of a new image unknown to the system. Such a classification, based on a set of examples of objects to be recognized, is called " supervised classification ".

Moreover, some methods have been developed to allow the system, without any previous knowledge of the objects to recognize, to assess the number of distinct objects present in the image, and to propose a classification for all the objects found. This type of classification is called " unsupervised ".

The way the computer analyses the image is often a pixel centered approach, and sometimes an approach based on a region of pixels. We will focus here on the pixel centered approach, which classifies each pixel in the best way possible. This approach is completely decentralized and there is no available information on whether a pixel belongs to any large homogeneous region of the image. The amount of information about one pixel is very low, and in this way, it is more difficult to determine to what class it must be assigned. But, at the same time, this approach is the simplest to implement and the most economical in computing time.

We will deal in this document with classifications performed according to a supervised mode: the classes to recognize are known, and we can use examples representative of each class. Classification is performed at pixel level. The information available at this level does not allow a correct distinction to be made between two classes. For instance, two classes can have some identical grey levels (overlapping of grey level histograms). Therefore, the classification of a single image (only one spectral band) is very poor. To improve classification, more information must be provided for each pixel to be able to correctly choose the class to which it should be assigned. The use of all the images available - i.e. spectral bands and out-image data related to the geographical context of the region considered - completes the information on each pixel and helps assign each pixel to its correct class.

The integration of all the information dealing with one pixel is called " information fusion ", and is dealt with below.