In satellite image classification, the data are of course provided by the satellite images, but also by the geographical context of the studied area (road maps, river maps, town locations, elevations, slopes, orientation, etc.).

Information-fusion methods are executed in two ways :

• combining each source in a parallel way: in this case, all data to be merged are available at the time of fusion,
• or reviewing information : data are available at different moments, and the decision to be taken is updated as soon as new information comes in.

Parallel processing uses all the available images at the same time. On the other hand, reviewing information means processing newly arrived satellite images to update the information base already built from the previous images. This process implies taking into account numerous problems due to the difference in dates between the shots. For instance, vegetation does not have the same appearance during summer and winter, or some elements can disappear (a forest can be replaced by fields), or appear (a town can spread).

We will detail here with the fusion of data sources in parallel processing. If all sources are reliable, it is possible to use a conjunctive fusion. But if some sources are reliable and some are not, or if their reliability is unknown, then it is better to use disjunctive fusion. A weighted logical combination can also be applied to merge data sources that have different degrees of reliability.

Several mathematical frameworks exist for modeling uncertainty :

• probability theory,
• belief theory,
• possibility theory.

Probability theory is the most commonly used. In this field, the a posteriori probabilities for an object to be member of a class are computed according to Bayes decision rule. A severe fusion operator then computes the product of these probabilities.

This probabilistic approach is criticizable for several reasons arising from the modeling of expert opinion and from the problem of the information fusion. Initially the construction of a probability distribution requires much more information than an expert is able to provide. The choice of a parameterized family of distribution functions mainly results from a concern to simplify calculation. Consequently, the fit of the model to the expert opinion is debatable. An expert prefers to provide intervals rather than isolated values because his knowledge is of reliability limited and spoilt by inaccuracy. In fact, probability theory does not aim to model the inaccuracy [Dubois and Prade, 1992].

Belief theory handles inaccurate and uncertain information, and fusion is performed by means of the Dempster-Shafer's orthogonal combination rule.

In possibility theory, many data fusion operators exist, with different behaviour properties that vary from conjunction to disjunction, including several levels of operators with average behaviour. Recent operators dependent on possibility theory are able to adapt their behaviour, and act in a conjunctive or disjunctive way according to the degree of reliability of the data sources to merge. Other recent operators are able to take into account different priorities assigned to data sources, depending on whether they are reliable or not.