Stability maps for really exploitable automatic classification results

The paper describes a new methodology for producing really exploitable results from automatic classification algorithms. The output of these algorithms is usually constituted by an image with each region assigned to one out of n classes. If the end user, on the basis of results obtained from a control set provided with a ground truth, simply knows that classification over the whole dataset can be considered correct at, for example, 85% (s)he cannot know where correct and erroneously classified regions are really located in the whole dataset. Obviously, the result obtained can be exploited to effectively compute global indexes over the dataset, but it cannot be used as a thematic map. Thus, in addition to the assignment of a class to each region we propose an approach that provides a stability map, a binary image that separates regions (S) classified with high accuracy from those (U) whose classification result should be verified before being used. Two further benefits derive from the construction of the stability map: the control set can be used to set up a good threshold for binarizing the stability map (that is, a threshold by which all regions S are effectively correctly classified); unreliable regions U can help the end user to identify principal causes of (types of regions leading to) misclassification and corresponding (fuzzy, neural, rule based, etc.) approaches to overcome them.