Information mining opens new perspectives and a huge potential for information extraction from large volumes of heterogeneous images and the correlation of this information with the goals of applications. We present a new concept and system for image information mining, based on modelling the causalities which link the image-signal contents to the objects and structures within interest of the users. The basic idea is to split the information representation into four steps: 1 image feature extraction using a library of algorithms so as to obtain a quasi-complete signal description 2 unsupervised grouping in a large number of clusters to be suitable for a large set of tasks 3 data reduction by parametric modelling the clusters 4 supervised learning of user semantics, that is the level where, instead of being programmed, the systems is trained by a set of examples; thus the links from image contents to the users are created. The record of the sequence of links is a knowledge acquisition process, the system memorizes the user hypotheses. Step 4. is a man-machine dialogue, the information exchange is done using advanced visualization tools. The system learns what the users need. The system is presently prototyped for inclusion in a new generation of intelligent satellite ground segment systems, value adding tools in the area of geoinformation, and several applications in medicine and biometrics are also forseen.