Expert maps: an alternative for integrating expert knowledge in satellite imagery classification ‡ Manuel L. Campagnolo 1 and Mario Caetano 2 1 Dep. of Mathematics, I.S.A. , Lisbon University of Technology Tapada da Ajuda 1399 Lisboa Codex, Portugal. Tel: 351 1 3638161; Fax: 351 1 3630723; e-mail: mcampagnolo@isa.utl.pt 2 National Center for Geographic Information (CNIG) Rua Braamcamp, 82, 1 dto., 1200 Lisboa, Portugal. Tel: 351 1 3860011; Fax: 351 1 386287; e-mail: mario@helios.cnig.pt ‡ This work was partially suported by ISA/DGOT/JNICT contract 28/94. Abstract: A new classification method which permits the integration of expert knowledge in satellite image classification is presented. This approach differs from other knowledge based ones for the following reasons: 1) it is a low-cost procedure; 2) it allows that the available experts explicitly express all their knowledge on the area to be classified, and 3) it does not have to face the problem of rule generalization over the whole area. The method is applied to a real image data set and it is shown that the achieved classification is more accurate than the classification just based on spectral data. INTRODUCTION The use of expert knowledge has been widely used in the last decade for improving the accuracy of remote sensing derived land cover maps (see [6], [1]). The knowledge has been represented as a set of evidential rules which encompasses classes and surface attributes at a pixel level and also, in more recent studies, at a spatial context level. Knowledge based methods are usually used to improve, via Bayesian [4] or Dempster-Sheffer [5] evidences combination, land cover maps generated from spectral data. However, the proposed approaches for expert knowledge integration have performed worse than expected when applied to complex rural landscapes. We believe that this is related to two intrinsic features of the approaches that have been proposed: (1) it is assumed that the rules are valid within the whole study area, and (2) it is assumed that the expert is able to translate his knowledge into a rule format. As a consequence, relevant knowledge which does not verify the above assumptions cannot be inserted in the knowledge base. For exemple, if an expert states that “if the altitude is greater than 600 meters there is no eucalyptus forest” it is assumed that this is true in the whole area to be classified. As a consequence the expert will not take the risk of stating that the threshold is, instead, 500 meters even if this last value seems more likely to him, because this last statement could not be always true. He could, as an alternative, establish a probability distribution instead of a simple threshold but then there will arise two related problems. Even if the expert is familiar with the concept of probability distribution it will be difficult (and boring) for him to define that distribution. If he his not familiar with that concept it will be almost impossible to him to define it without the help of another expert in data analysis. Furthermore, the rules like the previous one are very hard to gather from experts, specially in the case where experts haven’t any kind of mathematical background. For instance, in land cover mapping from satellite imagery in rural landscapes, the useful experts are, in general, forestry an agricultural agents which have o good knowledge of the area but who are not able to express it in rule format. METHODOLOGY Expert Maps To overcome the limitations mentioned above we propose a new methodology for the integration of expert knowledge. Our goal is to extend the ability of the classification process to gather relevant expert knowledge. To obtain the additional input needed in this methodology, people acquainted with the study area are asked to spatially locate known land uses on a satellite image color composite. This map is obtained from the satellite imagery and can be completed with non spectral data like roads and rivers networks if available. These maps, which are given to experts of the study area, will be used as the frame for the encapsulation of their knowledge. The expert is asked to draw on that map the borders of the areas that he thinks that correspond to a particular land cover class in the field. We designate this kind of knowledge representation by experts maps. For integration with satellite imagery, expert maps are digitized into a Geographic Information System. The classification procedure that we propose has the following steps: (1) calculate from the expert maps and for each pixel a likelihood for each class; (2) calculate, for each pixel, the expert maps entropy, to evaluate the disparity among them; (3) integrate the likelihood for each class, calculated in step one, with spectral and ancillary data in a pixel-level classification procedure. Formal Statement of the Method Suppose we know that there m different land cover classes w 1 , ...,w m in the area to be classified and we dispose of a random sample of classified pixels (x 1 ,...,x N ). Each x i represents a vector of spectral values of the i-th pixel, x i =(x i1 ,...,x ip ), for p dimensional data. If we use a minimum Bayes risk classifier we will obtain, under certain hypoteses (see [2] or [3] for details), for each pixel i, a vector of probabilities of belonging to the different classes (p 1 ,...,p m ), where p k =P(w k /x i ). In hard classification approaches the pixel is then allocated to a unique class: the one for which the probability value is larger, that is, x is allocated to w k if p k =p j , for all j in {1,...,m}. In our approach we dispose, also, of one other probability value that a pixel belong to a class. If l of the total number of experts, that we will denote by t, have labeled a particular pixel (note that each expert don’t have to cover the whole area) we can estimate the experts’ probability q k of that pixel belongs to