Knowledge-driven Information Mining in Remote-Sensing Image Archives

attributes Involving a significant amount of high-level reasoning about the meaning and purpose of the objects or scenes depicted (e.g. illegal plantation). They are outside the scope of the current research activity. Figure 2. Logical representation of data and derived attributes extraction of primitive features and the reduction of the resulting data into primitive feature clusters. The primitive features to be extracted need to be carefully selected, since they mainly determine the quality and capabilities of the resulting system. SAR and optical images, for example, will have to be handled differently and texture primitive features will have to be extracted at various resolution levels, since different textures can dominate at different scales. The steps necessary to properly extract primitive features in the EO context are shown in Table 2. The steps need to be repeated iteratively for each band of the image. Primitive feature extraction generates a huge amount of data, which cannot be handled in practice and therefore has to be compressed somehow. This process is represented in the left part of Figure 3, which depicts the result of the scanning of two images (or of two bands of the same image). Each pixel of the image will be located in n-dimensional space in the position determined by the values of the contributing primitive features (their units are non-commensurable, e.g. texture and spectral features). The pixels will tend to group themselves into specific regions of this space. Through clustering (right part of Fig. 3), the ‘clouds’ of image primitive features are replaced by parametric models of their groups, which can be represented in more compact forms. This reduces the precision of the system, similar to a quantisation process, but permits its practical use thanks to the huge data reduction obtained. The primitive features are compressed into clusters using the Kmeans approach. The clusters (condensed representation of primitive features) have no direct meaning, since they group points in an n-dimensional space of non-commensurable variables. Still they represent characteristics of the image seen as a multi-dimensional signal. It is possible to associate meaning with these clusters through training. A user can tell the system that a specific, weighted combination of some clusters represents a derived feature of the image. By making this association, it is possible to select all images in the database that have that specific combination and may therefore contain the feature that the user is searching for. This step is discussed below in more detail. Second step: information mining The second step in KIM is aimed at assigning physical meaning to the primitive features, i.e. remote-sensing image archives 29 Table 2. Steps for primitive feature extraction in the EO context