SAR IMAGE RECOGNITION BY INTEGRATION OF INTENSITY AND TEXTURAL INFORMATION

Synthetic aperture radar (SAR) images exhibit many interesting pecularities but are characterized by considerable speckle noise, which noticeably affects image quality. In this case, conventional statistical classifiers employing intensity data yield poor results. In order to improve classification methods, we suggest the use of co-operative low-level techniques, driven appropriately by a knowledge-based structure. In addition to the intensity image, an important role is also played by textural information obtained by fractal analysis. The resulting textural image has proved useful for both segmentation purposes and region characterization, and integrates the textural information obtained by conventional statistical methods. The system architecture is based on a database (blackboard type), where intermediate results can be stored, and on a control structure with a rule interpreter. Segmentation provides 'elementary' regions which are characterized by some features, and represent the symbolic data manipulated by the high-level subsystem. The knowledge about areas or objects (urban areas, mountains, lakes, crops, flat lands, and valleys) is inserted in a semantic net. Preliminary results are promising, thus confirming the potentialities of the knowledge-based approach. The novelty of this work lies basically in the use of a fractal method for texture characterization and texture-based segmentation (using an adaptive technique developed by the authors), and in the integration of the fractal dimension and of other textural information with intensity features, under the control of a knowledge-based system.