Visual mining of large DEM and image data bases

The availability of Earth Observation (EO) imagery is relatively recent, i.e. less than 20 years. With the development of REMOTE Sensing (RS) technology, (e.g. high-resolution images, higher acquisition frequencies and new sensor types), the scope of RS Images allows users to perform a huge variety of specific applications, such as land cover mapping, thanks to increased levels of information and the possibility of synergy between data sets. Information in forms of geo-referenced DEMs / Images should be used for 3D real-time visualisation over large area at the country scale while preserving high resolution data/content and realistic rendering. Up to now, the exploitation / integration of large and complex databases integration in the overall interactive 3D visualisation process is currently extremely difficult. However, the demand for such technologies exploded in the last years and lead to a reinforcement of world acquisition programs (SRTM2). Therefore, RS and Virtual Reality communities are faced to several challenges to deal in real-time with such amount of data: e.g. Landsat images mosaic / ERS Tandem DEM over the whole Germany (gridding: 25m) Ô (40000² pixel images) × (R-G-B channels + elevation information). In addition to the rendering time constraints, realistic visualisations require to enhance / regularise the database. In this purpose, an important level of understanding and content extraction have to be reached in order to perform significant improvements in the data and particularly in the DEMs, which bring the geometry information. Indeed, DEM information is now recognised as one of the most important data structures used for geo-spatial analysis and 3D rendering. Nevertheless, despite high accuracies, EO DEM are still pervaded with errors and artefacts mainly due to the acquisition / generation techniques. As a consequence, the elevation data have to be analysed, filtered and enhanced. These pre-processing steps are determinant in order to cope the artefacts, generate a higher level of realism and simplify the data for the virtual reality enhancing processes (meshes simplification, hierarchical decomposition...). The article presents / evaluates several signal processing techniques and information extraction algorithms: Non-stationary approach (Bayesian approach, Gauss Markov Random Fields(GMRF)) Multi-resolution approach (fBm, Wavelets) Segmentation algorithms Object Extraction /Topology analysis Since world DEM coverage are available (SRTM, generated by SAR interferometry), a Bayesian filter has been developed to deal with non-stationary data such as DEM. It intends to reduce the thermal noise, smooth the Phase Unwrapping (PU) artefacts while preserving contours and objects included in the data. Despite the good results obtained in term of statistical analyses and rendering aspects, the filtering is still not enough to cope with important artefacts (specular reflexion, PU). Indeed, complementary information have to be added to reach very realistic flight simulation. The principle of the DEM regularization is the following: Based on the image data, relevant information are extracted. An interactive learning procedure is done through a Graphical User Interface is used to link the signal classes to the semantic ones, e.g. to include human knowledge in the system. The selected information, in form of objets are merged with the DEM by assigning regularisation constraints. In this purpose, a segmentation algorithm based on region growing method is presented and an object extraction algorithm. The extracted objects are classified and stored in a tree structure in the sense to preserve topology relations between the objects and reflect their dependencies. A database preparation line for Virtual Reality scenarios has been presented. It constitutes a key step for realistic 3D visualisations of EO data and for the rendering optimisation to manage flight simulation in real time of large areas. 1. DEM: Digital Elevation Models: digital representation of the Earths relief 2. SRTM : Shuttle Radar Topography Mission: Generation of World DEM coverage, largest homogenous DE