Data Mining via Geometrical Features of Segmented Images

Data Mining via Geometrical Features of Segmented ImagesLuca Galli, Antonella Petrelli and Andrea ColapicchioniAdvanced Computer Systems S.p.A., Rome, Via della Bufalotta 378, Italy.The availability of new high-spatial resolution satellite sensors (e.g. IKONOS and QUICKBIRD) permits to have a large amount of very detailed digital imaging of urban environment. But, due to the higher spatial resolution, it is more difficult to apply traditional digital image investigation methods: new analysis algorithms are required to fully exploit these new data.This paper investigates the use of an object-based image analysis approach rather than the more traditional pixel-based one. The object-based approach gives more satisfactory results because spatial information is taken into account. Moreover, the group of pixels contained in each region provides a good statistical sampling of data values for more reliable characterization based on multispectral feature values. In addition, the object shape description may give an additional advice for an appropriate classification. Geometry features become as much important as increase the satellite images spatial resolution.In general, the extraction of such object oriented primary features is a difficult problem: first the image must be accurately segmented in homogeneous areas (with respect to spectral distribution and texture), and then the statistics and the shape of these areas must be extracted.For what concern the segmentation step, we developed a new hierarchical image segmentation algorithm particularly tailored for multispectral satellite images. This algorithm is based on a fast multiscale iterated weighted aggregation method developed by [1]. The image is represented as afour/eight-connected graph, where the coupling associated at each edge represents the likelihood that the corresponding pixels are separated by an edge. Segments are detected finding the cuts that minimize the “normalized cut” global objective function. This is achieved through a recursive-coarsening step, which reproduces the same minimizing problem at different scales using fewer and fewer coarse-scale representative pixels and producing an irregular pyramid. This process is defined by an interpolation rule that relates the coarse to the fine scale pixels. This rule may be interpreted as the likelihood of a fine variable to belong to a coarse scale representative pixel. As pyramid is built from bottom to top, at an appropriate level each segment will be represented through a single node. Then, the location in the image of the segments is determined by applying a top-down procedure. This is achieved by applying the interpolation rule at each segment representative pixel from the highest level downward to the image pixel level, and assigning at each image pixel the segmentation label corresponding to the highest likelihood. Although the “normalized cut” problem represents an edged-based segmentation approach, the possibility to modify the coarse coupling according to the aggregate pixels statistics introduces also a region-based method. This method, combining both region-based and edge-based approach together results well suited for complex remote sensing images, which present heterogeneous multiscale properties.We generalized the weighted aggregation method to multispectral remote sensing data choosing an appropriate spectral homogeneous criterion. The criterion is highly dependent on both the choice of “spectral closeness” metric and the statistical model used to describe the multichannel data. To this purpose, we adopted the multivariate Gaussian model, and we preferred the Bhattacharyya distance with respect to the more popular Mahalanobis one. In fact, the Bhattacharyya distance is more general,and, also, for multivariate Gaussian distribution it has a closed expression, which is an analyticalfunction of the mean and covariance matrix. Such statistical moments can be easily handled within our algorithm: since they can be computed recursively during the bottom-up pyramid building, the overall linear complexity of the segmentation process is preserved. More details can be found in [2]. In the following figure is reported both an IKONOS image of San Pietro (on the left) and the corresponding segmented image (on the right).For what concern the shape geometrical characterization, we used moment descriptors approach. These moments describe special geometric features, such as form factor, roundness or modification ratio. In particular we used both Hu and Zernike moments, which are both invariant with respect to: region absolute position, rotation and scale. Moreover, statistical mean and variance are calculated for each segment in order to give a good statistical sampling of data values for more reliable indexing based on multispectral feature values.Our segmentation method together with the shape and statistical analysis have been successfully integrated into KIM/KES, a tool developed within an ESA research project for content-based image retrieval from remote sensing large archives [3-4]. Such object-based indexing features have been easily included into the standard pixel-based indexing features used so far in KIM/KES. In fact, for each pixel inside a region is associated a unique feature vector, which characterizes the entire region (shape descriptor, mean value, standard deviation, etc.). In this way, standard class files containing the feature parameters are generated and used as the standard pixel based feature class files [4].REFERENCESFigure: IKONOS image of San Pietro (left) and the corresponding segmented image (right).[1] E. Sharon, A. Brandt, and R. Basri, “Fast multiscale image segmentation”, CVPR, I: 70-77, 2000.[2] L. Galli and D. De Candia, “Multispectral image segmentation via multiscale weighed aggregation method”, SPIE Remote Sensing Symposium, Brugge 2005.[3] M. Datcu, K. Siedel, and M. Walessa, “Spatial information retrieval from remote sensing images-Part I: information theoretical perspective”, IEEE TGRS, 1998.[4] M. Datcu, H. Daschiel, A. Pelizzari, M. Quartulli, A. Galoppo, A. Colapicchioni, M. Pastori, K. Siedel, P.G. Marchetti, and S. Delia, “Information mining in remote sensing image archivi:system concepts”, IEEE TGRS, 2003.。