Research On Case-based Spatial Data Retrieval

With the rapid development of the spatial data acquiring technologies, the demand for new retrieval methods on the magnanimity of spatial data is brought forward. This thesis puts forward the case-based spatial data retrieval, which bases on the spatial relation content of spatial scenes. During the preprocessing of spatial data, a multi-sensor image registration algorithm is proposed to align two heterogeneous spatial images. Based on the formalization of spatial relations, many spatial relation invariants are computed to build a spatial relation eigenvector. The sketch based spatial data retrieval simply applies those invariants to measure the spatial similarity of two scenes. Account for the complexity of spatial scenes, the case-base spatial data retrieval is proposed to retrieve the similar spatial scene using statistic learning methods. A two-phase spatial data index schema is also proposed to support the cased-based spatial data retrieval. And the filter-refinement algorithm and relevance feedback are used to optimize the query process. A series of initiations and fruits are realized as follows:1. A spatial data retrieval framework based on spatial cases and sketches is proposed. The framework is comprised of the preprocessing of spatial data, the spatial data index and spatial data retrieval, which introduces the content-based retrieval to the fields of spatial data retrieval. It breaks through the exact matching query of the traditional spatial database, and affords users the query interface using sketch and spatial case, which provides a new method for spatial data retrieval.2. A multi-sensor image registration algorithm based on multi-resolution shape matching is proposed to align two heterogeneous spatial images. It adopts the wavelet decomposition to get the multi-resolution images which represent the main characters under different resolutions. An improved contour search algorithm is designed to acquire more closed contours uninterrupted in vision. A FFT-based multi-resolution shape matching algorithm is also brought forward to find two similar contours. Experiments show the algorithm has more precision than traditional ones.3. The formalization of spatial relations is also analyzed in this thesis. And a lot of spatial relation invariants are given out, including the spatial topological invariants, the spatial measurement invariants and the spatial direction invariants. Using these invariants, a sketch-base spatial dataretrieval system is proposed by computing the distance between the spatial eigenvectors of two spatial scenes. Experiments show the invariants really reflect the spatial relations of scenes.4. The case-base spatial data retrieval is proposed to retrieve the similar spatial scene using statistic learning methods. On the basis of extracting the spatial relations invariants of the spatial scene, the independence topological relation feature space are constructed by using Independent Component Analysis (ICA), and the spatial relationship features of the spatial scene are recognized by Fuzzy Support Vector Machines. The experiments show the typical scenes can be distinguished by this method.5. A two stage index mechanism is proposed for the case-based spatial data retrieval, which introduces the spatial relation features into the traditional R-Tree. This paper also gives out the insertion, query and splitting algorithms. Many experiments proved the index method not only keep the efficiency of traditional spatial queries, but also support the case-based query.6. The thesis proposes two filter-refinement algorithms using reverse searching and obverse searching, which respectively optimize the retrieval with non-spatial-attributes and without non-spatial-attributes. At the same time, the relevance feedback based user interface is designed to adjust the weights of the spatial relation features according to the feedback's model.