Researches On The Method Of RBF Networks And SVR Modeling Based On Spatial Adjacency

With various data collection means advancing, spatial data analysis has faced more complex space-time objects and more huge amounts of space-time data than past. Some traditional means of spatial data analysis have not meet dealing with the huge amounts of data in the geographic information system and geo-information science. Computational intelligence including artificial neural network (ANN) and support vector machine (SVM) plays a more and more important role in the research of related fields with geo-science. The developing tendency is irreversible, which of all kinds of complicated information fusion theories and techniques, integrated Geo-information science with intelligence science.In this paper, taking the integration of spatial data analysis with computational intelligence as start, with basis of spatial knowledge and spatial relations, making use of the basic principle and algorithms of traditional radial basis function neural networks (RBF) and SVR, according to the specialty of spatial data, some resuits are obtained:(1) It is established that several RBF and SVR models based on the spatial adjacency, viz. RBF-IF,RBF-HF1/2,RBF-OF and LS-SVR-IF,LS-SVR-HF1/2,LS-SVR-OF, and an analysis is made with an example that the data is from the crime rate information of 49 areas in Colombia, Ohio, USA;(2) Through comparison and analysis again and again, applying some predicted performance criterion to make results evaluation, it is proved that mutual integration of spatial data analysis with computational intelligence is feasible;(3) Through comparison between regional RBF and SVR and that between RBF-HF1 and LS-SVR-HF1/2, it is proved that the generalization ability of SVR is superior to that of RBF;(4) Least squares support vector machine regression (LS-SVR) is applied in SVR models, simplifying study algorithms, increasing study speed;Finally, taking RBF and SVR models based on spatial adjacency as study objects, in Visual C++ with MATLAB, the module is compiled, with analysis of an example made, achieving forecasting effect.