Based On The Optimal Band Selection Of Hyperspectral Feature Classification Method Research

Hyperspectral remote sensing has been attracted more and more attention,because its high spectral resolution plays a key role in object recognition.Currently hyperspectral remote sensing has been applied to mineral exploration,fine land change monitoring,precision agriculture,ocean color and SST monitoring,pest and disease monitoring and other fields that conventional remote sensing technology can't meet the needs.However,hyperspectral remote sensing still has some problems: Firstly,ultra-high spectral resolution hyperspectral remote sensing leads to an increase of data,which gives burden on rapid processing of data;Secondly,since the spectral characteristics of different features are same or similar within a wavelength range,which resulting in the spectral characteristics of adjacent bands with strong correlation and information redundancy;Furthermore,the sensor and external environment may influence the data production processing,and product lots of noise,which would reduce the accuracy of the feature recognition greatly during the data process;Finally,“Hughes” phenomenon exists.Overmuch bands take part in classification can't improve the classification accuracy,on the contrary,it may be reduce the classification accuracy in a way.Therefore,how to reduce the dimensionality effectively,extract the key bands and use the efficient classification arithmetic for classification,are three directions of hyperspectral remote sensing.On the basis of the characteristics of hyperspectral data,we proposed an arithmetic which based on subspace—rough set on band selection.The arithmetic based on traditional subspace division,and by using of reduction ideas to select the spectroscopic feature,then,by using of GLCM method to calculate the texture information of the primary bands and reduction the texture information bands,finally,we get the spectrum and texture information bands by reduction and overlay.In the part of terrain classification,we choose the different attribute sets based on the band selection.According to the integration theory,we use the BP neural network,SVM and wavelet neural network to classification for each attribute set.Finally,by using of majority vote to confirm the category pixel by pixel,ultimately get the final classification result.In this paper,we used Heihe data,Indiana band data and Dalian oil spill data test the band selection algorithm,and used the integration algorithm to test its availability.Experiments show that,by using of the subspace—rough set to select bands,not only reduce the dimension greatly,but also dig the texture information of the data,and achieve a comprehensive and effective use of the spectrum information and texture information.Compared with the traditional BP neural network,SVM,and wavelet neural network classification algorithm,the integrated classification algorithm could further improve the classification accuracy.