Hyperspectral Image Band Selection Based On Local Spatial Information

The hyperspectral image data contains hundreds of spectral segments with very high spectral resolution,and the spectral information in hyperspectral images can be used for accurate classification of features.But the high-dimensional form of data and the high degree of redundancy of information bring great challenges to the subsequent data processing.Therefore,how to preserve the useful information of hyperspectral data to the maximum extent,while reducing the data dimension becomes one of the important technical problems of hyperspectral image processing.Band selection is a commonly used hyperspectral dimension reduction method.By selecting some of the bands in the original band set and making no changes to the data,this method preserves the physical meaning of the band.In the classification of hyperspectral images,the combination of spectral information and spatial information is widely recognized,but in the field of band selection,the combination of spatial information has not been paid enough attention.Based on this,two kinds of wrappered hyperspectral band selection methods based on local spatial information are proposed in this paper.(1)Band Selection Method Based on Markov Model.The algorithm uses the Markov model to combine the local spatial information with the spectral information effectively,and use the value of the energy function to judge the discrimination of the current band.The data items in the model are composed of the classification probability of each pixel,and the smoothing term reflects the consistency of the adjacent pixel labels in the local space.The algorithm selects bands with high classification probability and high consistency of adjacent pixel labels.The energy function is smaller,the corresponding band quality is better.The minimization of the energy function is solved by the graph cuts algorithm.(2)Band Selection Method Based on Super-pixel Segmentation.The algorithm uses the super-pixel segmentation to get the local uniform region,and count the classification results of each pixel in the region.The statistical characteristics reflect the advantages and disadvantages of classification performance.The algorithm uses the principal component analysis(PCA)method to compress the hyperspectral image,selects the first three principal components to select the band,and then uses the SLIC segmentation algorithm tosegment the compressed image to obtain the segmentation graph.Training classifiers according to the mark sample and using the classifiers to classify the hyperspectral image to get the classification map.Finally,the global classification accuracy is predicted by combining the segmentation graph and the classification map,and the band selection is based on its value.The two algorithms proposed in this paper not only make use of the spectral information of hyperspectral data,but also improve the performance of band selection by combining with spatial information.