Hyperspectral Image Classification Combining Improved Local Binary Mode and Superpixel-level Decision

‍ ‍Hyperspectral images to be classified with a small number of labeled samples， in addition to using spectral features， spatial texture features are also indispensable. This paper proposed a support vector machine （SVM） classification method that uses multi-scale and multi-directional Local binary mode （LBP） descriptors to obtain texture features and combines super-pixel level to guide decision. First， the method in this paper improved the traditional LBP descriptor into a multi-scale and multi-directional LBP descriptor. On the one hand， the relationship between neighboring pixels was fully considered， and on the other hand， the horizontal and vertical directions and diagonal directions were considered in the calculation. Second， when using statistical histograms to obtain texture features， a combination of multiple size windows was used to obtain multi-range， high-precision texture features. Third， the traditional super pixel segmentation method， simple linear iterative clustering （SLIC） was improved， the spectral distance was redefined and the texture feature distance was introduced to obtain a more accurate super pixel segmentation map. Finally， the super-pixel segmentation map combined with the majority voting strategy was used to further guide and correct the classification results. Experiments show that the proposed method can extract texture features more effectively， combined with the guiding decision of superpixel segmentation maps， and further improve the classification performance of hyperspectral images.