Adaptive Scalable Kernel For Hyperspectral Image Classification

Hyperspectral images are widely used in urban mapping,environmental management,crop analysis and mineral detection.However,when dealing with high resolution hyperspectral images,it is often difficult to obtain high classification accuracy directly.There are two main factors affecting the classification results of hyperspectral images.One is that the high dimensionality of spectral information will produce Hughes phenomenon,which will significantly reduce the classification accuracy.On the other hand,the classification accuracy will be greatly affected by the great improvement of spatial resolution.In this paper,a new framework of hyperspectral image classification based on adaptive scalable kernel is proposed,which has good performance in removing irrelevant texture details and protecting key features.The method consists of three steps.Firstly,we propose an algorithm based on band fusion for dimensionality reduction of hyperspectral images.The fused image has more detailed information,such as edge and structure,and can expose potential texture features,which also has a great role in promoting the next step of texture separation.Then,we use recursive filters to extract image features,but in the process of using recursive filtering,the performance of the guide graph is the key factor to determine the filtering result.In order to ensure the integrity of the detail and corner structure of the guidance graph,we propose the concept of adaptive kernel scale.Traditional guidance maps are obtained by fixed-scale Gaussian smoothing,which will lose many small-scale structures such as corners.And our algorithm can make up for this defect very well.It can protect the small-scale structure very well and protect the area away from the direction of x and y very well.In the classification step,we use a large margin distribution machine to classify the spectral structure feature maps.The traditional SVM classifier maximizes the minimum margin,but does not consider the effect of margin distribution on classification accuracy,the large margin distributor can optimize the margin distribution and get better results.We also improve the two-class large margin distribution machine to a multi-classifier through the OAO algorithm.Finally,we have done a lot of experiments on three data sets and selected three evaluation indicators to evaluate the performance of the algorithm.Five hyperspectral image classification algorithms are selected in the comparative experiments.The results show that the proposed algorithm achieves a fairly high level of classification accuracy and computational efficiency.