Adaptive Neighborhood Based Spatial-spectral Kernel Method For Hyperspectral Classification

With the further development of hyperspectral remote sensing processing technology,hyperspectral image classification has become a research hotspot in recent years.Hyperspectral image classification technology is facing many challenges,such as ultra-high spectral resolution,limited training samples and image noise.Based on the existing spatial spectral kernel classification method,this paper makes full use of the characteristics of the distribution of objects in hyperspectral images and the correlation between neighbor pixels,and further improves the classification accuracy and efficiency of hyperspectral images under the condition of small samples.The main contributions of this paper are as follows:(1)Two kinds of hyperspectral image classification methods based on adaptive neighborhood are proposed.One is edge-modified superpixel based spatial-spectral kernel,which constrains the homogeneous regions of the pixels at the edge of the superpixel through a square window,satisfactory classification performance is achieved.The other is low-rank induced spatial-spectral kernel.This method assumes that the adjacent pixels are more likely to be the same kind of object.It is assumed that the same object have similar spectral characteristics,that is,there is a potential low-rank condition to screen out these similar pixels to obtain homogeneous regions,therefore the low-rank induced spatial-spectral kernel could be constructed for hyperspectral image classification.(2)A generalized form of composite kernel and spatial-spectral kernel is proposed.Composite kernel and spatial-spectral kernel can be regarded as extracting spatial spectral information in original space(Euclidean space)and feature space(manifold space)respectively,while there is an assumption that manifolds and European space are similar locally,based on this,a generalized spatial-spectral kernel is proposed?By implementing super-pixel as the local region and combines this with multi-scale super-pixel strategy,an adjacent super-pixels based multi-scale spatial spectral kernel method is proposed.This method further improves the classification accuracy without adding additional parameters,and achieves very high classification efficiency.The classification results on real hyperspectral datasets verify the effectiveness of the proposed method.