Hyperspectral Image Classification With Spectral–spatial Feature Learning

Hyperspectral image classification is a challenging task as there are two key issues impacting on the classification performance: the high spectral dimension and the high spatial resolution,and they can be addressed by incorporating the spatial information into the spectral information.After extensive research on spectral-spatial feature information of hyperspectral data,in this paper,we propose an effective hyperspectral image classification framework which based on spectral-spatial feature learning.Firstly,we present a novel spectral feature learning algorithm based on weighted average fusion for extracting spectral features of the original hyperspectral datas.After weighting all the spectral bands,we divide all of bands into several subgroups,then the bands in each subset are fused together by averaging.It is a simple yet quite powerful spectral feature extraction method.Secondly,we propose two methods based on spatial feature learning to extract the spatial feature of the fused bands,which can incorporate the spatial contextual information into the classification process.They are albedo recovery method and the structure extraction method based on total variation,respectively.In the albedo recovery method,we use local neighborhood constraint to recover the intrinsic albedo feature from hypespectral images.For the affinity matrix that represents the neighborhood attribute information,we propose to use domain transform to solve it to effectively reduce the running time of the algorithm.In the structure extraction method based on total variation,a novel two-scale decomposition-based relative total variation method is proposed to extract structure features of hyperspectral images.By integrating the information of base scale into the detail layer to extract structure can help to get results more reasonable.In addition,we also develop a novel diffusion function that can well preserve the structure information of the hyperspectral images while having a more higher diffusion velocity.Next,recent theoretical results disclosed that optimizing the margin distribution achieves a better generalization performance than maximizing the minimum margin.Thus,we use large margin distribution machine instead of support vector machine as classifier to classify the learned spectral-spatial features.Large margin distribution machine is intrinsically a binary classifier,in order to apply it to hyperspectral image classification,we develop a multiclass large margin distribution machine from a set of binary large margin distribution machine classifiers by using the ”One Aaginst One” parallel strategy.Finally,we conduct lots of experiments to demonstrate the superiority of the proposed classification framework(including three hyperspectral datasets,four quality metrics and six comparing algorithms),and also analyze the related parameters settings in detail.In addition,by using the proposed spectral-spatial feature learning methods,the accuracy of the large margin distribution machine classifier can be improved significantly,which further demonstrate the effectiveness of the proposed spectral-spatial feature learning algorithms.