Research On Spatial-Spectral Dimensionality Reduction And Land-Cover Classification For Hyperspectral Imagery

Hyperspectral imagery(HSI)covers a rich spatial-domain and spectral-domain information,which contributes to the detailed recognition and land-cover classification of HSI.However,a large number of bands with strong correlation leads to redundancy,which easily causes a“Curse of dimensionality”with the traditional methods to process the hyperspectral data directly.Therefore,one of the hotspots of remote sensing field is how to effectively extract discriminat features from hyperspectral image for improving the classification performance.Inspired by the structure characteristics of HSI data,this paper focuses on the feature extraction and classification of HSI from three aspects:spatial-spectral combination,manifold learning and hypergraph learning.The main works are listed as follows:(1)The current research of HSI processing was introduced and analyzed.We summarized several representative dimensionality reduction and classification methods for HSI data.Meawhile,the HSI data sets used in all experiments and the evaluation parameters reflecting the classification performance of different algorithms are also introduced.(2)A spatial-spectral combined manifold reconstruction(S~2MRC)method was proposed for land-cover classification of HSI.In general,conventional HSI classification are supervised learning methods and they only consider spectral-domain features.S~2MRC exploits a few labeled pixels to seek its unlabeled spatial neighbors for semi-supervised learning based on the spatial consistency in HSI,and it computes the reconstruction error of test samples on each sub-manifold,which characterizes the similarity between pixels and imprpves the land-cover classification performance.The experiments on the Indian Pines scene and Pavia University scene data sets show that the classification accuracies of the proposed method are better than other algorithms under various conditions,especially when the number of training samples is small,it I more suitable for practical application.(3)A spatial-spectral combined regularization sparse hypergraph embedding(SSRSHE)algorithmwas proposed for dimensionality reduction of HSI.Many traditional graph-based dimensionality reduction(DR)algorithms mostly use spectral information to represent the first-order relationship between pixels.In view of this,SSRSHE exploits sparse representation to reveal the correlation between pixels for adaptively select neighbors,and then constructs the sparse intra-class and inter-class hypergraphs based on class labels of samples,which effectively characterize complex multi-relationships between pixels.Then,it utilizes the spatial consistency of HSI to compute local spatial neighborhood scatter to perserve the local neighborhood structure of samples.Furthermore,the overall scatter is explored to reveal the global structure of HSI data.Finally,in the low-dimensional embedded space,the within-class samples is compacted as close as possible,while the samples from different classes is kept away as far as possible,and the discriminat features are extracted to improve the land-cover classification performance.The experimenetal results on the Indian Pines scene,Pavia University scene,Kennedy Space Center scene HSI data sets demonstrate that the proposed method can improve the classification performance of HSI compared with the traditional spectral-based DR algorithms.In summary,this paper mainly conducted spatial-spectral combined dimensionality reduction and classification of HSI based on manifold learning and hypergraph learning.This paper proposed a spatial-spectral combined regularization sparse hypergraph embedding algorithm and a spatial-spectral combined manifold reconstruction land-cover classification algorithm for HSI data.The experiments on the public data sets show that the proposed algorithms can obviously obtain better classification accuracies for compared land-cover classification methods.