Land Cover Classification From Hyperspectral Remote Sensing Images Via Spatial-Spectral Collaborative Representation

Accurately monitoring and updating land cover information plays an important role in environmental protection,land resource planning and management,landscape pattern analysis,sustainable development,and so on.In recent years,large-scale land cover classification based on hyperspectral remote sensing images and machine learning algorithms has become a research hotspot in the field of remote sensing.The classification method of collaborative representation(CR)does not consider any prior distribution of sample data and does not need training process,thus avoiding the impact of the number of labeled training samples on the model fitting performance,and providing a new idea and method for hyperspectral imaging classification.In view of the imbalance in the number of training samples for each class,the spectral shift caused by the adjacency effect,and the inadequacy of spatial-spectral features in hyperspectral images,this paper applies local nearest neighbor,spatial weighting mechanism,spatial structure information mining and other methods to carry out research on land cover classification via hyperspectral images based on CR models to further improve the classification performance of CR models for ground objects in hyperspectral images,especially in the case of small-scale labeled samples.The main works and conclusions of this paper are as follows:(1)The imbalance of the number of training samples for each class in the dictionary of existing CR models affects the classification performance.To address this problem,this paper proposes land cover classification methods for hyperspectral images based on local nearest neighbor collaborative representation,namely LNNCRC and LNNCRT methods.And the performance and running time of land cover classification in hyperspectral images with unbalanced training samples are compared between the proposed methods and several related methods.The experimental results show that the proposed LNNCRC and LNNCRT methods have more advantages than other methods in classification performance and running time.And the proposed LNNCRT method achieves the best classification performance,with overall accuracy(OA),average accuracy(AA),and Kappa coefficient(Kappa)reaching 93.04%,90.31%,and 0.9071,respectively.The proposed methods not only further eliminate the interference of training samples and classes unrelated to test samples in CR dictionary,and effectively eliminate the influence of unbalance of training samples on the classification performance of CR model,but also effectively reduce the computational complexity of CR model.(2)To solve the problem of spectral shift caused by the adjacency effect in hyperspectral images,this paper proposes land cover classification methods for hyperspectral images based on weighted spatial-spectral kernel collaborative representation,namely WSSKCRT and WSSDKCRT methods.And the performance of land cover classification in hyperspectral images with small-scale labeled samples is compared and analyzed between the proposed methods and several related methods.The experimental results show that the proposed WSSKCRT method achieves the best classification performance,with OA,AA,and Kappa reaching 95.69%,95.56%,and 0.9429,respectively.Furthermore,the proposed WSSKCRT and WSSDKCRT methods obtain more than 94%OA value in the case of smallscale labeled samples,which not only can effectively alleviate the impact of spectral shift caused by adjacency effect on the classification performance of CR model,but also can improve the land cover classification performance of CR model in the case of small-scale labeled samples.(3)To fully mine the spatial-spectral features in hyperspectral images,this paper proposes a land cover classification method for hyperspectral images based on weighted spatial-spectral joint kernel collaborative representation,namely WSSJKCRC method.The performance of land cover classification is compared and analyzed between the proposed method and several related methods on three sets of real hyperspectral images with smallscale labeled samples.The experimental results show that the proposed WSSJKCRC method achieves the best classification performance on the three sets of hyperspectral scenes,and the values of OA,AA and Kappa for WSSJKCRC on the three sets of hyperspectral scenes reach more than 95%,96%and 0.95,respectively.This indicates that the proposed WSSJKCRC method can effectively mine the spatial-spectral features in hyperspectral images,and can further improve the land cover classification performance of CR model in the case of small-scale labeled samples.