Feature Extraction And Classification Algorithms Of Hyperspectral Images Based On Quaternions And Moments

Hyperspectral images can provide rich spectral information and spatial information at the same time,which provides favorable conditions for accurately identifying and classify-ing objects.However,it also brings new challenges to the classification of hyperspectral images.First,there is a lack of labeled samples in hyperspectral images.Second,the high dimensionality of the data can greatly increase the computational cost.Third,due to the impact of internal and external factors?such as atmosphere condition,material distribution,sensor parameters,etc.?,the same material may present spectral discrepancy and different materials may have similar spectral signatures,which causes the spatial variability of spec-tral signatures,inevitably degrading the classification performance of algorithms.Faced with the problems encountered in the classification process of hyperspectral images,this thesis carried out in-depth research on feature extraction and classification algorithms of hyperspectral images based on quaternion theory and image moment theory,and mainly s-tudied how to enhance the representation ability and robustness of features.In addition,this thesis also studied how to enhance the discriminative power of the algorithm,and effectively improved the discriminability of the algorithm through l₂regularization.The main works of the thesis are as follows:?1?A multiscale quaternion Weber local descriptor histogram algorithm is proposed.Considering the high dimensionality of hyperspectral images,the algorithm first performs dimensionality reduction on hyperspectral images and use quaternions to represent hyper-spectral images.This greatly reduces the computational cost and the dimensionality of feature vectors,and the algebraic structure of quaternions not only unifies the feature ex-traction process,but also preserves the spatial structure of data.Then,in order to establish the spatial relationship between the pixel and its neighboring pixels,the quaternion Weber local descriptor?QWLD?is constructed.QWLD has strong descriptive ability and good robustness,which effectively alleviates the spatial variability of spectral signatures.In ad-dition,due to the diversity of scales and shapes of homogeneous regions,multiscale feature histograms are further extracted.This not only captures more intrinsic spatial information,but also the statistical properties of histograms enhance the reliability of features.Finally,we combine the extracted spatial features with spectral features and achieve high-precision classification of hyperspectral images.?2?An adaptive weighted quaternion Zernike moments?AWQZM?algorithm is pro-posed.This algorithm inherits the advantage of Zernike moments,which not only effective-ly describes the shape features and internal information of homogeneous regions,but also increases the invariance of features.At the same time,it can use adaptive weights to enhance the similarity of pixels from the same class and the distinctiveness of pixels from different classes,and flexibly adjust the contribution of each pixel in the spatial neighborhood.In addition,AWQZM is constructed in spatial neighborhoods of different scales,which more intrinsically characterizes homogeneous regions.The introduction of AWQZM phase in-formation also makes the extracted features more informative and more distinctive.The experimental results show that the AWQZM not only achieves good classification results,but also has good robustness.?3?A l₂regularization-based discriminative collaborative representation?RDCR?algorithm for hyperspectral image classification is proposed.The algorithm effectively increases the distinctiveness of samples from different classes and the fidelity of discrim-inant results through l₂regularization.Moreover,the RDCR algorithm can directly solve the objective function while maintaining the stability of the solution.Both theoretical analysis and experimental results show that the RDCR algorithm has high stability and good classification performance.