Research On Object Recognition Of Hyperspectral Remote Sensing Based On Two Dimensional Quaternion Valued Singular Spectrum Analysis

With the exploration of the land surfaces,ocean resources,and even extraterrestrial planets,it is becoming more and more important to recognize different objects over wide space.However,due to the complex and changeable nature environment,the exploration of land surface and ocean resources will lead to the great amount of cost of instruments and resources.Therefore,more and more scientists have adopted hyperspectral remote sensing technology to analyze and recognize different objects.However,the quality of hyperspectral imaging is easily affected by the external environment.And the complex and changeable imaging environment often leads to a serious decline in the quality of hyperspectral image.Therefore,the recognition accuracies will be vastly low without the further process of the hyperspectral images.As an adaptive decomposition algorithm for processing the hyperspectral images,singular spectrum analysis,has attracted the attention of scholars in the field of hyperspectral images processing.However,the conventional singular spectrum analysis algorithms are mainly used for single plane processing,or single sequence processing problem,resulting in the loss of the spatial information or spectral information when processing the hyperspectral images.Finally,it will lead to a lower recognition accuracy.Therefore,to overcoming the shortcomings of these conventional methods,this thesis proposes a two dimensional quaternion valued singular spectrum analysis based object recognition method for hyperspectral images.Here,the main innovative achievements are shown as follows:Theoretical innovation:（1）By extracting local information from a color plane by moving the singular spectrum window,the conventional two dimensional real valued singular spectrum analysis decomposes a two dimensional signal into the sum of various singular spectrum components,which representing the signal trend,oscillation or noise.Due to this perfect characteristic,it is widely used in the field of two dimensional signal processing.However,the two dimensional real valued singular spectrum analysis has to process color planes one by one when processing the three dimensional signals（such as color images,hyperspectral images,etc.）,which leads to the neglection of the relationship among these color planes.On the other hand,a quaternion is a hypercomplex number with three imaginary components and one real component.There is a unique mathematical relationship among these components.Therefore,in order to overcome the shortcoming of conventional singular spectrum analysis,the novel two dimensional quaternion valued singular spectrum analysis is introduced,using the characteristics of a quaternion.It is convincing that the proposed method will make good use of the perfect characteristic of the singular spectrum analysis while retaining the relationships among each color planes when facing three dimensional signal problems.Application innovation:（2）There are many studies working on decomposing hyperspectral images and extracting useful components by singular spectrum analysis.However,there is one topic that scholars cannot escape is that conventional singular spectrum analysis shows bad performace on processing three dimensional signals.For examples,one dimensional real valued singular spectrum analysis can only decompose pixels one by one in the images which will neglect the correlation among spatial positions of pixels.That is,ignoring the spatial information.On the other hand,the two dimensional real valued singular spectrum analysis can only process spectral planes one by one in the images,which leads to the neglection of the correlation among spectral planes.Therefore,the proposed two dimensional quaternion valued singular spectrum analysis is employed for performing the object recognition of hyperspectral images.By applying the two dimensional quaternion valued singular spectrum analysis on the hyperspectral images,selecting the two dimensional quaternion valued singular spectrum analysis components,the correlation among both the pixels within each color plane and the pixels across different color planes are exploited.（3）It is worth noting that the components corresponding to larger singular values are of higher importance while the components corresponding to smaller singular values are seen as noises.Therefore,the singular spectrum analysis components corresponding to relatively larger singular values are usually used for reconstructing the signals.However,this approach is not always effective when facing the hyperspectral image recognition problem.Therefore,this thesis proposes a linear discriminant analysis based component selection method.It has been shown that the proposed method will effectively improve the recognition accuraces by maximizing the the interclass separation and minimizing the intraclass separation of the feature vectors.