Research On Methods For Anomaly Detection And Target Recognition In Hyperspectral Images By Combining Spatial Relationship And Spectral Characteristics

With the development of technology of remote sensing,not only the spectral resolution but also the spatial resolution of hyperspectral images(HSIs)are improved.Thus,using the double-high hypersepctral images to realize target detection and recognition becomes a research hotspot in hyperspectral remote sensing community.Anomaly detection is the main way of realizing target detection.The mainstream anomaly detection methods only rely on the spectral difference,but ignore the relationship between pixels.On the other hand,in the double-high observation data,a target always presents as an agglomeration(containing multiple pixels)that is composited of several different materials.The main spectral matching methods being designed for solving the target recognition on single pixel or subpixel level could not be applied directly on those new data.Around the above two problems,this dissertation is dedicated to anomaly detection and target recognition for HSIs from the spatial relationships and spectral characteristics.The research achivements of this dissertation are listed as follows:1.A tensor decomposition based anomaly detection algorithm for HSI is proposed.Anomalies usually refer to targets with a spot of pixels(even subpixels)that stand out from their neighboring background clutter pixels in HSI.Compared to backgrounds,anomalies have two main characteristics.One is the spectral anomaly,i.e.,their spectral signatures are different from those associated to their surrounding backgrounds;another is the spatial anomaly,i.e.,anomalies occur as few pixels(even subpixels)embedded in the local homogeneous backgrounds.However,most of the existing anomaly detection algorithms for HSI only employed the spectral anomaly.If the two characteristics are exploited in a detection method simultaneously,better performance may be achieved.The third-order(two modes for space and one mode for spectra)tensor representation of HSI has been proved to be an effective tool to describe the spatial and spectral information equivalently.Basing on this advantage,we proposed a new anomaly detection method which is divided into three steps:(1)Three factor matrices and a core tensor are first estimated from the third-order tensor that is constructed from the HSI data cube by using the tensor Tucker decomposition.Each factor matrix of each mode is composited by the eigenvectors ordered by decreasing magnitude of the respective eigenvalues.And their major and minor eigenvectors are likely to correspond to the main and secondary principal components(PCs)of the image,respectively.(2)A fast and adaptive method for determining the first largest PC number of the three matrices(i.e.,K1,K2 and K3)is proposed.The aim is to eliminate the background information as much as possible.Firstly,an energy function,whose independent variables are K1,K2 and K3,is constructed.Such energy function is used to measure the spatial anomaly degree and spectral anomaly degree of the secondary component of the HSI.Thereby,determining K1,K2,and K3 is converted into an energy optimization problem.Then,we compute the slice-variance curves of the core tensor along the three modes and determine the initial value of K1,K2 and K3 according to the largest changing slope criterion.Finally,an exhausitive method is applied to find out the optimal value of K1,K2 and K3 in the neighboring region of the initialization.(3)According to the optimal K1,K2 and K3,we can obtain a most anomalous secondary component of the HSI which only contains the information of anomaly target and noise.Under the assumption of the noise is a Gaussian noise,we apply the CFAR algorithm to accomplish anomaly detection.Multiple HSIs have been experimented and analyzed.Comparing with the mainstream anomaly detection methods,the experimental results have verified the superiority of our method in the detection performance.2.An anomaly extraction method based on combining spectral reconstruction accuracy and spatial constraint from HSIs is proposed.The existing anomaly detection methods for HSI only concern the locations of anomaly targets,but omit their spectral signatures and sizes.However,such information is crucial for further identifying whether the detected anomaly is a target of interest or natural clutter.This dissertation combines the anomaly targets detection and their spectral signatures and sizes estimation to one problem to propose a concept of the anomaly extraction.To realized it,we designed a new method which is divided into four steps:(1)Firstly,a pixel-spectral-component label matrix with the same size of the HSI in spatial domain is introduced.Each element of the matrix is a binary number which descripts the composition of spectral component of the corresponding pixel.The length of the number is one more than the number of the background regions of the whole HSI.If the pixel contains a certain kind of background,then the corresponding bit is 1;otherwise,the bit is 0.If the pixel contains the anomaly spectrum,then the last bit is 1;otherwise,the last bit is 0.In this case,the problem of anomaly detection is converted into the problem of solving the label matrix.(2)Then,an energy function,whose independent variable is the label matrix,is introduced.In this dissertation,we firstly prove the label matrix must satisfy the sum-8 spatial constraint.Then,the energy function is defined as the sum of the spectral reconstruction accuracies of all pixels,under such spatial constraint.Each pixel’s spectral reconstruction accuracy is computed by comparing the observation spectrum and the reconstructed spectrum from the label matrix.Thus,the problem of estimating the label matrix is converted into an energy optimizing with constraint problem.(3)Thirdly,a method for initializing the label matrix is designed.Firstly,we find out all pure background pixels(PBPs)through the local linear fitting(LLF)method.Then,according to the spectral similarity and spatial connectivity,we divided those PBPs into different background regions.We compute the mean spectrum of each background region and assign a certain bit to label it.In addition,we use the labeled PBP to fit the unlabeled pxiels,if a pixel can be well-fitted,then label it as a mixed background pixel(MBP);otherwise,label it as a anomaly pixel.Finally,we change the pixels’ label to satisfy the sum-8 spatial constraint.(4)Finally,a iterative strategy for optimizing the label matrix is designed.Firstly,we find out the pixel whose energy is minimum.Then,change it to another possible label under the spatial constraint and compute the new energy of the HSI.If the energy increased,then accept the alteration;otherwise,change the label to another possible one.Repeating this iterative procedure until the energy of the HSI is convergent.Multiple HSIs have been experimented and analyzed.Comparing with the mainstream anomaly detection methods,the experimental results have verified the superiority of our method in the detection performance.Furthermore,our method can also estimate the spectral signatures and sizes of the detected anomaly targets reliably.3.A tensor block term decomposition based target recognition method for HSI is proposed.The traditional spectral mathcing based target recognition methods only exploited the spectral information of the target.However,in the double-high hypersepctral images,the spatial features of the target represent remarkably.Under this case,only spectra information could not be used to descript the target adequately.The tensor block term decomposition can integratively descript the spatial distributions and spectral characteristics of different materials of the target.Basing on this advantage,we proposed a new target recognition method which is divided into two steps:(1)A set of terms can be obtained when apply the tensor block term decomposition on the HSI.Each term is composited of a matrix and a vector,where the matrix represents the spatial distribution of a certain material and the vector represents the material’s spectral characterisctic.Thus,a target with multiple materials is decomposed into a collaborative representation of different materials’ spatial distributions and spectral characterisctics.(2)Define the similarity measure based on the above mentioned target description model.The measure combines both of the correlation of the spatial distribution and the correlation of the spectral characteristic of each material.Two HSIs have been experimented and analyzed.Comparing with the spectral matching based target recognition method,the experimental results have proved the validity of our method.