Research On Hyperspectral Image Umixing Based On Tunable Archetypal Analysis

Mixed pixels in the image have become the key factor which restricts the accurate identification of hyperspectral image.Variation of the Field of View(FOV)of imaging equipment,height of imaging platform,number of class types and their distribution ranges leads to the sensed image data with complicated distribution of pure and mixed pixels.To undertake spectral unmixing adaptive to the characteristics of mixed data,research of spectral unmixing covers a wide range from linear unmixing models to nonlinear unmixing models,from single endmember based unmixing to multiple endmembers based unmixing.However,most models has limitations in adaption to the mixed characteristic of analyzed data,thus they neither can be used as a common methodology.As the spread of remote sensing field,it brings great challenge for spectral unmixing by various mixed characteristic of mixed data.The Nonnegative Matrix Factorization(NMF)used for spectral unmixing can achieve estimation of endmembers and abundances simultaneously.For the sake of flexible modification of NMF model,it has been a hot study of spectral unmixing with NMF models in recent years.The adaptability of the NMF model to the mixed data characteristics depends on the relationship between the generated endmembers and the mixed data.The model interpretation is of great importance especially when both pure pixels and mixed pixels exist and spectral variability become dominant.However,few NMF unmixing models are explored for the direct relationship between generated endmembers and the mixed data.They are neither studied for the common use referring to the characteristic of mixed data.In this dissertation,based on nonlinear kernel mapping and linear sparse representation,we focus on study of tunable NMF unmixing models which has flexibility to adapt to characteristics of mixed data and can be used as common methodologies.More specifically,the main aspects included are as follows:(1)To explore the relationship between the characteristics of mixed datasets and unmixing methods which generate endmembers,Archetypal Analysis(AA),a type of NMF model with explicit indication set of combinations in generated features,is introduced for unmixing.For the highly mixed data with absent of pure pixel,AA model with added relaxed factor is explored to solve the unmixng problem of such kind of dataset.The experimental results demonstrate that,the relaxed AA model with proper parameter is more effective than other state-in-art unmixing methods in the endmember and abundance estimation of highly mixed hyperspectral image.(2)To indirectly deal with the the hyperspectral image with nonlinear interaction effect and spectral variability,Kernel Archetypal Analysis(KAA)is investigated.Our study proves KAA has the fuctions of both finding the principle convex hull and clustering.Then,the parameter setting of kernel width depending on data distribution is provided which should be based on PCA dimension reduction and given general information of the mixed level of analyzed data.Besides,two fast KAA unmxing methods,namely Nystr?m Fast KAA(NFKAA)and Random Nonlinear AA(RNAA)are developed based on Standard Nystr?m method and random nonlinear feature mapping.The experimental results shows that KAA achieves more promising unmixing results than other linear unmixing methods in unmixing the data with nonlinear interaction and spectral variability effects.Through tuning the kernel parameter as the way proposed,the KAA model can be effectively tuned to adapt to the mixed data with spectrl variability.The experimental results also verify that NFKAA and RNAA improve KAA in terms of unmixing speed.As for the image data wth large homogeneous area,these two fast KAA methods get close unmixing accuracy as KAA.In particular,their sparse abundances have potential value for hyperspectral image classification.Moreover,with proper sampling number,NFKAA achieves higher unmixing accuracy while RNAA is better than or approaching to NFKAA in terms of unmixing efficiency.(3)In order to use the abundance features for unmixing and incorporate the spatial information as well,the Multilayer KAA(MLKAA)unmixing framework is established.The multilayer unmixing structure realizes to transform the original input features into abundance high-level features based on which the unmixing results get further optimized.Meantime,bilateral filtering function is used as nonlinear activation function operated on abundance features between neighbor layers.This incorporates the effect from spatial neighbor pixels and makes the abundances refined.The experimental results tell that to set the number of layers in MLKAA as two is reasonable.In comparison between TLKAA and KAA,TLKAA obtains abundance maps more approaching to the real distributions of classes.With wellpreserved mixed edges,TLKAA obtains more smooth abundances in homogeneous aera.(4)To deal with the image datasets mixed at different level directly,tunable Sparse Archetypal Analysis(SAA)model is developed based on sparse representation for spectral unmixing.Extra sparse constraints respectively on the combinations of generated endmemebrs and mixed data are added into AA model,it achieves hierarchical sparse solution by coping with the sparse constraints and convexity constraints in sequence.Therefore the spectral variability summarized in the generated endmemebrs can be controlled by tuning the sparse factors in the model.Furthermore,the problem of spectral variability can be dealt with by the way of sparse unmixing with multiple endmembers which are extracted from the combinations of generated endmembers.Unmixing results of the experiments show that SAA can be tuned to adapt to the mixed datasets in a larger range of mixed level.For seriousely mixed data,SAA surpasses KAA and achieves a close unmixing performance as AA.For the mixed data collected by low-altitude remote sensing,SAA obtaines approaching unmixing performance as KAA and performs like two-layer multiple endmemebrs based sparse unmixing.In analysis of hyperspectral images of high spatial resolution collected in laboratory,KAA gets higher accuracty than SAA in idetification task with more complex data.On contrary,SAA obtains much better performance than KAA in idetification task with simple data.