Research And Application Of Spectral Unmixing Method For Hyperspectral Remote Sensing Imagry

Recently,as the development of remote sensing technique,the space resolution and spectral resolution of hyperspetral images are both improved and the processing methods of them are also better than before.Hyperspectral images not only include spectral information,but also the space information of the observed area.This is the characteristic of hyperspectral images,which is called “graph and spectra in one image”.The advantages of hyperspectral images attract the research interest of many literatures.During the process of obtaining hyperspectral images,remote sensing sensor records the reflection,dispersion,and interactional reflection of the observed materials.The space resolution of the images will decrease because remote sensing sensor platform usually records information from a far distance.Moreover,the nature materials are complicated,which will mix the spectra.All the reason mentioned above makes mixed pixels in hyperspectral images is very common.A mixed pixel is a kind of pixel that includes more than one spectral signature.Correspondingly,a pure pixel is a kind of pixel that includes only one spectral signature.Thus,existence of mixed pixels makes obtaining pure spectral signature of materials not available,which restricts the accurate analysis and application of hyperspectral images.And also,influences the development of remote sensing technique.Hyperspectral unmixing technique is used to resolve the mixed pixels.It unmixes the original images data into a composition of endmembers and abundance,which makes accurate spectral application available.Thus,hyperspectal unmixing technique is an important condition for quantity research and application of hyperspectral remote sensing.The primary works and achievements of this dissertation are as follows:1.Based on linear mixing model,an endmember extraction method is proposed for the assumption of pure pixel exists in the image.The method applies QR factorization using Givens Rotations to original hyperspectral data,and obtains the orthogonal projection of the data.The method is called Endmember Extraction based on QR factorization using Givens Rotations.And then,the Fully Constrained Least Square method is applied to the extracted endmembers to estimate the abundance of them.Synthetic hyperspectral data and real hyperspectral imager are both applied to the proposed method.Experiments result shows that the accuracy of extracted endmember is better than the art-of-the-state endmember extracted algorithms that realized with projection criterion.Moreover,the proposed algorithm is very conveniently to realize with high performance computation tools,which makes real-time computation of the algorithm is available.That is also the research point in the future.2.A Sparse Constrained Graph Regularized Nonnegative Matrix Factorization Algorithm is introduced for no pure pixels exist in the images.Both synthetic data and real hyperspectral image are used to analyze the algorithm.Two metrics,called spectral angle distance(SAD)and abundance angle distance(AAD)are used to compare the algorithm with the art-of-the-state methods that are realized with nonnegative matrix factorization.The experiments demonstrate that unmixing result of the proposed algorithm is good.3.An unmixing method is applied to remove haze for hazed multi/hyper-spectral imagery.First,extract the haze endmember and corresponding abundance.Secondly,Remove the energy of the haze endmember and adjust abundance of rest endmembers,and then estimate the haze removal image with a model.The effect of the haze removal is good.Comparing with haze removal method with filter,the unmxing haze removal method reserves the background information.And,the result is more close to real image.