High Resolution Remote Sensing Image Analysis

With the rapid development of the image spectrometer technology,more and more high spectral and spatial remote sensing images have been obtained.Faced with the hyperspectral images,since their spatial resolution is quite low,mixed pixels are widely existed.In order to make full use of these data,hyperspectral unmixing,which aims to decompose the mix pixel into a collection of constituent materials named endmembers and their corresponding fractional abundance,is very important,and it's a preprocessing procedure of hyperspectral image object recognition and abnormal detection.As to the high spatial resolution remote sensing(HSR)images,since one of them covers a much large ground surface and the mapsheet is huge,it's difficult to utilize these data directly and scene classification becomes very significant for the sake of subsequent researches like change detection and image retrieval.Above all,unmixing and scene classification are of great importance in remote sensing image analysis,which make the applications of remote sensing data in both military and civil areas more convenient.Based on this,several approaches are proposed in this thesis for hyperspectral unmixing and high spatial resolution image scene classificaiton,and it involves:For hyperspectral images with relatively low spatial resolution,the distribution of ground substances is very complex.One type of substance distributes not only in a collection but also around the whole image.In order to conquer this problem,a refined method named substance dependence constrained sparse nonnegative matrix factorization(SDSNMF)is proposed in this thesis.The proposed SDSNMF method makes fulluse of not only the spectral information but also the substance dependence of hyper-spectral data,which improves the accuracy and robustness of hyperspectral unmixing effectively.Most existing HSR image scene classification methods are based on the handcrafted features like SIFT and GIST.These handcrafted features are not adaptive and usually involve a lot of engineering work to develop,which makes them not applicable for remote sensing images.Hence,an unsupervised feature learning method is proposed in this thesis,which learns features from the original raw pixels and overcomes the limitation of the hand-crafted features.In HSR images,the interclass similarities are widely existed and there are several the matic classes in one scene.In order to alleviate these difficulties,a weighted deconvolution model based system is proposed in this thesis.The proposed method firstly utilizes the weighted decomvolution model to learn a set of feature maps which are be aggregated by the spatial pyramid model.The aggregated feature vectors are treated as the final representation of the HSR image scenes and input into the classifier.The proposed method is verified in two challenging data sets and improves the accuracy greatly.