A Tensor Based Image Noise Reduction, Feature Extraction And Classification Framework For Remotely Sensed Imagery

The development in the sensor technology makes it possible to provide a large amount of spatial and spectral information for accurate ground analysis applications. The rise of very-high-resolution (VHR) and hperspectral (HSI) remote sensing image is confronted with opportunities as well as challenges. Firstly, due to the limits of current sensor techniques, the acquired images are easily disturbed by various kinds of noise, which not only degrades the visual quality of the image but also limits the precision of the subsequent image interpretation and analysis. Secondly, due to sensor design considerations, the wealth of spatial information in hyperspectral data is often not complemented by extremely spectral resolution. Consequently, objects from the same class become spectrally heterogeneous while objects from different classes become spectrally similar. That means spectral information alone is not adequate for accurate feature presentation; Finally, the improvement in resolution will lead to larger data volume and higher data dimension, which raise a higher requirement for previous image processing algorithms.In order to circumvent the above problems, this thesis presents a new data interpretation framework for remote sensing imagery, including tensor based noise reduction; tensorial feature extraction; support tensor machine classification and tensorial feature based change detection. By combining the state-of-the-art image processing algorithms with the multilinear technique, the proposed framework can fully take the advantage of rich information contained in the image data, thus, improve the interpretation performance.The second chapter briefly defines tensor and introduces multilinear algebra. Then follows the formulation of global and local tensorial description for remotely sensed imagery. The main contributions of this thesis are as follows:(1) Aiming at solving the noise contamination, two novel noise reduction approaches is established based on Tucker3 tensor decomposition and PARAFAC tensor decomposition, respectively. The multi-bands image is considered as a three-order tensor that is able to jointly treat both the spatial and spectral modes. The spatial and spectral information can be simultaneously taken into account during the denosing procedure, hence, a better denoised result can be obtained while the spectral consistency can be fully preserved. Compared with the conventional denoising methods and the newly developed SSAHTV, the proposed tensor based denoising algorithm achieves a significant improvement in terms of both visual interpretation and quantitative measurements.(2) Previous research for feature extraction is mainly based on extracting spatial relationships from each band individually, which will lead to a loss in spatial-spectral contextual information. In order to address this issue, a 3D wavelet discrete transform(3D DWT) based texture feature extraction algorithm and a volumetric texture extraction(VTE) is investigated. In this context, the local imagery patch is considered as a cube, and, hence, is capable of representing the imagery information in both spectral and spatial domains. The notable characteristic of the proposed methods is the ability to decompose an image into a set of spectral-spatial components, and simultaneously exploit the spectral and spatial information. To meet the balance in accurate local description and computational load, we propose three localized approaches:pixel-wise, non-overlapping, and overlapping cube, respectively,. Experiments reveal that the proposed textures achieve much better results than the widely used spectral-spatial classification methods.(3) From the perspective of image classification, the vector-based feature alignment of the conventional classification models will lead to information loss for multibands image representation, which is intrinsically related with tensor-based data structures. In this thesis, a new multiclass support tensor machine (STM) is specifically developed for remotely sensed image classification, which processes the image as a data cube, and then identifies the information classes in tensor space. For the purpose of alleviating potential Hughes phenomenont, a multilinear principal component analysis (MPCA) is used as preprocessing of our tensor-based processing chain, in order to reduce the tensorial data redundancy while, at the same time, preserving the tensorial structure information in the high-order subspace. Experiments with four hyperspectral data sets, covering agricultural and urban areas, are conducted to validate the effectiveness of the proposed classification framework. Our experimental results show that the proposed STM and MPCA-STM can achieve better results than the standard SVM classifier.(4) From the perspective of change detection, a weighted spatial-spectral change tensor analysis(SSCTA) is investigated, which employs the 3D DWT to simultaneously exploit the spectral and spatial change information. Afterwards, the change information from spectrum and contextual is fused through a weighted strategy. To deal with the tensor data and keep the data structure in high-order feature space, support vector machine has been extended to support tensor machine by the multilinear algebra. Experiments conducted on ZY3 multitemporal images reveal that the SSCTA yields much better results than the standard CVA and 3D DWT gives a better representation for the rich spatial-spectral information from remotely sensed imagery.