Tensor FCM And Its Application On Hyperspectral Image Classification

With high spectral resolution and the integration of image with spectrum, hyperspectral images has been effectively utilized to the object recognition and classification. However, traditional image classification methods are subject to some limitations due to the high resolution, high dimensionality, and extremely large volumes of hyperspectral data. As such, the research on tensor fuzzy c-means algorithm and its application in hyperspectral image classification has been carried out in this thesis, whose main contents are as follows:(1) According to structure feature, hyperspectral images are mapped into tensor space, in which we discuss their tensor processing algorithms. Firstly, tensor definition and its related multilinear algebra theory are introduced, and then the advantage of tensor computation is presented. Finally, we give the tensor description of spatial-spectral features for hyperspectral images.(2) In the original high-dimensional space, hyperspectral image classification has the problem of the curve of dimensionality, which not only increases the computational complexity and storage space, but also decreases the performance of classification. To overcome the above issue, the multilinear principal component analysis (MPCA) is utilized to perform dimensionality reduction, that MPCA dimensionality reduction can improve the classification speed and accuracy of hyperspectral imagery. Traditional fuzzy c-means clustering (FCM) algorithm has been applied to the hyperspectral image classification, for which it just exploits the spectrum information of test samples, but neglects the spatial information existed in samples, leading to poor classification performance.(3) To improve the classification accuracy, tensor FCM (TFCM) algorithm is proposed to simultaneously take into the spatial-spectral information account. Experiments show that the overall classification accuracy can be substantially increased. Nevertheless, the objects with similar spectrum cannot be well distinguished. To further improve the classification effect, we also proposed the weighted tensor FCM (WTFCM) algorithm. An unsupervised weight mean is assigned to each sample, which represents the relation between the current sample with other samples. The WTFCM is obviously superior to TFCM, whose performance, however, is influenced by the degree of fuzziness, and is unstable. To address this issue, a new weighted tensor FCM (NWTFCM) is further presented. Unlike that in WTFCM, this weight is defined as the relation between the current sample with the centroid of cluster. Experimental results verify the validity of the proposed NWTFCM algorithm not only in classification performance but also in stability.