Dimensionality Reduction Of Hyperspectral Image Data With Mixed Pixels

Hyperspectral remote sensing data contains a lot of spatial and spectral information. The spectra reflects the properties of different materials, and the image reveals the geometric characteristics of the material, so it is beneficial to the classification and identification of the landcovers in the Earth's surface. Because hypespectral data has so many bands, it results in difficulty in the selection of modal and training. Thus it is more and more necessary for extract features in hyperspectral data processing.Due to the limited spatial resolution, the hyperspectral image contains “mixed pixel”, such that the spectral vectors of the same class may be different and the spectral vectors of the different class may be the same. However the previous methods often ignore the phenomenon of mixed pixels, resulting in lower discriminative samples and misclassification problems. So the paper designs several planes, based on discriminative subspace projection. The study mainly includes the following aspects:Firstly, the paper proposes fuzzy signature based discriminative subspace projection for hyperspectral data classification. First, define fuzzy signature, and make the assumption that similar samples owns similar fuzzy signature takes the place of the assumption that similar samples owns the same labels,constructing Laplace regularizer; second, the labeled samples generate discriminate term, maximizing the margin between neighbors in different classes and neighbors in same class after the projection. The two terms are combined to obtain the fuzzy signature and project matrix. Without the classifier, the paper realizes dimensionality reduction and classification of HIS simultaneously. With rare labeled samples in the Indian Pines, Salinas-A, Pavia University,Botswana and KSC hyperspectral datasets, experiment results show that the algorithm can alleviate the problem of mixed pixels and realize better classification.Secondly, the paper proposes mixed signature based semi-supervised dimension reduction of hyperspectral image. First, segment the image by turbopixels algorithm, fully obtaining the spatial information. Then separate the mixed pixels. Second, based on adjusted superpixels and mixed pixels, generate superpixel based fuzzy signature and pixel based fuzzy signature. Third, combining with maximizing margin criterion, obtain projection matrix and classification result. With rare labeled samples in the Indian Pines, Salinas-A and Pavia University hyperspectral datasets, experiment results show that the algorithm maintains the spatial data consistency and improves the classification rate.Thirdly, the paper proposes self-spaced learning based semi-supervised dimension reduction of hyperspectral image. Due to mixed pixels bad for dimension reduction and classification for hyperspectral data and double convex optimal problem often gets stuck in local optimization, introduce self-paced learning, add more discriminate samples into training and optimize parameters at each iteration rather than training all samples, such that alleviate the problem of mixed pixel. With rare labeled samples in the Indian Pines, Salinas-A and Pavia University hyperspectral datasets, experiment results show that the algorithm can alleviate the problem of mixed pixel and avoid getting stuck in local optimization.