Local One Dimensional Embedding for Hyperspectral Image Classification

Abstract

Hyperspectral image (HSI) classification aims to allocate a unique label to each pixel in the HSI dataset. However, limited number of labeled pixels and high dimensionality of the HSI dataset makes HSI classification a challenging task. To address these issues, we propose local one dimensional embedding (L1DE), which makes full use of the characteristics of HSI dataset. First, pixels have similar spectral signatures should be the same class. Second, pixels in the same spatial area are more likely to be the same class. Advantages of L1DE can be summarized as follows. The L1DE maps the high dimensional feature vector into 1D space. And it ensures that if the coordinates of two pixels are adjacent in the 1D space, they are more likely to be close to each other in the original spatial domain, and have similar spectral signatures. Additionally, the local strategy used in L1DE can reduce the computation burden significantly. By use of L1DE, we propose two frameworks for HSI classification: multiple L1DE interpolation (ML1DEI) and multiscale spatial information fusion (MSIF). Analysis of the computational cost for these frameworks are given in this report. Experimental results on four widely used HSI datasets indicate that, the proposed frameworks show outstanding performance when compared with other state-of-the-art methods, even when very limited labeled pixels are available.

Biography

Prof. Hong Li is a Professor in School of Mathematics and Statistics, Huazhong University of Science and Technology. Her research interests include Approximation and Calculation, Wavelet Analysis and its Application, Machine Leaning and Artificial Intelligence, Pattern recognition and image processing, Time-Frequency Analysis and Signal Processing.