JM变换优化的高光谱图像自适应降维

Hyperspectral remote sensing images, which collect rich spectral and spatial information of observed targets, usually contain dozens to hundreds of narrow bands with wavelengths ranging from the visible light region to the near-infrared spectra. With such an abundant number of spectral features, hyperspectral images (HSI) allow us to distinguish different of objects or targets by rule and line. Unfortunately, such high-dimensionality data pose a challenge in data transmission, storage, and processing. Specifically, those HSIs with high redundancy information and strong correlation are prone to a Hughes phenomenon during the image classification process. Therefore, dimensionality reduction is necessary for targets classification. Moreover, without using prior label samples, unsupervised dimensionality reduction can effectively simplify the HSI feature space, and prevent the Hughes phenomenon in the targets classification. In this paper, the Jeffries-Matusita (JM) modified adaptive band selection (JM²ABS) method is proposed to extract proper features from HSI datasets. Generally speaking, a band that contains many information and demonstrates strong independence is a very important feature that helps unsupervised band selection methods to classify targets. The JM²ABS method considers both the information capacity and independence of HSI bands. Given the significant differences in the measurements of a band's information capacity and its independence, we introduce the JM transform function to normalize the distributions of the information capacity and the independence of HSI data. Thus JM²ABS shows that both the information capacity and the independence are equally important in unsupervised dimensionality reduction. We also compare our proposed JM²ABS method against three typical methods, namely, the modified adaptive band selection method, the Laplacian score feature selection method, and the infinite feature selection method. By using random training samples, we perform supervised classification experiments on two kinds of HSI public datasets (linear and planar arrays). The results demonstrate that JM²ABS outperforms the other three typical methods in terms of Kappa value, overall classification accuracy, and average classification accuracy. Moreover, under a small number of bands, JM²ABS can reach a high and stable level regardless of the different datasets and different classifiers used. The proposed JM²ABS can select the proper features of HSI datasets for their classification application. The JM transform function is a kind of nonlinear distribution function that can standardize variables from different scales to the same. To demonstrate the feasibility of JM transform optimization, we set the same weight for the information content and the independence. In our future work, we will explore the similarities and differences between the information capacity and the independence in dimensionality reduction.