Spectral-spatial classification of hyperspectral images based on multifractal features

Abstract

Classification of hyperspectral images is an important step of hyperspectral image interpretation. Different studies demonstrate that spatial features can provide complementary information for increasing the accuracy of hyperspectral image classification. In this study, we propose a method of spectral-spatial classification of hyperspectral images that is based on the use of specific multifractal features as the spatial features. The proposed method of hyperspectral image classification consists of the following steps. First, informative multifractal features are extracted from first few principal components of spectral features. For construction of the multifractal features, in the windows centered on each element of principal component images, using a generalized local-global multifractal image analysis, various 1D and 2D multiracial characteristics can be calculated including our early introduced 2D multifractal characteristics of global scaling exponents. After that, obtained multifractal features are stacked with spectral features into high-dimensional feature vectors. Finally, the resulting high-dimensional vectors of spectral and multifractal features are classified by a support vector machine classifier. The multifractal characteristics that are used to construct multifractal features have a lot of advantages: these characteristics provide a good textural separability of image objects, demonstrate an invariance to image scaling and rotation, and they are also insensitive to image noise. The experiments performed on several widely known test hyperspectral images have demonstrated that proposed method exhibits better performance than competitive methods of spectral-spatial classification of hyperspectral images, in terms of the overall accuracy and kappa statistic. In addition, it is shown that the introduced classification method can outperform some deep learning methods of hyperspectral image classification, which in recent years have attracted great interest in hyperspectral image classification. In particular, it was established that the proposed method can achieve good classification results over deep learning methods if we use small training samples for classification. In the future, we will focus on developing methods for object-oriented classification of hyperspectral images, which are based on the use of multifractal features. The study has been supported by the Ministry of Education and Science of the Russian Federation (Project No. МК-3477.2019.5) and by the Russian Foundation for Basic Research (Project No. 19-05-00330 А).