Learning high-level spectral-spatial features for hyperspectral image classification with insufficient labeled samples

Abstract

Hyperspectral image (HSI) classification research is a hot area, with a mass of new methods being developed to improve performance for specific applications that use spatial and spectral image material. However, the main obstacle for scientists is determining how to identify HSIs effectively. These obstacles include an increased presence of redundant spectral information, high dimensionality in observed data, and limited spatial features in a classification model. To this end, we, therefore, proposed a novel approach for learning high-level spectral-spatial features for HSI classification with insufficient labeled samples. First, we implemented the principal component analysis (PCA) technique to reduce the high dimensionalities experienced. Second, a fusion of 2D and 3D convolutions and DenseNet, a transfer learning network for feature learning of both spatial-spectral pixels. The achieved experimental results are comparatively satisfactory to contrasted approaches on the widely used HSI images, i.e., the University of Pavia and Indian Pines, with an overall classification accuracy of 97.80% and 97.60%, respectively.