Spectral-Spatial Classification of Hyperspectral Imagery with Convolutional Neural Network

CNNs have been stated to use quite often in the classification of hyperspectral images. In 3D CNN-based methods, the hyperspectral data is convoluted with 3D kernels, while spectral and spatial information is preserved. In contrast, spectral information is lost, as only spatial data will be preserved in the hyperspectral data that is convoled with 2D kernels. Combination of 3D and 2D CNN make the hyperspectral image easier to analyze so the computation time is reduced. While both spatial and spectral information are used together with the 3D CNN structure, spatial information is reinforced with the 2D CNN structure. In this study, a method for the effective use of 3D and 2D CNN structure is presented. The effect of the proposed method on classification performance was tested using Aviris Indian Pines, Rosis Pavia University and Salinas Scene datasets and the results are compared with other deep learning-based methods. It is seen that the proposed approach provides good results in terms of classification performance.