HyperspectraI Image Classification With Diverse Region Multiscale Feature Extraction and Spectral Imaging

Abstract

We propose a dual-branch multiregion Gaussian pyramid-based multiscale feature extraction and spectral feature extraction network. We use the principal component analysis algorithm to reduce the dimensionality of the data before feature extraction, but the degree of data downscaling is different for the two-branch network. For the spatial information of hyperspectral images (HSIs), we use a multiregion piecewise Gaussian pyramid downsampling method to generate multiscale and multiresolution image data and use an improved resnet network to extract spatial information, so that the network can extract specific contextual features of HSIs. For the spectral information of HSIs, we use the method of imaging spectral information, first reducing the dimension of the spectral data and then expanding the spectral data into an image of N×N. By expanding into an image, the resnet network is introduced to extract spectral information, but the number of network layers is different from that of the resnet network for obtaining spatial information, which can solve the problem of low correct classification rate due to changes in similar spectral data. Finally, the spatial and spectral features after the dual-branch feature extraction network are combined into a fully connected network for classification, and the fusion of the two features can improve the classification accuracy. Our experiments on three commonly used datasets show that the method can improve the accuracy of the classifier.

Keywords

• Deep learning
• Gaussian pyramid
• hyperspectral image (HSI) classification
• multiscale feature extraction