Selective HybridNET: Spectral-Spatial Dimensionality Reduction for HSI Classification

Abstract

Hyperspectral images are remote sensing images containing more than a hundred spectral bands of the same ground space with various wavelengths. It has multiple applications but the random nature of latent data such as correlation, variability, and the number of spectral bands turned classification into a challenging task. These natures can be made to be less discriminatory by using a stand-alone preprocessing approach (dimensionality reduction techniques) with a classifier. A model performs poorly when redundant features are present and spatial-spectral concerns are ignored. A 2D Convolutional Neural Network (CNN) model is treated as a good method for hyperspectral image classification whereas accuracy depends on both spectral-spatial properties. Therefore, 3D CNN can be used as an alternative variant but has high computational complexity due to the large size of the volume and spectral dimension. A selective spectral-spatial HybridNET model that embeds dimensionality reduction and deep learning convolutional approaches are provided for both feature selection and extraction in order to solve these sorts of difficulties. In which both 3D and 2D convolutional networks have been adjusted to make a composite network with selective data preprocessors. Thus, this model is able to resolve time complexity issues as well as handle large amounts of data. Experiments have been performed using selective HybridNET on two available datasets such as Indian Pines and Pavia University, to confirm the stability of the proposed selective HybridNET over different state-of-the-art methods.