A 3D-Spatial-Spectral Feature Network for Hyperspectral Remote Sensing Image Classification

Abstract

Hyperspectral images (HSIs) are commonly applied in environmental monitoring, urban mapping, crop study, and mineral identification.These applications recurrently call for the distinguishing of the class of each pixel.Although several convolutional neural network (CNN) models have been proposed by recent researchers, none of them have been established as promising in terms of classification accuracy because of the wealth of information involved in these sorts of images for the classification of hyperspectral remote sensing images.Also, the high-dimensionality of the information, the problem of inseparability, and the limited availability of training samples are still an open challenge.This research proposes a novel convolutional neural network 3D spatial-spectral network (Model3DSN) model for the classification of hyperspectral remote sensing data, i.e., Indian Pines, Salinas Scene, and PaviaU.First, we deployed the principal component analysis (PCA) technique for low-dimensionality reduction of pixels and then 2-D and 3-D convolutions for discriminative spectral-spatial feature learning.We compared Model3DSN's efficiency against the existing spatial-spectral state-of-the-art (SOTA) models.The high accuracy achieved with the Model3DSN demonstrates its efficiency as a SOTA method for HSI remote sensing image classification, thus providing an in-depth interpretation of HSI images.