Extraction of Information from Hyperspectral Imaging Using Deep Learning

Abstract

Keeping pace with the rapid change of time, very powerful deep learning techniques have become available. There are many optical analysis techniques that are used to identify different objects from large-scale images. Hyperspectral is one of such techniques, which is used mainly when the images from the satellite are captured and are used to identify different objects. Hyperspectral images are made up of a large number of bands naturally. Thus, extracting information from the satellite comes with many problems and challenges. But with the use of powerful deep learning (DL) methods, the earth’s surface can be precisely explored and analysed. The combination of spatial and spectral information helps to track and find out remotely sensed scrutinized data everywhere. However, the presence of high-dimensional features along with various bands hinders the accuracy in hyperspectral imaging (HSI) analysis. Thus, a large number of bands and dimensions in hyperspectral images must be reduced using some linear or nonlinear dimensionality reduction (DR) techniques. By eliminating all these redundant information, data can be classified and interpreted in a far better way. Mainly three hyperspectral datasets have been extensively used for this purpose, namely Pavia University, Indian Pines, and Salinas Valley. In this paper, the main focus is on the use of principal component analysis (PCA), a technique for DR, followed by that the use of a deep neural network, such as 3D convolutional neural network (CNN) for classification of the hyperspectral datasets. Subsequently, all these classified datasets are then compared with the respective ground truth images, analysing the performance of the DL model based on various metrics.