Hyperspectral Image Analysis Using a Custom Spectral Convolutional Neural Network

Abstract

In recent time, the most applied classification method for hyperspectral images is based on the supervised deep learning approach. The hyperspectral images require special handling while it consists of hundreds of bands / channels. In this article, the experiments are conducted using one of the widespread deep learning models, Convolutional Neural Networks (CNNs), specifically, Csutom Spectral CNN architecture (CSCNN). The introduced network is based on the data reduction and data normalization. It firstly ommits the unnecessary data channels and retains the meaningful ones. Then, it passes the remaining data through the CNN layers (convolutional, rectified linear unit, fully connected, dropout,…etc) until reaches the classification layer. The experiments are applied on four benchmarcks [hyperspectral datasets], namely, Salinas-A, Kenndy Space Center (KSC), Indian Pines (IP), and Pavia University (Pavia-U). The proposed model achieved an overall accuracy more than 99.50 %. In last, a comparison versus the state of the art is also introduced.