---

Dimension Reduction Using Supervised LDA and Unsupervised PCA and 3D CNN for Classification of Hyperspectral Image

Corresponding Author : T. Jannat (jannat22tasmia@gmail.com)

Authors : T. Jannat (jannat22tasmia@gmail.com)

Keywords : Hyperspectral Image, Feature Extraction, Dimensionality Reduction, PCA, LDA

Abstract :

Hyperspectral Image (HSI) classification is becoming extremely appealing now-a-days because it contains huge information about a scene. The main problem of dealing with hyperspectral image is “Hughes Phenomenon” caused by the large number of dimensions. To address this, dimensionality reduction techniques need to be applied as a preprocessing step. In this paper, we used supervised LDA and unsupervised PCA for dimension reduction. Then for classification we used 3D CNN. Performance of classification depends on both spatial and spectral information for which 3D CNN is a good choice. Still, it’s not popular because of its computational complexity. But dimension reduction makes things easier by removing irrelevant spectral or information. So, we propose an approach in which various feature extraction methods are performed independently to the dataset Indian Pine and then classified using a 3D Convolutional Neural Network. The results are compared with Support Vector Machine and other Convolutional Neural networks such as 2D CNN and 3D CNN. The experimental outcomes show that our method is best among these three with about 99

Published on July 31st, 2021 in Volume 3 Issue 1, Computer Science, Electrical and Electronics