An Effective Approach for Hyperspectral Image Classification Based on 3D CNN with Mish Activation Function

Abstract

Hyperspectral image (HSI) has been grown in popularity for its’ capacity to capture details about a scene. Convolutional Neural Networks (CNN) perform well in HSI classification, as they do in other areas of image processing. However, the majority of CNN techniques are two dimensional. In HSI, both spectral and spatial data are required to analyse. Considering that 3D CNN uses both kind of information, it is one of the best choice in this case. Although hyperspectral images have the advantage of conveying a lot of information, dealing with them might cause dimensionality-related issues like the "Curse of Dimensionality" due to the huge number of dimensions or spectrum. This is where feature extraction can help. In this paper, both supervised such as Linear Discriminant Analysis (LDA) and unsupervised Principal Component Analysis (PCA) have been used to discard irrelevant information. The 3D CNN along with activation function mish is then used for classification. Due to its computational complexity, 3D CNN is not widely used on its own, but things become simpler when dimension reduction techniques are used prior to CNN. In this paper, a method is proposed that has applied dimension reduction techniques PCA and LDA on the Indian Pine and Pavia University dataset separately as part of preprocessing. Then the transformed dataset was classified using a custom-made 3D CNN based model with an activation function called "mish". The experimental results suggest that, with an approximately 99% accuracy rate for both dataset, our approach is one of the best.