Supervised Linear Discriminant Analysis for Dimension Reduction and Hyperspectral Image Classification Method Based on 2D-3D CNN

Hyperspectral Image (HSI) is widely popular nowadays for its ability to contain huge information about a scene. Unlike RGB (Red-Green-Blue) image, HSI is an image with hundreds and more spectral bands, due to which it's possible to identify the smallest differences between objects. But the huge number of dimensions or spectrum leads to problem like "Hughes Phenomenon" or "Curse of Dimensionality". It causes if the amount of data doesn't increase with the increase of dimension. In this paper we used supervised Linear Discriminant Analysis (LDA), a popular dimensionality reduction approach as a preprocessing step to remove irrelevant spectral information. Then for classification, we used a combination of 2D-3D CNN with mish activation function. With the rise of convolutional neural network, an immense improvement in image processing technology specially in the hyperspectral image processing sector has been brought. Most of the methods include 2D CNN which makes inter-band correlations of HSI underutilized. But HSI needs both spatial and spectral information to be monitored. For spectral information analysis, 3D CNN is the best option but it's costly. So we propose a method that has applied feature extraction method LDA on the dataset Indian Pine and University of Pavia and after that, classified the data using 2D-3D Convolutional Neural Network with mish activation function. The results are compared with other available alternatives. The experimental results indicate our model is best among these with the average accuracy of approximately 94% for Indian Pine and 98% for Pavia University dataset using 5% training data.