Hyperspectral Image Classification Using Fisher's Linear Discriminant Analysis Feature Reduction with Gabor Filtering and CNN

Abstract

Deep learning-based approaches for hyperspectral image (HSI) feature extraction and classification have gained popularity in recent years. Effective extraction of spectral and spatial information is desired for classifying HSI using a convolutional neural network (CNN) to avoid overfitting. Previous research suggests that Fisher's linear discriminant analysis (LDA) is a better alternative for HSI feature reduction compared to principal component analysis (PCA). In this work, an LDA approach is studied as a dimensionality reducer along with a Gabor filter for extracting spatial features and classification using CNN. The efficacy of the proposed approach is compared with a similar classification scheme with the PCA. Experimental results from two benchmark HSI datasets show the benefits of using LDA with notable improvements in class and overall accuracies.