Spectral and Spatial Feature Extraction Using Folded PCA and Convolutional Neural Network for Effective Hyperspectral Image Classification

• Published in: 2021 5th International Conference on Electrical Information and Communication Technology (EICT) pp. 1 - 6
• Main Authors: Roy, Monoshi Kumar, Mishu, Sadia Zaman, Mamun, Md. Al
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17.12.2021
• Subjects: Computational efficiency; Convolutional Neural Network; Convolutional neural networks; Feature extraction; Folded Principal Component Analysis; Hyperspectral Image Classification; Memory management; Neural networks; Principal Component Analysis; Support Vector Machine; Support vector machine classification; Task analysis
• DOI: 10.1109/EICT54103.2021.9733696

Hyperspectral Image (HSI) consists of a hundred to thousand channels. These channels (also can be referred to as features) should be extracted so that the minimum number of channels can express the original information. Those countless numbers of bands are cumbersome for convolutional models, and their computational costs are on the higher side. Principal component analysis (PCA) is one of the most used feature extraction techniques. However, principal component analysis does not consider the local variance of all the bands. That is why a novel approach has been taken to exploit both local and the global variance of the bands. This approach is named as folded principal component analysis (FPCA). In FPCA, each band is folded into a matrix. From this matrix, both the global and local structures of the bands are properly explored. In addition, FPCA is computationally and in terms of memory usage is more efficient than PCA. After feature extraction, Non-Linear radial basis function (RBF) kernel support vector machine (SVM) and convolutional neural network (CNN) architecture have been used on the Indian Pines dataset for the classification task. Spectral features extracted by folded PCA technique show better classification results than features extracted by conventional PCA. In CNN, both spectral and spatial features are being fused, which has been resulted in better classification accuracy than state-of-the-art method.