Information-theoretic feature selection with segmentation-based folded principal component analysis (PCA) for hyperspectral image classification

• Published in: International journal of remote sensing Vol. 42; no. 1; pp. 286 - 321
• Main Authors: Uddin, Md. Palash, Mamun, Md. Al, Afjal, Masud Ibn, Hossain, Md. Ali
• Format: Journal Article
• Language: English
• Published: London Taylor & Francis 02-01-2021; Taylor & Francis Ltd
• Subjects: Accumulation; Classification; Entropy; Entropy (Information theory); Feature extraction; Feature selection; Finite element method; Hyperspectral imaging; Image classification; Image segmentation; Information theory; Land cover; Principal components analysis; Redundancy; Thematic mapping; Variance analysis; Wavelength
• ISSN: 0143-1161; 1366-5901
• DOI: 10.1080/01431161.2020.1807650

Hyperspectral image (HSI) usually holds information of land cover classes as a set of many contiguous narrow spectral wavelength bands. For its efficient thematic mapping or classification, band (feature) reduction strategies through Feature Extraction (FE) and/or Feature Selection (FS) methods for finding the intrinsic bands' information are typically applied. Principal Component Analysis (PCA) is a frequently employed unsupervised linear FE method whereas cumulative-variance accumulation is used for selecting top features from PCA data. However, PCA can fail to extract intrinsic structure of HSI due to global variance consideration and domination by visible and near infrared bands while cumulative-variance accumulation has no capability to exploit non-linear relationships among the transformed features produced by PCA-based FE methods. Consequently, we propose an information theoretic normalized Mutual Information (nMI)-based minimum Redundancy Maximum Relevance (mRMR) non-linear measure to select the intrinsic features from the transformed space of our previously proposed Segmented-Folded-PCA (Seg-Fol-PCA) and Spectrally Segmented-Folded-PCA (SSeg-Fol-PCA) FE methods. We extensively analyse the effectiveness of the proposed unsupervised FE and supervised FS combinations Seg-Fol-PCA-mRMR and SSeg-Fol-PCA-mRMR with that of PCA-based existing linear and non-linear state-of-the-art methods. In addition, cumulative variance-based top features pick-up strategy is considered with all FE methods and Renyi quadratic entropy-based FS is used with Kernel Entropy Component Analysis (Ker-ECA). The experimental results illustrate that SSeg-Fol-PCA-mRMR and Seg-Fol-PCA-mRMR obtain highest classification result e.g. 95.39% and 95.03% respectively for agricultural Indian Pines HSI, and 96.58% and 95.30% respectively for urban Washington DC Mall HSI while the classification accuracies using all original features of the HSIs are 70.28% and 91.90% respectively. Moreover, the proposed SSeg-Fol-PCA-mRMR and Seg-Fol-PCA-mRMR outperform all investigated combinations of FE and FS using the real HSI datasets.