An adaptive semantic dimensionality reduction approach for hyperspectral imagery classification
Résumé
Hyperspectral imagery (HSI) is widely used for several fields of remote sensing such as agriculture, land cover monitoring, and deforestation. However, the HSI classification is a challenge task due to the large number of spectral bands, unavailability of training samples, and the high correlation inter-bands. To address these challenges, we propose in this work a semantic reduction dimensionality approach based on the principal component analysis (PCA) and mutual information-based band selection (MI). Firstly, we project the original HSI using PCA to obtain a novel subspace with lower dimensions. Using the obtained components, a set of rules can be generated to find the relevant spectral bands based on score contribution coefficient. Moreover, the mutual information (MI) is used to select the spectral bands that contain a higher information based on the entropy criterion. We propose then to exploit the selected bands for HSI classification using SVM technique. Experiment results demonstrate that our proposed approach is effective and perform for HSI classification compared to other dimensionality reduction approaches.