Dimensionality reduction and spectral-spatial features for hyperspectral image classification

Abstract

The aim of this work is to classify three hyperspectral datasets using the Principal Component Analysis as a tool for dimensionality reduction and then combine the spatial and spectral features in order to enhance the resultant accuracies. The proposed method is generated by combining all spatial features obtained from these techniques with the principal spectral bands. Then, a support vector machine with optimal hyperparameters is utilized to evaluate the classification performance. Ex- periments are conducted on three remote sensing hyperspectral datasets: Indian Pines (rural), Jasper Ridge (rural) and Pavia University (urban). The results of the spatial techniques are reasonably high, especially of the GLCM-based approach. However, our proposed method achieves a higher performance. Our findings suggest that exploiting the spatial correlation between the pixels using di?erent techniques is more efficient. The highest classification gain reached 30% in comparison with the PCA-based classification. Overall, our approach is effective enough to generate rich spectral-spatial information than several state-of-the-art methods.