Spatial-spectral analysis method using texture features combined with PCA for information extraction in hyperspectral images

Abstract

This work proposes a new method to treat spatial and spectral information interactively. The method extracts spatial features, ie, variogram, gray-level co-occurrence matrix (GLCM), histograms of oriented gradients (HOG), and local binary pattern (LBP) features, from each wavelength image of hypercube and principal component analysis (PCA) is applied on this spatial feature matrix to identify wavelength-dependent variation in spatial patterns. Resultant image is obtained by projecting the score values to the original data. Three datasets, including a synthetic hyperspectral image (Dataset 1), a set of real hyperspectral images of salmon fillets (Dataset 2), and remote-sensing images (Dataset 3), were utilized to evaluate the performance of the proposed method. Results from Dataset 1 showed that the spatial-spectral methods had the potential of reducing baseline offset noise. Dataset 2 revealed that spatial-spectral methods can alleviate noisy pixels with strong signal and reduce shadow effects. In addition, substantial improvements were obtained in case of classification between white stripe and red muscle pixels by using the HOG-based approach with correct classification rate (CCR) of 0.97 compared with the models directly built from raw and standard normal variate (SNV) preprocessed spectra (CCR = 0.94). Samson image of Dataset 3 suggested the flexibility and effectiveness of the proposed method by improving CCR of 0.96 using conventional PCA on SNV pretreated spectra to 0.98 using GLCM-based approach on SNV preprocessed spectra. Overall, experimental results demonstrated that the spatial-spectral methods can improve the results found by using the spectral information alone because of the spatial information provided.