Feature Selection Technique for Hyperspectral Imagery Classification with Noise Reduction Preprocessing

The rich information available in hyperspectral imagery has posed significant opportunities for material classification and identification. The main problem encountered with the classification process is the high dimensionality of hyperspectral data and the low-sized training dataset. Hence, dimensionality reduction is often adopted to avoid the "curse of dimensionality" phenomenon. However, noise generated by various sources (primarily the sensor and the atmosphere) inevitably decrease the precision of the classifier. In this paper, two wavelet-based methods, wavelet shrinkage and discrete wavelet transform, are applied to preprocess the hyperspectral imagery in sequence, denoising the spatial images and spectral signatures, respectively. After that, affinity propagation, which is a recently proposed feature selection approach, is used to choose representative features from the noise-reduced data. Experimental results demonstrate that the features acquired by the new scheme make the classification results more accurate than those without noise reduction preprocessing.