Band selection-based Gabor wavelet feature extraction for hyperspectral imagery classification

Recently, we have introduced 3-D Gabor wavelets to extract the discriminative features from the hyperspectral imagery for classification. High classification accuracies have been achieved even with small training sample set. However, the computational load of the convolution operator between the original hyperspectral data and the 3-D Gabor wavelet filter is quite high. Furthermore, more than fifty Gabor wavelet filters are convolved with the original data, which needs huge amount of space to store the generated feature sets, making the following feature fusion and classification procedures not practical for hyperspectral imagery covering large spatial area. In this paper, we firstly choose the representative bands from the whole hyperspectral data using affinity propagationbased clustering algorithm, then the Gabor wavelet filters are convolved with the selected bands. Experimental results show that the obtained classification accuracies are not much affected, whereas the computational cost and storage requirement are largely decreased.