3D-Gabor-Based Feature Selection Via Enhanced Fast Density-Peak-Based Clustering

Hyperspectral image(HSI) has hundreds of bands witch provide hyperspectral data to distinguish different ground objects. However, owing to the low number of samples and strong relativity between different bands, its always tough to get a precise classification result. In this paper, 3D-Gabor-based feature selection algorithm name 3DG-EFDPC is proposed to address this problem. 3D Gabor filter can preserve intrinsic information of hyperspectral image and extract distinguished features with joint spatial/spectral domains for increasing accuracy of classification, via capturing different orientation, scale, and wavelength properties of hyperspectral image data. However, a feature selection process must be taken due to redundance might exist in 3D Gabor feature. For this purpose, we introduce enhanced fast density-peak-based clustering(EFDPC) algorithm. Through calculating the normalized local density and the intracluster distance of each feature, EFDPC regard features which both values are anomalously large as cluster center which contain discriminative and valuable information. Furthermore, in order to confirm the number of chosen features decided by the cutoff threshold, the EFDPC algorithm propose an exponential-based learning rule. We demonstrate the performance of the 3DG-EFDPC features for classification on real hyperspectral remote sensing data sets.