An effective feature extraction method via spectral-spatial filter discrimination analysis for hyperspectral image

Multi band, high spatial resolution and information redundancy are the most significant characteristics of hyperspectral image. These remarkable characteristics are mainly caused by the high-dimensional image data. In view of these characteristics, feature extraction of hyperspectral image has naturally become one of the research hotspots in this field. However, it is impossible to fully describe the intrinsic geometric structure of hyperspectral image only by using spectral information. To improve the subsequent forecasting accuracy, spatial information should be mined to further describe the geometric structure of hyperspectral image. Therefore, an effective feature extraction method (SSF_HM) was proposed via using harmonic mean and spectral-spatial filter. This investigation divides the SSF_HM into three steps. First, the p i principal components were extracted by using principal components analysis (PCA) skill and subsequent the p i spatial filtering features were obtained via using area median filter (AMF) method. Then, the original spectral features and extracted spatial filtering features combine to form the fusion feature matrix, and then the scatter matrix S_b^HM (based on harmonic mean (HM) spectral-spatial filter inter-class) and scatter matrix S_w^HM (based on harmonic mean (HM) spectral-spatial filter intra-class) can be established in the fusion feature space, respectively. Finally, combining the Fisher discriminant analysis model and regularization technique, a new feature extraction method SSF_HM is developed. The proposed SSF_HM method combines spectral information and spatial information. At the same time, the range of feature extraction is expanded from spectral space to spectral-spatial fusion space. The experimental results on three real-world hyperspectral image data sets show the better performance of SSF_HM in comparison with other feature extraction methods in small sample size situation by using maximum likelihood classifier (MLC).