Two-stage feature extraction algorithm using linear and nonlinear transformation for hyperspectral image classification

Study and analysis of Hyperspectral data is one of the major research areas in remote sensing technology. Hyperspectral imaging collects information in the form of several hundreds of spectral bands with narrow bandwidths and provides full spectral information regarding the data. Hyperspectral image classification has been a vital topic which is growing rapidly in the field of research. Several applications can be built by making proper use of this data, i.e. from the abundance of spatial and spectral information which provides strong binding to build new ideas and new algorithm. Few approaches have concluded that the linear features directly extracted from hyperspectral data for classification provides distinct classification output and few other approaches conclude that the features selected using nonlinear features help to diminish dimensionality of the data, for better modelling of the intrinsic nonlinearity of classification. By combining the spectral features extracted using Root Mean Square (RMS) feature extraction technique and spatial features extracted using Extended Multi-Attribute Profile (EMAP) feature extraction technique we can get better classification output compared to existing algorithms. The proposed algorithm utilizes both linear and nonlinear features along with a Sparse Multinomial Logistic Regression (SMLR) classifier to provide a better classification. The developed algorithm is tested over four widely used hyper spectral data sets Hydice, Aviris Indian Pines, Pavia University, and Salinas. Test results obtained from the proposed work justify that the algorithm provides better and accurate Hyperspectral image classification.