Data Mining based on Spectral and Spatial Features for Hyperspectral Classification

Hyperspectral remote sensing technique provides fine and detailed spectral information by contiguous and narrow spectral channels. For the traditional classification algorithms, most of them are based on spectral information; spatial information which is useful for the hyperspectral data analysis is paid a little attention to. So, hyperspectral image classification based on effective combination of spectral and spatial information needs further investigation. In this paper a new classification method for hyperspectral remote sensing is proposed in order to mine spatial information which hidden in the image. Firstly, some spectral features which are statistics indexes (MinBand, MaxBand, AvgBand, StdBand, etc.) are extracted from OMIS image data. Secondly, spatial structure information and spatial information such as Area, Length, Compact, Convexity, Solidity, Roundness, and so on are extracted. Then the spectral and spatial attributes are computed and used for the following classification. Lastly, several kernel functions which using joint spectral and spatial information such as Linear, Polynomial, Radial Basis Function (RBF) and Sigmoid kernel are adopted for SVM classification model. The experiments proved that the classification algorithm which joint spectral and spatial information can work more effectively compared to the traditional classification methods, and this new approach is useful for hyperspectral classification and imaging analysis, and has the potential ability in hyperspectral remote sensed data processing.