When supervised classification of hyperspectral images, the traditional supervised learning method for hyperspectral data classification needs to obtain enough samples marked as training samples, which can effectively avoid Hughes effects. Hyperspectral data under actual conditions with more bands and relatively small training set a challenge to the traditional remote sensing image classification. Therefore, this paper presents an approach based on a weighted combination of features and characteristics of hyperspectral image classification algorithm for texture analysis more difficult reality, the use of a first-order statistical characteristics describe the image histogram texture features within class scatter matrix by inverse matrix method as a feature weighting matrix structure combined kernel function hyperspectral spectral characteristics and spatial characteristics integrate, while taking advantage of features to improve the small weighted training samples for supervised classification accuracy. Experimental results show that the method proposed in this paper for a small sample of hyperspectral data classification with good results.