Obtaining labeled training sets for hyperspectral image data classification is often time consuming and expensive. Therefore, classification of hyperspectral data with insufficient training samples catches the attention of researchers lately. A spectral and texture feature co-training algorithm is proposed based on a semi-supervised classification scheme. The two views of spectral and spatial information of hyperspectral imagery are combined by using Co-training mechanism. Our algorithm is well suited for the hyperspectral image data classification problem, especially when the size of training samples is small. Experimental results on real data demonstrate that the algorithm can yield good results in land-cover classification with hyperspectral image data.