A novel target spectrum learning method for small target detection in hyperspectral imagery is proposed to obtain a more accurate target spectrum for better supervised target detection. This method is composed of two components: adaptive weighted learning method and self-completed background dictionary. Given a complete background dictionary, the former component refines the target spectrum through sparse coding and gradient descent algorithm. The latter component guarantees the background dictionary completeness by gradually size enlarging. Both experimental results on simulated and real hyperspectral data show that the proposed method has an advantage in extracting the accurate target spectrum, which enables better detection results.