Due to the high mixed degree of hyperspectral remote sensing images, the basic component extracted by the traditional Nonnegative Matrix Underapproximation(NMU) algorithm is still "impurity", moreover, it is susceptible to noise. To overcome the above shortcomings, a method named L1/2-regularized soft-thresholding NMU for hyperspectral unmixing was proposed. Firstly, the L1/2 regularization term for abundance was introduced to improve the distinguishing ability, which can further improve the purity of the extracted components. Secondly, the soft-threshold penalty function was introduced to replace the residual nonnegative constraint in NMU. By adjusting the penalty factor, the number of non-negative elements could be well controlled, which could improve the anti-noise ability. Experimental results on the simulational and real datasets show that the proposed algorithm can obtain better separation results even under noisy conditions.