A novel mixed noise estimation method based on the spectral and spatial information of hyperspectral images was reported. Firstly, homogeneous image blocks were automatically detected using data masking. Then signal value and noise value of each pixel in homogeneous blocks were split with a multiple liner regression model. Meanwhile, rough approximations of SD and SI noise were obtained. Finally, likelihood function was built based on the mixed noise model, where parameters of the noise model were calculated by maximum-likelihood estimation approach. The proposed method is demonstrated to be accurate and robust by experiments with both synthetic images and real hyperspectral images.