Hyperspectral images are able to provide the detailed spectral information necessary for the discrimination of different land targets in various kinds of remotely sensed images. The linear spectral mixing model is a widely used discrimination method for modeling the variety of multiple land targets in hyperspectral images. Basically, the linear spectral mixing model is solved by least squares adjustment to acquire the minimal-error solutions. It is believed that the reduction of the noise that inherently exists in each spectral band of the hyperspectral image can increase the discrimination accuracy. In general, this noise is usually a consequence of noise recorded by the hyperspectral sensor, which is caused by the absorption or emission of the radiance of atmospheric particles during the energy transportation process. A noise filtering preprocessing process based on empirical mode decomposition (EMD) is proposed. The purpose is to reduce the inherent noise and further minimize the residuals of least squares solutions for hyperspectral images. Given EMD's fully data driven characteristics, the original data can be adaptively decomposed into several components. These components are then filtered by dropping those with noise. The simulation test results indicate that the EMD noise filtering process can effectively decrease the residuals of least square adjustment and mean abundance errors. Moreover, the improved accuracy of the classification demonstrates that the EMD process should be a valuable for noise reduction in hyperspectral image.