For resolving the problem that there are lots of mixed pixels in the Synthetic Aperture Radar （SAR） images

against the flaw that the traditional Independent Component Analysis（ICA） can not solve the decomposition of mixed pixels effectively

we propose a new algorithm: Supervised Learning ICA algorithm（SL-ICA）. Adding supervised learning restrictive conditions to the negentropy objective function

we implement negentropy and restrictive conditions in a unified objective function

which minimizes the error while maximizing the negentropy. At the same time

we optimize the objective function using a new dual-gradient descent algorithm iteratively

which accelerates the computing speed. By testing SL-ICA and Principal Component Analysis （PCA）. on artificial simulated SAR images and ENVISAT-ASAR （Advanced Synthetic Aperture Radar） images of Beijing

the results show that SL-ICA can get more precise results than the PCA.