A novel independent component feature separation based on Spatial Down Sample(SDS) was presented for solving the long run-time defection of traditional Independent Component Analysis(ICA). Small windows were obtained by gridding the two-dimensional spatial space of a hyperspectral image. In each window, the distance between the central pixel and around pixels was measured by spectral similarity and the around pixels whose distances were smaller than the threshold value were discarded. The projection matrix was calculated by FastICA with the central and the around pixels whose distances were larger than the threshold value. The feature separation ICA components were achieved by projecting the original hyperspectral image using a project matrix. The performance of traditional ICA and SDS_ICA were compared. The influences of threshold values, window size values and the initial projecting matrix on the feature separation performance and run-time of SDS_ICA were studied. Experiment results show that SDS_ICA has the similar feature separation performance with the traditional ICA and its run-time has reduced above 30% under moderate threshold values and insensitivity window sizes. The novel method can be widely applied in the fields of hyperspectral feature extraction, data reduction, target detection etc.