Abstract: Hyperspectral remote sensing image is important in image spatial information and spectral information.However,the traditional dimension reduction method based on optimal band selection only considers the feature space information based on gray statistical characteristics,but ignores the image spatial information,thus causing a large amount of calculation.This paper,combined with the feature space information and image space information of hyperspectral remote sensing,proposes a new dimension reduction method combining multi-features of hyperspectral images and applies the method in mineral mapping.In this method,band correlation is first analyzed statistically and divided into different feature subspace.Then,fractal dimension of each band is computed and the band with the smallest fractal dimension is chosen as a candidate band in each subspace.Finally,calculation is made of the correlation coefficient between the spectrum of ground objects to be identified in the candidate bands so as to select quickly the optimal band combination.The result shows that,the image of the optimal band combination selected with this method is clear and shows an obvious correlation between different altered minerals.Besides,the mineral alteration information extracted by the method of mature spectral angle mapping(SAM)is roughly the same as the information extracted according to the feature selection.The above results indicates that the dimension reduction method combined with the image space and feature space can help to choose the ideal band combination.This reduces the dimension of hyperspectral image data and has a good effect of information extraction.