In the traditional linear spectral mixture model,a feature class of t he hyperspectral image is represented by a single endmember.However,the spectral variability within a n endmember class is usually large because of wide space range and the feature complexity of the hyperspectral image. Under these conditions,a single endmember is difficult to portray a feature cate gory accurately, leading to incorrect unmixing results.Classical multi-endmember spectral unmixing algorithms play a positive role in overcoming the intra-class spectral variability,but there are shortcomings on large amount of calculation,cumbersome endmembers pre selection and so on.For these issues,we propose a hierarchical multi-endmember spec tral mixture analysis algorithm.The first layer is to determine the feature category by solv ing the maximum unmixing abundance error,and the second is stratified to find the optimal numbe r of endmemers contained in the pixels on the basis of the first layer.Simulated data and real hyperspectral data experiments prove that the proposed algorithm is better than the fixed end member unmixing algorithms,and the average abundance error cuts down by 2.65% at most,while comp ared with MESMA algorithm,the proposed algorithm reduces the comp utation and improves computational efficiency greatly,with almost the same preci sion.