A Bayesian inference model was developed to select per-pixel endmembers set. Considering the uncertainty of endmembers’ spectra, based on Bayesian inference and linear spectral mixture model, the posterior probability of per-pixel endmembers set was obtained. Using the prior knowledge of endmembers’ coexistence, combined with the normal distribution function of endmembers’ spectra, the optimal endmembers set was selected based on maximal liklihood. The experiment on an ETM+ image including 147431 pixel showed that, compared with the MRES and PG algorithms provided by IDRISI, the Bayesian approach reduced 70% redundant endmembers, and the unmixing error induced by inaccurate endmembers was reduced to 28%. The result showed that, considering the uncertainty of endmemebers′ spectra and the coexistence of endmembers, Bayesian inference can increase the accuracy rate of endmembers′ selection, so the unmixing accuracy was improved significantly.