The phenomena of different objects having the same spectrum and the same objects having different spectrum bring inconsistency for the same endmember. The existing of endmember variability issue will lead the process of endmember selection and extraction more difficult and decrease the final unmixing accuracy. Aiming to minimize the intraclass variability and maximize the interclass variability, a new method named weighted coefficient of variation analysis (WCVA), which permitted the comparison of variants free from scale effects and made the weighting become more automatic, was proposed for multispectral data. It was on the basis of coefficient of variation (CV) and weighting theory. The proposed method was successfully indicated from theoretical and experimental parts. The comparison with the commonly used optimal index factor (OIF) was conducted in terms of visualizing the spatial distribution of all available band combinations, efficiency and the final unmixing accuracy by fully constrained least squares (FCLS) and post polynomial postnonlinear mixture (PPNM) with TM and GeoEye images in the same research area. In the experimental results, the unmixing accuracy (0.183 and 0.160) based on the feature combination selected by WCVA was higher than that by OIF. Meanwhile, the computation of WCVA was much less than that of OIF as well. The results showed that WCVA not only had benefits for solving endmember variability issue and enhancing the unmixing accuracy, but also had higher efficiency.