Due to differences in data acquisition and vegetation growth periods, vegetation recognition on low-and medium-resolution remote sensing imagery widely suffers from endmember variability. The endmember variability directly leads to large vegetation unmixing errors. To increase the vegetation recognition accuracy on the multispectral imagery, an intra-inter distance genetic algorithm (IIDGA) that accounts for the endmember variability was proposed. IIDGA can decrease the intra-distance and increase the inter-distance simultaneously, which enhanced the distinguishability of the endmembers through an automatic feature selection method. An optimal feature space for vegetation unmixing was constructed on the medium resolution imagery to improve the vegetation recognition accuracy based on the Landsat imagery. The importance of optimal feature selection was demonstrated by comparing the linear unmixing model accuracy based on the classical band combination features, the spectral and textural feature set and the proposed IIDGA. The results verified that feature selection was beneficial to improve the unmixing accuracy. The RMSE of IIDGA equalled 0.180 which was the lowest among the three methods. Meanwhile, the IID index-based Filter method, the standard GA-based Wrapper method and the proposed method were compared with their performances in automatic optimal feature selection. The results confirmed the superiority of the IIDGA in trading off accuracy and efficiency.