Abstract: It was the trend of landmine detection using vehicle-mounted ultra-band ground penetrating radar, which had the capability to detect landmines over large area. The difficulty in ultra-band ground penetrating radar was the extraction of steady features. In this paper, the two-dimension time-frequency image was extracted based on the five-dimension scatter function, which was constructed of range, azimuth, frequency, and angle of reection. And the feature selection and classification algorithm were based on the time-frequency images of targets and clutters. To decrease the training error and increase the generalization capability of classifier, the feature selection was involved in the design of classification. And to guarantee the probability of landmine detection, the false rate was regarded as cost function of classification through the training of classifier. It was proved by real data that the method was applicable to vehicle-mounted ultra-band ground penetrating radar. The classifier finally generated has a good generation capability and can offer constant probability of detection.