It is important in military field to detect landmine using vehicle-mounted ultra band ground penetrating radar, which has the capability to detect landmines over large area. The extraction of steady features has been the crucial factor in the practical use of ultra-band ground penetrating radar. In light of this, a new feature extraction and classification method was proposed based on the time-frequency atom decomposition. The method extracted the two-dimension time-frequency image based on four-dimension scatter function. After analyzing the character of time-frequency image, the time- frequency atoms were used to decompose the one-dimension range profile of landmine. Several time- frequency atoms which could describe the time-frequency character of landmine were sent to hierarchy classifier as features. It was proved by real data that the method was applicable to vehicle-mounted ultra band ground penetrating radar. Compared with the conventional feature extraction algorithms based on time or frequency field, the proposed method can extract steadier feature and improve the performance of landmine detection effectively.