Abstract: Synthetic aperture radar for earth observation has the characters of wide coverage, multi-polarization, multi-resolution, all time and all weather, which have been widely used in intelligent monitoring systems. With the improvement of SAR remote sensing image resolution, the fine-grained target classification becomes a challenge task. In this paper, we use the observation data of 10 types of vehicles with 0.3 m resolution, HH polarization, and multi-azimuth angle under spotlight imaging mode, aiming at traditional classifiers’ feature generalization ability limited by low intra-class variance problem for fine-grained vehicle type recognition of SAR remote sensing images, the multi-scale residual convolution neural network with high order feature representation is proposed, which can improve the feature extraction ability corresponding to remote sensing scenes and enhance accurate and robustness of vehicle recognition. Extensive experiments carried on MSTAR dataset show that proposed method performs the remarkable result comparing with the state-of-the-art intelligence classification methods, whose OA (Overall Accuracy) can reach 99.88% for vehicle recognition over 10 classes.