Numerous automatic recognition methods based on convolutional neural networks (CNNs) can achieve high calcu-lation efficiency owing to its end-to-end structure. However, the internal mechanism of CNNs is intransparent which is limiting or even disqualifying in SAR image interpretation. To provide a visual understanding of CNNs' mechanism, we propose a Group-Self-Matching class activation mapping (G-SM-CAM) inspired by the split strategy and Self-Matching CAM. In specific, feature maps are firstly “self-matched” with the input image as the renewed feature maps. Then these renewed feature maps are split into several groups. Finally the saliency map can be generated by these sub-feature maps. G-SM-CAM is efficient and effective on SAR images, which runs dramatically faster than Self-Matching CAM at minor cost of saliency map quality. Numerous experimental results demonstrate the validity and efficiency of G-SM-CAM based on a benchmark dataset MSTAR.