Ground-penetrating radar (GPR) is a commonly used high-resolution electromagnetic technique for subsurface imaging which has many applications such as landmine, cable and pipes detection. It is well known that the clutter presented in the GPR images deteriorates the image quality and decreases the performance of the target detection methods, especially for shallowly buried targets. To deal with this issue, we propose a new clutter removal method based on the low-rank approximation of the GPR image by non-negative matrix factorization (NMF). The proposed method is applied to real GPR images as well as to a realistic dataset provided by the new version of the electromagnetic software simulation tool gprMax. The visual and quantitative results obtained for different scenarios containing different soils, surfaces, burial depths and target types validate the effectiveness of the proposed method over the widely used state-of-the-art clutter removal methods.