Ground-based synthetic aperture radar (GB-SAR) has a wide range of applications in bridge health detection by monitoring dynamic deflection data. However, the collected dynamic deflection signals are easily subjected to interference by noises during GB-SAR monitoring due to ground motion and complex traffic factors. It is also difficult to accurately eliminate the influence of noises by using the traditional modal decomposition method. Therefore, we propose a cyclically shifted extreme-point symmetric mode decomposition (CS-ESMD)-based progressive denoising approach, which can accurately identify high/low-frequency noise information from dynamic deflection signals through a progressive process. First, CS-ESMD is used to construct virtual multi-channel signals for the following progressive denoising process. Second, ESMD is performed on multi-channel dynamic deflection data to separate useful and high-frequency noise information. Finally, the low-frequency noises and the residual high-frequency noises are further identified and removed by second-order blind identification (SOBI) and the fast Fourier transform (FFT) method. Through simulation and practical experiments, we show that the accuracy of the progressive denoising method can be increased by 37.2% compared with traditional methods, which shows its effectiveness in improving the precision of GB-SAR dynamic deflection signals.