Sampling GHz wideband signals at Nyquist rate is challenging. Some well-known sub-Nyquist sampling methods such as multi-coset sampling (MCS), have been proposed to overcome the sampling bottleneck. In the schemes based on MCS, inaccuracy of delay elements, clock offset and channel discrepancies cause undesired timing skews which consequently lead to measurement errors and degrade the spectrum reconstruction performance. In this paper, we propose an I-Q MCS architecture and a corresponding timing skew calibration approach. The new architecture performs two sets of MCS after implementing quadrature amplitude modulation on the received continuous-time signal. Unknown timing skews are estimated with the sub-Nyquist samples and used for calibrating the sensing model, which is built based on the relation between the sub-Nyquist spectrum and the Nyquist spectrum. Experiments show that the proposed method is able to reconstruct the spectrum with desirable performance even under significant timing skews. Robust spectrum sensing is accomplished at sub-Nyquist rates.