This study presents a novel modal parameter identification method enabling approximation of the mode shapes of linear systems using white noise or earthquake inputs. The majority of well established existing system identification methods perform successfully when the system is excited by broadband white noise excitation. However, they encounter serious limitations when analysing the vibrations triggered by non-stationary earthquake inputs. Thus, the presented technique extends the applicability of system identification and modal based structural health monitoring methods. The method operates in modal space and is based on mode superposition in short windows. The mode shapes are identified using an optimization algorithm minimizing the weighted sum of cross correlation of frequency response spectra. The technique is validated analytically using simulation results of a simple 3D structure representing a simplified model of a real bridge pier structure, which enables exact comparison to known properties. The results show the method provides relatively good identification accuracy of modal parameters of systems excited by white noise and earthquake inputs. The identified modal frequencies showed <1% error, where the mode shape coefficients were identified within 5% error. The method performs robustly even for high levels of simulated sensor noise and can be readily applied to more complex MDOF systems.