An effective identification method and an informative signal are two inseparable aspects for precise estimation of power systems' electromechanical modes. Informative signal is conventionally selected by power spectrum density (PSD) analysis from numerous amounts of ambient data. However, directly comparing PSD plots of different signals is casual and imprecise, especially for those non-obvious modes which could be overlooked in the PSD plots. In this paper, a signal selection approach based on PSD analysis is proposed to accurately evaluate whether a signal is informative to a mode. Simulation is carried out in an 8-generator 36-node benchmark system; mode estimation is employed to validate its effectiveness.