Detecting the onset of P-waves in seismic signals is a crucial objective in the development of early warning systems for earthquake-prone regions. This work presents a novel time-frequency based method to efficiently detect and pick the onset of P-wave in seismic signals with low SNR (signal-to-noise ratio). The proposed technique rests on a combination of time-series modelling of seismic noise and wavelet packet transform (WPT), in which the core idea involves tracking the difference between energies of data and one-step ahead model predictions over a select set of wavelet packets (frequency bands). Auto-regressive integrated moving average (ARIMA) models are used for modelling seismic noise, while the packets are selected using the prevailing understanding of P-wave frequency content. The proposed method is superior to the existing methods in two respects, (i) accuracy of detection and picking since it zooms into the frequency bands of interest (corresponding to P-wave onset) and (ii) robustness in the sense of minimal false alarms due to outliers and other sources, especially from low SNR seismograms. The performance of proposed method is illustrated on a simplified simulated process and real-time seismic data sets acquired under low SNR conditions. A comparative study with a recently developed maximum normalized cross correlation method is also presented to demonstrate the superiority of the proposed method. © 2019 The Society of Instrument and Control Engineers - SICE.