Accurate and reliable first arrival pickup is one of the key part in micro-seismic monitoring. Responding to the difficulty in picking up the first micro-seismic signals under the condition of low signal-to-noise ratio, S-transform is applied to the micro-seismic signal with strong noise, and then energy of each frequency point is calculated according to the S-transform spectrum to obtain the energy curve. The maximum frequency of the energy curve is used to determine the spectrum segmentation boundary, and the empirical wavelet transform is improved. After the micro-seismic signal is decomposed into several intrinsic mode functions according to frequency and energy distribution adaptively, the intrinsic mode functions outside the main frequency threshold are eliminated to reconstruct the de-noised signal. The average, power and Akaike information criterion value of de-noised signal are extracted as features. The signals are divided into effective signal cluster and noise cluster by Gath-Geva clustering, and the first time of effective signal cluster is taken as the first event arrival time. Taking the manual picking results as reference, and the picking deviation within 20 ms is regarded as correct. The picking accuracy is 71% when signal-to-noise ratio is as low as -10 dB, and 98.8% when signal-to-noise ratio is 5 dB, which is 2.5 percentage points higher than that of fuzzy c-means clustering algorithm. The actual micro-seismic picking shows that the algorithm is effective, and the recognition accuracy is up to 98%, which has a good applied value.