Abstract: Accurate picking of the first arrival of microseisms is the prerequisite for the estimation of source location. The traditional manual picking method is inefficient and time-consuming. The short time average long time average (STA/LTA) method, commonly used in automatic picking, has low picking accuracy for low signal-to-noise ratio signals. To address the above problems, a random forest-based automatic picking method of microseismic first arrival is proposed. Firstly, this study extracts the amplitude, energy and amplitude ratio of adjacent moments of microseismic data as features and mark each sample with feature categories. Secondly, a random forest model is constructed to identify microseismic first arrivals. Thirdly, the random forest model is used to calculate the probability of each test sample belonging to a certain category, and the first data sampling point with a probability of no less than 0.5 is defined as the microseismic first arrivals sampling point. In this experiment, microseismic monitoring data in deep boreholes of coal mine roadways is used. The results show that as the number of decision trees reaching 137 and the maximum depth reaching 6 in the random forest algorithm, the accuracy of the method for classifying microseismic data samples could reach 98.5%, and the average picking error for first arrivals of microseismic is 23.1 ms. Therefore, this method is better than the method of STA/LTA in terms of picking accuracy.