Arrival Picking for Distributed Acoustic Sensing seismic based on fractional lower order statistics

In recent years, fiber-optic distributed acoustic sensing (DAS) has been gradually applied to seismology because of its long-distance and dense observation capability. It is a great challenge to effectively process the massive seismic data recorded by DAS. At present, the seismic data processing methods based on deep learning have achieved great success, especially in the tasks of seismic detection and arrival-time picking. However, due to the differences between DAS and geophone, such as sensing principles, spatial and temporal sampling rates, and noise intensity. The seismic arrival time picking model based on deep learning, which is trained by geophone seismic data with low spatial and temporal sampling rates and low noise intensity, severely degrades in performance on DAS seismic data with high spatial and temporal sampling rates and high noise intensity. In addition, a new seismic arrival time picking model is trained by fully supervised learning, which usually requires a large number of seismic data with accurate labels. However, the huge cost of manual picking and the lack of effective automatic picking models make it very difficult to build large-scale DAS seismic data sets with accurate labels. Therefore, it is very difficult to build an arrival time picking model based on fully supervised learning for DAS seismic data.

In this study, we propose a DAS seismic arrival time picking method based on fractional lower order statistics. Based on the difference of probability density function between noise and seismic signal, the proposed method uses alpha-stable distribution modeling noise (generally follow a Gaussian distribution) and seismic signal (generally follow a non-Gaussian distribution), and uses fractional lower order statistics under the assumption of alpha-stable distribution as the characteristic function to pick the arrival time.

Synthetic and actual DAS data tests show that the proposed method has better performance and robustness to random noise than other methods based on characteristic functions, such as STA/LTA, AR-AIC and kurtosis. Since the actual DAS seismic data has no ground truth of arrival time, we have further the performance of the proposed method on the geophone seismic data set. The proposed method provides better results on geophone seismic data and the data after up-sampling them to the typical time sampling rate of DAS.