Weak Template Matching Based on GPU Parallel Acceleration and Its Application in Sichuan-Yunnan Region

In view of the critical problem that the template recognition class method takes too long to compute, the GPU parallel acceleration is carried out in the three main operation parts of the weak template matching method including cross-correlation, normalization and widening peak, which effectively improves the computing efficiency. Furthermore, the influence of filter frequency on the cross-correlation calculation results under discrete sampling is discussed, and a good filter frequency band is selected for weak template matching. The method is applied to Sichuan-Yunnan region. Using the precise location of earthquakes by double-difference seismic tomography as a template, we scanned the consecutive data of 60 days before and after the June 17, 2019 Changning earthquake in Sichuan Province, and detected 81, 704 earthquakes. According to the relationship between the frequency of earthquakes and the cross-correlation value, an appropriate threshold is selected to screen out 7 618 earthquakes as the final result, which is 3 times of the number of earthquakes detected by the deep learning automatic pick up algorithm. The micro-earthquakes obtained by the weak template matching method can construct a more detailed earthquake sequence, which reflects more clustering of earthquakes in time. Combined with the spatial location information, the micro-earthquakes can be used to study fault morphology, seismicity and its variation.