Reza Mansoori

Given the complexity of earthquake occurrence and the lack of clear patterns, accurately predicting them remains challenging. However, recent studies have shown that neural networks can play a significant role in analyzing earthquake data, particularly in determining the potential location of an earthquake. In this research, with the aim of improving earthquake location accuracy, data from 4300 earthquake events with a magnitude greater than 2.5 from the IR seismic network, recorded by 15 seismic stations in three components BN, BZ, and BE, were collected. The waveform of each earthquake was converted into an image. Then, these waveform images were labeled with spatial information (latitude, longitude, and depth), and the spatial distribution of earthquakes was modeled in 3D using a Gaussian function and entered as input data into a fully convolutional neural network (FCN) method. By running the FCN method on earthquake waveform images, a 3D image of the earthquake locations is produced. Additionally, to evaluate and assess the location accuracy, P and S phase information of this earthquake dataset was extracted and used with linear hypo71 and nonlinear NonLinLoc methods to estimate the location error of this data. Finally, the location results of all three methods, hypo71, NonLinLoc, FCN neural network, and IRSC catalog were compared for 400 selected earthquake events from the training dataset. The comparison results show that the FCN neural network method, which earthquake locations using waveforms, has higher accuracy compared to the other two methods, hypo71 and NonLinLoc, which estimate earthquake locations using P and S phases.