P-Wave detection using deep learning in time and frequency domain for imbalanced dataset

Early tsunami and earthquake warning systems need a good Automatic First Arrival Picking (AFAP) subsystem to determine the earthquake arrival time. This subsystem has a time-domain earthquake signal as the input and the arrival time of the primary earthquake wave (P-Wave) as the output. There are several methods of AFAP that are widely used nowadays, one of them is Short Term Average/Long Term Average (STA/LTA) fused with the AR-AIC method. Even though this method is real-time, its performance is still relatively low. With similar characteristics between the seismic signals and image data, utilising deep learning on AFAP can further increase its performance. The seismogram channels can be seen as the image height, and the signal at a certain window can be seen as the image width. Unfortunately, these image data will be considered an imbalanced dataset. In this research, deep learning with time domain and frequency domain are proposed as inputs with the Synthetic Minority Oversampling Technique (SMOTE) method. Deep learning is used because of its ability to generalise well on a huge dataset, while SMOTE is used to overcome the imbalanced dataset problem. With this proposed system, the accuracy is 99.3%, the Root Mean Square Error (RMSE) is 0.202 seconds, and the maximum execution time is 0.17 seconds with a periodic time of 0.4 seconds. With these results, the AFAP system has good results for estimating the first arrival earthquake time.