Application of deep learning-based neural networks using theoretical seismograms as training data for locating earthquakes in the Hakone volcanic region, Japan

Earth, Planets and Space

Table 1 Root Mean Square Errors (RMSE) of hypocentral parameters estimated by various networks

Method	Training	Estimation	Root Mean Square Error (RMSE)
Method	Training	Estimation	Longitude (degree)	Latitude (degree)	Depth (km)	Time (s)	Magnitude
1DCNN (ConvNetQuake^a)	Simulation data (80%)	Simulation data (20%)	0.0406	0.0365	1.679	–	0.0930
2DCNN	Simulation data (80%)	Simulation data (20%)	0.0213	0.0184	1.3267	0.5002	0.1018
3DCNN	Simulation data (80%)	Simulation data (20%)	0.0100	0.0078	0.6208	0.2095	0.0371
		Observed data	0.0301	0.0291	1.3062	1.3816	0.5830
3DCNN + CUTOUT	Simulation data (80%)	Simulation data (20%)	0.0075	0.0091	0.5298	0.2444	0.0311
		Observed data	0.0167	0.0313	1.7504	0.9427	0.6063

^aConvNetQuake (Perol et al. 2018) is a task to classify earthquake location into specific regions. The results show in this table for ConvNetQuake is obtained by added our task to perform regression to locate hypocentral parameters during the learning stage