[M-GI33] Data-driven geosciences

[MGI33-09] Testing on Earthquake Prediction by Deep CNN-Labeling Model

*mitsuhiro toriumi¹ (1.Japan agency of marine science and technology )

Keywords:prediction testing of earthquake, deep convolutional neural network, variational timestep and time-shift method

The global and regional correlated seismicity time series calculated from micro to small earthquakes seismicity data matrix by means of singular value decomposition method are characteristic in earth mechanics and seismicity. The correlated seismicity time series, then potential macroscopic features showing the mechanical behavior of plate boundary zones in the global and regional scales. In this consequence, the relative large earthquakes occurred in the plate subduction zones are possibly corresponded to the multivariate correlated seismicity time series by means of deep neural network type transformation techniques such as RNN, LSTM, CNN, state-space modeling and VAE modeling. This study is the first attempt for prediction testing by these methods with multiple labeling on large earthquake event and shows the available results on global and Japanese island region. The available methods of 1dCNN of various timestep and time-shift are suggested from the viewpoint of accuracy scores and small number of incorrect prediction result. The variational autoencoder (VAE) is also applied for investigation of possible pre-seismic signals appeared in the correlated seismicity time series.