[S-TT40] 最先端ベイズ統計学が拓く地震ビッグデータ解析

[STT40-05] 畳み込みニューラルネットワークによる地震波形古記録からの深部低周波微動の検出

金子 亮介^1,2、*長尾 大道^2,1、伊藤 伸一^2,1、鶴岡 弘²、小原 一成² (1.東京大学大学院情報理工学系研究科、2.東京大学地震研究所)

キーワード：深部低周波微動、地震古記録、深層学習、残差学習、畳み込みニューラルネットワーク

The establishment of the High Sensitivity Seismograph Network (Hi-net) in Japan more than 20 years ago has led to the discovery of deep low-frequency tremors. As such tremors are considered to relate to large earthquakes, it is an important issue in seismology to investigate tremors that occurred before establishing seismograph networks, which record seismic data in digital format. We use a deep learning model, convolutional neural network (CNN), to detect evidence of tremors from seismic waveforms recorded on papers more than 50 years ago. First, we construct a CNN based on the ResNet structure to detect tremors from seismogram images. Then, we conduct learning with synthetic images generated referring to the past seismograms recorded at the Kumano observatory, operated by Earthquake Research Institute, The University of Tokyo. The results show that the trained CNN can correctly determine whether tremors exist or not in the seismogram images. In addition, the Grad-CAM heatmaps to visualize the model predictions clearly indicate the tremor location in each image. Next, we conduct learning with seismogram images generated from the real data of Hi-net. Despite contaminations in the seismogram images due to a variety of noises, the CNNs trained through a fine-tuning successfully detect tremors. Finally, we apply the trained CNNs to the seismogram images at the Kumano observatory. The CNNs show a promising potential to detect tremors from the past seismogram images.