9:45 AM - 10:00 AM
[SCG51-04] Comparison of 1D and 2D CNNs for seismic event classifications ~ applications to volcanic events and tectonic tremor
Keywords:Machine learning, explosion earthquakes, slow earthquakes
Introduction: Machine learning methods have been applied to seismic signal detections and classifications. Ross+ (2018) developed a method based on 1D convolutional neural network (CNN) with waveform trace input, while Nakano+ (2019) and Takahashi+ (2021) developed a method using 2D CNN inputting running spectrum (RunSP) images. In the latter method, the signal frequency components that reflect the source characteristics are used for signal classifications. However, a convolutional (conv) layer works as a set of FIR filters, and 1D CNN may also learn frequency components effective for signal classifications. In this study, we compare the performances of these CNNs using the same data catalogs.
Method: We used 2D CNN same as Nakano+ (2019), in which two sets of conv and pooling (pool) layers are connected then put into two fully connected (FC) layers to classify signals into 3 classes as earthquake, tremor, or noise. We used 1D CNN like this one: Since 2D conv layer may be equivalent to two 1D conv layers, we connected 4 sets of conv and pool layers, then connected to two FC layers. We tried several 1D CNNs with different channel numbers.
The waveforms were first decimated to 25 Hz then cut into 163.84 s = 4096 pts data, as input of 1D CNN. Using the waveforms, we made RunSP images of 64 x 64 pixels (=4096 pts) by computing FFT of 5.12 s (128 pts) half overlap time windows. The RunSP images display frequency components between 0.2 and 12.5 Hz. Both input data are normalized by the maximum values.
We used two catalogs, one is for the Nankai trough shallow tremor used in Nakano+ (2019), and the other is for explosion events at the Sakurajima volcano in 2015 by the Japan Meteorological Agency (JMA). Seismometer records from DONET and observation network by JMA at the Sakurajima were used. The Nankai tremor catalog were separated into 70%, 15%, and 15% for train, trainval, and validation data, respectively. In the Sakurajima catalog, events in March to July are train, January and February are trainval, and August and September are validation. Since the Sakurajima catalog is strongly imbalanced, we allow duplication of samples from minority classes to balance the data with majority classes to create the minibatch. The balanced accuracy (BACC) is used for performance evaluations.
Results: Learning rate of the 2D CNN was faster than the 1D one, but the BACC for the validation data were similar. This result indicates that 1D and 2D CNN have similar performances. The difference in the 1D CNN models tried was trivial.
Discussion: Most AI methods are called a “black box” because it is very difficult to know the basis for signal classifications. So does CNN. In 1D CNN, the conv layer is a set of FIR filters and their responses could be the basis for the signal classifications at the first step. We tried a simple model with only 4 channels in the first conv layer, and computed the filter responses after training. They constitute low-pass and high-pass filters, with a common stopband in 4-8 Hz for tectonic tremor and volcanic signals. Although other characteristics are also used, the signals in the stopband commonly appear in the input signals and are not useful for the signal classifications.
Acknowledgments: This research was supported by JSPS KAKENHI Grant Number JP19K04050 and 21H05205 in Grant-in-Aid for Transformative Research Areas (A) “Science of Slow-to-Fast Earthquakes”. We used JMA catalog for explosion events in Sakurajima, and data obtained from DONET and JMA seismic network in Sakurajima.
Method: We used 2D CNN same as Nakano+ (2019), in which two sets of conv and pooling (pool) layers are connected then put into two fully connected (FC) layers to classify signals into 3 classes as earthquake, tremor, or noise. We used 1D CNN like this one: Since 2D conv layer may be equivalent to two 1D conv layers, we connected 4 sets of conv and pool layers, then connected to two FC layers. We tried several 1D CNNs with different channel numbers.
The waveforms were first decimated to 25 Hz then cut into 163.84 s = 4096 pts data, as input of 1D CNN. Using the waveforms, we made RunSP images of 64 x 64 pixels (=4096 pts) by computing FFT of 5.12 s (128 pts) half overlap time windows. The RunSP images display frequency components between 0.2 and 12.5 Hz. Both input data are normalized by the maximum values.
We used two catalogs, one is for the Nankai trough shallow tremor used in Nakano+ (2019), and the other is for explosion events at the Sakurajima volcano in 2015 by the Japan Meteorological Agency (JMA). Seismometer records from DONET and observation network by JMA at the Sakurajima were used. The Nankai tremor catalog were separated into 70%, 15%, and 15% for train, trainval, and validation data, respectively. In the Sakurajima catalog, events in March to July are train, January and February are trainval, and August and September are validation. Since the Sakurajima catalog is strongly imbalanced, we allow duplication of samples from minority classes to balance the data with majority classes to create the minibatch. The balanced accuracy (BACC) is used for performance evaluations.
Results: Learning rate of the 2D CNN was faster than the 1D one, but the BACC for the validation data were similar. This result indicates that 1D and 2D CNN have similar performances. The difference in the 1D CNN models tried was trivial.
Discussion: Most AI methods are called a “black box” because it is very difficult to know the basis for signal classifications. So does CNN. In 1D CNN, the conv layer is a set of FIR filters and their responses could be the basis for the signal classifications at the first step. We tried a simple model with only 4 channels in the first conv layer, and computed the filter responses after training. They constitute low-pass and high-pass filters, with a common stopband in 4-8 Hz for tectonic tremor and volcanic signals. Although other characteristics are also used, the signals in the stopband commonly appear in the input signals and are not useful for the signal classifications.
Acknowledgments: This research was supported by JSPS KAKENHI Grant Number JP19K04050 and 21H05205 in Grant-in-Aid for Transformative Research Areas (A) “Science of Slow-to-Fast Earthquakes”. We used JMA catalog for explosion events in Sakurajima, and data obtained from DONET and JMA seismic network in Sakurajima.