← Back to Gallery

Session: Urban Seismology

Type: Oral

Date: 10/10/2024

Time: 12:00 PM

Room: Stanley Park Ballroom

Data Augmentation Techniques to Improve Automatic Detection in DAS Records

In addition to regional earthquakes, Mexico City is also affected by local seismic events. To better understand these local phenomena, it is important to monitor local earthquakes and create a comprehensive catalog that includes events with very low signal-to-noise ratios as well as those with higher amplitudes.

Using data collected during a DAS experiment conducted from May 2022 to June 2023, we have detected nearly a hundred local seismic events using classical algorithms such as STA/LTA. However, manual inspection of some records revealed that STA/LTA missed many low-amplitude events, likely due to the noise produced by the city.

To address the limitations of classical detection algorithms, we propose the implementation of a convolutional neural network (CNN), which has been proven effective in similar situations. The challenge, however, is that a labeled dataset of only a hundred samples is insufficient to properly train this type of neural network. To overcome this limitation, we performed data augmentation on the original dataset to generate several thousand labeled samples. The advantage of this approach is that it eliminates the need for simulations or modeling, which would be time-consuming and computationally intensive given the complexity of Mexico City's geology.

Once trained, the proposed CNN can process the entire 120 terabytes of data acquired during the DAS experiment in just a couple of days, enabling the detection of hundreds of seismic events each month.

Presenting Author: Alfonso

Additional Authors