← Back to Gallery

Session: Urban Seismology [Poster]

Type: Poster

Date: 10/9/2024

Time: 07:00 AM

Room: Stanley Park Ballroom

Deep Learning Models Applied to Localization of Mexico City Microseismicity Recorded with DAS

Localizing earthquakes is not always easy. In particular, we show that most of the localizations of local earthquakes in Mexico City can be seriously questioned. Although the seismic network has been densified for the 20 past years, the highly heterogeneous media in the city and the rapidly evolving ratios Vp/Vs through depth (about a factor of 100 in the first 80m of depth and then about 1.7 at 1km depth) make obsolete the general tools for localization. Worst, the first observed shear arrival does not always follow a direct path, and phase identification with a scattered seismic network in such a geological context is uncertain.Then, the high spatial and temporal resolution measurements in Mexico City, such as the ones associated with Distributed Acoustic Sensing (DAS) technology, are of significant interest. We experimented from 2022 to 2023 and detected a hundredth of local earthquakes. We show that it is easy to follow several phases through a large number of channels. However, DAS measurements conducted with dark fibers in urban environments don’t allow for observation of compressional wave.

We will also show that localizations can be improved by considering more than the direct P and S arrivals. However, identifying each phase is out of the question. Instead, we would apply convolutional neural networks (CNN), which are well-suited for this task due to their ability to capture spatial and local patterns within the data. The CNN architecture was designed to include multiple convolutional layers with appropriate kernel sizes to extract features from data, followed by fully connected layers to map these features to the 3D coordinates. Data preprocessing involved denoising, normalization, and segmentation of DAS signals to focus on relevant temporal windows associated with microseismic events. The model was trained using a supervised learning approach, where the known positions of microseismic events served as ground truth. A comprehensive evaluation using a separate validation set was conducted to assess the model’s performance and generalization capability.

Presenting Author: Kevin

Additional Authors