← Back to Sessions

Improving Earthquake Detection and Localization in Hawaii With Deep Learning and High-performance Computation

Description: 

Earthquake and volcano monitoring depends on earthquake detections and locations. The accuracy of earthquake detection and localization is of great importance for understanding detailed structures and dynamics. With the continuous expansion of seismic networks and the surge in earthquake data volume, traditional methods are time-consuming, labor-intensive, and require continuous human involvement when processing large amounts of data. Thus, more and more deep-learning-based approaches have been developed and achieved better performance than traditional methods. ArrayConvNet, a convolutional neural network model trained by 1,843 analyst-reviewed earthquakes in Hawaii, can seamlessly detect and localize events and achieved 99.4% detection accuracy and predict hypocenter locations within a few kilometers of the U.S. Geological Survey catalog. Furthermore, application to continuous records results in more detection of earthquakes. Because of the enhanced detection sensitivity, localization granularity, and minimal computation costs, ArrayConvNet is a potentially valuable model, particularly for real-time earthquake monitoring.
To further enhance the deep learning model, we collected a comprehensive high-precision relocated earthquake catalog and compiled a large-scale seismic dataset, which includes 598,400 earthquake events in Hawaii after data augmentation. By using multiple GPU cards in High-Performance Computational Clusters, we are able to utilize such a huge amount of seismic data in model training and accelerate the process. The two-order of magnitude increase in the data leads to an improved deep learning model. The result shows that the new model outperforms the initial one in terms of localization accuracy on a test dataset, reducing the mean distance error by one order of magnitude and the standard deviation error by 18%~25.8% in horizontal (north-south/east-west) and vertical (depth) directions, as well as reducing the event origin time residuals by one order of magnitude.

Session: Building and Decoding High-resolution Earthquake Catalogs With Statistical and Machine-learning Tools [Poster]

Type: Poster

Date: 4/15/2025

Presentation Time: 08:00 AM (local time)

Presenting Author: Zhangbao

Student Presenter: Yes

Invited Presentation: 

Poster Number: 34

Authors