← Back to Sessions

Towards a Deep Learning Approach for Short-term Data-driven Spatiotemporal Seismicity Rate Forecasting Using Standard and High-resolution Earthquake Catalogues

Description: 

Recent advances in earthquake monitoring from a combination of dense, low-cost, broadband networks and automated data processing methods, have enabled the generation of high-resolution catalogues with a lower magnitude detection threshold and greater precision in relative location. These catalogues, generated using techniques such as template matching and machine learning, are created much faster and contain far more seismic activity data than traditional catalogues compiled by human analysts. When such detailed catalogues are used, statistical and physics-based earthquake forecasting models have been shown to perform better in terms of forecasting power. Combining high-resolution catalogues with machine learning algorithms for earthquake forecasting, which themselves have significantly developed over the last few years due to increased data availability and computational power, offers a promising method for discovering hidden patterns and laws in earthquake sequences. This study aims to develop short-term, data-driven spatiotemporal seismicity forecasting models using deep learning, testing the hypothesis that deep neural networks can reveal previously uncovered relationships in earthquake data. In particular we compare between using standard and high-resolution catalogues to train the deep learning-based forecasting models and offer insights on the impact of high-resolution data on the forecasting power using metrics commonly used in both data science and earthquake forecasting. We test the performance of two different neural network architectures, a convolutional neural network and a transformer, and reach the conclusion that both architectures achieve similar forecasting skill. Overall, the findings indicate that deep learning algorithms are a promising tool for producing short-term seismicity forecasts, provided they are trained on a dataset that accurately reflects earthquake sequence properties.

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: Foteini

Student Presenter: Yes

Invited Presentation: 

Poster Number: 35

Authors