PLAN: A Graph Neural Network Revolutionizing Earthquake Monitoring

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jan 11 2024

In a paper published in the journal Communications Earth & Environment, researchers introduced a groundbreaking graph neural network for earthquake monitoring. Unlike previous methods, it simultaneously handled phase picking, association, and location tasks using multi-station seismic data. This innovative approach, considering station relationships, yielded high accuracy and consistency. Testing in the Ridgecrest region and Japan showcased its superiority over existing techniques, offering a self-consistent system for automated earthquake monitoring's next generation.

Related Work

Past work in earthquake monitoring has traditionally involved several distinct steps: phase picking, association, and event location. Similar to computer vision's detection problems, phase picking has recently improved through deep learning methods, primarily employing convolutional neural networks (CNNs). Subsequent phases involving association and location utilize traditional or deep learning-based algorithms to link seismic phases and obtain earthquake hypocenters.

However, these tasks are interconnected, where the accuracy of one impacts the others. Efforts to enhance earthquake monitoring have faced challenges in leveraging inter-task and inter-station constraints, often conducting these tasks separately or on a station-by-station basis. While some recent graph-based approaches have shown promise in handling irregularly spaced stations, effectively integrating inter-task and inter-station constraints while performing all tasks remains challenging.

Graph-Based Earthquake Monitoring Study

This study utilized a graph-based neural network approach for multi-station earthquake monitoring, aiming to tackle the challenges of irregularly spaced stations during phase association and event localization. The innovative graph-based network structure transformed data from a matrix to a graph format, representing each station as a node and leveraging three-channel data alongside station locations as node features.

Unlike traditional single-station methods, this approach processed all three-channel data from multiple stations per event as a single sample, allowing efficient information aggregation during network training. Through meticulous evaluation, the transform graph convolution (TransformGCONV) module, employing attention mechanisms for message aggregation, emerged as the most suitable method among various graph aggregation techniques, enabling enhanced representation of inter-station relationships and significantly improving method accuracy and efficiency.

The network architecture comprised four interconnected modules: seismic phase picking, location, and association network (PLAN). The waveform feature extraction module utilized an encoder-decoder architecture to derive relevant features from seismic data, ensuring the integration of multi-station information. The earthquake location module used station coordinates and waveform features to predict each station's earthquake depth and epicentral distance. The physics-informed module employed TransformGCONV for feature aggregation to address multi-station phase picking, aligning waveform features across stations to improve accuracy and robustness.

Central to the approach was the multi-station association module, pivotal in synchronizing P/S-wave picks across stations by estimating time shifts. This module improved multi-station picking accuracy by aligning waveform features and leveraging graph convolutions in a temporally aligned space. Comparative analyses between models with and without this module consistently showcased PLAN's superior performance, affirming the efficacy of the multi-station association module in enhancing network robustness.

Researchers devised a dedicated workflow to adapt the PLAN model for processing continuous waveform data. This workflow involved initial prediction, shift-and-stack strategies for event detection, station selection based on alignment criteria, and catalog generation, resulting in a preliminary earthquake catalog from continuous waveforms. The process ensured stability and adaptability to variations in picking positions within continuous data by incorporating retraining and data augmentation techniques.

Researchers employed multi-task learning during training, using three distinct loss functions tailored to phase picking, association, and earthquake localization. The training involved optimizing the model using the adaptive moment estimation (ADAM) method, gradual learning rate decay, and selecting subsets from the training dataset to enhance efficiency. They introduced evaluation metrics to assess algorithm performance better, considering precision, recall, and F1 scores across multiple thresholds. The data used in this study and the developed model are available to facilitate further research and advancements in earthquake monitoring methods and applications.

Multi-Station Earthquake Monitoring: Advancements

The multi-station multi-task PLAN presents a comprehensive approach to earthquake monitoring, leveraging a graph neural network (GNN) backbone. This innovative model integrates waveform feature extraction, earthquake location, multi-station association, and physics-informed phase picking. Its GNN architecture handles irregularly spaced seismic data by representing seismic stations as nodes and their information as feature vectors. Through dynamically learned linking weights among stations, the model infers station relationships, allowing adaptability to variations in station numbers and locations.

The system's workflow involves encoding waveform features via a shared encoder-decoder CNN and extracting geographic features using two multi-layer perceptrons (MLPs) shared among nodes. Subsequent modules predict event depth and station-event offset, aiding event location triangulation. The critical multi-station association module aligns waveform features from various stations before phase picking, ensuring accurate aggregation. The network optimizes performance across interconnected modules by joint training using regression loss functions for phase picking, association, and earthquake location.

Performance evaluations in the Ridgecrest and Japan regions showcased PLAN's superiority over existing methods in phase picking and location accuracy. PLAN's phase-picking accuracy in Ridgecrest surpassed established methods like phasenet and earthquake transformers (EQTransformer). Additionally, its location estimation outperformed aggregated GNN, demonstrating concentrated residuals and minor errors. Similarly, in Japan, PLAN excelled in phase picking, particularly in mRecall and mF1 scores, showcasing its precision in detecting P/S waves accurately.

The model's application to continuous waveform data for the 2019 Ridgecrest earthquake sequence exhibited its potential for real-time earthquake monitoring. While acknowledging limitations related to processing multiple events within specific time windows, PLAN-generated earthquake catalogs are comparable or slightly superior to existing state-of-the-art methods. The catalog demonstrated robustness and high-quality event detection, aligning closely with established earthquake catalogs, reaffirming its credibility and reliability in earthquake detection and localization from continuous waveform data.

Conclusion

In conclusion, the multi-station multi-task PLAN revolutionizes earthquake monitoring with its sophisticated graph neural network architecture. Demonstrating superior accuracy in phase picking and event localization across Ridgecrest and Japan regions, it showcases adaptability to irregular seismic data and the potential for real-time monitoring. While there are limitations in handling multiple events within specific time windows, its generated earthquake catalogs exhibit robustness and reliability, aligning closely with established methods in earthquake detection and localization from continuous waveform data.