Extracting Slow Slip Events from Noisy Geodetic Time Series Using Spatiotemporal Graph Neural Networks

Core Concepts

A spatiotemporal graph neural network-based method, SSEdenoiser, is developed to efficiently denoise geodetic time series and extract slow slip events from the denoised signals.

Abstract

The paper presents a deep learning-based method, SSEdenoiser, for denoising geodetic time series and extracting slow slip events (SSEs) from the denoised signals. Geodetic data, such as from Global Navigation Satellite Systems (GNSS), are affected by various noise sources that are spatially and temporally correlated, making it challenging to separate the signals of interest, like SSEs, from the noise.
The key aspects of the methodology are:

Synthetic data generation: The authors generate realistic synthetic GNSS noise and SSE signals to train the model, accounting for the complex spatiotemporal characteristics of the data.

Graph-based recurrent neural network: SSEdenoiser uses a graph-based recurrent neural network to extract spatial and temporal features from the multi-station GNSS time series, learning the graph connectivity from the data.

Spatiotemporal Transformer: A spatiotemporal Transformer module is used to attend to the learned spatial and temporal features, enabling the model to focus on the relevant space-time relationships.

The proposed method is evaluated on both synthetic and real GNSS data from the Cascadia subduction zone. On synthetic data, SSEdenoiser outperforms traditional signal processing techniques as well as other deep learning baselines. On real data, the denoised displacements from SSEdenoiser show good correlation with independent seismic tremor observations, validating the ability of the method to extract the weak SSE signals from the noisy geodetic time series.

Stats

The average signal-to-noise ratio (SNR) is defined as:
SNR = 1 / |S'| * Σ_j∈S' Σ_k (10 log10(Σ_t |ξ^k_j(t)|^2 / Σ_t |n^k_j(t)|^2))
where S' is the set of stations that recorded a non-zero displacement.

Quotes

"Denoising geospatial data is, therefore, essential, yet often challenging because the observations may comprise noise coming from different origins, including both environmental signals and instrumental artifacts, which are spatially and temporally correlated, thus hard to disentangle."
"The challenge in GNSS data denoising lies in developing a method able to learn how to decorrelate these signals by separating what we consider as noise and the different signals from each other."

Key Insights Distilled From

Denoising of Geodetic Time Series Using Spatiotemporal Graph Neural Networks: Application to Slow Slip Event Extraction

by Giuseppe Cos... at arxiv.org 05-07-2024

https://arxiv.org/pdf/2405.03320.pdf

Deeper Inquiries

How can the proposed spatiotemporal graph neural network architecture be extended to handle other types of geospatial data beyond GNSS, such as satellite imagery or seismic data

The proposed spatiotemporal graph neural network architecture can be extended to handle other types of geospatial data beyond GNSS by adapting the graph representation and incorporating domain-specific features. For satellite imagery, the network can be modified to process image data by treating each pixel as a node in the graph and considering spatial relationships between neighboring pixels. This would involve using convolutional layers to extract spatial features and capture patterns in the imagery. Additionally, for seismic data, the network can be tailored to analyze waveform signals by representing seismic stations as nodes in the graph and incorporating temporal dependencies in the data. By adjusting the input data format and the network architecture, the spatiotemporal graph neural network can effectively denoise and extract relevant information from various geospatial datasets.

What are the potential limitations of the synthetic data generation approach, and how could it be further improved to better capture the complexity of real-world geophysical processes

The synthetic data generation approach may have limitations in capturing the complexity of real-world geophysical processes due to simplifications in the underlying models and assumptions made during data generation. One potential limitation is the reliance on synthetic slow slip event models, which may not fully represent the variability and nuances of actual slow slip events observed in nature. To improve the synthetic data generation process, more sophisticated physical models can be incorporated to simulate a wider range of slow slip event scenarios, including variations in slip behavior, duration, and spatial extent. Additionally, introducing more realistic noise models that mimic the complexities of environmental and instrumental noise in geospatial data can enhance the fidelity of the synthetic data. By refining the synthetic data generation techniques to better reflect the intricacies of real-world geophysical processes, the denoising model can be trained on more representative data, leading to improved performance and generalization.

Could the learned graph connectivity from SSEdenoiser provide insights into the underlying tectonic structure and processes in the Cascadia subduction zone beyond the extraction of slow slip events

The learned graph connectivity from SSEdenoiser can provide valuable insights into the underlying tectonic structure and processes in the Cascadia subduction zone beyond the extraction of slow slip events. By analyzing the adjacency matrix and the strength of connections between GNSS stations, researchers can infer spatial relationships and dependencies between different regions within the subduction zone. This information can offer clues about fault interactions, stress distribution, and potential seismic hazards in the region. Furthermore, the importance values assigned to individual stations can indicate their relevance in monitoring tectonic activity and deformation. By leveraging the learned graph connectivity, researchers can gain a deeper understanding of the geophysical dynamics in the Cascadia subduction zone and potentially uncover hidden patterns and correlations in the geospatial data that go beyond slow slip event detection.