Abstract
The continuously growing amount of seismic data collected worldwide is outpacing our abilities for analysis, since to date, such datasets have been analyzed in a human-expert-intensive, supervised fashion. Moreover, analyses that are conducted can be strongly biased by the standard models employed by seismologists. In response to both of these challenges, we develop a new unsupervised machine learning framework for detecting and clustering seismic signals in continuous seismic records. Our approach combines a deep scattering network and a Gaussian mixture model to cluster seismic signal segments and detect novel structures. To illustrate the power of the framework, we analyze seismic data acquired during the June 2017 Nuugaatsiaq, Greenland landslide. We demonstrate the blind detection and recovery of the repeating precursory seismicity that was recorded before the main landslide rupture, which suggests that our approach could lead to more informative forecasting of the seismic activity in seismogenic areas.
Introduction
Seismic networks worldwide generate massive, continuous data streams, but analysis capabilities have not kept pace. Traditionally, seismologists rely on human-expert-intensive supervised methods or "template matching," which are designed to identify signals resembling known earthquake patterns. This creates two major bottlenecks: scalability, as humans cannot manually inspect every second of data from thousands of stations, and bias, as supervised models may ignore subtle, novel, or "non-standard" signals like volcanic tremors or slow-slip events. This paper introduces an unsupervised deep learning framework that automatically detects and clusters seismic signals within continuous, unlabeled data, bridging the gap with a "model-free" approach.
Methods
The framework consists of a three-stage pipeline designed to be both powerful and physically interpretable:
- Deep Scattering Network (DSN): Instead of standard Convolutional Neural Networks (CNNs), a DSN uses a tree of wavelet transforms. Wavelets are mathematically constrained to specific frequency bands and timescales, making the network's internal logic easier for geophysicists to interpret.
- Learnable Wavelets: The authors use Hermite cubic splines to allow the network to "learn" the optimal shape of the mother wavelet at each layer, driven by the clustering task itself.
- Dual-Loss Optimization: The model is trained using a Clustering Loss (minimizing negative log-likelihood of a Gaussian Mixture Model in a 2D latent space) and a Reconstruction Loss (ensuring the network can reconstruct the original signal).
Case Study: The Nuugaatsiaq Landslide
The approach was validated using real-world data from the NUUG seismic station in Greenland, focusing on the June 2017 Nuugaatsiaq landslide, a massive rockfall that triggered a tsunami.
The study utilized 24 hours of three-component seismic data leading up to the main rupture and a 17-day long-term set to test stability against environmental noise. By forcing the model to use wavelets, the authors ensured extracted features were relevant to signal processing (frequency, duration, amplitude).
Key Results and Insights
The most striking result was the "blind" recovery of an accelerating sequence of precursory seismic events starting 9 hours before the landslide. These events were too weak for raw data visualization and previously required specific, supervised template matching to identify.
The unsupervised model identified 171 precursory events, capturing 97% of those previously found via supervised methods. Additionally, in the 17-day dataset, the model successfully separated different types of ambient noise, such as ocean-radiated microseismic energy, matching known oceanographic models. During training, the clustering loss improved by a factor of ~4.5x while maintaining signal integrity.
Conclusion and Future Work
While successful, the model currently uses linear projection for dimensionality reduction; future iterations could employ non-linear Autoencoders. Additionally, while the model is unsupervised, certain hyperparameters like network depth still require manual definition. Future work aims for a fully automated "deep-hierarchical" clustering system. This framework represents a significant step toward autonomous seismic observatories capable of alerting researchers to unusual activity, even without prior templates.
Summary updated