TL;DR
This paper introduces a novel workflow for automatically detecting and classifying seismic events like rockfalls and landslides from continuous sensor array data. By combining advanced multi-channel detection with a graph-based feature weighting approach, they significantly improve event identification and classification accuracy, especially in noisy environments.
What Problem Does This Paper Address?
Slope instabilities, triggered by factors like heavy rainfall or earthquakes, pose a significant threat, causing mortality and infrastructure damage. Early detection and classification of seismic events associated with these instabilities are crucial for mitigation. However, analyzing continuous, noisy seismic data from multiple sensors presents several challenges:
- Noise: Seismic recordings are often contaminated by background noise, making it difficult to distinguish genuine events.
- Data Volume: Continuous monitoring generates massive amounts of data, requiring efficient processing.
- Feature Importance: Understanding which seismic signal characteristics are most important for classifying different event types (rockfalls, slide quakes, earthquakes, noise) is complex.
- Interpretability: Many deep learning approaches, while powerful, act as "black boxes," lacking transparency in their decision-making.
- Data Scarcity: Annotated seismic datasets, especially for specific event types like landslides, can be limited and imbalanced.
This paper aims to address these challenges by developing an interpretable and effective system for seismic event detection and classification from multi-channel sensor arrays.
Methods
The proposed methodology consists of a multi-stage workflow:
- Multi-Channel Event Detection:
- Multi-channel Coherency Migration (MCM): This technique is used to analyze the linear coherence across multiple seismic channels within a time window, helping to identify potential seismic signals.
- Neyman-Pearson Lemma: This statistical tool is applied to filter out windows with low signal-to-noise ratios (SNR), effectively reducing background noise.
- Optimal Channel Selection: For each detected event, the channel with the highest SNR is automatically selected. This stage is summarized in Algorithm 1.
- Feature Engineering and Classification:
- Feature Construction: A comprehensive set of 119 temporal, spectral, cepstral, and acoustic features are extracted from the detected seismic events.
- Graph-Based Feature Weight Optimisation: This is a core novel contribution. It adapts a graph-based semi-supervised learning approach to learn the importance of each feature. The method iteratively optimizes feature weights and classification labels, allowing the model to prioritize features that best discriminate between event types. This process is detailed in Algorithm 2.
- Graph-Based Classification: The optimized feature weights are then used to train graph-based classifiers, such as the Graph Laplacian Regularisation (GLR) classifier.
Datasets
The study utilizes publicly available raw seismic recordings from the SZC array of the French Landslide Observatory OMIV. The data was collected during specific periods in 2013, 2014, and 2015. For labeling and benchmarking, the authors used:
- A catalog of manually verified events from [12] for initial validation.
- A period of continuous data (24-28 November 2014) that was manually labeled by an expert for more realistic scenario testing.
Key Insights and Ideas
- Multi-channel coherence is key: Combining data from multiple sensors using MCM significantly improves event detection, especially for low-amplitude signals and in noisy conditions, outperforming traditional single-channel methods.
- Graph-based feature weighting offers interpretability: Unlike black-box deep learning models, the graph-based approach explicitly learns and assigns weights to features, revealing which seismic characteristics are most indicative of different event types. This provides valuable insights into the physics of seismic signals.
- Feature selection is crucial for performance: The study demonstrates that intelligently selecting features, particularly using the proposed graph-based method, leads to more focused feature sets and enhanced classification accuracy, often with fewer features than traditional methods.
- Graph-based classifiers benefit from optimized features: While standard feature sets yield similar performance across various classifiers, graph-based classifiers show significant improvements when paired with features optimized by the proposed graph-based weighting approach.
- A comprehensive feature list for slope failure events: The paper provides a valuable resource by identifying and recommending key features that best characterize rockfalls, slide quakes, earthquakes, and noise, aiding future research.
Key Results
- Improved Detection: The proposed multi-channel detection method identified 614 more seismic events than a standard STA/LTA approach on continuous data over a five-day period. It also detected more cataloged events and missed fewer compared to a single-channel method.
- Enhanced Classification: Feature selection, particularly via graph-based feature weight optimization, reduced the number of features by more than half while improving classification performance. For example, the GLR classifier's sensitivity for rockfall and slide quake events increased to 92% and 88%, respectively, from 89% and 85%.
- Robustness: The graph-based approach demonstrated resilience to imbalanced datasets and noisy conditions.
- Feature Importance: The study identified specific temporal and spectral features that are highly discriminative for different seismic event types, providing valuable insights for future analysis.
Limitations and Future Work
- Thresholding Sensitivity: The effectiveness of the current detection thresholding strategy is dependent on the consistency of background noise. Low-amplitude events might still be missed.
- Noise Categorization: The classification of Natural/Anthropogenic noise events, due to their complex and changing nature, requires further enhancement.
Future work suggested by the authors includes:
- Developing adaptive background noise removal techniques to further increase SNR.
- Identifying concealed microseismic activity.
- Implementing post-processing steps to improve classification outcomes, especially for noise-type occurrences.
Summary updated