Applying Compression Metrics to Seismic Data to Assist Analysts

Angela Zhou, R. A. Field, Laura E. Matzen, Christina Ting2023 · DOI

Abstract

This work introduces a novel compression-based signal processing method that can assist analysts with triaging and analyzing seismic data.

Introduction

Seismic analysis is a fundamental component of global nuclear monitoring and national security. However, the current workflow is heavily dependent on manual "picking"—the process where human analysts identify the precise arrival times of seismic waves amidst background noise. Automated systems designed to alleviate this burden frequently suffer from high false-positive (~40%) and false-negative (~20%) rates, forcing analysts to spend significant time correcting machine errors. There is a critical need for algorithms that are not only accurate but also provide reliable uncertainty quantification (UQ) to guide human intervention.

Methodology: Sliding Information Distance (SLID)

To address these challenges, researchers at Sandia National Laboratories developed the Sliding Information Distance (SLID) metric. Unlike traditional methods that rely on physical waveform modeling, SLID leverages principles from information theory.

The core premise is that data similarity can be measured through compressibility: if two segments of data are similar, they can be compressed more efficiently together than separately. The SLID algorithm employs two adjacent "sliding windows" that move across a seismic signal. When both windows contain only background noise, the information distance remains low. As the leading window encounters a seismic arrival, the similarity between the windows drops sharply, resulting in a peak in the SLID score.

Crucially, the shape of this peak serves as a built-in UQ measure. A sharp, distinct peak indicates high confidence in the arrival time, while a broad or noisy peak signals high uncertainty, alerting the analyst that manual review is required.

Experimental Results

The SLID method was evaluated using a dataset of 444 waveforms from 95 distinct seismic events, with expert human picks serving as the ground truth.

Performance Metrics

SLID demonstrated high precision, achieving a mean difference of only 0.54 seconds compared to expert picks. In comparative testing, SLID outperformed the industry-standard Akaike Information Criterion (AIC). While the two methods agreed within 0.1 seconds in 88% of cases, SLID proved more accurate in instances where their results diverged.

Human-in-the-Loop Integration

A key finding of the study involved the "anchoring effect." Traditional automated tools often present a single "best guess" line, which can bias human analysts toward the machine's choice even when it is incorrect. By providing the full SLID curve and UQ information, the researchers empowered analysts to understand why the algorithm might be uncertain. This transparency led to more accurate human picks and a stated preference among participants for UQ-rich visualizations over simple automated markers.

Discussion and Future Work

The application of compression metrics as a proxy for geophysics allows for robust detection without the need for complex modeling of every possible noise source. While highly effective at the individual waveform level, future research will focus on "event-level" analysis. By correlating high-confidence detections across multiple seismic stations, the system could automatically refine lower-confidence picks at noisier locations.

One remaining consideration for large-scale deployment is the computational overhead associated with real-time compression algorithms compared to simpler statistical models. Nevertheless, SLID represents a significant step toward a more efficient, "self-aware" seismic monitoring pipeline.

Summary updated

Abstract

This work introduces a novel compression-based signal processing method that can assist analysts with triaging and analyzing seismic data.
BibTeX
@misc{zhou2023applying,
  author = {Angela Zhou and R. A. Field and Laura E. Matzen and Christina Ting},
  title = {{Applying Compression Metrics to Seismic Data to Assist Analysts}},
  year = {2023},
  doi = {10.2172/2431251},
}
Figure