📄 earth_science

Multifractal signatures reveal source-dependent differences between tectonic earthquakes and glacial tremors

This study introduces a robust classification framework combining Multifractal Detrended Fluctuation Analysis (MFDFA) with machine learning to effectively distinguish tectonic earthquakes from glacial tremors by leveraging their distinct multifractal signatures, achieving up to 97.35% accuracy in complex, noisy environments.

Original authors: Walid E. AboElnasr, M. A. Zahran, Mohamed M. Abdelsalam, Elshaimaa Amin

Published 2026-07-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Walid E. AboElnasr, M. A. Zahran, Mohamed M. Abdelsalam, Elshaimaa Amin

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing in a noisy room where two different groups of people are making sounds. One group is a chaotic crowd of people shouting, bumping into each other, and creating a complex, messy noise (these are tectonic earthquakes). The other group is a single person snapping a dry twig or a block of ice cracking (these are icequakes).

To the human ear, and even to standard computer microphones, these two sounds can sometimes look very similar on a graph. They both start suddenly and have a certain "crack" to them. This makes it very hard for automated systems to tell them apart, which is a problem for scientists trying to map real earthquake risks in icy places like Alaska.

This paper introduces a new way to listen to these sounds that goes beyond just "volume" or "pitch." Here is how they did it, explained simply:

1. The Problem: The "Look-Alike" Confusion

In places like the Columbia Glacier in Alaska, the ground shakes from two main sources:

Real Earthquakes: Caused by massive tectonic plates grinding against each other deep underground.
Icequakes: Caused by glaciers cracking, calving (breaking off), or shifting near the surface.

Standard computers often get confused because both events look like a sudden spike on a graph. If a computer mistakes an icequake for a real earthquake, it might think the region is more dangerous than it actually is.

2. The Solution: The "Fractal Fingerprint"

The researchers realized that while the sound might look similar, the texture of the vibration is fundamentally different. They used a mathematical tool called Multifractal Detrended Fluctuation Analysis (MFDFA).

The Analogy: The Rough vs. The Smooth
Think of the earthquake signal as a jagged, stormy mountain range. It has deep valleys, sharp peaks, and everything in between. It is incredibly complex and "rough" at every level you look at it. This is because a real earthquake involves a messy, chaotic release of energy deep underground.

Think of the icequake signal as a smooth, rolling hill or a simple, clean crack. It is much more uniform and "smooth." It lacks the deep, complex layers of the earthquake.

The researchers didn't just look at the height of the waves; they measured the complexity of the shape. They found that:

Earthquakes have a "wide" and complex signature (like a stormy mountain range).
Icequakes have a "narrow," simple signature (like a smooth hill).

3. The Test: Teaching the Computer to "See" the Texture

Once they measured these "textures" (which they called multifractal features), they fed this data into a smart computer program called a Support Vector Machine (SVM).

Think of the SVM as a very strict bouncer at a club. Instead of asking "How loud is the sound?", the bouncer asks, "How complex is the texture of this sound?"

If the sound is complex and jagged (Earthquake), the bouncer lets it in.
If the sound is simple and smooth (Icequake), the bouncer stops it.

4. The Results: A Highly Accurate Filter

The team tested this on over 28,000 recorded sound waves from Alaska.

The Score: The computer got it right 93.95% of the time on average.
The Best Case: At one specific listening station (KNK), the computer was right 97.35% of the time.

Even at the stations where the signal was a bit noisier, the system still managed to get it right over 90% of the time.

5. Why This Matters (According to the Paper)

The paper claims this method is a powerful tool for cleaning up seismic catalogs (the official lists of earthquakes).

By automatically filtering out the "ice noise," scientists can create a cleaner, more accurate list of real earthquakes.
This ensures that when they calculate the risk of future earthquakes (seismic hazard), they aren't accidentally counting ice cracks as dangerous tectonic events.

In summary: The researchers built a system that doesn't just listen to what the ground is shaking, but analyzes how complex the shake is. By recognizing that real earthquakes are "messy and complex" while ice cracks are "simple and smooth," they created a highly accurate filter to separate the two, helping scientists keep their earthquake records clean and reliable.

Technical Summary: Multifractal Signatures in Seismic Discrimination

Problem Statement

Accurate discrimination between local tectonic earthquakes and cryo-seismic events (icequakes) remains a critical challenge for automated seismic monitoring, particularly in glaciated regions like Alaska. While automated detection algorithms have improved efficiency, classifying these events is difficult because they share similar time-domain characteristics, including overlapping magnitude ranges and impulsive onsets. Traditional linear spectral methods often fail to distinguish between event types with subtle structural similarities, and deep learning approaches frequently lack physical interpretability or require massive labeled datasets that are unavailable for specific local phenomena like icequakes. Consequently, seismic catalogs in these regions risk contamination by non-tectonic events, which can artificially inflate seismicity rates and compromise probabilistic seismic hazard assessments.

Methodology

The study proposes a classification framework that combines Multifractal Detrended Fluctuation Analysis (MFDFA) with Support Vector Machines (SVM) to quantify the intrinsic non-linear scaling properties of seismic sources.

Data Collection and Preprocessing

Dataset: A comprehensive dataset of 28,022 waveforms was compiled from the Columbia Glacier region, Alaska, spanning January 1, 2005, to January 1, 2022.
Event Selection: The dataset includes 2,650 labeled icequakes and a balanced subset of 3,000 tectonic earthquakes (magnitude $M_L \le 3.0$ ) from May 2018 to April 2022 to ensure class parity.
Processing: Vertical-component (BHZ) waveforms were extracted in 5-minute windows, resampled to 50 Hz, and instrument responses were removed. Signals were detrended and band-pass filtered (0.5–23 Hz) to isolate the frequency band of interest.

Feature Extraction (MFDFA)

The authors applied MFDFA to extract 10 robust multifractal features that characterize the signal's complexity, regularity, and geometric structure. Key features include:

$\alpha_0$ (Peak Singularity): Represents the dominant regularity of the signal.
$\Delta\alpha$ (Spectrum Width): Quantifies the multifractality; a wider width indicates a more heterogeneous process.
$D_1$ and $D_2$ : Information and correlation dimensions, respectively, measuring information content and correlation properties.
$A$ (Asymmetry): Describes the skewness of the multifractal spectrum.
$\alpha_{min}$ and $\alpha_{max}$ : Correspond to the most singular (rough) and most regular (smooth) components.

Classification

A Support Vector Machine (SVM) classifier was employed using a Radial Basis Function (RBF) kernel. The model was optimized via a grid search with 5-fold cross-validation, identifying optimal hyperparameters ( $C=1000$ , $\Gamma=0.01$ ). The final evaluation utilized a rigorous 10-fold stratified cross-validation scheme across nine seismic stations.

Key Contributions

Large-Scale Diverse Dataset: The compilation of a balanced dataset spanning a wide range of Signal-to-Noise Ratio (SNR) conditions, enabling a robust assessment of feature stability across varying signal qualities.
Robust Feature Extraction: The implementation of a reproducible workflow extracting mathematically grounded multifractal features ( $\Delta\alpha$ , $D_q$ , $\alpha_0$ ) that capture intrinsic non-linear scaling behavior, moving beyond qualitative visual assessment.
Physical Interpretation: A rigorous linking of multifractal signatures to source physics. The study demonstrates that tectonic earthquakes exhibit higher structural complexity and persistence (heterogeneous fault rupture), whereas icequakes display simpler, quasi-monofractal, and anti-persistent signatures (brittle tensile fracture).
High-Accuracy Classification: The framework achieved a global average accuracy of 93.95%, validating that physics-based multifractal features can effectively discriminate between event types while maintaining interpretability.

Results

Discriminatory Power: Earthquake waveforms consistently displayed broader multifractal spectra ( $\Delta\alpha$ ) and higher information/correlation dimensions ( $D_1, D_2$ ) compared to icequakes. Earthquake spectra were also found to be left-skewed (indicating sensitivity to large fluctuations), while icequake spectra were more symmetric.
Performance Metrics:
- Global Average Accuracy: 93.95%.
- Station Performance: All nine stations exceeded 90% accuracy. The highest performance was observed at station KNK (97.35%), attributed to high SNR and clear spectral separation. Even the lowest-performing station in the high-fidelity subset (KLU) maintained 90.24% accuracy.
- Sensitivity and Specificity: The framework achieved an average sensitivity of 93.49% and specificity of 93.87%, ensuring both the preservation of true tectonic events and the effective exclusion of cryo-seismic noise.

Significance

The paper claims that this multifractal framework provides a physically grounded and highly accurate tool for purifying seismic catalogs in complex glaciated environments. By effectively filtering cryo-seismic noise, the method ensures that Probabilistic Seismic Hazard Assessments (PSHA) are based on crustal deformation processes rather than catalog contamination. This reduces uncertainty in regional hazard models and strengthens risk mitigation strategies. The study emphasizes that the approach offers a direct, practical solution for automated seismic catalog cleaning, reducing the burden of manual analysis in data-rich environments.

The authors note that while the results are robust, future work is required to validate the transferability of these multifractal features to different geological settings (e.g., subduction zones) and to address computational latency for real-time early warning systems.