GreenPhase: A Green Learning Approach for Earthquake Phase Picking

GreenPhase is an efficient, interpretable, and sustainable deep-learning model based on the Green Learning framework that achieves state-of-the-art earthquake detection and phase picking performance on the STEAD dataset while reducing computational costs by approximately 83% through its unique feed-forward, multi-resolution architecture that eliminates backpropagation.

The Gist

Imagine you are trying to find a specific needle in a massive, noisy haystack. But this isn't just any haystack; it's a haystack made of sound waves from the Earth itself, and the "needles" are the tiny, split-second moments when an earthquake starts (the P-wave) and when the stronger shaking begins (the S-wave).

For decades, scientists have tried to find these needles. First, they used human experts (who got tired and made mistakes), then they used simple math rules (which got confused by noise), and recently, they started using Deep Learning (AI).

The problem with the current AI models is that they are like giant, hungry monsters. To learn how to find the needles, they need to eat massive amounts of data, run for days on powerful computers, and consume a lot of electricity. They are also "black boxes," meaning even the scientists aren't exactly sure how they found the needle, they just know they did.

Enter GreenPhase. Think of GreenPhase not as a hungry monster, but as a smart, efficient detective.

Here is how GreenPhase solves the problem, broken down into simple concepts:

1. The "Zoom Lens" Strategy (Multi-Resolution)

Imagine you are looking at a map of a huge country to find a specific house.

• Old AI: It looks at every single tree, rock, and blade of grass on the entire map at the highest zoom level. It's thorough, but it takes forever and uses a lot of energy.
• GreenPhase: It uses a three-step zoom lens.
  • Step 1 (The Binoculars): It looks at the whole country from far away (low resolution). It doesn't look at every tree; it just spots the general neighborhood where the house might be.
  • Step 2 (The Street View): It zooms in on just that neighborhood.
  • Step 3 (The Doorstep): It zooms in one last time to look right at the front door.

By only doing the "super detailed" work on the tiny area where the earthquake actually happened, GreenPhase saves a massive amount of time and energy. It ignores the rest of the "haystack" because it already knows the needle isn't there.

2. The "No-Backtracking" Rule (Feed-Forward Design)

Traditional Deep Learning is like a student taking a test, getting a bad grade, and then having to rewind time, re-learn everything from scratch, and try again. This "rewinding" (called backpropagation) is what makes them slow and energy-hungry.

GreenPhase is different. It's like a well-organized assembly line.

• Station A: Looks at the raw sound and picks out the important patterns.
• Station B: Takes those patterns and filters out the junk.
• Station C: Makes the final decision.

Once Station A is done, it never goes back to change its mind. It passes the work to Station B, which passes it to Station C. Because the work flows in one direction without constant rewinding, it is incredibly fast and uses very little electricity. It's "Green" because it doesn't waste energy.

3. The "Math Detective" (Interpretability)

When a traditional AI says, "I found an earthquake," it's like a magician pulling a rabbit out of a hat. You know the rabbit is there, but you don't know the trick.

GreenPhase is like a detective showing you the evidence. Because it uses clear mathematical steps (like a filter bank and specific statistical tests), scientists can look at the "detective's notebook" and see exactly why the model made a decision. It's transparent and trustworthy.

The Results: Fast, Cheap, and Accurate

The paper tested GreenPhase against the best AI models currently available (like EQTransformer).

• Accuracy: GreenPhase found the earthquake waves just as well as the big, hungry AI models (almost perfect scores).
• Efficiency: It used 83% less computing power to do the same job.
• Sustainability: If you ran the old AI models on a standard computer, it would take a long time and generate a lot of carbon emissions. GreenPhase did the same job in a fraction of the time with a tiny carbon footprint.

The Bottom Line

GreenPhase is a new way to listen to the Earth. It proves that you don't need a giant, energy-guzzling AI to solve big problems. By being smart about where it looks (the zoom lens) and how it learns (the assembly line), it can detect earthquakes faster, cheaper, and more clearly than ever before. It's a "Green" revolution for seismology.

Technical Summary

Here is a detailed technical summary of the paper "GreenPhase: A Green Learning Approach for Earthquake Phase Picking."

1. Problem Statement

Earthquake detection and seismic phase picking (identifying P-wave and S-wave arrival times) are critical for seismology but face significant challenges:

• Data Complexity: Seismic signals often suffer from low signal-to-noise ratios (SNR), waveform variability, and overlapping events.
• Limitations of Traditional Methods: Early automated tools (e.g., STA/LTA, AIC) are highly dependent on signal quality and lack adaptability across different geological regions.
• Drawbacks of Deep Learning (DL): While recent DL models (e.g., EQTransformer, PhaseNet) achieve high accuracy, they rely on computationally expensive backpropagation. This results in:
  • High energy consumption and carbon footprints (sustainability concerns).
  • "Black-box" nature, lacking mathematical interpretability.
  • Heavy reliance on massive datasets and long training times.

2. Methodology: GreenPhase

The authors propose GreenPhase, a multi-resolution, feed-forward model based on the Green Learning (GL) framework. Unlike DL, GL eliminates backpropagation, using a structured, mathematically interpretable pipeline.

A. Core Architecture

The model operates on three resolution levels (Level 3: Coarse $\to$ Level 1: Fine) using a coarse-to-fine strategy:

1. Multi-Resolution Design:

   • Level 3 (Coarse): Analyzes a downsampled sequence (375 samples) to identify candidate Regions of Interest (ROIs).
   • Level 2 (Intermediate): Refines predictions within the ROIs identified by Level 3 (750 samples).
   • Level 1 (Fine): Performs final precise picking within narrow windows (1500 samples).
   • Benefit: Computation is restricted to candidate regions, avoiding full-resolution scanning of the entire waveform.

2. Three-Stage Module (Per Resolution Level):

   • Unsupervised Representation Learning: Uses the Saab (Subspace Approximation via Adjusted Bias) transform. This is a PCA-like variation that extracts spatial-spectral features (DC and AC components) without labels. It also includes a custom Window Energy Feature to detect abrupt energy changes.
   • Supervised Feature Learning:
     • Relevant Feature Test (RFT): Selects the most discriminative features by minimizing weighted Mean Squared Error (MSE) across feature bins.
     • Statistics-based Feature Generation (SFG): Generates new, higher-order features from selected subsets using Least-Square Normal Transform (LNT) to enhance discriminative power.
   • Supervised Decision Learning: Uses an XGBoost regressor/classifier to predict arrival probabilities or classify events. This module is trained independently without backpropagating gradients to previous stages.

B. Data Processing

• Dataset: Stanford Earthquake Dataset (STEAD).
• Preprocessing: Band-pass filtering (1–45 Hz), absolute value transformation, and min-max normalization to $[0, 1]$.
• Label Generation: Continuous pseudo-labels are generated using a symmetric window around ground-truth arrivals. Labels peak at 1.0 at the arrival time and decay smoothly, providing a regression target.
• Sampling: A stratified subsampling strategy balances the dataset between noise ($y=0$), intermediate, and high-confidence ($y \ge 0.8$) samples to prevent bias.

C. Detection Workflow

1. Phase Picking: P-wave and S-wave arrival times are predicted hierarchically. S-wave picking is masked by the predicted P-wave time to ensure physical consistency.
2. Event Detection: A lightweight XGBoost classifier determines if the waveform is an earthquake or noise, using features derived from the predicted arrival times and probabilities of both P and S waves.

3. Key Contributions

• Backpropagation-Free Training: By replacing end-to-end gradient descent with a feed-forward, module-wise optimization, GreenPhase drastically reduces training time and energy consumption.
• Interpretability: The model is mathematically transparent. Each stage (Saab, RFT, SFG, XGBoost) has a clear statistical purpose, unlike the opaque layers of deep neural networks.
• Computational Efficiency: The coarse-to-fine ROI selection reduces the search space significantly, leading to a massive reduction in FLOPs (Floating Point Operations) during inference.
• Data Efficiency: The model achieves high performance with significantly fewer training samples compared to DL counterparts.

4. Results

Evaluated on the STEAD dataset (1.2M training samples, 120k test samples):

• Performance Metrics:

  • Detection: F1 Score of 1.0 (Perfect).
  • P-wave Picking: F1 Score of 0.98 (Precision: 0.96, Recall: 0.99).
  • S-wave Picking: F1 Score of 0.96 (Precision: 0.93, Recall: 0.99).
  • Comparison: These results are comparable to the state-of-the-art EQTransformer (F1 0.99 for P, 0.98 for S) and significantly outperform traditional methods (e.g., AIC, FilterPicker) and other DL models (PhaseNet, GPD).

• Efficiency & Sustainability:

  • Inference Cost: GreenPhase requires 22M FLOPs per sequence, a ~6x reduction compared to EQTransformer (129M FLOPs).
  • Training Cost: Trained on a 16-core CPU in 8 hours.
  • Carbon Footprint: Inference on 120k waveforms produced 20.59g CO2e (GreenPhase) vs. 31.11g CO2e (EQTransformer). Training produced 1.80 kg CO2e vs. 72.04 kg CO2e (EQTransformer), representing a ~40x improvement in training sustainability.
  • Scalability: The model maintains high performance (F1 > 0.95) even when trained on only 60k samples (5% of the full dataset).

5. Significance

• Sustainable Seismology: GreenPhase demonstrates that high-accuracy seismic analysis does not require massive computational resources, offering a "green" alternative for large-scale, real-time monitoring.
• Trustworthy AI: The mathematical interpretability of the GL framework allows geophysicists to understand why a phase was picked, addressing the "black-box" issue prevalent in deep learning.
• Resource-Constrained Deployment: The low FLOP count and CPU-only training capability make this model suitable for deployment on edge devices or in regions with limited computing infrastructure.
• Future Outlook: The modular nature of GreenPhase provides a foundation for extending these capabilities to other tasks like magnitude estimation and event classification, paving the way for transparent, efficient, and scalable earth science AI.