TL;DR
This paper introduces a novel approach to earthquake location by applying the quadratic Wasserstein metric to seismic waveform inversion. This method offers improved convexity and robustness to noise compared to traditional L2-norm based methods, leading to more accurate and efficient earthquake parameter estimation.
What Problem Does This Paper Address?
Accurately locating earthquakes is a fundamental challenge in seismology. Traditional methods often struggle with low accuracy, especially when seismic wavelengths are comparable to the scale of wave propagation. Waveform inversion methods, while more accurate, are susceptible to the "cycle-skipping problem," requiring precise initial data. This is further complicated in earthquake location by the highly singular nature of the seismic source. The authors aim to address these limitations by leveraging the properties of the Wasserstein metric, which has shown promise in overcoming cycle-skipping and providing a more global comparison of seismic signals, particularly when stations are widely distributed.
Methods
The core of this paper lies in applying the quadratic Wasserstein metric (W₂) to measure the misfit between observed and synthetic seismic waveforms for earthquake location. Key methodological contributions include:
- Derivation of a concise analytic expression for the Fréchet gradient of the W₂ misfit function. This simplifies the implementation of the adjoint method, crucial for efficient inversion.
- Reformulation of the distance metric to handle non-positive seismic signals by squaring them. This ensures positivity and allows for numerical observation of the misfit function's convexity.
- Introduction of a new control parameter to mitigate the impact of noise, which can be problematic after squaring signals.
- Application of the Levenberg-Marquardt-Fletcher (LMF) method for solving the resulting optimization problem. The LMF method is chosen for its effectiveness in least-squares problems and its ability to handle potential issues with noise.
- Development of a modified LMF algorithm specifically designed to handle noisy data, where the objective function is not strictly guaranteed to decrease at each iteration.
Datasets
The paper does not explicitly mention or introduce new datasets for training. Instead, it uses simulated seismic data generated from various velocity models (two-layer and subduction plate models) to test and validate the proposed method. These simulations are crucial for evaluating the performance of the Wasserstein metric-based inversion under different scenarios, including noise-free and noisy conditions.
Key Insights and Ideas
- Convexity is Key: The quadratic Wasserstein metric exhibits superior convexity properties for earthquake location compared to the L² metric. This means the misfit function has a single global minimum, making it easier for optimization algorithms to find the true earthquake parameters without getting stuck in local minima.
- Trace-by-Trace Comparison: For widely spaced seismic stations, a trace-by-trace comparison of waveforms using the W₂ metric is an effective strategy.
- Noise Resilience: While noise is a challenge, the W₂ metric, especially with the proposed modifications and the LMF algorithm, demonstrates a remarkable ability to handle noisy seismic data, leading to reasonably accurate location results.
- LMF Method's Efficacy: The LMF optimization algorithm, particularly its modified version for noisy data, proves to be a robust and efficient choice for solving the earthquake location problem formulated with the W₂ misfit.
Key Results
- Improved Accuracy: Numerical experiments show that the proposed method can accurately locate earthquakes, even when starting from initial estimates far from the true hypocenter.
- Efficiency: The method, aided by the derived Fréchet gradient and the LMF algorithm, requires only a few iterations to converge to the correct earthquake parameters.
- Robustness to Noise: Even with significant noise levels (up to 20% of the signal amplitude), the method yields location errors generally less than 1 km, which is considered a good result.
- Comparison with Other Metrics: The paper demonstrates that the W₂ metric (both with squared and shifted signals) generally shows better convexity and performance in optimization landscapes compared to the L² distance and the KR norm.
Limitations and Future Work
- Amplitude Determination: The paper acknowledges that due to signal normalization, the method cannot determine the amplitude of the seismogram at the source. The authors suggest that unbalanced optimal transport theory might be a way to address this.
- Noise Estimation: While the method is robust to noise, it requires estimating noise parameters (mean and variance) to optimize a parameter (λ). The authors note that statistical methods exist for this, and the estimation doesn't need to be perfectly accurate.
- Computational Cost: While efficient, the computational cost of calculating the Wasserstein metric and its gradient can still be significant, especially for large datasets or complex models.
- Applicability to Micro-earthquakes: The authors mention that the techniques developed might be applicable to micro-earthquake inversion and are actively working on these topics.
Summary updated