Abstract
Seismic refraction traveltime tomography is a powerful tool for estimating subsurface physical properties. However, as a non-linear inverse problem, its results are often sensitive to the choice of the initial model and survey geometry. This paper evaluates a Monte Carlo-based uncertainty analysis method that uses initial model randomization to quantify the reliability of tomographic solutions. By performing hundreds of inversions with randomized starting points, we derive statistical maps of mean velocity and standard deviation. Case studies using numerical simulations and field data from the Kitakami Lowland, Japan, demonstrate that this method provides objective reliability metrics, highlighting geological features such as fault zones and structural boundaries.
Introduction
In deep crustal exploration, integrated analysis using seismic reflection data often relies on refraction traveltime tomography to provide a broad velocity structure. A significant challenge in this workflow is the inherent uncertainty of the resulting velocity model. Geologists must distinguish between genuine physical structures and artifacts caused by numerical instability or poor ray coverage. Traditional methods often lack quantitative measures of reliability. This study explores a statistical approach to address these spatial uncertainties through initial model randomization.
Methodology: Initial Model Randomization
The core of the proposed method is a Monte Carlo uncertainty analysis. Instead of a single inversion, the workflow involves generating a large ensemble of potential solutions to assess the stability of the result.
As shown in Figure 1, the process follows these steps:
- Randomization: Hundreds of 1D initial velocity models are generated, with velocities at specific depths varied randomly within a plausible range.
- Tomographic Inversion: Each initial model undergoes a standard inversion process using the Simultaneous Iterative Reconstruction Technique (SIRT) and Linear Travel-time Interpolation (LTI).
- Statistical Synthesis: The results from all inversions are aggregated to calculate the mean velocity and the standard deviation.
This statistical approach allows for the identification of stable features (low standard deviation) versus those highly dependent on the starting model (high standard deviation).
Field Application and Results
The method was applied to a high-density seismic survey line in the Kitakami Lowland, Northeast Japan. The survey targeted the Ide fault zone, utilizing 712 shot points over a 40 km line. For this analysis, 500 randomized initial models were used to ensure statistical convergence.
The resulting average velocity model (Figure 5) reveals a detailed subsurface structure. Key findings include:
- Structural Correlation: The estimated velocity distribution aligns well with known geological structures and reflection seismic sections, confirming the validity of the averaged model.
- Uncertainty as a Diagnostic: High standard deviation values were observed near sharp velocity contrasts and fault zones. This suggests that uncertainty maps can serve as a secondary tool for geological interpretation, pinpointing areas of structural complexity.
- Convergence: Analysis showed that the statistical results (mean and standard deviation) stabilize after approximately 200 to 300 iterations, providing a balance between computational cost and reliability.
Conclusion
Initial model randomization provides a robust, quantitative framework for evaluating the reliability of seismic tomography. While computationally intensive—requiring significant CPU time—the method is highly parallelizable and offers an objective basis for geological interpretation. These high-reliability models are essential for advanced seismic processing techniques, such as Full-Waveform Inversion (FWI) and pre-stack depth migration, where an accurate starting model is paramount.
Summary updated