Simultaneous Source for non-uniform data variance and missing data

Eldad Haber, Matthias ChungarXiv 2014 · OpenReview

TL;DR

This paper tackles a critical limitation in geophysical inverse problems: the breakdown of efficient "simultaneous source" methods when data has varying reliability (non-uniform variance) or is incomplete (missing data). The authors introduce and compare several novel techniques to overcome these challenges, enabling efficient and accurate parameter estimation in such scenarios, demonstrated on the Direct Current Resistivity problem.

What Problem Does This Paper Address?

Geophysical inverse problems aim to infer subsurface properties (like conductivity) from indirect measurements. A powerful technique called "simultaneous source" allows researchers to process data from multiple sources simultaneously, significantly speeding up computations for large datasets. However, this efficiency is lost when the data quality varies across measurements (non-uniform standard deviation) or when some measurements are simply missing. These real-world issues, common in fields like electromagnetics and electrical impedance tomography, render standard simultaneous source methods ineffective. This paper addresses this gap by developing and evaluating new methods to retain the computational advantages of simultaneous sources even with non-uniform and missing data.

Methods

The paper proposes several techniques to adapt the simultaneous source framework for non-uniform and missing data:

  • Data Completion: For missing data, this method treats the problem as a "missing data" problem. It uses a reduced forward model (e.g., a simpler 1D or 2D model) to "invent" plausible data for the missing entries, effectively regularizing the solution towards a simpler model.
  • Low Rank Approximation of C: When the "variance matrix" (which captures data reliability) is not of a simple form, it can be approximated by a low-rank matrix. This allows for a stochastic representation of the misfit function, enabling efficient computation similar to the original simultaneous source method, but with a computational cost dependent on the rank.
  • Stochastic Approximation of the Data Matrix: For cases where the variance matrix is large rank and data completion is not feasible, this approach approximates the misfit function by moving the expectation operator inside the norm. This leads to a different type of stochastic programming problem.
  • Working with a Subset of Sources: This method, inspired by randomized Kaczmarz methods, treats the problem as a stochastic programming problem where the expectation is taken over the index of the source. This is easy to implement but can converge slowly.

The paper then discusses solving these stochastic programming problems using the Sample Average Approximation (SAA) method, combined with an adaptive strategy for increasing the sample size and a continuation process for the regularization parameter.

Datasets

The paper uses a synthetic dataset based on the SEG salt model for the Direct Current (DC) Resistivity inversion problem. This model features a large salt body embedded in a layered media with varying conductivity. The experimental setup involves 800 dipole line sources and 900 receivers, resulting in a large number of data points (1,440,000). The authors systematically vary the properties of the variance matrix C to simulate different scenarios of non-uniformity and missing data.

Key Insights and Ideas

  • The "simultaneous source" advantage is lost but recoverable: The core insight is that while non-uniform variance and missing data break the standard simultaneous source method, new techniques can restore its efficiency.
  • Data completion as regularization: Inventing data for missing entries using a reduced model acts as a form of regularization, pulling the solution towards a simpler, plausible model.
  • Low-rank approximation is a powerful shortcut: Approximating the variance matrix with a low-rank representation offers a computationally efficient way to handle complex data reliability structures, maintaining the simultaneous source framework's benefits.
  • SAA with adaptive sample size is key: The combination of Sample Average Approximation with an adaptive strategy for increasing sample size and a continuation for the regularization parameter is crucial for efficiently solving the resulting stochastic programming problems.
  • Data redundancy is beneficial: Reducing the amount of data still yields reasonable results, suggesting that geophysical datasets often contain redundancy, which can be exploited.

Key Results

The paper presents numerical comparisons in Table 1, evaluating different methods across various scenarios of data non-uniformity and missing data.

  • Best Recovery: The nonlinear Kaczmarz iteration (subset approach) generally yields the best recovery accuracy, closely followed by the low-rank approximation.
  • Computational Cost: The low-rank approach is substantially cheaper than the subset approach.
  • Data Completion: Data completion using a reduced model also provides reasonable results.
  • Impact of Data Reduction: As the amount of data decreases, fewer iterations are needed, and the resulting models remain reasonable, indicating data redundancy.
  • Non-uniform variance matters: Assigning appropriate standard deviations to each datum leads to the best results but is also the most expensive. Conversely, assuming a constant standard deviation for all data leads to lower resolution reconstructions.

Limitations and Future Work

  • Computational Cost of Best Methods: While the subset approach offers the best recovery, it comes with the highest computational cost.
  • Data Completion Bias: The data completion method introduces "invented" data, which could potentially bias the solution if the invented data is far from the true data.
  • Convergence Speed: The subset approach, while effective, may converge slowly if sources are not combined effectively.
  • Future Work: The authors suggest that their approach is broadly applicable to other inverse problems in geophysics and medical physics, promising significant savings. The question of how to effectively collect data with less redundancy is also raised as an area for future investigation.

Summary updated

Abstract

The use of simultaneous sources in geophysical inverse problems has revolutionized the ability to deal with large scale data sets that are obtained from multiple source experiments. However, the technique breaks when the data has non-uniform standard deviation or when some data are missing. In this paper we develop, study, and compare a number of techniques that enable to utilize advantages of the simultaneous source framework for these cases. We show that the inverse problem can still be solved efficiently by using these new techniques. We demonstrate our new approaches on the Direct Current Resistivity inverse problem.
BibTeX
@misc{haber2014simultaneous,
  author = {Eldad Haber and Matthias Chung},
  title = {{Simultaneous Source for non-uniform data variance and missing data}},
  year = {2014},
  eprint = {1404.5254},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1404.5254},
}
Figure