Applying Waveform Tomography to Refraction Seismic Data - Inversion Strategies and Resolving Power

A number of studies on waveform inversion of synthetic refraction seismic data have proven its capability to resolve complex subsurface structure at sub-wavelength scale. However, these data often have an unrealistic bandwidth, were generated with unrealistic models (e.g. lacking attenuation), assuming optimal geometry (e.g. no topography), and unrealistic signal-to-noise ratio (mostly ∞). This study investigates the practical limitations of resolution when applying frequency domain full waveform inversion to exploration scale surface refraction data, and the possibilities to push these limits through appropriate inversion strategies. Instead of computing and inverting more realistic synthetic data --- certainly a valid approach in order to bridge the gap to application --- actual field data from two different surveys were inverted, which are similar in acquisition geometry but differ strongly in terms geological heterogeneity. Exploring the parameter space of these inversions allows for evaluating the importance of the various factors that affect the results (the background model, the data properties and preconditioning, the forward modeling, the inversion strategy). The most crucial problem in this context is the variability of observed signal amplitudes: Near-surface layers often exhibit strong and strongly varying attenuation due to weathering, different levels of lithification, different porosity and the like, which may, in addition to receiver coupling, alter the signal amplitude on the order of magnitudes. When attempting to minimize an objective function that has been posed as the (squared) residual of recorded and computed seismograms --- the most common approach in waveform inversion --- the use of true amplitudes is most likely to fail in the face of non-negligible, but unknown, attenuation variations. Finding a strategy to address this problem is a requirement for a successful application. Attempting to simultaneously reconstruct an attenuation model (along with velocities) does not alleviate this problem. Using normalized amplitudes to reconstruct velocities and logarithmic amplitudes to subsequently image attenuation is a much more promising approach. As a consequence of amplitude normalization, the introduction of weighting factors becomes essential in order to mitigate the impact of noise in the data. Another problem that effectively limits the resolution for one of the surveys is the lack of a surface in the model despite strong elevation variations.