qhjqhj00 stf-extraction-eval
Learning earthquake sources using symmetric autoencoders — Bharadwaj et al. (2023) (arXiv:2304.02404, 2023)
What this evaluates
Evaluates a model's ability to recover coherent source time functions (STFs) from scattered, noisy seismic wavefields without relying on traditional deconvolution or labeled seismograms.
Datasets
- Synthetic Scattering Simulation — total 35; splits: train (30), test (5)
Metrics
maximum normalized cross-correlation (MNCC)(primary) — range: [0, 1]- Normalized cross-correlation between predicted and true STF, computed as the maximum correlation coefficient over time lags after normalizing both signals by their L2 norms.
Input / output format
Input: Pressure wavefield seismograms sampled at 100 receivers from a 2D attenuating acoustic medium containing circular scatterers.
Output: Reconstructed source time function (STF) and its envelope.
Scoring recipe
def compute_mncc(pred, gold):
pred = pred / np.linalg.norm(pred)
gold = gold / np.linalg.norm(gold)
corr = np.correlate(pred, gold, mode='full')
return np.max(corr)
Common pitfalls
- Wavefield stacking fails for real earthquakes due to polarity differences, requiring envelope stacking instead.
- High-frequency information loss due to attenuation is irreversible, limiting STF smoothness.
- MNCC is computed on envelopes for real earthquakes, not raw waveforms.
Evidence (verbatim from paper)
We use maximum normalized cross-correlation (MNCC) as a robust metric to quantify optimal alignment and similarity between two signals. The plots Figs. 8(f)-(j) show a higher MNCC between SymAE and true STFs as opposed to the wavefield stacking and true STFs.
Citation
@misc{bharadwaj2023learning,
title={Learning earthquake sources using symmetric autoencoders},
author={Bharadwaj et al. (2023)},
year={2023},
note={arXiv:2304.02404}
}
- arXiv: 2304.02404