Coherent Source Subsampling For Ambient Noise Interferometry
Coherent Source Subsampling (CSS) is a data-driven approach that selects coherent cross-correlation windows associated with stationary‑phase source regions and averages only those windows, restoring causal–acausal symmetry in empirical Green’s functions and improving dispersion measurements when source distributions are non‑uniform.
TL;DR
- What: CSS selects and averages only the coherent windows that contribute stationary‑phase energy between station pairs.
- How: A Symmetric Variational Autoencoder (SymVAE) learns a discrete latent variable (λ) that identifies source states; windows assigned to the stationary‑phase state are averaged to form conditional empirical Green’s functions.
- Why it helps: Restoring causal–acausal symmetry improves interstation coherence and reduces bias in surface‑wave dispersion picks.
Overview
Ambient-noise based tomography assumes equipartition of ambient sources. In the real world, sources (storms, shipping, traffic) are spatially and temporally uneven. CSS provides a principled way to isolate the coherent contributions that approximate equipartitioning, enabling more reliable Green’s functions and tomographic measurements.
Method (brief)
- Preprocess: windowing (e.g., 30 min), normalization, cross-correlation, split causal/acausal branches.
- SymVAE: encoder ingests windows and outputs a categorical posterior over λ (source state), shared coherent features g, and per-window nuisance features n.
- Subsampling: assign windows to states (one‑hot selection); compute conditional average using only windows labeled stationary.
Key results (highlights)
- Empirical Green’s functions show markedly improved causal–acausal symmetry after CSS.
- Dispersion picks become more robust; tomography bias is reduced in scenarios with non‑uniform sources.
