dv/v processing parameters and best practices by application
A community-literature survey of how the dv/v knobs are set, and why
What this page is A literature-derived synthesis of the processing parameters used in ambient-noise / coda seismic velocity-change (\(\delta v/v\)) monitoring — frequency band, coda (lapse-time) window, \(\delta v/v\) method, station geometry, and depth sensitivity — organized by science application (volcanoes, earthquakes/faults, landslides, groundwater, cryosphere, geothermal). It distills 103 studies into recommended value ranges and rules-of-thumb, so a new study can choose sound parameters and know which choices bias the result.
This survey backs the measurement-design scoring in codameter’s uq_measurement module: the parameters tabulated here are exactly the knobs scored there. Full citations with DOIs are on the references page; the complete per-study parameter table (CSV + Markdown, one row per paper) lives in the literature/ folder of the repository.
Citations below link to the reference list at the foot of this page (each with a clickable DOI); the full, per-study citations with DOIs are on the references page.
n/rin the per-study table means a value was not found in the source as scanned — it was never guessed.
Provenance. The four measurement fields (frequency band, coda window, estimator, uncertainty treatment) were re-checked against the full text for 82 of the 103 studies (open access, plus AGU/Wiley titles read through a UW Wiley text-and-data-mining token). Those rows are flagged full text in the table’s measurement_source column, and an n/r there means the value is genuinely not stated in the paper.
The remaining rows carry abstract scan (19) or abstract only (paywalled) (2): these are the studies published outside the Wiley token’s coverage (Elsevier, AAAS/Science, Springer, SEG/SSA, one EGU abstract), for which full text was not available here. Their fields come from the abstract, so an n/r means “not found in the abstract,” not “not reported” — the value is very likely stated in the methods and awaits a manual full-text check (institutional VPN or a supplied PDF). Treat the coda-window and frequency-band gaps on these rows as a lower bound on what the literature reports, not a finding about the field.
Cross-cutting methodology rules
These hold across all applications (sources: (Snieder et al. 2002; Sens-Schönfelder and Wegler 2006; Bensen et al. 2007; Hadziioannou et al. 2009; Hadziioannou et al. 2011; Clarke et al. 2011; Weaver et al. 2011; Obermann et al. 2013, 2016; Zhan et al. 2013; Lecocq et al. 2014; Mikesell et al. 2015; Stehly et al. 2015; Daskalakis et al. 2016; Wang et al. 2017; Obermann and Hillers 2019; Jiang and Denolle 2020; Wang and Yao 2020; Yuan et al. 2021)).
- Coda time shift equals −dv/v for a homogeneous change, and sensitivity grows with lapse time (
dt = −t·dv/v). Later coda amplifies tiny velocity changes — the physical basis for using coda over direct waves (Snieder et al. 2002; Obermann et al. 2013). - A fully reconstructed Green’s function is NOT required — temporally stable noise sources are. Partial GF retrieval is acceptable provided the noise-source distribution and correlation waveforms are stable in time (Hadziioannou et al. 2009).
- Use the canonical 4-phase preprocessing: time-domain normalization (one-bit or running-mean) + spectral whitening → correlate → stack → SNR-based QC (Bensen et al. 2007), as implemented in MSNoise (Lecocq et al. 2014) and NoisePy (Jiang and Denolle 2020).
- Stretching is more robust than MWCS/cross-spectral methods at low SNR and for uniform changes; cross-spectral methods cycle-skip at large dv/v. DTW and wavelet methods help at low SNR / large perturbations (Mikesell et al. 2015; Yuan et al. 2021; Mao et al. 2020).
- Quantify dv/v uncertainty explicitly: coherence-weighted per-window phase errors for MWCS (Clarke et al. 2011); the analytic RMS formula based on coherence loss, bandwidth, and coda-window length for stretching (Weaver et al. 2011). SNR is a usable quality proxy.
- Depth is set by frequency band and coda lapse time — not assumed. Early coda / higher frequency → shallow surface-wave sensitivity; later coda → deeper body-wave sensitivity. Select band and window for the target depth using lapse-time-dependent 3-D kernels (Obermann et al. 2013, 2016).
- Reference-stack choice and stacking-window length bias the long-term trend and seasonal amplitude. Use a long, representative reference and a consistent stacking scheme for cross-station comparability (Wang et al. 2017).
- Beware spurious dv/v from non-stationary noise. Temporal changes in the noise-source spectrum (Zhan et al. 2013) and seasonal source fluctuations (Daskalakis et al. 2016) create apparent velocity changes; verify spectral stationarity, restrict to stable bands, and normalize the CCFs.
- Boost temporal resolution via denoising/clustering, not just longer stacks: curvelet denoising (Stehly et al. 2015) and waveform clustering by noise-source condition (Hadziioannou et al. 2011) reach daily resolution from short stacks.
- Cross-check with multiple estimators and reproducible, parameter-transparent software (MSNoise, NoisePy); each method has distinct failure modes (Obermann and Hillers 2019).
Volcano
| Parameter | Typical / recommended | Notes |
|---|---|---|
| Frequency band | ~0.1–2 Hz (deep/edifice) up to 1–4 Hz, >5 Hz for coda | Multi-band stacking is the emerging standard for depth discrimination (Feng et al. 2020; Donaldson et al. 2019; Takano et al. 2017). |
| Coda / lapse window | ~5–120 s | Short (5–35 s) for near-field single-station; up to ~100–120 s for station pairs (min lag set by inter-station distance). |
| dv/v method | Stretching (default); MWCS for error bars; wavelet for environmental separation | CWI of repeating earthquakes when noise unstable during crises (Hotovec-Ellis). |
| Station config | Pairs/arrays where dense; single-station (cross-comp > autocorr) for sparse networks | Single-station survives when only one station remains during an eruption (De Plaen). |
| Depth sensitivity | Shallow, upper ~0.3–3 km; occasionally mid-crustal magma (~3–10 km) | dv/v dominated by compliant, crack-rich shallow edifice. |
Key rule: environmental (rainfall) correction is necessary — hydrological dv/v is comparable in amplitude to the pre-eruptive signal (Rivet et al. 2015).
Earthquake / Fault
| Parameter | Typical / recommended | Notes |
|---|---|---|
| Frequency band | 0.1–2 Hz (crustal); 0.06–0.9 Hz for mid/deep crust; 4–12 Hz for shallow damage | Band selects the depth probed. |
| Coda / lapse window | ~5–30 s (crustal noise); ~3 s for high-freq aftershock autocorrelations | Later coda → deeper/larger volume but lower SNR. |
| dv/v method | Stretching (tolerates coseismic waveform change); MWCS / wavelet for frequency-resolved precision | Wavelet handles cycle-skipping at large dv/v (Mao et al. 2020; Sheng et al. 2022). |
| Station config | Single-station autocorr for dense mapping; pairs/arrays for deep & slow-slip | Autocorr maximizes coverage; pairs give long-period sensitivity. |
| Depth sensitivity | Coseismic damage shallow (top ~100 m–few km); deep change needs long-period / tiltmeters | Strong-motion softening saturates near surface (Rubinstein & Beroza; Nakata; Sawazaki). |
Key rule: coseismic velocity reductions are dominantly a shallow (top ~100 m) nonlinear site effect — separate near-surface from genuine fault-zone change before interpreting (Rubinstein and Beroza 2005).
Landslide
| Parameter | Typical / recommended | Notes |
|---|---|---|
| Frequency band | ~2–20 Hz (clayey precursors cluster 4–12 Hz) | Far higher than volcano/tectonic — shallow bodies. |
| Coda / lapse window | ~0.05–2 s | Inter-sensor distances of tens of m; usable coda arrives within ~1–2 s. |
| dv/v method | Stretching dominates; SPAC / MASW-interferometry for structure | MWCS comparatively rare. |
| Station config | Short-distance pairs for a single failure plane; dense / borehole arrays for 4-D imaging & early warning | Liu 2025 (28-stn array); Whiteley 2026 (10-stn array); de Wit 2026 (borehole). |
| Depth sensitivity | Very shallow, top few m to ~40 m | Failure surfaces and pore-pressure changes are near-surface. |
Differs from volcano/tectonic: higher frequencies, shallower depth, sub-second coda, denser/closer geometry, larger dv/v (several % to ±10 %), and rainfall/freeze-thaw forcings with multi-day lags that must be removed before precursor detection.
Groundwater / Hydrology
| Parameter | Typical / recommended | Notes |
|---|---|---|
| Frequency band | ~2–4 Hz (shallow aquifer); ~0.1–2 Hz multi-band for deep/depth-resolved | Lower frequency → greater depth. |
| Coda / lapse window | ~2–8 s (single-station autocorr); ~15–100 s (coda-wave) | Low-coherence early lags discarded. |
| dv/v method | Stretching (default); MWCS (Gräfenberg); multi-band coda inversion for depth | Mao 2022/2025; Fokker 2023. |
| Station config | Single-station autocorr/cross-comp for cheap dense aquifer monitoring; arrays for depth/spatial resolution | — |
| Depth sensitivity | Upper ~50–500 m (high-freq) to ~200–700 m (low-freq multi-band) | Enables explicit shallow-vs-deep-aquifer separation. |
Separating hydrologic from thermoelastic/tidal: fit/remove a thermoelastic component (Tsai 2011 framework) and model the hydrologic part as drained-poroelastic precipitation / pore-pressure diffusion. Diagnostic = seasonal phase lag (thermoelastic lags temperature; hydrologic lags/anticorrelates with precipitation & groundwater level); multi-year trends → net groundwater storage change (Clements and Denolle 2018; Clements and Denolle 2023; Wang et al. 2017).
Cryosphere (permafrost, glaciers, freeze-thaw)
| Parameter | Typical / recommended | Notes |
|---|---|---|
| Frequency band | ~1.5–30 Hz (active layer / rock glacier 4–14 Hz); lower for ice sheets | Targets shallow few meters. |
| Coda / lapse window | ~0.3–0.8 s for shallow high-freq arrays | — |
| dv/v method | Stretching; spectral-resonance/modal as complement | Guillemot 2021. |
| Station config | Pairs; single-station autocorr at sparse permafrost/polar sites; dense arrays | Lindner 2021 (single 3-comp); Luo 2023 (polar autocorr). |
| Depth sensitivity | ~0–10 m (active layer / firn) | Large seasonal swings (e.g. +3 % to −8 %, James 2019). |
Geothermal / Reservoir / CO2 / mining
| Parameter | Typical / recommended | Notes |
|---|---|---|
| Frequency band | ~0.25–3.5 Hz (mining as low as 0.6–1.2 Hz) | Lower band for depth penetration to reservoir. |
| Coda / lapse window | ~20 s coda windows for km-scale reservoirs | Tie window to depth-sensitivity (lapse-time) analysis. |
| dv/v method | Stretching; MWCS for larger borehole arrays (Salton Sea) | — |
| Station config | Pairs for reservoir localization; dense arrays for tailings dams / mine slopes | — |
| Depth sensitivity | Hundreds of m to a few km; coda-wave kernels attribute location | Caveat: surface arrays at CO2 sites (Ketzin) sense the shallow subsurface, not the deep plume (Gassenmeier et al. 2014). |
Key rule: ambient noise can reveal aseismic reservoir response invisible to standard microseismic monitoring (Obermann et al. 2015; Hillers et al. 2015).
Survey compiled from a structured scan of the published literature (2026). Scope is dv/v passive monitoring only; fluvial seismology and landslide detection (power-spectral methods) are out of scope. See the references page for full citations and the repository literature/ folder for the machine-readable table and provenance.