18.5. Correct Static Shift#

This tutorial shows a conservative static-shift correction workflow for EDI surveys. It is written for the common profile-processing case: you have already loaded and inspected a survey, you suspect that some apparent-resistivity curves are shifted vertically, and you want to estimate station-level correction factors before preparing inversion files.

Static-shift correction should be handled carefully. It changes impedance amplitudes and apparent-resistivity levels. It does not fix bad phases, frequency-dependent source effects, wrong station coordinates, incorrect component orientation, or a poorly converted EDI file. Always inspect before and after correction.

This tutorial deliberately works through three real survey types – AMT, MT, and CSAMT – because static shift does not behave the same way on all three. AMT gives a clean, worked “apply a correction” example. A long-period MT line shows large, real factors that should still not be trusted blindly because the station spacing cannot constrain them. A CSAMT line shows that per-station factors and detect_near_surface() classification can disagree station by station, so a single average factor is not automatically the right answer. Seeing all three outcomes – correct, question, and investigate further – is the point, not just the mechanics of one function call.

18.5.1. What You Will Learn#

After this tutorial you should be able to:

diagnose when static-shift correction is worth trying
estimate station-level correction factors with AMA
inspect the meaning of delta_log10_rho, fac_rho, and fac_z
apply the correction without mutating the original survey
compare before and after correction using static-shift QC plots
read the static-shift radar diagnostic and interpret its period/frequency axis correctly
export corrected EDI files and a correction manifest
choose between AMA, LOESS, reference-median, and bilateral correction styles
run static-shift correction from a pipeline configuration when needed
recognize when a survey’s own geometry or distortion mix means the correction should be treated with more caution than the AMT case

18.5.2. Prerequisites#

Run first-pass survey inspection before correcting static shift:

>>> from pycsamt.api import read_edis
>>> from pycsamt.emtools.qc import build_qc_table, station_confidence_table
>>> survey = read_edis(
...     "data/AMT/WILLY_DATA/L18PLT",
...     recursive=False,
...     strict=False,
...     progress=False,
... )
>>> sites = survey.collection
>>> qc = build_qc_table(sites, api=True)
>>> confidence = station_confidence_table(sites, method="composite", api=True)
>>> print(qc)
>>> print(confidence)

Static-shift correction is usually a second step. If the survey did not pass basic loading, station-order, frequency-coverage, and quality checks, solve those problems first. See Inspect and QC a Survey.

This tutorial uses the bundled WILLY L18PLT line so the printed output and figures are reproducible. Replace the path with your own EDI directory when you repeat the workflow on another survey.

18.5.3. When Static Shift Is a Good Candidate#

Static shift is plausible when:

one station is shifted up or down relative to neighboring stations;
the apparent-resistivity curve shape is similar to neighboring curves;
phase does not show an equivalent anomaly;
the offset is approximately frequency independent over the selected band;
the station belongs to a profile where neighboring stations can define a local spatial trend.

Static-shift correction is risky when:

the anomaly is frequency dependent;
phase changes strongly at the same station;
the station order or coordinates are wrong;
several neighboring stations are shifted together because the geology is genuinely different;
the data contain strong CSAMT source effects or near-field behavior;
the EDI file came from another software package and still needs recomputation or component normalization.

For the scientific background, read Static Shift.

18.5.4. Three Real Surveys, Three Different Pictures#

Every number and figure in this tutorial comes from one of three bundled, real field surveys:

data/AMT/WILLY_DATA/L18PLT: 28-station AMT line, 53 frequencies per station. This is the deep-dive dataset: it is used for the full estimate/inspect/apply/compare workflow below because it has a station spacing and a frequency band that make AMA’s neighbour-trend assumption reasonable.
data/MT/kap03lmt_edis: 26-station KAP03 long-period MT profile from the Southern African Magnetotelluric Experiment (SAMTEX), roughly 60 km long, 18-20 frequencies per station spanning periods from about 25 s to nearly 5 hours. It reappears later in this tutorial as a survey where AMA returns large, real factors that are still not safe to apply without independent constraints – see Condition an MT Line With Tipper and Rotation for the fuller conditioning workflow this line goes through.
data/CSAMT: 10-station Tongkeng CSAMT line, 50 m station spacing, a real grounded- dipole controlled-source survey. It reappears later as a survey where the AMA factors and the near-surface/static-shift classifier disagree from station to station, which is itself the useful CSAMT lesson – see Map Groundwater Geology From CSAMT for the fuller single-component processing workflow.

Warning

Static-shift factors are numeric output from a real estimator running on real data, not fixed constants. They depend on the installed pyCSAMT version, the exact ordering and windowing parameters, and the data itself. Regenerate the numbers on your own installation rather than trusting cached values copied from another environment or an older documentation build.

18.5.5. Create a Review Folder#

Keep factors, plots, and corrected files together.

>>> from pathlib import Path
>>> review_dir = Path("static_shift_review")
>>> plot_dir = review_dir / "plots"
>>> corrected_dir = review_dir / "corrected_edis"
>>> plot_dir.mkdir(parents=True, exist_ok=True)
>>> corrected_dir.mkdir(parents=True, exist_ok=True)

18.5.6. Estimate AMA Correction Factors#

The standard interactive correction path uses pycsamt.emtools.ss.estimate_ss_ama(). AMA means adaptive moving average: each station is compared with its neighbors in log apparent-resistivity space.

>>> from pycsamt.emtools.ss import estimate_ss_ama
>>> factors = estimate_ss_ama(
...     sites,
...     half_window=3,
...     weights="tri",
...     pband=None,
...     max_skew=6.0,
...     robust_freq="median",
...     robust_overall="median",
...     recursive=False,
...     api=True,
... )
>>> factor_df = factors.to_pandas(copy=True)
>>> factor_df.to_csv(review_dir / "static_shift_factors_ama.csv", index=False)
>>> print(factor_df.head(5).to_string(index=False))
station  delta_log10_rho  fac_rho    fac_z  n_used
18-001A         0.174137 0.669673 0.818336      28
18-002U         0.172052 0.672895 0.820302      25
18-003A        -0.195724 1.569366 1.252743      19
18-004A         0.017444 0.960630 0.980117      16
18-005U         0.072309 0.846625 0.920122      18
>>> print(factor_df["n_used"].describe().to_string())
count    28.000000
mean     14.178571
std       6.543606
min       3.000000
25%       8.750000
50%      16.000000
75%      18.250000
max      28.000000

On L18PLT, the conservative skew filter still returns a factor for every station, but the number of surviving frequency samples varies a lot from station to station. n_used drops as low as 3 out of 53 frequencies at some stations, which means the skew mask (max_skew=6.0) is rejecting most of the band there. For this line, the first-pass QC tutorial already showed high skew across the profile. A strict skew mask leaves too few samples at the weakest stations for a comfortable automatic correction, so the factors above should be treated as review evidence, not applied blindly.

The factor table contains:

station: Station identifier used to match the factor back to the EDI object.
delta_log10_rho: Estimated log10 apparent-resistivity offset before correction. Positive values mean the station is high relative to its local trend. Negative values mean it is low.
fac_rho: Apparent-resistivity scale factor. This is the multiplicative correction for apparent resistivity.
fac_z: Impedance scale factor. pyCSAMT applies this to the impedance tensor because apparent resistivity is proportional to impedance amplitude squared.
n_used: Number of frequency samples used to estimate the station factor.

18.5.7. Choose the Spatial Order#

AMA needs the profile order to make physical sense: it compares each station against its half_window nearest neighbours in that order, so a wrong order makes the algorithm compare stations that are not actually adjacent in the field. The sort_by parameter controls that order, and leaving it at its default, sort_by=None, is normally the right choice.

sort_by=None (default): Inherit whatever ordering ensure_sites() already applied while loading sites – by default, pycsamt.api.PYCSAMT_ORDERING’s mode="auto" policy, which computes a coordinate-derived profile chainage and uses it only when the station coordinates pass a conservative single-line geometry check (linearity and cross-track thresholds); otherwise it falls back to input order rather than guessing. This robust chainage handles oblique and gently curved lines correctly, unlike a raw longitude or latitude sort. This tutorial does not override sort_by for exactly this reason.
sort_by="lon" / sort_by="lat": Force a raw longitude or latitude sort, bypassing the chainage check. Only reach for these when you have specifically verified the line is nearly due east-west or north-south – on an oblique line they can silently reorder stations incorrectly, which is why they are no longer the default as of pyCSAMT 2.2.
sort_by="name": Use station name order. Useful when station names already encode line order and coordinates are missing or unreliable.

If you are unsure which order was actually used, inspect the coordinates (or call ensure_sites() yourself with an explicit order_by) before trusting a large correction factor.

18.5.8. Use a Period Band#

Sometimes only part of the spectrum is appropriate for estimating static shift. Use pband=(min_period, max_period) in seconds to restrict the factor estimate.

>>> factors_mid_band = estimate_ss_ama(
...     sites,
...     sort_by="name",
...     half_window=3,
...     weights="tri",
...     pband=(0.01, 10.0),
...     max_skew=6.0,
...     recursive=False,
...     api=True,
... )
>>> factors_mid_band.to_pandas(copy=True).to_csv(
...     review_dir / "static_shift_factors_ama_0p01s_10s.csv",
...     index=False,
... )

Choose a band where the curves appear shifted but not strongly distorted. Avoid using a band dominated by dead-band noise, transmitter source effects, or obvious phase jumps.

For the remaining figures in this tutorial, we use max_skew=None as an exploratory setting so every frequency contributes to the demonstration. In a production workflow, keep the stricter setting unless your dimensionality review justifies relaxing it.

>>> factors_full_band = estimate_ss_ama(
...     sites,
...     sort_by="name",
...     half_window=3,
...     weights="tri",
...     pband=None,
...     max_skew=None,
...     recursive=False,
...     api=True,
... )
>>> factor_df = factors_full_band.to_pandas(copy=True)
>>> print(factor_df.head(6).to_string(index=False))
station  delta_log10_rho  fac_rho    fac_z  n_used
18-001A         0.265869 0.542164 0.736318      53
18-002U         0.178793 0.662532 0.813961      53
18-003A        -0.324956 2.113273 1.453710      53
18-004A        -0.126610 1.338473 1.156924      53
18-005U         0.145856 0.714734 0.845419      53
18-006A        -0.159680 1.444375 1.201822      53
>>> print(factor_df[["fac_rho", "fac_z", "n_used"]].describe().to_string())
         fac_rho      fac_z  n_used
count  28.000000  28.000000    28.0
mean    1.430922   1.091054    53.0
std     1.434726   0.499430     0.0
min     0.108562   0.329488    53.0
25%     0.537734   0.733285    53.0
50%     1.053300   1.026299    53.0
75%     1.594623   1.259177    53.0
max     6.514305   2.552314    53.0

With every frequency contributing (max_skew=None), every station reaches n_used=53, and the fitted offsets range from fac_rho=0.109 to fac_rho=6.514 across the line.

The tutorial figures are generated by docs/scripts/generate_tutorial_static_shift.py. The first figure shows the estimated station offsets. Blue bars have positive delta_log10_rho and therefore receive a downward resistivity correction; red bars move upward.

AMA static-shift factors estimated for the L18PLT tutorial line.

18.5.9. Inspect Large Corrections#

Large static-shift factors should be reviewed before they are applied.

>>> large = factor_df[
...     (factor_df["fac_rho"] < 0.5)
...     | (factor_df["fac_rho"] > 2.0)
...     | (factor_df["n_used"] < 5)
... ]
>>> print("stations requiring manual review")
stations requiring manual review
>>> print(large[["station", "delta_log10_rho", "fac_rho", "fac_z", "n_used"]].to_string(index=False))
station  delta_log10_rho  fac_rho    fac_z  n_used
18-003A        -0.324956 2.113273 1.453710      53
18-012A        -0.307242 2.028811 1.424364      53
18-015U         0.430200 0.371364 0.609397      53
18-018A        -0.813868 6.514305 2.552314      53
18-019U        -0.701074 5.024286 2.241492      53
18-020A         0.339255 0.457873 0.676663      53
18-021B         0.552005 0.280540 0.529661      53
18-021U         0.964321 0.108562 0.329488      53
18-022U        -0.489865 3.089337 1.757651      53
18-023V         0.321711 0.476748 0.690470      53
18-024U        -0.396803 2.493463 1.579070      53
18-025A        -0.326893 2.122723 1.456957      53

12 of the 28 stations fall outside the 0.5-2.0 band or have too few usable samples. A factor outside that range is not automatically wrong, but it should make you ask why the station is so different from its neighbors. 18-021U is the most extreme case on this line (fac_rho=0.109, nearly a factor of 9), and it reappears below as the worked before/after example.

18.5.10. Apply Factors Manually#

The safest workflow is to estimate factors first, inspect them, then apply them to a copy of the site collection with pycsamt.emtools.ss.apply_ss_factors().

>>> from pycsamt.emtools.ss import apply_ss_factors
>>> corrected = apply_ss_factors(
...     sites,
...     factors_full_band,
...     key="fac_z",
...     inplace=False,
...     recursive=False,
...     verbose=1,
... )
>>> applied_delta = (-factor_df.set_index("station")["delta_log10_rho"]).round(6)
>>> print(applied_delta.head(6).to_string())
station
18-001A   -0.265869
18-002U   -0.178793
18-003A    0.324956
18-004A    0.126610
18-005U   -0.145856
18-006A    0.159680

Because apparent resistivity is proportional to |Z|^2, applying fac_z changes determinant apparent resistivity by approximately -delta_log10_rho in log10 space, confirmed directly above rather than merely asserted.

Use inplace=False during exploration. This returns a corrected copy and keeps sites unchanged for comparison. Use inplace=True only inside a controlled pipeline or after you have saved the original state.

18.5.11. Apply AMA in One Step#

For routine processing, pycsamt.emtools.ss.correct_ss_ama() combines factor estimation and application:

>>> from pycsamt.emtools.ss import correct_ss_ama
>>> corrected = correct_ss_ama(
...     sites,
...     sort_by="name",
...     half_window=3,
...     weights="tri",
...     pband=None,
...     max_skew=None,
...     robust_freq="median",
...     robust_overall="median",
...     inplace=False,
...     recursive=False,
...     verbose=1,
... )
>>> type(corrected).__name__
'Sites'

The one-step call returns a corrected Sites collection directly. The one-step function is convenient, but the two-step factor workflow is better when you are still learning the dataset.

18.5.12. Compare Before and After#

After correction, compare the original and corrected collections. A good correction should reduce station-level vertical jumps without creating frequency-dependent artifacts.

>>> import matplotlib.pyplot as plt
>>> from pycsamt.emtools.ss import (
...     plot_ss_delta_profile,
...     plot_ss_delta_psection,
...     plot_ss_station_curves,
... )
>>> ax = plot_ss_delta_profile(
...     sites,
...     corrected,
...     pband=None,
...     robust="median",
... )
>>> ax.figure.savefig(plot_dir / "delta_profile.png", dpi=200, bbox_inches="tight")
>>> plt.close(ax.figure)
>>> ax = plot_ss_delta_psection(
...     sites,
...     corrected,
...     axis_y="logperiod",
...     pband=None,
... )
>>> ax.figure.savefig(plot_dir / "delta_psection.png", dpi=200, bbox_inches="tight")
>>> plt.close(ax.figure)
>>> worst_station = str(
...     factor_df.loc[factor_df["delta_log10_rho"].abs().idxmax(), "station"]
... )
>>> worst_station
'18-021U'
>>> ax = plot_ss_station_curves(
...     sites,
...     corrected,
...     station=worst_station,
...     pband=None,
... )
>>> ax.figure.savefig(plot_dir / "station_curve_before_after.png", dpi=200)
>>> plt.close(ax.figure)

The generated determinant-resistivity comparison below is intentionally simple: the top panel reduces the correction to one median value per station, and the bottom panel shows that the applied static-shift factor is period independent.

Static-shift correction by station and by period for the L18PLT line.

Before and after apparent-resistivity curve for the largest corrected station.

The profile plot shows one correction value per station. The pseudo-section shows whether the correction is approximately uniform across period. The station-curve plot is useful for manual review of individual stations. worst_station here is 18-021U, the most extreme station already flagged above (fac_rho=0.109): its raw determinant apparent resistivity spikes to about 25,000 $\Omega\cdot$m near $10^{-3}$ s, roughly 20 times the level of the surrounding periods, and the correction pulls that spike back down to a curve shape consistent with its neighbours – exactly the frequency-independent-offset signature described in Static Shift.

18.5.13. Read the Static-Shift Radar#

The profile and pseudo-section above summarize the correction numerically. pycsamt.emtools.ss.plot_ss_radar() gives a complementary, single-station view: apparent resistivity plotted on a polar axis, where the radius is $\log_{10}\rho_a$ (or $\rho_a$ with radial="rho") and the angle encodes period or frequency rather than a compass direction. The xy and yx off-diagonal components are drawn as two separate curves that trace out roughly one revolution per decade of period.

>>> import numpy as np
>>> import matplotlib.pyplot as plt
>>> from pycsamt.emtools.ss import plot_ss_radar
>>> fig, axes = plt.subplots(1, 2, figsize=(9.6, 4.8), subplot_kw={"polar": True})
>>> _ = plot_ss_radar(sites, station=worst_station, rotate="none", ax=axes[0])
>>> _ = axes[0].set_title(f"{worst_station} (raw)")
>>> _ = plot_ss_radar(corrected, station=worst_station, rotate="none", ax=axes[1])
>>> _ = axes[1].set_title(f"{worst_station} (AMA-corrected)")
>>> r_all = np.concatenate([line.get_ydata() for ax in axes for line in ax.lines])
>>> r_all = r_all[np.isfinite(r_all)]
>>> for ax in axes:
...     _ = ax.set_ylim(r_all.min(), r_all.max())

Because the angle in this plot is period, not direction, its tick labels are period values ($10^{-3}$ s, $10^{-2}$ s, …) rather than compass degrees, and a short caption under the plot states the convention explicitly (angle = log10(period), clockwise from top). rotate="none" is used here so the xy/yx comparison is not also confounded by a phase-tensor rotation; the default rotate="pt" is usually more informative when inspecting dimensionality rather than static shift specifically. The shared set_ylim call matters: matplotlib polar axes otherwise autoscale each panel’s radius independently, which would visually hide a uniform shift by letting both panels fill their own frame the same way.

Static-shift radar for station 18-021U before and after AMA correction, with period-labeled angular ticks.

Both panels share one radial scale, so the shift is visible directly: the raw curve (left) sits at a systematically larger radius than the corrected curve (right) at almost every angle, which is exactly what a frequency-independent multiplicative offset should look like – a uniform inward pull at every period, not a change of shape. Compare this against (4) and (8) in Static Shift: the radial contraction here is the same fac_rho/fac_z relationship shown algebraically there, just viewed across the full period band at once instead of at one frequency.

18.5.14. Use One-Call QC Wrappers#

The ss_qc_* helpers estimate, apply, and plot in one call. They are useful for fast experiments. Use return_sites=True when you want the corrected collection returned with the plot.

>>> from pycsamt.emtools.ss import ss_qc_psection, ss_qc_profile
>>> ax, corrected_preview = ss_qc_profile(
...     sites,
...     method="ama",
...     return_sites=True,
...     half_window=3,
...     max_skew=6.0,
... )
>>> ax.figure.savefig(plot_dir / "ama_preview_profile.png", dpi=200)
>>> plt.close(ax.figure)
>>> ax = ss_qc_psection(
...     sites,
...     method="ama",
...     half_window=3,
...     max_skew=6.0,
...     axis_y="logperiod",
... )
>>> ax.figure.savefig(plot_dir / "ama_preview_psection.png", dpi=200)
>>> plt.close(ax.figure)

These wrappers are excellent for exploration. For a formal processing record, also save the factor table produced by estimate_ss_ama.

18.5.15. Try Alternative Estimators#

AMA is the recommended starting point, but pyCSAMT also exposes other factor estimators:

>>> from pycsamt.emtools.ss import (
...     estimate_ss_bilateral,
...     estimate_ss_loess,
...     estimate_ss_refmedian,
... )
>>> loess = estimate_ss_loess(
...     sites,
...     half_window=3,
...     poly=1,
...     it=2,
...     pband=(0.01, 10.0),
...     max_skew=6.0,
...     api=True,
... )
>>> bilateral = estimate_ss_bilateral(
...     sites,
...     half_window=4,
...     pband=(0.01, 10.0),
...     max_skew=6.0,
...     summary="median",
...     api=True,
... )
>>> refmedian = estimate_ss_refmedian(
...     sites,
...     pband=(0.01, 10.0),
...     max_skew=6.0,
...     api=True,
... )
>>> loess.to_pandas(copy=True).to_csv(review_dir / "factors_loess.csv", index=False)
>>> bilateral.to_pandas(copy=True).to_csv(
...     review_dir / "factors_bilateral.csv",
...     index=False,
... )
>>> refmedian.to_pandas(copy=True).to_csv(
...     review_dir / "factors_refmedian.csv",
...     index=False,
... )
>>> comparison = factor_df[["station", "fac_z"]].rename(columns={"fac_z": "ama"})
>>> comparison = comparison.merge(
...     loess.to_pandas(copy=True)[["station", "fac_z"]].rename(columns={"fac_z": "loess"})
... )
>>> comparison = comparison.merge(
...     refmedian.to_pandas(copy=True)[["station", "fac_z"]].rename(columns={"fac_z": "refmedian"})
... )
>>> comparison = comparison.merge(
...     bilateral.to_pandas(copy=True)[["station", "fac_z"]].rename(columns={"fac_z": "bilateral"})
... )
>>> print(comparison.head(6).to_string(index=False))
station      ama    loess  refmedian  bilateral
18-001A 0.736318 1.178057   0.766587   0.828447
18-002U 0.813961 0.894079   0.877973   0.959148
18-003A 1.453710 1.488648   1.463350   1.445438
18-004A 1.156924 1.007281   1.076534   1.064187
18-006A 1.201822 1.194864   0.978901   0.993871
18-007U 1.037921 0.577156   0.607560   0.839317
>>> len(comparison), len(factor_df)
(20, 28)

Comparison of static-shift factor estimators for the L18PLT tutorial line.

Notice 18-005U is missing from comparison.head(6) even though it was present two rows up in factor_df.head(6): the merge calls above are inner joins, so a station has to survive max_skew=6.0 masking in every one of AMA, LOESS, reference-median, and bilateral to appear at all. 18-021U – this tutorial’s own worked before/after example – is one of the eight stations dropped entirely from this particular comparison. That is the same masking behaviour already seen in the factor table above, just compounded across four estimators instead of one; always check len(comparison) against len(factor_df) before concluding that four methods “agree,” since disagreement can also look like silent absence.

Use the alternatives as checks:

estimate_ss_loess: Useful when you want a local smooth trend rather than a simple moving neighbor median.
estimate_ss_refmedian: Useful when the whole profile can be compared to a survey-wide reference.
estimate_ss_bilateral: Useful when you want spatial weighting that also respects value similarity.

If all methods point to the same shifted stations, confidence in the correction increases. If the methods disagree strongly – or drop out entirely – inspect the data before applying a large correction.

18.5.16. A Long-Period MT Line Where Large Factors Are Not Enough (KAP03)#

L18PLT has a station spacing and profile geometry where a 3-station neighbourhood is a physically reasonable proxy for “the regional trend near this station.” That assumption gets much weaker on data/MT/kap03lmt_edis: KAP03 spans roughly 60 km with only 26 stations, so each AMA neighbourhood covers many kilometres, not tens of metres.

>>> from pycsamt.api import read_edis
>>> from pycsamt.emtools.ss import estimate_ss_ama
>>> kap_survey = read_edis("data/MT/kap03lmt_edis", recursive=False, strict=True, progress=False)
>>> kap_sites = kap_survey.collection
>>> kap_factors = estimate_ss_ama(
...     kap_sites, sort_by="name", half_window=3, max_skew=None,
...     recursive=False, api=True,
... ).to_pandas(copy=True)
>>> cols = ["station", "delta_log10_rho", "fac_rho", "fac_z", "n_used"]
>>> print(kap_factors[cols].head(8).to_string(index=False))
station  delta_log10_rho   fac_rho    fac_z  n_used
kap103        -0.538979  3.459229 1.859900      20
kap106        -0.427549  2.676388 1.635967      20
kap109        -0.078186  1.197253 1.094191      16
kap112         1.096054  0.080158 0.283122      20
kap115        -1.125988 13.365590 3.655898      20
kap118         0.968727  0.107466 0.327821      20
kap121         0.418260  0.381715 0.617831      20
kap123        -0.430698  2.695864 1.641909      20
>>> print(kap_factors["fac_z"].describe().to_string())
count    26.000000
mean      2.964259
std       7.395291
min       0.255704
25%       0.637926
50%       1.031243
75%       1.941398
max      38.308807

These factors are real numbers produced by the same AMA implementation used above, and they are large: fac_z ranges from about 0.26 to over 38 across just 26 stations, with a station-to-station standard deviation larger than the mean. The radar view of one of them, kap112 (fac_z=0.283, i.e. AMA thinks this station’s impedance needs to shrink by roughly a factor of 3.5), shows a textbook-looking, roughly period-independent gap between the xy and yx curves:

>>> from pycsamt.emtools.ss import plot_ss_radar
>>> ax = plot_ss_radar(kap_sites, station="kap112", rotate="none")
>>> _ = ax.set_title("kap112 (KAP03 MT, raw)")

Static-shift radar for KAP03 station kap112 with period-labeled angular ticks spanning tens of seconds to hours.

Notice the angular ticks now span roughly $10^{1.4}$ s to $10^{3.9}$ s – three orders of magnitude further out in period than the AMT radar above, because KAP03’s frequencies extend to nearly 5 hours. Reading the same kind of plot across such a different band is exactly why the redesigned ticks show physical units instead of degrees: a fixed “0-360 degrees” label would give no hint that this station’s period range is completely different from 18-021U’s.

The shape looks like a plausible static-shift candidate on its own. It is still not something to correct automatically here, for reasons the radar cannot show: Condition an MT Line With Tipper and Rotation runs this same diagnostic as part of a fuller KAP03 conditioning workflow and explicitly rejects applying it, because the ~60 km spacing and known DC-railway/mine cultural-noise contamination around the middle of the line mean the “neighbouring stations” AMA is trending against are not a trustworthy regional reference. A large, well-behaved-looking factor is necessary but not sufficient justification for a correction – the spatial-constraint question has to be asked separately, and answered with independent evidence (remote reference processing, known geology, or a documented decision to exclude the trial) rather than with the factor table alone.

18.5.17. A CSAMT Line Where Factors And Classification Disagree (Tongkeng)#

The Tongkeng CSAMT line (Map Groundwater Geology From CSAMT) is a single-line, grounded-dipole survey with only 10 stations at 50 m spacing. Its impedance tensor is effectively single-component: the xy term carries the real signal, while yx and the diagonal terms are governed by that page’s documented rank-1 degeneracy rather than an independent measurement. AMA and detect_near_surface() still run on the xy/determinant response without any tensor inversion, so both diagnostics remain valid here.

>>> tk_survey = read_edis("data/CSAMT", recursive=False, strict=False, progress=False)
>>> tk_sites = tk_survey.collection
>>> tk_factors = estimate_ss_ama(tk_sites, recursive=False, api=True).to_pandas(copy=True)
>>> cols = ["station", "delta_log10_rho", "fac_rho", "fac_z", "n_used"]
>>> print(tk_factors[cols].to_string(index=False))
station  delta_log10_rho  fac_rho    fac_z  n_used
 csa000         0.085430 0.821429 0.906327       9
 csa050        -0.006867 1.015936 1.007937      11
 csa100         0.032394 0.928125 0.963392      13
 csa150        -0.327626 2.126307 1.458186      14
 csa200        -0.230850 1.701571 1.304443      15
 csa250        -0.048445 1.118008 1.057359      15
 csa300        -0.000545 1.001255 1.000627      11
 csa350         0.198267 0.633480 0.795915      12
 csa400         0.083238 0.825585 0.908617      12
 csa450        -0.079181 1.200000 1.095445      13

Every station gets a different, non-trivial factor here (fac_rho from 0.63 to 2.13) – AMA does not refuse to produce numbers just because the survey is short and near-field-dominated. Whether those numbers represent real galvanic static shift is a separate question, and detect_near_surface(), tuned to this survey’s own 0.125-8196.7 Hz band with f_split=32.0 (the default f_split=1.0 assumes a broadband MT-style survey and leaves most stations’ low-frequency group empty here), gives a station-by-station answer instead of one summary number:

>>> from pycsamt.emtools.ss import detect_near_surface
>>> tk_ns = detect_near_surface(tk_sites, f_split=32.0, recursive=False)
>>> ns_cols = ["station", "n_hf", "n_lf", "ns_index", "ss_delta_log10", "ns_flag", "ss_flag", "distortion_type"]
>>> print(tk_ns[ns_cols].to_string(index=False))
station  n_hf  n_lf  ns_index  ss_delta_log10  ns_flag  ss_flag distortion_type
 csa000     5     4  4.147385        0.085430     True    False    near_surface
 csa050     7     4  1.070355       -0.006867    False    False           clean
 csa100     7     6  0.314632        0.032394    False    False           clean
 csa150     8     6  5.047407       -0.327626     True     True           mixed
 csa200     9     6  7.209748       -0.230850     True     True           mixed
 csa250     9     6  1.620579       -0.048445    False    False           clean
 csa300     7     4  4.717898       -0.000545     True    False    near_surface
 csa350     6     6  0.642146        0.198267    False     True          static
 csa400     6     6  0.622055        0.083238    False    False           clean
 csa450     7     6  1.435197       -0.079181    False    False           clean

Only csa350 classifies as pure static. csa000 and csa300 have real AMA offsets but classify near_surface (frequency-dependent high/low-frequency slope mismatch dominates); csa150 and csa200 classify mixed – both effects present together. Five stations, more than half the line, come back clean despite AMA still reporting a non-unity factor for every one of them, because a small AMA offset within the classifier’s ss_threshold=0.1 tolerance is not treated as a flagged static shift. This is the practical CSAMT lesson from Static Shift’s CSAMT-specific cautions made concrete: a CSAMT line close to a grounded-dipole transmitter can show near-field and source-related, frequency-dependent effects that a single static-shift factor should not be used to “fix,” and the right response depends on which distortion type a station actually has, not on the AMA factor alone.

>>> ax = plot_ss_radar(tk_sites, station="csa000", rotate="none")
>>> _ = ax.set_title("csa000 (Tongkeng CSAMT, raw)")

Static-shift radar for Tongkeng CSAMT station csa000, showing a visible xy curve and a near-zero, degenerate yx curve.

The ρa_yx entry is present in the legend but not visible as a curve: csa000’s yx amplitude is at or near the numerical floor described above, so $\log_{10}(\text{near-zero})$ falls far below the plotted range rather than tracing a second, comparable curve. That is the same single-component degeneracy documented in Map Groundwater Geology From CSAMT, visible here on the radar rather than hidden by it – not a plotting bug, and not evidence about static shift specifically, since plot_ss_radar only ever compares xy against yx on one station, and this survey does not have two independent horizontal components to compare in the first place.

18.5.18. Export Corrected EDI Files#

Once the correction is accepted, export the corrected collection as new EDI files. Do not overwrite the raw EDI directory during tutorial or exploratory work.

>>> from pycsamt.site.export import write_sites
>>> written = write_sites(
...     corrected,
...     corrected_dir,
...     template="{station}_static_shift_corrected.edi",
...     exist_ok=True,
...     manifest_csv=review_dir / "corrected_edi_manifest.csv",
... )
>>> print(f"wrote {len(written)} corrected EDI files")

The manifest records which station was written to which file. Keep it with the factor table and QC plots.

18.5.19. Run Through a Pipeline#

For repeatable production work, put static-shift correction in a pipeline configuration. SS001 is the AMA correction step.

name: static_shift_review
input: data/AMT/WILLY_DATA/L18PLT
output_dir: results/static_shift_review

steps:
  - name: qc_before
    code: QC001
  - name: static_shift
    code: SS001
    params:
      half_window: 3
      weights: tri
      max_skew: 6.0
  - name: qc_after
    code: QC001

Inspect the registered static-shift steps from the command line:

pycsamt pipe steps --category static_shift

The pipeline registry currently includes:

SS001: AMA correction with pycsamt.emtools.ss.correct_ss_ama().
SS002: LOESS factor estimation followed by factor application.
SS003: Reference-median factor estimation followed by factor application.
SS004: Bilateral factor estimation followed by factor application.

Use the pipeline when the same correction recipe must be shared, reviewed, or re-run on several lines.

18.5.20. Checklist Before Inversion#

Before using corrected data for inversion preparation, confirm that:

the original EDI files are still available;
the factor table has been saved;
large correction factors have been manually reviewed;
phase behavior is still geologically reasonable;
correction plots show station-level offsets, not frequency-dependent distortion;
corrected EDI files were exported into a separate directory;
the inversion input preparation step points to the corrected directory, not the raw directory by accident;
if correction was rejected, deferred, or applied only to some stations (as in the MT and CSAMT cases above), the reason is written down alongside the factor table, not left implicit.

18.5.21. Troubleshooting#

All factors are exactly one: The selected band may not contain usable samples, all stations may already be aligned, or the estimator could not build a neighbor trend.
One station receives an extreme factor: Check whether it has enough frequencies, whether its coordinates are correct, and whether the station is an actual geologic outlier.
The pseudo-section correction is not period independent: The problem may not be static shift. Inspect source effects, dead-band behavior, phase jumps, and frequency-dependent noise.
Corrected curves look worse: Revisit station order, period band, half_window, and skew threshold. The default sort_by=None already applies a robust, coordinate-checked chainage order (“Choose the Spatial Order” above); do not override it with a raw sort_by="lon"/"lat" sort unless you have verified the line is genuinely close to due east-west or north-south. Try sort_by="name" only when coordinates themselves are missing or unreliable.
Correction changes interpretation too strongly: Treat the correction as provisional. Run one inversion with raw data and one with corrected data, then compare residuals and geologic plausibility.

18.5.22. See Also#

Inspect and QC a Survey: Inspect station quality before correction.
Condition an MT Line With Tipper and Rotation: Fuller KAP03 MT conditioning workflow, including the static-shift trial referenced above and why it is rejected.
Map Groundwater Geology From CSAMT: Fuller Tongkeng CSAMT processing workflow, including the single-component tensor degeneracy referenced above.
Prepare an Occam2D Inversion: Prepare corrected data for inversion.
Static Shift: Scientific background and interpretation cautions.
Pipeline Steps: Pipeline static-shift step catalogue.
EDI Recompute Workflow: Recompute and rewrite EDI files before correction when imports are inconsistent.
pycsamt.emtools: EMTools API reference.