steadfastasart/geoscience-skills/gprpy

Auditable GPR processing

GPRPy operates on arrays shaped (samples, traces). Inspect acquisition metadata before assigning physical axes: a trace index is not distance, and a radar vendor's SEG-Y interval field may need a documented time scale. Preserve unknown units explicitly. Do not obtain a site velocity from a generic material lookup and present it as measured.

Read and select operations

GPRPy 1.0.14 natively reads paired MALA .rad/.rd3, Sensors & Software .DT1/.HD, GSSI .DZT, and its documented BSQ inputs. Native .gpr files are Python pickle sessions, not a MALA exchange format; read only trusted sessions. GPRPy does not natively read or export SEG-Y.

import gprpy.gprpy as gp
profile = gp.gprpyProfile()
profile.importdata('profile.rad')  # requires matching profile.rd3
raw = profile.data.copy()
profile.dewow(window=64)  # argument is sample count, not ns

Select operations to answer the user's question. dewow(window) and agcGain(window) take sample counts; remMeanTrace(ntraces) takes traces. Background subtraction may remove real continuous reflectors. Time-power gain changes amplitude scaling and is not a calibrated attenuation correction. Keep the raw arrays and record every applied operation. Window edge behaviour and the effective interior support need particular care; see the processing reference.

setVelocity(velocity) uses m/ns and computes the zero-offset constant-velocity scenario depth = twtt * velocity / 2; it does not estimate velocity or migrate reflectors. Resolve time zero and antenna-offset assumptions first. Retain native two-way time alongside any assumed depth, and state which independent measurements support velocity and its uncertainty.

Reproducible helper

The processing helper supports tested .rad/.rd3 and .sgy/.segy routes. The latter explicitly uses segyio before real GPRPy processing. It writes profile.npz and processing.json and reads them back; NPZ loading uses allow_pickle=False. Outputs include raw amplitudes, intermediate arrays, final amplitudes, original sample/trace indices, available physical axes, source checksums, parameters and selected encoded header fields. A new output directory prevents silent replacement of a previous result.

python scripts/process_gpr.py profile.rad --metadata sampling.json \
  --output processed --dewow-samples 64

Resolve paths relative to the installed skill directory. sampling.json contains source_note, sample_interval_ns, time_origin_ns, trace_spacing_m, profile_origin_m, horizontal_crs and vertical_reference. Use JSON null for unknown time or spacing; explain provenance in source_note. Uniform physical sampling requires a finite origin and a positive interval. Irregular positions require a separately validated geometry route; do not replace them with a fabricated regular line.

Only supply --gain-power when nonnegative physical time is known. Only supply --velocity-m-ns for an explicitly labelled zero-offset constant-velocity depth scenario. Both are rejected when time calibration is unresolved. No processing step is enabled by default. The helper fails on missing/nonfinite amplitudes, unsupported formats, non-live SEG-Y trace flags, variable trace sampling or a failed requested export instead of reporting partial work as success.

Other APIs and verification boundary

For plotting, use showProfile() or printProfile(...); neither accepts the previously documented ax= export pattern. GPRPy 1.0.14 has no bandpassFilter, exportASCII, exportSEGY or gprpyCMP API. Its CMP/WARR class is gprpyCW; consult the installed source before using its stacked amplitude methods. Topographic correction uses topoCorrect(topofile, delimiter=',') after a separately chosen velocity, not a velocity= keyword. These advanced branches require their own acceptance checks.

The checked path uses fixed GPRPy source version 1.0.14, native generated MALA bytes, and one licensed observed SEG-Y line. The field line's time scaling and coordinate units remain unresolved, so only sample-index processing is claimed. Synthetic timing and impulse depth checks are separate from field accuracy. See source installation and limits for the tested commit and dependency distinction.