Sample dataset guide

Every tool in this workbench carries a Load demo button, and all of those buttons draw on the same three files. This page documents them completely — the scenario they represent, every field's meaning and provenance, and the expected results at each stage of the workflow — so that you can run the entire pipeline on data with a known answer before trusting it with yours. The dataset is a teaching construction: physically consistent, deliberately covering every deformation regime the tools distinguish, and absolutely not a real survey. Citing its numbers as measurements would be a misuse; using them to rehearse the workflow is exactly their purpose.

The scenario

A simulated highway embankment slope monitoring campaign over a site in the Pacific Northwest of the United States (Washington State, UTM zone 10N), observed by Sentinel-1 C-band over a six-month small-baseline window from March to September 2026. The scene contains 87 persistent scatterers distributed across four deformation regimes — stable ground, slow steady subsidence, accelerating deformation, and isolated outliers of both signs — so that the conversion, classification and time-series tools each have meaningful work to do. Points span roughly 0.14° of longitude and 0.07° of latitude, a few kilometres across, with slopes between 3° and 26° and a range of aspects.

File 1 — the InSAR point table (87 points)

sample-insar-ps-points.csv follows the CSV specification exactly. The columns: point_id (P001–P087); lon, lat (decimal degrees, WGS84); easting, northing (UTM 10N, metre-level rounding); los_vel (mm/yr, negative away from the satellite — the standard convention); incidence_angle (degrees from vertical, 36–40°, consistent with a Sentinel-1 mid-swath); azimuth (the ground-to-satellite LOS azimuth in degrees); slope, aspect (local terrain geometry in degrees); date_start, date_end (the six-month window). The velocity regimes were assigned before noise was added: roughly forty stable points near ±1.5 mm/yr, twenty-five slow subsiders between −6 and −15 mm/yr, fifteen accelerating points between −18 and −34 mm/yr, and seven outliers — four strongly positive and three strongly negative — placed to exercise the classifiers' velocity safeguards and the time-series step detection.

File 2 — the ground observation table (12 points)

sample-ground-gnss.csv simulates a GNSS campaign: twelve benchmarks selected from the PS locations, each with a vertical velocity equal to the point's true vertical rate (LOS divided by the cosine of its incidence angle) plus realistic observation noise of about 1.5 mm/yr, a survey date (2026-09-10) and an accuracy column between 1.2 and 2.5 mm/yr. The noise is what gives the validation tool something honest to measure.

File 3 — the time series (6 points × 12 epochs)

sample-timeseries.csv is a long-format table covering October 2025 to September 2026 in monthly epochs: two stable points (flat noise), two steady subsiders (linear), one accelerating point (quadratic steepening), and one outlier point whose otherwise linear record contains a mid-record discontinuity. Each mechanism fires exactly one detection in the tools, which is what makes the series a calibration target for your threshold settings.

Expected results at every stage

LOS Converter. Dividing by cos 38° amplifies magnitudes by about 27%: the slow-regime LOS velocities of −6 to −15 mm/yr convert to roughly −7.5 to −19 mm/yr vertical. The single-point form gives −12.69 mm/yr vertical for a −10 mm/yr input at 38° incidence — a number you can verify by hand, which is the point. Hazard Classifier (highway package, 100 m radius). The expected distribution is approximately seventy Stable, one Moderate, thirteen High Hazard and three Critical: the three critical points are the strong negative outliers promoted by the velocity safeguard, and the accelerating cluster supplies most of the High Hazard count. GNSS Validation (200 m radius, k = 2). Twelve pairs match; the expected statistics are a bias near +0.9 mm/yr and an RMSE near 1.9 mm/yr, with residual spread of the same order as the injected observation noise — no pair exceeds the 2σ tolerance. Time-series Calculator (3 mm/yr acceleration threshold, 15 mm step threshold). The accelerating point flags accelerating, the discontinuous point flags step with its date, and the remaining four flag linear. Error Budget (screening defaults). The total sits near 7 mm with the atmospheric term dominating the pie — the demonstration of why atmospheric correction is the usual first investment.

Running the full chain

The intended rehearsal: load the point table in the LOS Converter and convert; move to the Hazard Classifier via the Data Tray and classify; open the GNSS Validation tool, load the observation table, and run the comparison; finish with the Time-series Calculator and the Report Generator, whose statistics injection pulls the numbers you just produced. Ten minutes end to end, and every figure in your own future report will then correspond to a step you have seen once on known data.

Regeneration & provenance

The files are generated by a fixed-seed script (scripts/gen_sample_data.py in the project repository), so every download of the dataset is bit-identical and every expected number above is stable across versions. The scenario packaging differs by language — this English page describes the North American highway campaign — but the underlying data is shared across the site's language versions by design; only the narrative wrapper is localised. If you want to teach with a different scenario, change the wrapper text, not the physics.

Using the dataset in training and review

The trio works well beyond a solo rehearsal. For onboarding a junior engineer, run the chain supervisor-style: they operate the tools, you hold this page and check each stage's numbers against the expected values above — a ninety-minute exercise that teaches both the software and the habit of verifying against a reference before interpreting. For tool regressions, the expected values double as a test suite: after any change to thresholds or kernels, rerun the demo and confirm the classification distribution, validation statistics and flags still match this page; deviations are either a bug or a deliberate behaviour change that belongs in the version notes. For client education, the demo answers the question every stakeholder eventually asks — "what exactly would we get?" — with a concrete, complete example rather than a feature list: here are the points, here is the classification, here is the agreement with ground truth, here is the report. Because the scenario is synthetic, you can also show the workflow's failure modes deliberately: narrow the validation radius until pairs go unmatched, or tighten the time-series acceleration threshold until seasonality starts masquerading as acceleration. Demonstrating what the tools refuse to tell you builds more trust than any success case.

What the sample deliberately does not include

Knowing the boundary of the demo prevents two opposite mistakes: over-trusting it and over-extending it. The dataset contains no atmospheric or orbital noise — real processor exports carry error terms whose signatures (spatially correlated residuals, track-dependent biases) the demo cannot reproduce, so passing the demo proves your mechanics, not your data quality. There is no seasonal signal in the time series: the synthetic curves are built from linear rates, a deliberate acceleration ramp and one step, so seasonality-related behaviours must be rehearsed on your own data. No unwrapping artifacts are planted either — the single step in the series is a real step by construction, whereas in production data each flagged step is a genuine question of mechanism versus artifact. And the geometry is simplified: the LOS table carries a single ascending-track geometry rather than an ascending/descending pair, so cross-track separation exercises need real campaign data.

If you build your own training data, mirror the demo's disciplines rather than its files: fix a seed or otherwise pin the values so the exercise is reproducible, write the expected results down before running the tools, and make at least one point deliberately anomalous so that the flagging path gets exercised. A teaching dataset whose every point is boring teaches the least of all. Keep the synthetic files clearly labelled as such in any shared folder — the disclaimer below exists because synthetic numbers, once loose in a project drive, have a way of being rediscovered and cited as if they were measurements.

Where to go next

Start the chain at the LOS Converter, with the CSV format specification open in a second tab; the workflow page provides the map when you are ready to replace the demo files with your own campaign.