Aperture Radar — radar & optical satellite monitoring

1. Acquisition. Query a public satellite catalog for all Sentinel-2 scenes intersecting the AOI with cloud cover under 5 %, keep one per month with the lowest cloud value. Each scene is downloaded only over the AOI and reprojected to Web Mercator at 10 m / pixel.
2. Detection. Dry sand reflects more red light than the surrounding gravel pavement, so dunes appear as bright blobs. Per frame: high-percentile threshold, connected-component labelling, keep blobs of area 5 000–200 000 m² and eccentricity ≤ 0.95 — about 700-1 400 dune polygons per scene.
3. Tracking. A greedy nearest-neighbour matcher links each detection to the closest detection in the next frame (max link 60 m, area-consistency |ΔA|/max < 0.3). Tracks surviving ≥ 6 observations across ≥ 1 year are kept.
4. Velocity fit. Per-axis ordinary least-squares regression on UTM coords vs time → (vx, vy) in m / day, plus a heading and R².
5. Forecast. Each dune is projected forward at its measured velocity, with radius derived from its observed area. The collision detector samples the predicted trajectory in 0.1-year steps and reports the first time the dune body intersects any segment of the user-drawn pipeline. Dunes already touching the pipeline at t = 0 are skipped.

Straightness = ‖net displacement‖ / cumulative path length, on a 5-month smoothed trajectory. 1.0 = perfectly linear motion. Near 0 = wandering (the matcher kept hopping between neighbouring dunes). Wind-direction-agnostic — it doesn't assume the SSW wind.

R² = quality of the linear velocity fit. Low R² means residuals are large compared with the regression line.

Min observations = how many monthly samples a dune must have to count. The default (6) drops short-lived tracks.

Future prediction shifts each dune forward at its measured 2020-2026 velocity. Uncertainty grows roughly with √time, so very long horizons are illustrative only.

Raise the filters for strict / cleanest tracks; lower them to see every candidate. Predictions update instantly.

Track-level distributions on the unfiltered set of 1 274 trajectories:

• Migration rate: 25.6 m/yr median, MAD 10.3 m/yr, IQR [16.0, 36.8] m/yr.
• OLS fit quality (R²): median 0.50, MAD 0.24, IQR [0.25, 0.73]. Wide spread — many tracks are short or noisy.
• Straightness (5-month smoothed): median 0.88, MAD 0.12 — most tracks are physically monotonic after smoothing, the long tail near 0 is what the slider filters out.
• Trajectory length: median 10 monthly observations across a 1.6-year span.

Filter survival. Default thresholds (S ≥ 0.7, R² ≥ 0.4, n ≥ 6) retain ~660 tracks. Strict (S ≥ 0.9, R² ≥ 0.8) retains 169. Stricter filtering raises the median rate (the dropped tracks were disproportionately near-stationary mismatches), which is the expected behaviour of a quality filter.

Known gaps. We have not yet performed manual ground-truth digitisation in Google Earth Pro for an independent error measurement; the published 13–29 m/yr SSW range for the same region is our external sanity check rather than a per-dune validation. Per-dune velocity confidence intervals from the OLS fit (Var(β̂) = σ²(X^⊤X)⁻¹) are computable but not yet surfaced in the viewer.

1. Acquisition. Sub-metre stripmap SAR scenes are calibrated and geocoded by the provider. We unzip each one, extract the VV amplitude, and store the raw float32 data locally.
2. Tile pyramid. Each scene is converted to an 8-bit percentile-stretched PNG, then run through libvips dzsave to build a Deep-Zoom-Image pyramid: 512-px JPG tiles at quality 85, 15 zoom levels. The viewer pulls only the tiles it needs at the current zoom and viewport — bandwidth per pan/zoom action is bounded by viewport size, not by scene size.
3. Co-registration. SAR amplitudes can drift by a pixel or two between passes due to satellite-orbit variations. For each pair of consecutive scenes, normalised cross-correlation (cv2.matchTemplate with TM_CCOEFF_NORMED) on a sliding ±6 px window finds the best-fit local shift before comparing radiometry.
4. Log-ratio change. On the co-registered pair, |mean(log A) − mean(log B)| per 64-px tile. The log domain is natural for SAR: it converts multiplicative speckle into additive noise and makes the metric symmetric under bright/dark inversion. We subtract the per-pair median to remove global radiometric drift.
5. Particle flow. The dense scalar change field is promoted to a vector field by gradient + 90° rotation, then 4 000 particles are integrated through it on a canvas overlay aligned to the SAR pyramid. Particles trail along regions of high change.

Why SAR, not optical, for industrial monitoring. Radar backscatter depends on surface roughness and dielectric properties, not illumination — so a 6 a.m. winter pass and a 1 p.m. summer pass produce comparable images. Optical comparisons are confounded by sun angle, cloud cover, dust plumes, and seasonal vegetation. For a 24/7 operational site, SAR is the only modality that gives a regular baseline independent of weather and daylight.

Direct transfer to pipeline-corridor monitoring. The same processing chain — sub-metre stripmap acquisitions, NCC co-registration, log-ratio change with per-pair drift normalisation, particle-flow visualisation — applies unchanged to a Saudi pipeline right-of-way. At 3-day cadence it would surface dune-crest movement onto the alignment, third-party encroachment on the easement, new excavation activity, and surface-deformation patterns indicative of subsidence — all in conditions where optical, ground crews and aerial overflights are limited (night, dust storms, restricted-access terrain).