SBAS time series
SBAS (Small Baseline Subset) is an InSAR technique that inverts a network of tens-to-hundreds of pairs — each with small perpendicular baseline (B⊥) and short temporal separation (ΔT) — into cumulative surface displacement (mm) and an annual velocity field (mm/yr). Where single-pair DInSAR loses mm-scale signal under atmospheric phase (APS) noise, SBAS averages APS down over time and covers distributed scatterers such as forest and flatland, making it the practical standard for urban subsidence monitoring. The two decisive gates are reference-point correction — pinning a stable site to zero to cancel the per-pair APS bias of up to ±37mm — and unwrap-error identification; trustworthy velocity only emerges after a one-year-plus stack and ERA5 atmospheric correction.
Where SBAS fits — versus DInSAR and PS-InSAR
- Single-pair DInSAR uses one pair and suits cm-scale before/after snapshots of events like earthquakes or landslides, but it cannot resolve the mm/yr signal of slow urban subsidence.
- SBAS links tens of pairs with B⊥<100m and short ΔT to build time series over distributed scatterers (including forest and flatland), reaches mm/yr after averaging, and needs at least 15-20 scenes.
- For a mixed forest/flatland/urban AOI, full-coverage SBAS is the first choice, while high-resolution point-wise displacement in dense urban areas is best handled by PS-InSAR using permanent scatterers.
| Single DInSAR | SBAS | PS-InSAR | |
|---|---|---|---|
| Pairs | 1 | Tens | Tens |
| Target | Before/after event (cm) | DS time series | Point-scatterer structures |
| Detection limit | cm level | mm/yr (averaged) | sub-mm/yr |
| Min scenes | 2 | 15-20 | 25-30 |
Pipeline and pair-network design
- SBAS builds a short-baseline pair network from a coregistered SLC stack, runs per-pair DInSAR, then inverts the time series (MintPy WLS) into an mm/yr velocity field and time series.
- Which pairs to connect dominates overall quality; the Small Baseline strategy (B⊥<100, ΔT<48; roughly 40-60 pairs for 20 scenes) is the SBAS standard, securing balance and redundancy.
- By the connectivity rule every scene must join at least one pair with no disconnected islands.
Reference-point correction — canceling ±37mm APS
- In the 9-pair POC the per-pair global APS mean varied by up to ±37mm, and this date-by-date atmospheric bias is larger than the mm-scale relative displacement signal inside the AOI.
- The fix pins the pixel of a site believed structurally stable (the 582m forest at Sirubong on Gwanggyosan) to zero in every pair, and reads all other pixels as values relative to that reference.
- A reference point is only a site believed stable, not provably the most stable one, so using an AOI spatial average is more robust than a single pixel.
- raw phase at pixel x and date t
- stable reference pixel, forced to zero in every pair
- phase relative to the reference
Confidence and traps
- The 62-day POC (9 pairs) is trustworthy only to a cumulative ±5mm trend; extrapolating 62 days to 365 produces unrealistic velocity such as +89mm/yr at Hyeongjebong, an artifact of unwrap error and APS read as a short slope.
- A single PyAPS+ERA5 correction cut POC9 std from 35.31 to 9.61 mm/yr (~3.7x), and comp0 (82 pairs, one-year-plus) reached std 2.51 mm/yr, entering the noise floor.
- MintPy coherence-weighted least squares auto-down-weights bad pairs — forcing pairs to drop leaves velocity nearly identical — so there is no need to branch debugging on suspected pair quality.
On a short stack (the 62-day POC), trusting a velocity (mm/yr) extrapolated from 62 to 365 days yields extreme values even over stable terrain — like +89mm/yr at Hyeongjebong — almost entirely caused by unwrap error, atmospheric-phase residual, and too few pairs. Trust velocity numbers only after a one-year-plus stack with ERA5 atmospheric correction and reference-point correction; before that, read only the cumulative displacement trend and do not use them for official agency reporting.