Two honest, negative results from testing a simple idea: can free, public satellite radar (Sentinel-1 InSAR) independently measure how much a piece of safety-critical infrastructure is moving, well enough to contradict the official or self-reported record that says it's fine?
This repo is part of a deliberate practice of publishing failures, not just wins. Both case studies here were killed by the evidence. They're published anyway, with the full reasoning, the reusable code, and the lesson, because a fast, well-documented "no" is a real research output.
| Case study | Independent measurement | Official record it tried to beat | Verdict |
|---|---|---|---|
| Tailings dams | Sentinel-1 burst InSAR → MintPy SBAS, dam-wall deformation | Self-reported Global Tailings Portal / operator stability declarations | NO-GO |
| Levees | OPERA DISP-S1 deformation, levee-centerline subsidence | US National Levee Database (LSAC safety ratings) | INCONCLUSIVE |
The pattern, borrowed from a validated climate project (an independent CF4-emissions measurement that beat a self-reported inventory): find a quantity where (1) a free satellite measurement can serve as ground truth, (2) the official record uses a crude or self-reported proxy, and (3) no funded institution already publishes the cross-referenced result. Then publish the discrepancy.
Everything here is framed as measured observation vs the self-reported record, a discrepancy, never a prediction or an accusation, and validated on already-failed structures first (Brumadinho 2019; the Pajaro levee breach 2023). No live naming of operators.
Brumadinho Dam I (Vale, Brazil) failed catastrophically in January 2019, and published Sentinel-1 reanalyses documented a pre-collapse deformation precursor (~−39 mm cumulative line-of-sight at the crest, accelerating in the final ~2 months). The kill-test: can a free, simple pipeline (ASF HyP3 burst InSAR → MintPy SBAS) reproduce that precursor, the prerequisite for ever screening other dams against their self-reported stability?
It can't, at the verified dam. At the literature-correct crest, the free pipeline measures only −16 mm (no atmospheric correction) to −4 mm (with it) vs the published −39 mm, at marginal coherence (R² 0.2–0.6), with no acceleration and only ~12–24% of the steep dam face measurable. The adjacent open-pit mine measures cleanly (−90 to −102 mm, R² 0.97), so the pipeline works where coherence is good; the dam wall is simply too incoherent for cheap C-band SBAS. The published result used a heavier method (ISBAS + 20 m + dual-track) that recovers partially-coherent pixels standard SBAS discards.
A cautionary detail worth its own note: an earlier version of this analysis produced a false positive, a "−37.6 mm match" that turned out to be an edge-of-search-box pixel ~850 m off the actual dam. An adversarial methodology audit caught it before it became a published "success." See writeups/tailings.md.
The Pajaro River levee (California) breached in March 2023; the National Levee Database had rated it only "Moderate." The kill-test: does free OPERA DISP-S1 InSAR show levee-specific deformation the official rating missed?
The pipeline runs and measures real motion (~10–28 mm/yr of subsidence along the levee), but it isn't separable from the regional field. A localization check against 148 off-levee control points: levee-centerline median −16.5 mm/yr vs regional −16.3 mm/yr, excess −0.2 mm/yr, p = 0.66. The Pajaro Valley subsides regionally from groundwater pumping; the levee just moves with it, and the breach point shows no hotspot. A 30 m pixel can't isolate a narrow levee from its setting. See writeups/levees.md.
In both cases, and in three sibling projects that were also killed, the attribution/data half was fine (clean owner→structure joins, a public self-reported record to beat, a genuinely unoccupied slot). The independent-measurement half was the systematic killer: the dam wall is too incoherent, the levee signal too regional. The takeaway, now a standing rule: on any "independent measurement vs official record" idea, prove the cheap measurement can decisively reproduce a known positive before building anything else. That ordering kills the dead ends fastest.
lib/, a source-agnostic InSAR analysis core.insar_core.py(deformation metrics: LOS velocity, acceleration ratio, a regional-control Mann-Whitney separability test) works on any(time, lat, lon)millimetre LOS stack;opera.pyandhyp3.pyare the OPERA DISP-S1 and ASF HyP3 loaders that normalize to that contract.tailings/andlevees/, the per-case probes (fetch_disp.py,run_mintpy.py,analyze_dams.py,pajaro_probe.py,nld.py) plus the cloud-processing runbook (CLOUD_RUNNER.md,ec2_bootstrap.sh,mintpy_bootstrap.sh).
All inputs are free. You'll need a (free) NASA Earthdata login for Sentinel-1 / OPERA / HyP3 access (export EARTHDATA_BEARER_TOKEN). ASF Cumulus direct-S3 is us-west-2 in-region only, so the heavy pulls run on a small EC2 instance, see the CLOUD_RUNNER.md runbooks. pip install -r requirements.txt; MintPy installs via conda-forge (tailings/mintpy_bootstrap.sh). Note: pin numpy<2, MintPy's network inversion crashes under NumPy 2.x.
The case-study scripts are the originals from the research project; their internal paths assume the original layout and they expect data/credentials not included here. They're published as the record of what was actually run, not as a turnkey tool.
Code under MIT; the writeups under CC-BY-4.0. See CITATION.cff. Archived on Zenodo (DOI in CITATION.cff). Author: Zach Dissington (ORCID 0009-0007-0196-1371).
Part of a series of open, honest research kill-tests, most are negative results, published so others can build on or disprove them rather than duplicate the work:
- InSAR accountability kill-tests: free Sentinel-1 deformation vs official/self-reported records (tailings dams, levees). DOI 10.5281/zenodo.21045290
- Nightlights vs official GDP: does the autocracy over-reporting gradient reproduce on VIIRS? DOI 10.5281/zenodo.21045541
- Space-economy solo-wedge assessment: is there a foundable solo wedge in the orbital attribution layer? DOI 10.5281/zenodo.21045736
- Climate Solutions Research Kit: spatial-prior method + the validated CF4 smelter prior. DOI 10.5281/zenodo.20617485
Author: Zach Dissington (ORCID 0009-0007-0196-1371).