Demo · Brumadinho, 2019 · hindcast, not a live feed

Reading a disaster before it happened

On 25 January 2019, the Dam I tailings dam at Corrego do Feijao in Brumadinho, Brazil collapsed without public warning, releasing about 12 million cubic metres of mine waste and killing 270 people. In the year before it fell, the dam face was already moving, and that movement was measurable from space.

Everything on this page is our own measurement: we reprocessed the raw Sentinel-1 radar over the dam ourselves, ending with the final satellite pass three days before failure. No ground sensors, no site visit, one independent read from orbit. The public radar record shows the dam moving for more than a year before it fell.

Our own processing · real data

We measured it ourselves

This is our own InSAR result, produced by the same Rheality pipeline that monitors the Three Gorges Dam today. We built 37 burst interferograms from raw Sentinel-1 radar over Corrego do Feijao, processed them through ASF HyP3, and inverted the stack into a displacement series of 20 epochs from June 2017 to 10 January 2019, the final satellite pass before the collapse.

How this pipeline works, step by step →

Median velocity over the dam area, line of sight
Lowest point of the series
Epochs, one per satellite pass
From the last measured pass to the collapse

Measured line-of-sight displacement, our pipeline

One point per satellite pass. Down means the ground is moving away from the satellite. The orange series is the median over the whole dam area; the red series is the fastest-subsiding tenth of its pixels. The dashed line is the fitted trend; the red vertical line marks the collapse. Press replay to watch the final year play out, pass by pass, from 1 February 2018 to 31 January 2019. The alert computed from this series was already live before the year began.

Dam-area median Fastest 10% of dam pixels

The final eight weeks, zoomed. Same data, December and January only. The last satellite pass lands on 22 January, three days before the collapse; the timeline runs on to 31 January.

Dam-area median Fastest 10% of dam pixels

    The published study · production-grade monitoring

    What a fully funded monitor sees

    After the disaster, a specialist team reprocessed the same free radar with a heavier method (Grebby et al. 2021): every coherent point on the dam, tracked individually. At that level the final acceleration is easy to see, and an alert fires 40 days before failure, peaking at −119 mm/yr on the lower embankment. A funded service adds tasked commercial radar (ICEYE, TerraSAR-X) that revisits daily instead of every 12 days. That live, per-point service is what Rheality is raising to build.

    −36 mm/yrBaseline face velocity, prior 12 months
    −119 mm/yrPeak velocity, final month
    40 daysAlert lead time before failure
    1,240Coherent measurement points on the dam

    Where it was moving

    Replayed from the published analysis: every dot is a radar measurement point, colour is how fast it moves. The red cluster sits low on the embankment, where the wall later failed.

    The worst point, tracked alone

    Per-point tracking over the final 13 months. Down means sinking. The shaded band is where slow creep turned into acceleration; amber is the alert, 40 days out; red is the collapse.

    How this compares with the published study

    The two reads agree on the substance: this dam was moving, fast, for the whole period. Our measurement shows the fastest tenth of the dam sinking at about 5 centimetres a year throughout 2018, roughly 8 centimetres down by the final pass, while the dam area as a whole subsides at a third of that pace. A healthy embankment does not move like that; on our alerting thresholds this signal is flagged long before the failure. What our coarse 25-metre quick-look does not resolve is the final-weeks acceleration the study found: that took per-point tracking of the single worst spot on the face, tuned after the event. Production monitoring runs that per-point tracking on every pass, in advance, that is the subscription.

    See the other demo: Three Gorges Dam, monitored live →

    What else we monitor

    Beyond this dam

    Brumadinho is a tailings dam, but the same monitoring engine watches any ground that can move: slopes and landslides, reclaimed land and runways, ports and rail, and the data centres and LNG terminals going up across Asia now. The AI build-out is putting billions of dollars of data centres onto fast-reclaimed land in Johor, Batam and Jakarta; the insurers and lenders financing them, and the contractors building on still-settling ground, need independent proof the site is stable from the reclamation phase onward.

    Your Asset · Live Archive Check

    Now point it at your own.

    You just read a dam failure measured from orbit. The same archives cover your site today: search for it or click it on the map. The check is instant. The measurement itself takes about a day.

    What this is. The coverage numbers are live queries against the Copernicus Sentinel archives (ASF for radar, Earth Search for optical); nothing is canned. The ground motion itself is measured, never simulated: the first deformation history arrives about a day after an order, built the same way as the series above.

    Demo note. The measurement charts on this page (the displacement series, the zoom, and the coherence circles) are our own processing results, built from raw Sentinel-1 radar through the same pipeline that runs our live monitoring today. The two panels under "What a fully funded monitor sees" are a stylised replay of the published post-event analysis of Dam I (Grebby et al. 2021, Communications Earth & Environment), shown to illustrate what per-point production monitoring resolves; their values come from that study, not from our pipeline.

    Disclaimer. Rheality provides decision-support monitoring data, not a safety guarantee. The duty of care for any asset remains with its owner and their engineer. See Terms & Disclaimer and How it works.

    Header image: Kīlauea InSAR interferogram, ASI/NASA/JPL-Caltech (public domain), via Wikimedia Commons.