Caught on CCTV: The First Video of a Fault Tearing the Ground Apart

Twenty Metres East of the Fault

A camera is watching a gate.

Bolted to a post at a solar-panel compound in central Myanmar. Pointed southwest across a dirt yard, at a fence and a footpath. For months it records nothing worth keeping. Then, on 28 March 2025, it films the first earthquake fault ever caught in motion.

Time it off the video's own clock. Shaking starts at 9.5 seconds. It worsens at 12 seconds. Then, at 14.1 seconds, something separates. A feature on the far side of the yard moves against a feature on the near side. Not shaking with it. Going away, northward. It runs until 15.4 seconds [1].

1.3 seconds. The western half of the picture travels two and a half metres past the eastern half. Broken soil heaves up along the parting line. A mole track, one to two metres wide. The fault is the Sagaing, 1,400 kilometres of right-lateral boundary between the Burma plate and Sundaland, and it is letting go of a magnitude 7.7 earthquake whose hypocentre lies 120 kilometres north, near Mandalay [14, 17].

It would rupture more than 400 kilometres of fault. Over 200 of them went supershear. About 3,600 people died [15, 16]. The camera knows none of this. It sits twenty metres east of the trace. For 1.3 seconds it is the best-placed scientific instrument on Earth.

How You Turn a Security Camera Into an Instrument

We have played those thirty-one frames more times than is good for anybody. The twentieth viewing teaches you that your eye is worthless here. It cannot hand you metres. It cannot hand you seconds you can trust. A video is not data until somebody makes it data. The making takes five steps. Every one of them can go wrong.

Step one: find something to track. You cannot measure a whole frame. You measure a patch. The authors cut 25 overlapping windows of 26 × 26 pixels out of the picture, each one holding objects 70 to 80 metres from the lens, and set a standard algorithm, normalised cross correlation, hunting each patch down in the next frame. It slides the patch around and reports where the match is best. Do that 24 times a second. Out comes a displacement time series in pixels [1].

Step two: subtract the camera. Now the problem nobody outside the field thinks of. The camera shakes too. Every wobble of the post it is bolted to throws the whole scene sideways. So you track a second patch on the near side of the fault, the camera's own side. That patch rides with the camera. Subtract it. In the published record the far-side target moves +33 pixels horizontally while the near-side reference moves −1. The difference is fault slip. The common part is the tripod complaining.

Step three: find a ruler in the frame. A pixel is not a length. You need an object whose real size you know. The authors used a fence. An aerial photograph shows 24 posts along a 64-metre enclosure edge, which puts them 2.7 metres apart, and that is your scale bar, except that the fence stands 21 metres from the camera while the tracked target sits 79 metres away. A pixel therefore spans almost four times more ground at the target. Multiply by 79/21. Now you have metres at the right depth.

Step four: rotate into the fault. Pixels move in image coordinates. Faults slip in fault coordinates. Post-earthquake satellite imagery pins the strike at 353°. Project the displacement vector onto the fault-parallel direction and you have strike slip. Project it onto the perpendicular within the fault plane and you have dip slip.

Step five: differentiate, carefully. Slip you measure. Slip velocity you compute. Difference consecutive frames, divide by 1/24 of a second. The dangerous step, and the authors treat it as one. They run a 0.2-second moving average over the velocity before quoting a peak. §7 is about why that choice matters more than it looks.

Five steps. Two of them, the scale bar and the projection, are geometry problems with more than one answer. The authors say so themselves: because of "the inherent nonuniqueness in transforming pixel displacements into real-world coordinates, the exact orientation of the net slip vector cannot be precisely resolved" [1]. Remember that sentence. It comes back in §9.

The Ledger

Published numbers in the top block. Ours in the bottom one. Different kinds of claim, and the table keeps them apart. Nobody should quote our arithmetic back at us as though the camera produced it.

Everything in the lower half of that table, and everything in §4 through §7, is a student reconstruction built to sit consistent with the published measurements. It confirms nothing independently. It never touched the video.

Bench Notes: The Pulse Would Not Fit

We wanted to draw the slip pulse. Three numbers came off the paper: 2.5 metres, 1.3 seconds, 3.2 metres per second. We assumed that was plenty. Three numbers will not pin down a curve. They are exactly enough to test one. The test is a rectangle.

Draw the velocity pulse. Box it in. Height, the peak velocity. Width, the duration. The box has area \(v_{\max}T\). The area under the pulse is the total slip \(D\), since velocity integrated over time gives distance. The fraction of the box the pulse fills is then a pure, dimensionless shape number:

A flat-topped boxcar has \(k = 1\). A triangle has \(k = 0.5\). A half-sine hump has \(k = 2/\pi = 0.637\). A sharp crack-tip spike, a tall thin blade with a long low tail, runs well below 0.5. Put the published numbers in:

That sits between a triangle and a boxcar. The real pulse has a shoulder. It gets up to speed, holds near the top for a while, then lets go.

Somebody had brought the textbook function along. We tried it on the way out. Seismology's standard kinematic slip function is the regularised Yoffe function [9]. Start from Yoffe's 1951 crack-tip solution, which rises as \(1/\sqrt{t}\) and goes infinite the instant the rupture arrives [8]. Convolve it with a triangle and the peak turns finite. Two knobs: the Yoffe duration \(\tau_R\) and the smoothing width \(\tau_S\).

We swept the whole family at a fixed 1.3-second duration and printed the results. A ceiling sat in the middle of the page. At \(\tau_S/T = 0.10\) the peak velocity runs 8.66 m/s and \(k = 0.222\). At \(\tau_S/T = 0.98\) the smoothing triangle grows so wide that the pulse is essentially a triangle, and the peak still reaches 3.89 m/s at \(k = 0.494\). It never flattens past a triangle. The widest possible smoothing triangle is a triangle. So the flattest member of the family still peaks about 20% above the reported 3.2 m/s.

Call that a contradiction and you have gone too far. We want to be careful here. The claim covers three central values, each carrying an error bar. Two honest exits open up, both inside the published uncertainties:

• Keep D = 2.5 m and v_max = 3.2 m/s and the duration has to be at least 1.56 s (quoted: 1.3 ± 0.2 s);
• Keep T = 1.3 s and v_max = 3.2 m/s and the slip has to be at most 2.08 m (quoted: 2.5 ± 0.5 m).

A nice little result for a Tuesday evening. The published numbers gently prefer a flatter-topped pulse than the textbook crack function gives. Heaton's self-healing slip pulse should look exactly like that [10, 11].

So we dropped Yoffe. In came a two-parameter smoothed ramp. It rises as a power of elapsed time and falls as a power of remaining time:

Its peak sits at \(u = a/(a+b)\). Its shape factor \(k\) can be anything you like. Solve the two conditions, peak at 0.5 s and \(k = 0.601\), and out come a = 1.016 and b = 1.626. That shape hits all three published numbers exactly, by construction. Say it plainly: fitting a curve to three numbers is not a discovery. It buys you everything the three numbers imply but never state.

Those implications, from our model. Mean slip velocity, 1.92 m/s. The peak therefore runs only 1.66× the mean. A stubby pulse. Rise time from 5% to 95%: 0.89 s. Velocity full-width-at-half-maximum: 0.81 s. Of the 2.5 m, 1.09 m (43%) lands before the velocity peak and 1.41 m after. Peak slip acceleration reaches 12.7 m/s², about 1.29 g. One side of the fault gains speed on the other harder than gravity.

The Turn

Now the part that made the paper.

Ask a geologist what a strike-slip fault does. The answer comes back clean. One side goes one way, the other side goes the other, horizontally along the strike. Now ask the same geologist about the scratches on an exposed fault surface. The answer slows down and turns careful. Slickenlines are the grooves gouged into rock when one block drags across another, and very often they curve [3, 7]. They start oblique. They bend towards horizontal. Field geologists have photographed that curve for decades and argued about what it means. The argument never resolved. A slickenline is a single frozen mark with no timestamp on it. You can see that the slip direction changed. You cannot see when, or how fast, or whether the mark was cut during the earthquake at all rather than by afterslip, or by a block of unstable ground sliding downhill a week later.

The video has a timestamp on every frame.

The first metre of slip runs oblique. The far side drops relative to the near side, at a maximum rake of about 35°. The remaining metre and a half is very nearly pure strike slip. The dip-slip component grows to about 0.3 m at 0.5 s, exactly when the velocity peaks. Then it partly recovers, to about 0.2 m, as the pulse dies [1]. Draw that in the plane of the fault. The path starts tilted, sweeps round, straightens out. A curve.

Every curved slickenline anyone has ever photographed was a still frame of a motion nobody had watched. Here the motion runs.

We wanted to draw it, so we let the slip velocity vector rotate. Let \(\theta(t)\) be the instantaneous rake, with 0° pure right-lateral strike slip and positive down-to-the-west, and the path becomes the velocity magnitude pushed along a direction that keeps sliding underneath it. Our rake law carries two terms. A stretched exponential starts at 35° and drops. A bump grows, peaks and fades, pulling the rake briefly negative so the dip slip can peak and then recover, and the four free parameters behind all of it were fitted by grid search against four published targets, each weighted by the paper's own stated uncertainty.

The fit lands at

with a normalised misfit of 0.022 across all four targets. It reproduces peak dip slip, 0.278 m against a target of 0.30. Its timing, 0.499 s against 0.50. The final dip slip, 0.219 m against 0.20. A near-zero rake at arrest, −0.27°. Every one of them sits comfortably inside the uncertainties. Total rake swept: 39.0°.

One number in that fit stopped the room. Of the 39.0° of turning, 35.1° happens before the velocity peak and 0.2° after. Ninety-nine percent of the curve gets drawn in the first half-second, while the fault is still speeding up. Mean absolute rake over the first metre of slip: 19.1°. Over the remaining 1.46 m: 1.7°. The fault turns hard, then runs straight.

That timing is no artefact of our fitting. Dynamic rupture simulations have predicted it for years [4, 12]. Transient stresses inside the cohesive zone, right at the rupture front, drive the rotation. That zone passes the site during the acceleration phase, and only then. Once the front is past and the fault simply slides, nothing is left to deflect it.

How Big Is a Curve?

Curvature is a shape. Shapes are harder to measure than lengths. What follows is the arithmetic that decides whether anybody could have caught this before somebody left a camera running.

Take our fitted trajectory. Measure it two ways. Arc length, the distance the fault actually travelled along its curved path: 2.5000 m by construction, since that is the slip we fed in. Chord, the straight line from start to stop: 2.4620 m. The curve costs an extra 38.0 mm of path, about 1.5%.

Now think about a geologist with a tape measure across a fresh rupture. Not the arc. The chord. Start point and end point survive. The route between them is gone. And 38 millimetres sits a full order of magnitude inside the ±0.5 m uncertainty on the offset itself. You cannot detect this curve by measuring a distance. Which is precisely why the question stayed open for forty years.

One quantity does capture it. The sagitta, an old surveying word for how far a curve bows off its own chord at the widest point. Ours: 19.3 cm, at t = 0.446 s. Tightest radius of curvature anywhere along the path: 1.02 m. A genuinely tight turn for a block of ground weighing thousands of tonnes.

But the camera does not see centimetres. It sees pixels. The paper's own scaling, 33 pixels of tracked motion for 2.5 m of slip, gives 13.2 pixels per metre, or 7.58 cm of ground per pixel. So our 19.3 cm sagitta comes to 2.55 pixels.

Two and a half pixels. The entire discovery, in the units the instrument actually works in.

How contingent is that? We re-ran the whole path calculation for rotation amplitudes from 3° to 45°, holding slip and duration fixed. Sagitta scales almost linearly with the amount of turning:

Read the pixel column as a detection threshold. Below about 0.3 px you sit inside the noise. The curve goes invisible no matter how many frames you have. A 5° rotation bows 0.37 px, right at the edge and unpublishable. A 3° rotation bows 0.22 px. As far as this camera is concerned, it does not exist.

Honest summary of §5 and §6 together. The Sagaing Fault happened to turn through a large enough angle, in front of a camera at close enough range, for a 2.5-pixel signal to clear a 0.25-pixel floor. Shift any of that and the paper never happens.

Why Differentiating Video Is a Trap

Read this section before you ever try to get a velocity out of a video. The practical lesson of the whole exercise lives here, and it has nothing to do with earthquakes.

You measure positions. You want speeds. So you subtract consecutive positions and divide by the frame interval. Fine. Except every position carries an error. Subtract two noisy numbers and you get a number with more noise. Then you divide it by a small quantity. Give each frame's position a random error of standard deviation \(\sigma\) and the finite-difference velocity carries

Put real numbers in. Sub-pixel cross correlation is good to roughly 0.25 px. At this camera's scale, 1.89 cm. At 30 frames per second, \(\Delta t = 1/30\) s, so

13% of the peak velocity you are trying to measure, from a position error under two centimetres. Differentiation is an amplifier. Its gain is 1/Δt.

We ran 4,000 seeded Monte Carlo trials at each noise level. Practice turns out worse than the formula suggests. A second effect gets in, one the formula knows nothing about. You do not report a random velocity. You report the maximum. Taking the maximum of a noisy series is a rigged game. You select for whichever frame the noise happened to shove furthest up. Raw differencing at 0.25 px noise recovers a peak of 3.77 ± 0.21 m/s against a true 3.20, a +18% bias that neither wanders nor averages away. Push the noise to 1.00 px. The recovered peak hits 5.97 m/s. An 87% overestimate, and you would publish it.

Smoothing cures it. The cure has a price. Apply the paper's 0.2-second moving average, six frames at 30 fps, and the random error drops by a factor of 3.1. The square root of 6 predicts 2.45. The surplus comes from the smoother flattening the peak-selection effect as well. Recovered peak: 3.24 ± 0.07 m/s, an error of +1.2%. Excellent.

But run the same smoother on a perfect, noise-free curve. It still returns 3.16 instead of 3.20. A moving average cannot help flattening a real peak. Averaging does that. So smoothing trades a random error you can quantify for a systematic underestimate you are stuck with. Here, −1.3%, small only because the pulse is broad. Smooth a sharp pulse the same way and you shave off a great deal more.

One more counterintuitive result fell out, and we like it enough to state it. A faster camera does not buy you a more precise velocity. Raw differencing noise worsens at high frame rate because Δt shrinks: 0.32 m/s at 24 fps, 3.21 m/s at 240 fps. But a fixed-duration smoother averages proportionally more frames, and the two effects very nearly cancel. RMS velocity error stays around 0.10 m/s from 15 fps all the way to 240 fps. Frame rate buys time resolution. How sharply you localise the onset. How many samples you get through the curved part. It does not buy velocity precision. At the real video's 24 fps, the acceleration limb where all the curvature lives runs only 12 frames long.

Questions We Kept Asking Each Other

Steelman the Skeptics

Suppose you do not want to believe this. Being difficult is easy and useless. The useful move is to build the version of the world where every measurement in the paper is competently made and the headline conclusion is still wrong. Here is the strongest one we managed between us.

Objection one. A second team analysed the same video and got different numbers. The most concrete problem, and genuinely unresolved. Latour and colleagues worked independently from the same CCTV footage, published in Science two weeks after the paper we are analysing, and report a slip duration of 1.4 s, a cumulative slip of ≈3 m and a peak slip velocity of ≈3.5 m/s [2]. Kearse and Kaneko report 2.5 m and 3.2 m/s [1]. Those sit 20% and 9% apart, with overlapping quoted uncertainties; Latour et al. explicitly carry a ±20% scale factor. The disagreement is about converting pixels to metres. Fence at 21 m, target at 79 m. It has nothing to do with the physics. Two competent groups, one video, 20% apart. Call that the working precision of the technique today.

Objection two. The authors concede the slip direction is not uniquely resolvable. Their own sentence: the exact orientation of the net slip vector "cannot be precisely resolved" because of nonuniqueness in the pixel-to-world transformation [1]. Everything in §5 and §6 rests on decomposing the motion into strike slip and dip slip. The non-uniqueness attacks exactly that decomposition. Their defence: the curvature persists across every target they tracked, whichever projection you choose. A good defence, because a shape can survive a coordinate change that a direction cannot. Still a defence rather than a dismissal.

Objection three. n = 1, and one camera cannot see depth. No second view. No baseline to triangulate against. The dip-slip measurement carries an uncertainty of ±0.25 m on a peak value of 0.30 m, a signal barely clear of its own error bar, and we built a four-parameter model on it. It fits well, misfit 0.022. But fitting four parameters to four loosely constrained targets is not a demanding test.

Objection four. The slip-weakening distances do not agree. Inside the paper: \(d_c' = 1.2\) m estimated from the on-fault slip velocity, against \(d_c'' = 2.4\) m estimated at a strong-motion station 2.7 km away by the standard near-fault method [13]. A factor of two, on a parameter that controls fracture energy in every dynamic rupture model, and the authors blame free-surface effects contaminating the off-fault record, which is plausible and also means one of the two numbers everyone has used for twenty years is off by 2×. Latour et al., running a different framework on the same video, land near 2.67 m [2], closer to the strong-motion value than to the on-fault one.

Objection five. Ours, not the literature's. Our model imposes a rake law whose shape we chose. Nothing in §5 tests whether the fault had to turn that way. It tests whether a curve of that family can hit four published numbers, and of course one can, and if the true rake history carried two bumps, or overshot and came back, we would never know. The whole exercise also inherits the 20% scale ambiguity of objection one. Make the real slip 3 m rather than 2.5 m. Every number in our lower table moves.

Now the rebuttal, because a steelman you never answer is just hedging. Objections one and four argue about magnitudes. The paper's central claim is about a shape. Both independent analyses agree on a brief pulse of order one second with a peak near 3 m/s and an asymmetric tail. The pulse-like rupture result survives the scale factor untouched. The multiple-target consistency check meets objection two. More cameras answer objection three, which is the good kind of objection, since the world is now full of them. Objection five is why we labelled every one of our numbers and published the code.

What survives is a first observation, strong on shape and soft on scale. A normal place for a first observation to sit.

A Short History of Scratches

The idea under test here is old. It came from people on their knees in ditches.

When a fault breaks the surface, the two walls grind against each other and leave grooves, which geologists call slickenlines and which, for most of the twentieth century, served one purpose: reading off which way the fault moved. A groove is a direction. Measure its trend, write it in the notebook, move on.

Trouble started when people looked closely. The grooves were not straight. In 1998 Spudich and colleagues took striations from the 1995 Kobe earthquake, added dislocation models, and argued that the marks carried information about the stress field during rupture rather than merely afterwards [7]. A slickenline stopped being a direction indicator and became a recording. A bad recording, smudged and untimed. But a recording.

The 2016 Kaikōura earthquake in New Zealand gave the field its best specimens. More than ten metres of strike slip on the Kekerengu Fault tore open fresh rock faces where the rupture crossed bedrock canyons, exposing hundreds of square metres of surface inscribed with striae up to six metres long, all of them cut that night. Kearse and colleagues measured them. The curvature came out systematic, with a consistent sense of convexity. Then they proposed something that sounds outrageous until you see the models: the direction the rupture propagated is recoverable from the shape of the scratch [3]. Dynamic simulations showed why. Transient stresses at the rupture front push the slip vector off course, in a sense that depends on which way the front is coming from [4].

That prediction has since run backwards on faults nobody watched. In 2024 the same idea reached curved slickenlines preserved from prehistoric earthquakes on New Zealand's Alpine Fault, reading rupture directions off events centuries gone [5]. Other work mapped how the geometry of the curve depends on rupture direction and on two properties of the surface itself: how long its asperities survive, and how much it roughens as it slides [6]. A small, careful literature, built on an inference nobody could check.

Which is the whole significance of a security camera at a solar compound. For twenty-seven years the argument ran: the scratches curve, therefore the slip curved. Never circular. Never closed either. A scratch is cut by a mechanism nobody has watched, over some unknown fraction of the slip, and possibly not during the earthquake at all.

Now somebody has watched it happen, with a clock running.

How Would We Break It?

Five things we would want done, roughly in order of how much they would settle.

Test 1. Put two cameras on a fault and wait. One camera cannot resolve depth. That single weakness roots both the scale ambiguity and the projection non-uniqueness. Two cameras with overlapping views and a surveyed baseline would give true three-dimensional displacement, with no fence-post arithmetic at all, and nothing about that is exotic: a pair of cameras and a GPS survey. Faults run through inhabited ground everywhere. Find the CCTV already sitting near major traces and archive it.

Test 2. Settle the scale factor. Objection one in §9 is solvable, not philosophical. The scene still exists. A field survey of the real fence spacing and the real target distance, with the lens parameters measured instead of assumed, would collapse most of the 20% spread between the two published analyses [1, 2]. Somebody should go and measure the fence.

Test 3. Find the slickenlines at this exact site and compare. The paper's own logic demands this one. The rupture is on the ground. The mole track was described as it formed. If the curvature preserved in the physical scratches there matches the curvature in the video frame-for-frame, the forty-year inference closes in the strongest possible way. If it does not match, if the scratches run straighter or bow the other way, then slickenlines record something other than the coseismic slip path, and a chunk of paleoseismology needs revisiting [3, 5, 6].

Test 4. Check the rupture-direction prediction. The dynamic models say the convexity flips with the direction the rupture front came from [4, 12]. The Mandalay rupture passed this site heading south. The observed sense of curvature matches what those models predict for that direction. One confirmation. The prediction earns its keep only when somebody checks it at a site where the rupture ran the other way, which in a bilateral rupture like this one [16] means a camera north of the hypocentre.

Test 5. Do the population statistics. One pulse is an anecdote about pulses. Melgar and Hayes assembled slip-pulse properties from more than 150 rupture models to establish how rise time and peak velocity scale with magnitude [11]. A video-derived pulse drops a direct, model-free point into that space. Ten of them would make a distribution. Hazard models need distributions.

Call it a fault moving 2.5 metres if you like. The finding worth keeping is that a quantity we had only ever inferred, the shape of the path a point on a fault takes while it is failing, turns out to be observable by an instrument costing about as much as a pair of shoes, provided it happens to be pointed the right way.

For four decades the curve was a reasonable guess drawn from marks in rock. Now it is a measurement with a timestamp. The camera was watching a gate. The ground opened underneath it. For 1.3 seconds a piece of consumer electronics did seismology better than any instrument ever purpose-built for the job [18].

References