- The paper applies affine image stabilization and weighted subpixel cross-correlation to CCTV and handheld footage from the 2018 Hualien, 2025 Mandalay, and 2026 Kumamoto earthquakes.
- The analysis finds local surface motion lasted about 2 seconds in all three events despite magnitudes from 6.4 to 7.7, while peak velocities ranged from at least 0.84 to 2.1 m/s.
- The paper identifies multiple velocity peaks in the Kumamoto record and suggests that fault geometry, mixed slip, and near-surface deformation may shape motion histories, although tracking and calibration limits remain.
- What image-processing steps are most important for separating fault displacement from camera motion and ground shaking?
- Why might local surface-motion duration remain similar across earthquakes with substantially different magnitudes and whole-fault rupture durations?
- What geological or rupture-dynamic mechanisms could explain the three distinct velocity peaks observed in the Kumamoto footage?
- How can video-derived displacement measurements be integrated with seismic, satellite, and field-survey data to improve estimates of shallow fault slip?
- Find recent papers about video-based measurement of co-seismic fault displacement.
Overview
This paper presents a unified image-tracking analysis of CCTV and handheld video footage that captured co-seismic surface displacement during three earthquakes: the 2018 Hualien earthquake (M6.4), the 2025 Mandalay earthquake (M7.7), and the 2026 Kumamoto earthquake (M6.8). The central motivation is that video recorded at or near a surface rupture provides temporal information about fault motion that seismograms recorded several hundred meters or more from the fault cannot resolve, and that field surveys capture only the final static offset. By applying a common processing pipeline to all three events, the study extracts displacement trajectories, durations, and peak velocities of near-surface deformation, enabling direct inter-event comparison.
The three cases differ substantially in tectonic character: Hualien involved complex surface deformation near the horsetail splay of the Milun fault; Mandalay produced a clean strike-slip rupture; and Kumamoto occurred on the ~75°-dipping southern Hinagu Fault with combined normal and right-lateral strike-slip components, in a segment that had not ruptured in the 2016 sequence.
Methodology
The core challenge is that the footage contains strong ground motion, camera shake, and—in the Kumamoto case—handheld motion from a secondary recording of a playback screen rather than the original CCTV feed. The author addresses this by selecting two 50×50 pixel reference windows on the near side of the fault and, for each frame, translating, scaling, and rotating the image so that both reference windows remain fixed, following the approach of Gao et al. This affine stabilization suppresses both original camera motion and secondary recording artifacts. A third 50×50 pixel window on the far side of the fault is then tracked with subpixel precision via cross-correlation, yielding the apparent relative displacement across the fault.
Frames corrupted by motion blur produce low correlation coefficients; rather than discarding them, the analysis assigns them lower weights. Velocities are obtained not by frame-to-frame differencing, which is unstable, but by weighted local second-order polynomial fitting (a Savitzky–Golay-type filter) using the correlation coefficients as weights, with antisymmetric reflection at record boundaries. The filter half-widths are n=8 frames at 30 fps (Kumamoto, Mandalay) and n=5 at ~11 fps (Hualien). Pixel displacements and velocities are converted to physical units by matching the final tracked displacement to field-survey measurements.
A methodological caveat is that only in-plane x and y displacements are measured; these cannot be mapped uniquely onto fault-parallel and vertical components, so velocity calibration relies on the assumption that the two-dimensional resultant in the image corresponds to the surveyed resultant displacement.
Results by earthquake
The 2018 Hualien footage, despite a low effective frame rate of ~11 fps, shows an approximately linear, broadly unimodal trajectory completed in roughly 2 s. A slight pause shortly after motion onset is comparable to expected tracking scatter and is not interpreted as a resolved feature. Importantly, the tracked motion—closing of the gap between a building and the camera-side structure—is interpreted as reflecting ground compression (the site recorded 0.80 m of compression and 0.27 m of uplift) rather than fault slip alone.
The 2025 Mandalay results are consistent with prior analyses of the same footage. The trajectory is smooth after suppression of the strong ground motion that preceded slip, with a slight curvature near the end as the y-velocity decreased, consistent with earlier reports. Notably, despite this being by far the largest event, the slip duration at the recording site was approximately 2 s—the same as the smaller events.
The 2026 Kumamoto footage shows an approximately linear trajectory, consistent with a preliminary report by Ando, but with a distinct overshoot near the end of the horizontal component. The displacement amplitudes in the two image components are comparable, consistent with surveyed offsets of 1.05 m horizontal and 0.90 m vertical. The most striking result is the velocity history: even after ground-motion effects are minimized, it contains at least three local peaks (at approximately 2.0, 2.7, and 3.3 s). The author argues this complexity is genuine rather than a tracking artifact, because the fluctuating velocity persists independently of the intervals of low correlation associated with motion blur.
Duration and velocity comparison
The headline quantitative finding is that surface motion lasted approximately 2 s in all three cases despite magnitudes spanning 6.4 to 7.7, while USGS whole-fault rupture durations differ by more than an order of magnitude (8 s for Hualien, ~16 s for Kumamoto, 85 s for Mandalay). The paper concludes that there is no evident scaling between local surface-motion duration and earthquake magnitude, while acknowledging that whole-fault duration and local motion duration are not directly equivalent quantities.
Peak velocities, calibrated against field-survey displacements, are summarized below:
| Earthquake |
Surveyed displacement |
Tracked displacement (px) |
Peak velocity |
| 2018 Hualien (M6.4) |
≥ 0.84 m (incl. 0.80 m compression) |
42.1 |
≥ 0.84 m/s |
| 2025 Mandalay (M7.7) |
≥ 1.93 m (strike-slip) |
35.5 |
≥ 2.1 m/s |
| 2026 Kumamoto (M6.8) |
≥ 1.38 m (resultant) |
26.2 |
≥ 1.3 m/s |
The Mandalay and Kumamoto estimates fall within the 1–8 m/s range of peak slip velocities inferred from seismic and in-situ elastic wave observations. The author notes this correspondence is nontrivial: the seismological estimates represent slip-weighted maxima associated with a seismologically resolvable centroid, not surface displacement, so the similarity in order of magnitude warrants further investigation rather than being taken as a validation.
For Mandalay, the ~2 m of slip visible in the video contrasts with ~4 m of displacement discontinuity inferred from satellite imagery. The paper proposes that satellite-resolvable deformation is distributed across a finite-width zone surrounding the principal slip plane; if roughly 2 m of the 4 m total is accommodated on the principal slip plane, the video constrains how shallow deformation is partitioned between the two. This is presented as a hypothesis, not an established result.
Interpretation of the Kumamoto complexity
The fluctuating Kumamoto velocity history is the paper's most consequential observation. It suggests that compliant near-surface materials do not necessarily undergo purely dissipative, stable sliding—a claim that runs against a common assumption in treating shallow fault zones. The author attributes the complexity to event-specific structural factors: the steeply (~75°) dipping Hinagu Fault accommodating both strike-slip and normal components, and rupture propagating in the depth direction as well as laterally, unlike the simple horizontal pulse inferred for Mandalay. Since the three events involved demonstrably different surface-displacement mechanisms, the paper suggests—but does not demonstrate—that mechanism differences control the complexity of displacement time histories.
Limitations and open questions
Several limitations are conceded explicitly. The Hualien measurement records ground compression rather than fault slip, so its velocity is only an order-of-magnitude proxy and may underestimate slip velocity. The velocity calibration assumes image-plane resultants correspond to surveyed displacement resultants, an assumption that cannot be verified from two-dimensional tracking alone. The Kumamoto footage is a secondary handheld recording, and although the affine stabilization is designed to suppress this, residual artifacts cannot be fully excluded. The Mandalay vertical displacement is unavailable from the field survey, so its velocity estimate rests on photographic evidence that the vertical component is small. Whether the three-peak Kumamoto velocity history reflects genuine fault motion, near-surface deformation partitioning, or residual processing effects remains open, as does the mechanism behind the agreement between surface velocities and seismologically inferred slip velocities.
Conclusion
By applying a single correlation-based tracking and stabilization methodology to three videos of co-seismic surface displacement, this paper establishes that local surface-motion duration was approximately 2 s across events differing by more than a unit of magnitude, that peak surface velocities were on the order of 1 m/s, and that the complexity of the displacement time history varies with the faulting mechanism—most notably in the multi-peaked Kumamoto record. The results position video analysis as a complementary observational constraint on shallow fault behavior, while leaving open the quantitative relationship between surface displacement, principal-slip-plane motion, and distributed near-surface deformation.
