Shadows in the Data: How Satellites Finally Caught a Slow Andean Landslide in the Act

High in the Quebrada del Humahuaca of northwest Argentina, a mountainside has been quietly sliding for years, creeping downhill at rates that would be imperceptible to anyone standing on it. Now, a team of researchers has reconstructed a decade of that motion from space, and in doing so exposed a subtle but pervasive problem: the very satellite images used to track slow-moving landslides can fabricate seasonal signals that look strikingly like the real thing. The study, published in Earth Surface Dynamics, disentangles these artificial oscillations from genuine landslide behavior and reveals when, and why, a giant slope in the Río Del Medio catchment lurches into faster motion.

Slow-moving landslides, typically creeping at one to one hundred meters per year, are among the most widespread hazards in mountainous terrain. They threaten roads and villages, redirect sediment through river networks, and can occasionally fail catastrophically. Because they are often inaccessible and cover large areas, satellites have become the monitoring tool of choice. The standard technique is deceptively simple: take two optical images acquired on different dates, cross-correlate patches of pixels, and measure how far the ground has shifted. Repeat this across dozens or hundreds of image pairs, and a redundant network of displacement measurements emerges. A mathematical procedure borrowed from radar interferometry, known as time-series inversion, then stitches these pairwise measurements into a coherent displacement history.

The trouble is that optical images are not acquired in a vacuum. As the seasons turn, the sun’s elevation and azimuth change, casting shadows that migrate across the landscape. When images from different seasons are correlated, these illumination differences introduce systematic measurement errors. Once the network is inverted, the errors manifest as a regular, year-long oscillation superimposed on the displacement record. That oscillation is a nightmare for landslide scientists, because slow-moving landslides are genuinely known to respond to seasonal climate: rainfall and snowmelt raise pore-water pressures and modulate slope velocity. An artificial wiggle produced by shadows can therefore masquerade as, or bury, the true climatic fingerprint.

Ariane Mueting, Laurane Charrier, and Bodo Bookhagen, from the University of Potsdam and the University of Grenoble Alpes, tackled this problem using three satellite systems with very different characteristics: Landsat-8 with 15-meter panchromatic imagery, Sentinel-2 at 10 meters, and the PlanetScope constellation at 3 meters. Their target was the Del Medio landslide, a large active slope on the eastern drainage divide of the Río Del Medio catchment, where peaks reach 4,500 meters above sea level. The basin is bounded by inherited normal faults later inverted under compression, and its slopes expose pervasively sheared phyllites of the Puncoviscana Formation and quartzites of the Mesón Group. A catastrophic rockfall tore through the site during the rainy season of 2009, and prior work had shown the area was still moving at two to five meters per year as recently as 2023.

The team correlated 43 Landsat-8, 41 Sentinel-2, and 87 PlanetScope scenes, generating more than two thousand image pairs, and upsampled all imagery to a common 3-meter grid. That upsampling proved to be one of the study’s key technical insights. In controlled tests, the researchers found that seasonal measurement errors grew substantially as the spatial resolution of the input imagery decreased, an effect they attribute to the subpixel refinement step of the correlation workflow. At coarser resolutions, pinpointing the match within a larger pixel is simply less precise. Resampling everything to a common grid both reduced the bias and made the three satellite systems directly comparable, though at the cost of longer processing times and larger data volumes.

Even after this correction, a seasonal oscillation with an average amplitude of roughly one meter persisted over the landslide, and reached five meters or more in heavily shadowed terrain. Crucially, the oscillation appeared across stable ground as well, proving it was an artifact rather than real motion. By binning measurement errors over stable terrain by sun-angle difference, slope, and aspect, the team showed that the bias grows with the difference in sun position between correlated images and peaks on southwest-facing slopes, exactly where shadows extend when the average sunlight arrives from the northeast. Steeper terrain, where shadowing is more pronounced, showed larger errors.

With the problem characterized, the researchers compared four mitigation strategies. Modeling and subtracting a fixed-frequency sine wave from stable terrain grouped by slope and aspect depends on an external digital elevation model, and the 30-meter Copernicus DEM proved too coarse to capture the shadows visible in 3-meter optical data. Subtracting a best-fit sine per pixel was the most effective at erasing the oscillation, but risks removing genuine seasonal responses if the landslide truly accelerates every rainy season. Restricting correlation to image pairs acquired under similar illumination, within five to six degrees of sun elevation and azimuth difference, largely eliminated the bias while preserving the raw measurements, but thinned the correlation network and reduced temporal resolution. Weighting measurements by the difference in solar incidence angle, which accounts for local topography, showed the least improvement.

Because no ground-based measurements were available to validate the corrections, the team also ran synthetic tests, artificially moving a patch of stable ground according to a known signal with a single acceleration event. In well-illuminated terrain, all methods performed well; in shadowed, complex terrain, reconstruction became far harder, and the per-pixel sine fit worked best only when the true response was irregular in timing. For their final interpretation of the Del Medio landslide, the researchers relied on the similarly illuminated pairs, which avoid any post-processing that could alter the true signal. The resulting decade-long record shows the landslide essentially dormant from 2014 to 2016, then reactivating in early 2017 and racing through roughly 20 meters of displacement by March or April 2018, an average velocity near 13 meters per year, before settling back to about 2.5 meters per year through October 2024. In total, the slope accumulated approximately 35 meters of displacement over ten years.

After bias removal, three clear acceleration phases emerged, all falling within the austral summer monsoon from December to March: the reactivation of 2016/2017, a steep acceleration in early 2018, and a smaller event in 2021/2022 confined to the upper scarp, plus two fainter hints in late 2018 and early 2020. Earthquakes were ruled out as triggers, leaving precipitation as the prime suspect. Satellite precipitation estimates record individual days exceeding 50 millimeters, including about 157 millimeters on 5 January 2018, and the January 2017 storm that coincided with reactivation also unleashed the most catastrophic debris flow in the neighboring Los Filtros basin in 40 years, with reported rainfall near 170 millimeters in two hours. The landslide did not accelerate every rainy season, suggesting that heavy rainstorms, rather than the monsoon itself, push an already fragile slope past a critical threshold, one preconditioned by inherited faults, fractured phyllites that channel rapid infiltration, and freeze-thaw cycles at high altitude. The 2009 rockfall, by removing support from the slope’s toe, may have further destabilized the upslope mass. For the growing community tracking Earth’s creeping slopes from orbit, the message is clear: before reading a landslide’s seasonal heartbeat, first silence the shadows.

Subject of Research: Correcting seasonal illumination biases in satellite-derived displacement time series of a slow-moving landslide in NW Argentina

Article Title: Challenges in reconstructing seasonally driven landslide motion from optical satellite data: insights from the Del Medio catchment, NW Argentina

Article References: Mueting, A., Charrier, L., & Bookhagen, B. (2026). Challenges in reconstructing seasonally driven landslide motion from optical satellite data: insights from the Del Medio catchment, NW Argentina. Earth Surface Dynamics, 14(5), 781-800. https://doi.org/10.5194/esurf-14-781-2026

Image Credits: AI Generated

DOI: 10.5194/esurf-14-781-2026

Keywords: slow-moving landslide, optical satellite imagery, image correlation, seasonal bias, time-series inversion, Del Medio catchment, Argentina, Andes, precipitation trigger, Landsat-8, Sentinel-2, PlanetScope