Satellite Radar Breakthrough Lets Ordinary PCs Track Sinking Cities in Stunning Detail

A team of Chinese researchers has unveiled a distributed parallel processing framework that allows wide-area, time-series interferometric synthetic aperture radar (InSAR) analysis to run on an ordinary personal computer, a feat that previously demanded high-performance computing clusters and weeks of dedicated processing time. The framework, described in a study published in Environmental Earth Sciences, was applied to the full extent of Jining City in Shandong Province, China, where it produced a complete and continuous map of surface deformation and exposed the dramatic imprint of coal mining on the ground below millions of people’s feet.

InSAR has long been one of the most powerful tools in the geodetic toolbox. By comparing the phase of radar signals reflected from the same patch of ground on different satellite passes, the technique can detect displacements of just a few millimeters, whether caused by earthquakes, landslides, groundwater extraction, or the collapse of abandoned mine workings. The European Space Agency’s Sentinel-1 constellation, with its free and open data policy, six-to-twelve-day revisit cycle and stable global coverage, has made the method more accessible than ever. Yet the sheer volume of data has become a bottleneck: monitoring tens of thousands of square kilometers over several years requires processing terabyte-scale stacks of imagery, and conventional serial workflows can take weeks on a single workstation while straining storage and memory far beyond what most laboratories possess.

The innovation of the new framework lies in its choice of the basic processing unit. A standard Sentinel-1 single-look complex product, containing both VV and VH polarization channels, weighs in at roughly 8 gigabytes. But each product is actually assembled from dozens of bursts, the fundamental acquisition units of the satellite’s Interferometric Wide Swath mode, divided across three sub-swaths. A single VV-polarized burst represents only about one fifty-fourth of a standard product, less than 0.15 gigabytes. Crucially, bursts do not suffer from the along-track spatial offsets that plague full SLC products between acquisition dates. By treating each burst as an independent task, the researchers turned a monolithic, serial job into thousands of small, self-contained computations that can be dispatched across multiple devices simultaneously.

The mathematics of the parallelization are straightforward but consequential. If a study area contains n bursts and each forms m interferometric pairs over time, a conventional workflow constrained to a single node can run at most n parallel tasks, while the multi-node mode of the new framework reaches m times n, with every interferometric pair co-registered, unwrapped and geocoded independently. When computing resources are plentiful, the multi-node mode maximizes throughput; when they are scarce, a single-node mode keeps computational demand in check. The team implemented the workflow primarily with the ISCE and MintPy software packages, setting a temporal baseline threshold of 36 days to avoid decorrelation over Jining’s extensively cultivated, low-coherence landscape.

Independent sub-tasks, however, produce fragments that must be stitched back together, and this is where a second key contribution comes in. Adjacent bursts overlap by roughly 1.5 kilometers and neighboring sub-swaths by about 2 kilometers, but deformation values in those overlaps often disagree because of differing reference points and the uneven propagation of phase-unwrapping errors. The researchers built a correction and mosaicking workflow that resamples each dataset onto a common grid, fits the deformation-rate differences along the long edge of each overlap with a quadratic polynomial, and removes the resulting surface as a systematic bias. A distance-weighted blending scheme then smooths residual discrepancies across the seam. The quadratic form was chosen deliberately: it captures the slowly varying nonlinear patterns typical of unwrapping errors without the overfitting risk of higher-order models, and it is applied in only one direction because the strip-like overlap zones are far too narrow for stable two-dimensional fitting.

The payoff is visible in the numbers. Before correction, deformation-rate differences between two adjacent bursts showed a root mean square error of 12.01 millimeters per year, with visible jumps of more than 10 millimeters across burst seams; after quadratic detrending and weighted blending, the RMSE fell to 6.14 millimeters per year and the discontinuities vanished. A conventional workflow also cannot recover anything at all where data are missing: when one burst over Jining had no acquisition on 13 December 2021, standard time-series InSAR would either lose that entire date or force analysts to discard every other burst from it. The burst-based method simply processes what exists, achieving full coverage of the city.

Computational benchmarks underscore the practical significance. On a modest machine with an Intel Core i5-11400 processor, 6 cores and 12 threads, and 48 gigabytes of memory, the burst-based approach processed an interferometric pair far faster than the conventional full-scene method, which also required vastly more memory. The authors project that parallel processing of twelve interferometric pairs on a single node would demand nearly 200 gigabytes of RAM conventionally, versus under 8 gigabytes with the new method. Individual burst tasks were successfully completed on machines with as little as 8 gigabytes of RAM and 256 gigabytes of storage. Just 149 decompressed Sentinel-1 scenes exceed the storage of most mainstream PCs, so the difference between feasible and impossible hinges entirely on how the data are partitioned.

To validate accuracy, the team ran a conventional full-scene analysis over most of Jining under identical parameters and compared the two velocity fields. The correlation coefficient was 0.877, with a root mean square error of 6.53 millimeters per year, and profile curves agreed well in low-to-moderate deformation ranges, with deviations appearing only where deformation gradients are steep, likely from pixel-level resampling errors during geocoding. The method does have a documented limit: in a test on the magnitude 6.8 earthquake that struck Tingri County, China, in January 2025, burst-wise phase unwrapping failed where extreme, abrupt deformation broke phase continuity within a single burst, whereas full-scene unwrapping exploited surrounding low-gradient areas as connectivity constraints. The authors caution that block-based strategies should be applied judiciously where large, high-gradient deformation is expected.

Applied to Jining, the framework delivered a striking picture of a city slowly sinking. Using Sentinel-1A imagery from 149 acquisitions between 2019 and 2023 along Track 142, the analysis classified the surface into stable, moderately and severely deformed zones, and found that Rencheng and Yanzhou Districts bear the heaviest subsidence burdens, while Yutai and Sishui Counties remain essentially stable. When deformation was cross-referenced with the boundaries of the Jining, Yanzhou and Juye coalfields, the correlation was unmistakable: moderately deformed area made up 3.456 percent of mining zones but only 0.039 percent of non-mining land, roughly an 89-fold difference, and severely deformed area was about 222 times more prevalent in mining zones. Coal extraction, not any natural process, dominates the deformation field. At the Yangying Coal Mine, which ceased operations in February 2016, the surface has since stabilized, showing how subsidence tracks the life cycle of mining.

The consequences reach into everyday infrastructure. Drawing on building footprints and road vectors, the researchers found that 98.18 percent of buildings sit in stable ground while 1.79 percent experience moderate deformation and 0.01 percent severe deformation; along roads the corresponding figures were 96.21, 3.74 and 0.05 percent, and water ponding in badly subsided mining areas means the true severe fraction is higher than measured. Time-series analysis at two representative points revealed characteristic evolution: at K1 in Liangshan County, subsidence progressed from steady settlement through rapid sinking before gradually stabilizing, while at K2 in Rencheng District, within the Yunhe Coal Mine area, an initial linear phase has given way to accelerating subsidence that was still ongoing at the end of the study period. Field surveys confirmed the human cost, documenting cracks in buildings around K1 and fractures in roads near K2. The authors argue that continuous, affordable monitoring of this kind is essential for guiding mitigation, and their framework, by collapsing the hardware barrier from supercomputer to desktop, could bring millimeter-scale deformation surveillance within reach of small research groups, local agencies and individual practitioners worldwide.

Subject of Research: A distributed parallel processing framework for Sentinel-1 wide-area time-series InSAR deformation monitoring, applied to mining subsidence in Jining City

Article Title: A distributed parallel processing framework for sentinel-1 wide-area time series InSAR: Application in Jining City

Article References: A distributed parallel processing framework for sentinel-1 wide-area time series InSAR: Application in Jining City. (n.d.). https://doi.org/10.1007/s12665-026-13147-1

Image Credits: AI Generated

DOI: 10.1007/s12665-026-13147-1

Keywords: InSAR, Sentinel-1, parallel processing, surface deformation, mining subsidence, phase unwrapping, remote sensing, geohazard monitoring, Jining City, data mosaicking, time-series analysis, infrastructure safety