Hazard from Strain: Predicting When Urban Ground Collapses Using Continuous Deformation Fields

Cities usually map sinkhole and subsidence risk as a static picture. Each grid cell gets a "low" to "high" tier based on things like pipe age and soil type. Nothing in that picture says when a hidden void will break through to the surface. Satellite radar (InSAR, which measures millimetre-scale ground motion from repeated satellite passes) and monitoring pins at construction sites do record motion over time. The models built on that data either classify a scene as "deforming or not" or forecast displacement at individual pins. Neither estimates a time to failure, and neither handles sites that have not failed yet.

This proposal brings over two ideas from other fields. From surgical video reconstruction we take continuous, gap-tolerant models of a surface that changes shape over time. From survival analysis we take hazard models, which estimate the instantaneous risk that an event happens now, given that it has not happened yet. The central claim is that collapse risk is driven by spatial gradients of ground motion and by its acceleration, not by the motion at isolated points. Two gradient-based quantities matter here: differential settlement (neighbouring patches of ground sinking by different amounts) and strain concentration (stretching or compression piling up in a small area). Both are local derivatives of a displacement field that no current ground-deformation model reconstructs. We also expect hazard to rise sharply once deformation thresholds are crossed. A model that assumes risk ratios stay constant over time will therefore miss the run-up to failure.

Core Claim

We fit a continuous space-time displacement field to sparse, gappy measurements. From the field we compute strain and acceleration analytically. These quantities then enter a time-dependent hazard model that does not assume constant hazard ratios. We claim this predicts time-to-collapse better than three alternatives:

  1. hazard models on static infrastructure covariates;
  2. hazard models on point-wise velocity histories;
  3. simple velocity-threshold alarms.

The new element is not any single piece. It is the specific chain from field reconstruction, to spatial derivatives, to a hazard model that allows acceleration, evaluated where the time of failure is actually observed or properly censored. Censored means a site was watched, did not fail, and dropped out of observation; the model must still learn from it.

Why This Gap Is Real

The ground-deformation side shows the gap directly. The Korean pipeline study scores 77 ground-sink events with Random Forest, Gradient Boosting and logistic regression over 100–300 m grids (Machine-learning-based ground sink susceptibility evaluation using underground pipeline data in Korean urban area). It finds pipe age above roughly 30 years to be the strongest predictor. It produces risk tiers only: it has no notion of time and no handling of censoring.

Three other lines of work also stop short of failure timing:

All four motivate the proposal: each stops before failure timing.

Mechanics gives a reason to expect gradients to matter. Excavations create an uplift zone, a settlement zone, and a transition zone with little vertical motion but severe horizontal motion (Effects of Foundation Excavation on Metro Tunnels at Different Locations and Performance of Corresponding Reinforcement Measures: A Case of Shenzhen Metro Line 11, China). Such a transition zone is invisible to a vertical-velocity summary at one point but obvious in a strain field.

The representation comes from deformable tissue reconstruction. Forplane factorizes a 4D space-time volume into three spatial planes and three space-time planes. This cuts cost from $O(N^4)$ to $O(N^2)$ and trains in about 10 minutes (Efficient Deformable Tissue Reconstruction via Orthogonal Neural Plane). Its pose-free windowed variant handles long sequences (FLex: joint pose and dynamic radiance fields optimization for stereo endoscopic videos). Deformation-driven Gaussian splatting gives an explicit alternative (Endo-4DGS: Endoscopic Monocular Scene Reconstruction with 4D Gaussian Splatting, HFGS: 4D Gaussian Splatting with Emphasis on Spatial and Temporal High-Frequency Components for Endoscopic Scene Reconstruction).

We reuse the factorized-plane idea for its compactness and its smooth, differentiable output. We drop the rendering objective and fit displacement directly. The surgical literature also supplies a constraint. Endo-4DGS reports that errors in the depth prior propagate into non-physical tearing and bulging. A strain-based hazard model would read such tearing as failure precursors. So the field must be regularized and the hazard model must be checked against artifacts.

The hazard side supplies the estimator:

Our hazard model is a time-dependent version of the extended hazard form. That is the minimal methodological step needed here.

Research Program

Field fitting. For a site, measurements $d_i$ at locations $x_i$ and times $t_i$ are fit by a factorized field $u_\theta(x,t)$. The field is a sum of products of features on spatial and space-time planes, decoded by a small network, with a smoothness penalty on its second derivatives. The measurements are InSAR line-of-sight displacements (vertical displacement when ascending and descending tracks are combined), pin readings, or camera-tracked surface points. The site's derived covariates at time $t$ are:

  • the maximum and spatial-quantile values of horizontal strain $\nabla u$;
  • differential settlement;
  • the temporal acceleration $\partial_t^2 u$ in a neighbourhood;
  • a field-uncertainty term, the ensemble spread over refits.

Hazard model. These covariates $Z(t)$, together with static covariates $X$ (pipe age, diameter, fines content, water-table height), enter a time-dependent extended hazard:

$$h(t \mid X, Z) = h_0\!\left(t\, e^{g_1(X, Z(t))}\right) e^{g_2(X, Z(t))},$$

Here $h_0$ is an unspecified baseline hazard, and $g_1$ and $g_2$ are small networks. $g_1$ stretches or compresses time; $g_2$ scales risk. Setting $g_1 = 0$ recovers a time-dependent Cox model, which serves as the proportional-hazards ablation. Estimation follows DeepEH's kernel-smoothed pseudo-likelihood, adapted to counting-process data with covariates that change over time. All derived covariates at time $t$ are computed only from measurements before $t$, so no future information leaks into the inputs.

Testbeds. The program has three settings, ordered by how cleanly failure times are observed.

The first is a controlled laboratory testbed, the anchor of the program. We replicate the sand-tank apparatus of Model test for the observation of cavity formation in sandy ground - with reference to ground water level and relative density -: a 70×25×50 cm chamber, standard sand, and a 5 mm bottom slot. We vary water-table height, relative density and fines content. We add a stereo camera on the surface and one on the transparent side wall, so the deforming surface is reconstructed with the tissue-reconstruction pipeline. Cave-in time is observed exactly, and trials stopped early are right-censored.

The second is construction monitoring on the Chengdu pit dataset. Here the event is a pin crossing its regulatory alarm level. The field for each pin is fit only from other pins and earlier times.

The third is urban sinks: Sentinel-1 persistent-scatterer time series before the documented Korean ground-sink events, with landmark analysis at several horizons.

Baselines

The naive reference has two parts. In the tank, it is the stratified Kaplan–Meier curve of time-to-collapse by water-table height and density, which reproduces the published regularities. The source paper reports collapse in about 5 minutes at a 30 cm water table for both 60% and 75% density. At 20 cm the cavity lasted 15.25 versus 9.50 minutes. A model that cannot beat a lookup on these two covariates has learned nothing from deformation. In the field settings, the naive reference is a velocity-threshold alarm, the rule most monitoring practice uses. Its threshold is tuned on the training split to the same false-alarm rate as our model.

The strongest static competitor is the Korean study's Gradient Boosting and Random Forest susceptibility models, refit on identical grid cells. We add a gradient-boosted survival model on the same pipeline covariates, so that the comparison isolates the value of time-varying deformation rather than of the survival loss. No reported result found on Lacuna for the susceptibility classifiers' accuracy that we could quote here. For the survival comparison we will report concordance index (the fraction of site pairs whose predicted risk ordering matches their actual failure ordering), integrated Brier score (squared error of predicted survival probabilities, averaged over time) and time-dependent AUC on a spatially blocked split.

The strongest dynamic competitor is tdCoxSNN-style dynamic prediction, in which a sequence encoder over point-wise velocity and displacement histories feeds a time-dependent Cox network. On construction pins we also run ICPO-BITCN-BIGRU, with its forecast trajectory converted to a predicted crossing time. tdCoxSNN reports winning on four of six accuracy measures on PBC2, but there is no reported result found on Lacuna for any of these methods on ground-failure timing. All dynamic models get the same inputs before $t$, the same split, and the same hyperparameter budget of 50 random-search trials each.

Two ablations isolate the key ingredients. The first swaps the factorized neural field for Gaussian-process (kriging) interpolation and recomputes strain from it. The second sets $g_1=0$ to impose proportional hazards. Beating the dynamic baseline means a confidence-interval-separated gain in concordance and integrated Brier score, plus longer warning lead time at a fixed false-alarm rate. We add the field-reconstruction reference points from the surgical literature (Forplane's PSNR of 37.31 and HFGS's 38.14 on EndoNeRF) only as sanity targets for the tank camera reconstruction, not as a competing claim.

Why not just…?

Why not just put pipe age and diameter into a Cox model? This is the cheapest extension of the Korean study, and it will probably beat the static classifiers at ranking cells. But static covariates fix a cell's risk ordering forever. They cannot explain why one 40-year-old pipe zone fails this year and another never does. The tank testbed sharpens the question: with water table and density held fixed, does deformation-field information still separate fast from slow collapses? If the stratified Kaplan–Meier curves explain most of the variance in collapse time, the dynamic contribution is small, and we will report that.

Why not just feed point-wise velocity histories into a time-dependent Cox network? This is the real threat. The Chengdu pin model shows that point histories are informative, and tdCoxSNN shows that dynamic risk updates help. Our argument is that failure follows gradients, such as the transition zone in the Shenzhen excavation results, and that point histories cannot see gradients. This is an empirical question. The comparison against the point-wise dynamic baseline, with matched inputs and budget, settles it. If the point-wise model ties, the strain claim is wrong for these data.

Why not just interpolate with kriging instead of a neural field? A Gaussian process may well produce strain features as good as ours at site scale, and we treat this as a live possibility, not a straw man. The factorized field is chosen because it scales to millions of scatterers over long time series and gives exact derivatives. The core claim is about strain and acceleration as hazard covariates, whatever field estimator produces them. If the Gaussian-process ablation matches, we will present the method with kriging and drop the neural representation.

Why not just keep proportional hazards? Physical collapse plausibly accelerates as thresholds are crossed. Under that behaviour, a fixed hazard ratio misstates risk late in the process. The $g_1=0$ ablation, together with Schoenfeld-residual tests, checks this directly. Schoenfeld residuals test whether a covariate's effect drifts over time, which would violate the proportional-hazards assumption. If no drift appears, we simplify to the Cox form.

Proposed Experiments and Decisive Checks

The first decisive check is whether precursors are visible at all. In the tank, the internal cavity may migrate upward with almost no surface signal until breakthrough. We will measure how much the surface-strain covariates add to the hazard model's concordance as a function of time-before-collapse, using side-wall cavity tracking as ground truth for the hidden process. If surface strain carries no signal until the final seconds, the urban application, which sees only the surface, is not viable. We would then report that as a negative result.

The second check guards against artifacts. We inject the tissue-reconstruction failure mode into the inputs: synthetic tearing artifacts, and atmospheric-like phase noise of the kind that causes false positives at high over-wrap in the InSAR work. We then measure how far the predicted hazard rises falsely. Including the field-uncertainty covariate should suppress this rise.

The third check is transfer. Models trained on Korean sectors are tested on held-out sectors, and lead time is measured at a 5% false-alarm rate. Scoring by lead time rather than concordance alone ties the result to what a city engineer could act on.