Satellite radar interferometry (InSAR) maps millimetre-scale ground motion over whole cities and mining districts, which removes the main weakness of leveling and GPS: those instruments only see the points where someone already decided to put them. But each radar image records only one number per pixel, the motion projected onto the line between satellite and ground. Because the satellites fly nearly north–south, motion along that direction is almost invisible, even when ascending and descending passes are combined. Learning a spatial prior over what subsidence bowls, fault slips and magma intrusions look like could fill in the missing direction. The catch is that the obvious way to make such a network orientation-robust is wrong. Rotating a radar image does not correspond to a physically rotated deformation, because the satellite's viewing direction stays fixed while the ground pattern turns.
We propose to build networks whose behaviour is guaranteed under the correct symmetry: rotating the ground pattern and the satellite's look direction together. Concretely, each pass's line-of-sight image is fed to the network as a scalar field, and the pass's look direction is fed as a vector field that rotates with it. The network then outputs the full east, north and up displacement as a field of vectors that rotates consistently. Under this joint symmetry, a model trained on one set of orbit geometries carries over exactly to any rotated geometry. It can also use spatial structure to recover the north–south component that per-pixel inversion has to assume is zero.
Core Claim
Write $d_k(x)$ for the line-of-sight displacement of pass $k$ at ground position $x$. Write $\ell_k$ for that pass's unit look vector, $u(x)$ for the true 3D displacement, and $R$ for a rotation about the vertical axis that acts on horizontal components only. Then $d_k(x) = \ell_k \cdot u(x)$. If the ground pattern rotates to $u'(x) = R\,u(R^{-1}x)$ and every look vector rotates to $R\ell_k$, each observed image simply rotates, $d'_k(x) = d_k(R^{-1}x)$. So the inversion map $f$ we want to learn must satisfy
$$ f\big(\{d_k \circ R^{-1},\, R\ell_k\}_k\big) = R\, f\big(\{d_k, \ell_k\}_k\big) \circ R^{-1}. $$
This is an equivariance (outputs transform predictably when inputs are transformed) over a joint input. Image-rotation equivariance on its own, with $\ell_k$ held fixed, is physically false: it would assert that rotating the interferogram rotates the motion, which ignores how the projection changes. We claim three things. First, enforcing the joint symmetry beats both unconstrained networks and naively rotation-equivariant networks at recovering horizontal motion. Second, it transfers zero-shot to look geometries absent from training. Third, it recovers the north–south component better than the classical $U_N \approx 0$ inversion, measured against independent leveling where it exists.
Relation to Prior Work
Two strands of work motivate the gap. On the geodetic side, Resolving 3D Ground Deformation from Multi Geometry InSAR, A Model for Complex Subsidence Causality Interpretation Based on PS-InSAR Cross-Heading Orbits Analysis and Multidimensional Small-Baseline Subset (MSBAS) separate vertical from east–west motion by solving a linear system pixel by pixel across passes. They use no spatial context, so the north–south blind spot is structural in their formulation. We keep their projection equation but replace the per-pixel solve with a learned spatial operator. Three-Dimensional Deformation Prediction Based on the Improved Segmented Knothe–Dynamic Probabilistic Integral–Interferometric Synthetic Aperture Radar Model recovers 3D mining subsidence from a single radar track. It does this by fitting a closed-form rock-mechanics model, tuned per site with particle swarm search and tied to an assumed working-face geometry. It is our strongest real-data competitor. We differ in learning the spatial prior across many source types instead of fixing one functional form.
On the learning side, Detecting Volcano Deformation in InSAR using Deep learning and Application of Deep Learning to Detect Ground Deformation in InSAR Data classify wrapped interferograms with standard AlexNet-style CNNs. Hephaestus benchmarks ResNet and Swin backbones on 216,106 labelled patches. That paper also notes that standard augmentations such as random rotation distort the physical relationships in phase data. We read that observation as a constraint on our design. It is exactly why naive rotation equivariance is the wrong bias, and none of these works conditions on the look geometry.
The method comes from the equivariance literature. Rotation Equivariant Vector Field Networks and FILTRA: Rethinking Steerable CNN by Filter Transform supply steerable convolutions that carry vector-valued features. These are filters built from rotation-aware basis functions, so feature maps rotate exactly under input rotation. We reuse them, with input and output field types chosen to match the geodetic projection. EqGINO motivates predicting physical vector fields in a frame tied to the inputs rather than in fixed Cartesian axes. Our look-vector input is that frame. Two papers report that augmentation-trained baselines degrade under rotations absent from training and lose directional filter sharpness: ZZ-Net and Learning rotation invariant convolutional filters for texture classification. We treat their results as motivation, not proof, because their setting has no projection operator.
Research Program
We first build a synthetic benchmark. Elastic and empirical deformation sources, including Mogi point sources, dykes, sills, probabilistic-integral mining bowls and dipping fault slips, are drawn at random strike, aspect ratio and depth. Each is projected onto look vectors sampled from a range of headings, incidence angles and right- or left-looking geometries. Atmospheric look-alikes are added: stratified delays correlated with topography from the 30 m DEMs used in the cited studies, plus turbulent noise. The benchmark deliberately holds out heading ranges and source orientations so that transfer can be tested.
The model is a steerable U-Net. It takes as input one scalar channel per pass ($d_k$ and the vertical component of $\ell_k$) and one vector channel per pass (the horizontal part of $\ell_k$, broadcast spatially). It outputs a vector field for $(U_E, U_N)$ and a scalar field for $U_U$. Reflection symmetry is included only for passes where flipping also flips the looking side. Discrete-grid aliasing at angles that are not multiples of 90°, flagged by FILTRA, is measured directly as an equivariance error rather than assumed away.
We then analyse what the north–south recovery actually comes from. The working hypothesis is that elastic and mining sources couple horizontal motion to vertical gradients, so spatial context partly determines $U_N$. A network can exploit this coupling, while per-pixel inversion cannot. We will report where recovery fails, for example for pure north–south rigid translation, where no spatial cue exists. Finally we apply trained models to real data.
Baselines
The naive reference is pixelwise least-squares inversion of ascending plus descending passes with $U_N$ set to zero, as in the multi-geometry decomposition papers above. It defines the floor for vertical and east–west accuracy and has, by construction, no north–south skill. MSBAS is run where time series exist.
The strongest learned competitor is a standard U-Net of equal parameter count and matched training FLOPs. Matching FLOPs matters because steerable layers are slower; RotEqNet reports roughly 1.5× forward cost. This U-Net receives the same look-vector channels as plain scalars and trains on the same simulator with physically correct augmentation, meaning sources are regenerated at random orientations and headings rather than images being rotated. Every model gets the same number of simulated samples and the same hyperparameter search budget. Beating it means lower root-mean-square error (RMSE) on $U_N$ and on overall vector error in held-out heading ranges, with no worse in-distribution error.
Two ablations isolate the key ingredient. The first is a steerable network that is equivariant to image rotation alone, without look-vector inputs. We predict it will be worse than the plain U-Net on horizontal components, which would confirm that naive symmetry is harmful. The second is the plain U-Net trained without look-vector channels, as a single-geometry model.
On real mining data the strongest existing method is the ISK-DPIM-InSAR model. It reports leveling-validated RMSE of 45 mm along strike and 32 mm along dip in the Huaibei district from two ALOS PALSAR scenes. Beating it means lower RMSE on the same 71 strike and 55 dip leveling points without per-site parameter search. For detection, the reference is the Hephaestus binary deformation baseline, which reports roughly 97.4% accuracy and 82.2% F-score. Given the class imbalance, F-score is the number to beat.
Why not just…?
Why not just simulate every orientation and heading and train a standard network? This is the real threat. With an unlimited simulator, a conditioned U-Net may learn the joint symmetry approximately. The matched-budget comparison settles it in three regimes. The first is in-distribution, where we expect a small or no gap. The second is held-out heading ranges, where only the equivariant model is guaranteed to transfer. The third is a reduced simulation budget, where tied filters should pay off, as the single-image-per-class results in the texture paper suggest. If the U-Net matches us in all three regimes, the claim shrinks to compute efficiency, and we will report that.
Why not just invert ascending and descending passes by least squares? It has no information about north–south motion and breaks where phase decorrelates in steep bowls, as the ISK-DPIM paper documents for mining centres. The naive baseline quantifies how much spatial priors add beyond it.
Why not just fit an analytical model such as the probabilistic integral method? On ideal single-panel mining it may well win, and Huaibei is its home ground. Its weakness is that it requires an assumed source geometry and per-site search. Our test is whether one trained model matches it there and also handles dykes, faults and irregular urban bowls that it cannot represent.
Why not just use an off-the-shelf rotation-equivariant CNN? Without the look vector its symmetry assumption is physically false. The image-only equivariant ablation tests that prediction directly.
Decisive Checks
The idea stands or falls on four results. First, the joint-equivariance error, defined as the output discrepancy between a rotated input with rotated look vectors and the correspondingly rotated output, should stay near grid-aliasing levels at arbitrary angles, while both baselines show large errors. Second, $U_N$ RMSE on held-out headings should drop clearly below the conditioned U-Net's. Third, the image-only equivariant ablation should underperform on horizontal components. Fourth, on Huaibei the model should approach or beat the 45/32 mm leveling RMSE without site tuning.
For urban multi-track sites such as Wuhan, Beijing and Nansha, no reported result was found on Lacuna for independent north–south validation. These sites will therefore support only internal-consistency checks: predicting one pass from the others.
References
- Accessing the Time-Series Two-Dimensional Displacements around a Reservoir Using Multi-Orbit SAR Datasets: A Case Study of Xiluodu Hydropower Station
- A Model for Complex Subsidence Causality Interpretation Based on PS-InSAR Cross-Heading Orbits Analysis
- Mapping the Recent Vertical Crustal Deformation of the Weihe Basin (China) Using Sentinel-1 and ALOS-2 ScanSAR Imagery
- Three-Dimensional Deformation Prediction Based on the Improved Segmented Knothe–Dynamic Probabilistic Integral–Interferometric Synthetic Aperture Radar Model
- Detecting Volcano Deformation in InSAR using Deep learning
- Application of Deep Learning to Detect Ground Deformation in InSAR Data
- Hephaestus: A large scale multitask dataset towards InSAR understanding
- Rotation Equivariant Vector Field Networks
- FILTRA: Rethinking Steerable CNN by Filter Transform
- EqGINO: Equivariant Geometry-Informed Fourier Neural Operators for 3D PDEs
- ZZ-Net: A Universal Rotation Equivariant Architecture for 2D Point Clouds
- Learning rotation invariant convolutional filters for texture classification