Source-linked AI summary
FiLM-GPNet: Geometry-Aware Pseudo-Supervised Phase Restoration with Zero-Shot Generalization for Large Temporal InSAR Stacks
Getnet Demil, Muhammad Farhan Humayun, Tomi Westerlund, Jukka Heikkonen, Mourad Oussalah
TL;DR
Dense commercial InSAR stacks make fixed geometry-blind filtering inadequate for heterogeneous acquisitions. FiLM-GPNet restores wrapped phase by conditioning a pseudo-supervised network on per-pair geometry and interferometric physics, with uncertainty estimation. Relative to Goldstein filtering, it reduces temporal residuals and closure errors, improves downstream Western Australia metrics, and transfers zero-shot to a distinct Los Angeles stack.
Problem
Heterogeneous commercial InSAR acquisitions challenge fixed geometry-blind filters, while clean reference interferograms are generally unavailable for supervision.
Method
FiLM-GPNet uses a FiLM-conditioned U-Net with a 7D geometry descriptor, Goldstein pseudo-targets, triplet-closure regularization, and per-pixel uncertainty prediction.
Results
FiLM-GPNet reduces temporal residual by 68% in Hawaii and 66% in Western Australia relative to Goldstein, with closure-error reductions of 10% and 13%, respectively.
Takeaways & Limitations
The model generalizes without retraining to a geographically and geometrically distinct stack, supporting geometry-conditioned restoration across heterogeneous commercial SAR archives.
Abstract
from arXiv · showhide
The growing availability of dense commercial Synthetic Aperture Radar (SAR) time series enables temporal Interferometric SAR (InSAR) analysis, but fixed classical filters fail under heterogeneous acquisition geometries, degrading phase quality and temporal consistency. We propose FiLM-GPNet, a geometry-conditioned network for wrapped-phase restoration that explicitly adapts to acquisition differences using Feature-wise Linear Modulation (FiLM) and a 7D per-pair geometry descriptor. The model is trained with pseudo-supervision from Goldstein-filtered interferograms and regularized by interferometric physics via triplet-closure consistency, while also estimating per-pixel aleatoric uncertainty. Experiments on three Capella Spotlight stacks from the IEEE GRSS 2026 Data Fusion Contest show that FiLM-GPNet reduces temporal residual by 68% (Hawaii) and 66% (Western Australia) relative to the Goldstein baseline, alongside closure error reductions of 10% and 13%, respectively. In Western Australia, it further improves unwrapping success rate by 7.7 percentage points and Digital Elevation Model (DEM) Normalized Median Absolute Deviation (NMAD) by 31%. The model also shows strong zero-shot generalization to a geographically and geometrically distinct third stack (Los Angeles) without retraining, supporting geometry-conditioned restoration as an effective alternative to fixed classical filtering across heterogeneous stacks.
I. INTRODUCTION
Dense commercial InSAR stacks challenge geometry-blind filtering because acquisition conditions vary substantially across interferograms. FiLM-GPNet addresses this with geometry-conditioned pseudo-supervised restoration, uncertainty prediction, and zero-shot transfer.
- Motivation: Geometry-blind filters apply nearly uniform smoothing despite variation in temporal baseline, perpendicular baseline, incidence angle, and viewing geometry.This can produce inconsistent interferograms, higher triplet-closure errors, unstable unwrapping, and weaker multi-temporal inversion.
- Approach: FiLM-GPNet is a geometry-conditioned pseudo-supervised network for phase restoration in large temporal InSAR stacks.It uses Goldstein-filtered interferograms as proxy targets and triplet-closure consistency for regularization.
- Approach: The FiLM-conditioned U-Net adapts restoration to per-pair acquisition geometry rather than applying one fixed filter across heterogeneous stacks.This is the central architectural response to geometry variation in the data.
- Approach: Per-pixel aleatoric uncertainty prediction provides an additional output for downstream confidence assessment.The uncertainty estimate complements the restored phase output.
- Generalization: The model is designed to transfer without retraining from one geographic stack to a geographically and geometrically distinct site.The study evaluates the approach on multiple stacks from the IEEE GRSS 2026 Data Fusion Contest dataset.
II. METHOD
The appendix introduces the overall processing pipeline and indicates that its principal components are described in the following sections.
- Pipeline overview: Figure A1 summarizes the overall processing pipeline.The figure provides the pipeline-level view before the individual components are described.
- Pipeline overview: The pipeline is organized around several main components.The passage directs the reader to subsequent descriptions of those components.
- Pipeline overview: The component descriptions follow the overall pipeline summary.This establishes the appendix’s organization from system overview to component-level explanations.
A. Geometry-Aware Pair-Graph Construction
The method constructs a geometry-aware pair graph by scoring acquisition compatibility, rejecting unsuitable pairs, and enumerating valid closure triplets for later physics-based supervision.
- Pair compatibility: Qij combines temporal-baseline, incidence-angle, and geometry-consistency factors to score pair compatibility.The additional factor accounts for look-angle, squint, sampling, and footprint-overlap mismatches when metadata are available.
- Pair compatibility: Geometry-consistency penalties incorporate look-angle mismatch, squint mismatch, range/azimuth sampling mismatch, and low footprint overlap.These metadata-dependent penalties make pair selection sensitive to acquisition compatibility.
- Graph construction: Accepted pairs and valid closure triplets are retained for later physics-based supervision.Closure triplets are enumerated from the filtered graph.
- Coregistration: Precise pairwise alignment precedes learning-based phase restoration because Capella Spotlight PFA SLC pairs can have geometry-dependent misregistration and variable interferometric quality.The alignment pipeline refines offsets using phase correlation, local normalized cross-correlation, subpixel refinement, and robust outlier rejection.
- Quality control: Pairs with final mean coherence below 0.3 are rejected from downstream processing.The workflow retains both raw wrapped interferograms and Goldstein-filtered versions for comparison.
C. Geometry-Conditioned FiLM-GPNet
FiLM-GPNet conditions a U-Net phase-restoration backbone on each interferometric pair’s seven-dimensional acquisition geometry. The geometry code modulates feature maps throughout the network, which predicts both restored complex interferograms and per-pixel uncertainty.
- Geometry-conditioned inputs: Each interferometric pair combines a three-channel input tensor with a normalized seven-dimensional geometry vector describing acquisition conditions.The input channels are the raw interferogram’s real and imaginary components plus coherence; the geometry includes temporal and baseline variables, angles, acquisition mode, look direction, and an SNR proxy.
- Geometry conditioning: The geometry vector is mapped by a two-layer MLP from 7-D to a 64-D conditioning code.
- Feature-wise modulation: FiLM injects the conditioning code into every encoder and decoder block through channel-wise affine modulation of intermediate feature maps.The modulation uses learned channel-wise scale and shift parameters applied with element-wise multiplication.
- Backbone: The four-level U-Net uses skip connections, FiLM-conditioned double-convolution blocks, channel widths [32, 64, 128, 256], and a 512-channel bottleneck.The design aims to suppress noise while preserving fringe structure.
- Outputs: Two output heads predict a restored two-channel complex interferogram and a per-pixel log-variance map for aleatoric uncertainty estimation.
D. Pseudo-Supervised Training Objective
FiLM-GPNet is trained as a geometry-conditioned pseudo-supervised restoration model using Goldstein-filtered interferograms as pseudo-targets. Its objective combines restoration, uncertainty, fringe-preservation, temporal-consistency, and triplet-closure terms.
- Pseudo-supervision: Goldstein-filtered interferograms provide pseudo-targets for geometry-conditioned phase-restoration regression.The approach learns a geometry-adaptive generalization of the classical Goldstein baseline without reference DEMs or externally provided deformation labels.
- Predictions: The model predicts a restored two-channel complex interferogram and a per-pixel log-variance map from raw interferogram and coherence channels.
- Objective: The total loss combines wrapped-phase regression, heteroscedastic uncertainty, fringe-preserving gradient, temporal-consistency, and triplet closure-consistency terms.The terms respectively match pseudo-targets, model uncertainty, preserve fringe gradients, promote stack-level temporal consistency, and enforce closure.
E. Uncertainty-Aware Downstream Processing
After tile-wise inference and overlap-add reconstruction, the denoised interferograms enter a standard InSAR back-end for evaluating unwrapping stability and elevation recovery.
- Reconstruction: Tile-wise predictions are reconstructed with overlap-add before standard downstream InSAR processing.
- Downstream processing: The phase is unwrapped with SNA-PHU, converted to elevation using the standard phase-to-height relation, and terrain-geocoded for comparison with external DEMs.These steps evaluate unwrapping stability and elevation recovery.
A. Data and Experimental Setup
The experiments use three Capella X-band Spotlight stacks with distinct temporal and geometric characteristics, reserving Los Angeles for zero-shot evaluation. Performance is assessed with four complementary metrics under an AOI-aware, tile-based test protocol.
- Dataset: The dataset contains 791 Capella Space X-band Spotlight SLCs across 39 AOIs, with Hawaii, Los Angeles, and Western Australia selected for evaluation.
- Stacks: Hawaii is the densest stack, Western Australia is more homogeneous, and Los Angeles provides a distinct ascending stack for cross-stack and zero-shot assessment.The stacks differ in acquisition counts, temporal spans, pass directions, and incidence-angle ranges.
- Protocol: Training uses an AOI-aware temporal split after strict pair-graph filtering, with metrics reported on processed test stacks using 256 × 256 tiles and stride 128.
- Metrics: Triplet Closure Error, Unwrap Success Rate, DEM NMAD, and Temporal Residual measure pairwise consistency, downstream usability, elevation quality, and stack-level temporal stability.
C. Results
FiLM-GPNet improves phase-restoration and downstream InSAR performance across evaluated AOIs, while its consistency gains are linked to physics-based losses and visual outputs are compared against the baseline.
- 68% and 66% temporal residual reductions are reported for Hawaii and Western Australia, respectively, relative to the Goldstein baseline.
- 10% and 13% triplet closure-error reductions are reported for Hawaii and Western Australia, respectively.
- 7.7 percentage points higher unwrap success rate and 31% lower DEM NMAD are reported for Western Australia.
- Removing closure and temporal-consistency terms degrades both closure error and temporal residual on the Hawaii stack.The ablation attributes most observed gains to physics-based losses rather than the pseudo-supervised target alone.
- Figures 2 and 3 compare baseline and FiLM-GPNet intermediate outputs and elevation-domain outputs before geocoding.
IV. CONCLUSION
FiLM-GPNet conditions phase restoration on acquisition geometry and uses physics-informed pseudo-supervision to improve consistency across heterogeneous temporal InSAR stacks. It also maintains downstream performance and transfers without retraining to a geographically and geometrically distinct stack.
- Geometry-conditioned restoration with physics-informed pseudo-supervision provides more consistent and robust processing than fixed classical filtering.
- The largest reported gains are in temporal residual reduction on the Hawaii and Western Australia datasets.
- Downstream phase-unwrapping and DEM-quality performance is maintained across the evaluated heterogeneous Capella Spotlight stacks.
- Zero-shot transfer to a geographically and geometrically distinct stack occurs without retraining.