Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PhyUnfold-Net: Advancing Remote Sensing Change Detection with Physics-Guided Deep Unfolding

Zelin Lei, Yaoxing Ren, Jiaming Chang

Abstract

Bi-temporal change detection is highly sensitive to acquisition discrepancies, including illumination, season, and atmosphere, which often cause false alarms. We observe that genuine changes exhibit higher patch-wise singular-value entropy (SVE) than pseudo changes in the feature-difference space. Motivated by this physical prior, we propose PhyUnfold-Net, a physics-guided deep unfolding framework that formulates change detection as an explicit decomposition problem. The proposed Iterative Change Decomposition Module (ICDM) unrolls a multi-step solver to progressively separate mixed discrepancy features into a change component and a nuisance component. To stabilize this process, we introduce a staged Exploration-and-Constraint loss (S-SEC), which encourages component separation in early steps while constraining nuisance magnitude in later steps to avoid degenerate solutions. We further design a Wavelet Spectral Suppression Module (WSSM) to suppress acquisition-induced spectral mismatch before decomposition. Experiments on four benchmarks show improvements over state-of-the-art methods, with gains under challenging conditions.

PhyUnfold-Net: Advancing Remote Sensing Change Detection with Physics-Guided Deep Unfolding

Abstract

Bi-temporal change detection is highly sensitive to acquisition discrepancies, including illumination, season, and atmosphere, which often cause false alarms. We observe that genuine changes exhibit higher patch-wise singular-value entropy (SVE) than pseudo changes in the feature-difference space. Motivated by this physical prior, we propose PhyUnfold-Net, a physics-guided deep unfolding framework that formulates change detection as an explicit decomposition problem. The proposed Iterative Change Decomposition Module (ICDM) unrolls a multi-step solver to progressively separate mixed discrepancy features into a change component and a nuisance component. To stabilize this process, we introduce a staged Exploration-and-Constraint loss (S-SEC), which encourages component separation in early steps while constraining nuisance magnitude in later steps to avoid degenerate solutions. We further design a Wavelet Spectral Suppression Module (WSSM) to suppress acquisition-induced spectral mismatch before decomposition. Experiments on four benchmarks show improvements over state-of-the-art methods, with gains under challenging conditions.
Paper Structure (24 sections, 39 equations, 8 figures, 9 tables)

This paper contains 24 sections, 39 equations, 8 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. The Proposed Method
  5. Memory-Enhanced ICDM with SVE-Gated Residual Injection
  6. Training Objective and Inference
  7. WSSM: Wavelet Spectral Suppression Module
  8. Experiments
  9. Datasets and Evaluation Metrics
  10. Implementation Details
  11. Comparison with State-of-the-Art Methods
  12. Ablation Study
  13. Conclusion
  14. Training Dynamics
  15. Convergence Analysis of ICDM
  16. ...and 9 more sections

Figures (8)

  • Figure 1: Paradigm comparison. (a) Cross-temporal attention for invariant suppression. (b) Token interaction with implicit difference boosting. (c) Our explicit difference-space decomposition $D=C+N$ with unrolled refinement for stable, controllable change modeling with explicit separation control.
  • Figure 2: Overview of PhyUnfold-Net. WSSM pre-suppresses acquisition-induced variations before forming $D=\Delta(F_1,F_2)$; ICDM unrolls a $K$-step solver to decompose $D=C+N$ and refine $(C,N)$; the segmentation head predicts $\hat{Y}$ from $C^{K}$, with training-time losses and staged regularization shown in the lower panels.
  • Figure 3: SVE evolution across UNet layers during training, including Encoder-4, Decoder-1, Decoder-2, and Decoder-3 for progressive change-noise separation.
  • Figure 4: Training-time SVE trajectories for changed and unchanged patches on raw $D$ and ICDM steps $k{=}1,2,3$. The marked end-gap $\Delta$ increases from $0.13$ to $0.36$.
  • Figure 5: Patch-grid SVE maps across ICDM steps: changed region (top row) becomes progressively more compact and higher-valued, while unchanged region (bottom row) becomes more diffuse and lower-valued over iterative refinement.
  • ...and 3 more figures