acceptodds
Under review as a conference paper at ICLR 2027

PhyRestore: Physics-Structured Latent-Factor Restoration

Abstract

Estimating temporal soil-loss change is challenging when physically meaningful input factors are noisy or corrupted, particularly because substantial changes are rare relative to the large number of locations exhibiting little change. We study this problem through the Revised Universal Soil Loss Equation (RUSLE) and introduce PhyRestore, a physics-structured latent-factor restoration framework. Rather than directly predicting soil-loss change or correcting a degraded physical estimate, PhyRestore restores corrupted physical factors and reconstructs temporal change through the known physical relationship. We evaluate PhyRestore in a watershed-scale bitemporal raster setting under isolated and simultaneous corruption of rainfall erosivity and cover management, comparing it with the degraded RUSLE estimate and Direct RF, XGBoost, MLP, and CNN models. Factor restoration improves high-magnitude recovery when the corrupted factors remain identifiable, but its advantage weakens under joint corruption, sparse positive extremes, and factor values outside the training support.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
physics-structured machine learning, scientific machine learning, latent-factor restoration, corrupted covariates, imbalanced regression, temporal prediction, soil erosion, RUSLE

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8GfnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies temporal prediction of RUSLE-derived soil-loss change under controlled corruption of physical input factors and proposes PhyRestore, a physics-structured framework that restores corrupted factors from bitemporal spatial context before reconstructing temporal change through the known RUSLE relationship. Using watershed-scale raster data from 2017 and 2022, the authors compare PhyRestore against direct prediction and formula-based correction approaches, with particular emphasis on rare high-magnitude changes. PhyRestore substantially reduces tail prediction errors under isolated factor corruption, while its advantages diminish under simultaneous corruption and for factor values outside the training support.

Critical strengths
  1. The experimental evaluation is systematic, covering multiple prediction strategies and model families, with particular attention to rare high-magnitude changes. The substantial reductions in TAIL95 MAE under isolated R2022 and C2017 corruption demonstrate the potential benefits of factor-level restoration.
Critical weaknesses
  1. Limited robustness to unknown and mixed corruption. PhyRestore assumes prior knowledge of corrupted factors and trains condition-specific restorers, limiting its applicability to unknown corruption patterns. Moreover, its performance falls below the strongest Direct baseline under mixed corruption, raising concerns about robustness when multiple factors are degraded.
  2. Limited experimental generalizability.The evaluation is restricted to a single watershed and synthetic Gaussian corruption with a fixed noise realization. The spatial split protocol is also insufficiently detailed, leaving the method's robustness to realistic degradation and its spatial generalizability uncertain.
Questions for authors
  1. How could PhyRestore handle unknown corruption patterns, and could jointly trained, multi-output factor restoration improve performance under mixed corruption?
  2. Could the authors clarify the spatial split protocol and assess robustness across different corruption realizations or more realistic noise patterns?
Reviewer_Sr76OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

PhyRestore learns to recover the “correct” center-pixel value from bitemporal 7x7 adjacent pixels through a supervised CNN, where the “correct” label are not independent measured data, and then reconstruct the temporal difference through the fixed RULSE. The experiments are conducted in one basin and a pair of years, it is uncertain whether it has practical application capabilities. Moreover, there was no direct comparison with other factor-level denoising methods.

Critical strengths

None.

Critical weaknesses
  1. Limited methodological novelty: The core of PhyRestore is supervised CNN denoising of known physical factors followed by substitution into an existing deterministic equation. It is not yet clear what physical structure is introduced beyond this two-stage formulation.
  2. Limited experiment conditions: The experiment only covered one basin, a pair of years; it did not examine distinct basin areas, real spatiotemporal correlated errors.
Questions for authors

The proposed approach takes noise-corrupted physical factors as input, restores their “correct” values, and subsequently calculates bitemporal changes. Compared with directly estimating the factors using machine learning or deep learning methods, what additional benefits does this approach offer, particularly when the ground truth labels are not derived from field measurements? Please add experiments.

Reviewer_gFq9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes PhyRestore, a physics-structured framework, to restore corrupted physical factors and temporal changes for better estimates of soil loss.

Contributions:

  • formulate controlled corruption for spatial evaluation
  • PhyRestore, physics-structured framework for restoring corrupted factors
  • analysis of when this helps

Validation:

  • empirical on data under controlled corruption
Critical strengths
  • physics-structured modeling is important to guide machine learning models, in this case to help predictions stay physically interpretable
  • results on multiple model families, corruption settings, spatial folds, metrics (changes AND overall)
  • discussion of what doesn't work as well
Critical weaknesses
  • how much of the improvement comes from the proposed approach versus simply denoising the corrupted input factors. baselines do not include simpler denoising methods for factor restoration and only predict or correct the final target
  • one watershed
  • results for rare positive changes in TAIL95 are only based on 61 positive samples, method still struggles to capture their variability in some settings
  • minor clarity - examples of factors and visualizations would help illustrate RUSLE and corruption challenges; many acronyms/abbreviations throughout that make it difficult to follow in places (e.g., Direct introduced in 181, referenced again in a list in 287 and in results table 1 - had to go back in paper to recall)
Questions for authors
  • how does PhyRestore compare to simpler factor-level denoising or regression approaches?

  • would similar improvements be expected with spatially correlated noise or systematic biases rather than independent Gaussian corruption?

  • how well does PhyRestore recover the small number of positive examples in TAIL95, particularly their direction and variability?

  • why is the restoration error for previously seen C2017 values so small?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.