acceptodds
Under review as a conference paper at ICLR 2027

Multimodal Physical Posterior Inference for Early Prediction of Surface Deformation from InSAR Time Series

Abstract

Early observations of mining-induced subsidence reveal surface motion but leave its eventual magnitude and growth rate uncertain. This ambiguity makes forecasting from short interferometric synthetic aperture radar (InSAR) sequences difficult. We introduce Multimodal Physical Posterior Inference (MPPI), which uses geological and mining context to constrain physical parameters that early deformation maps cannot fully identify. MPPI combines visual patterns with textual descriptions of observation statistics and geological conditions to infer a posterior over maximum subsidence and growth rate. A physics-based decoder converts these parameters into future deformation fields and propagates posterior samples into prediction intervals. We pre-train the inference network on simulated sequences and fine-tune it on real InSAR observations. Across six coal-mining regions in Shanxi Province, China, MPPI achieves 4 cm RMSE for both maximum-subsidence estimation and deformation forecasting. Ablations identify geological information as the strongest complementary source, with textual representation providing further gains. Probabilistic parameter learning halves forecast RMSE relative to deterministic regression, while posterior sampling improves the continuous ranked probability score. These results support combining multimodal evidence with physical-parameter inference to forecast deformation from incomplete observations.

Primary area:
learning on time series and dynamical systems
Keywords:
Vision-language model, Time series forecasting, Computer vision

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_EhzaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This study investigates the prediction of surface subsidence in mining areas using early InSAR observations and proposes a multimodal physics-based posterior inference framework. The framework fuses deformation maps, observation statistics, and geological and mining conditions to estimate the probability distributions of the maximum subsidence magnitude and its growth rate. Future deformation is then generated through temporal evolution and spatial subsidence models, and prediction intervals are constructed via parameter sampling.

Critical strengths
  1. The use of physical parameters to bridge multimodal information and future prediction constitutes a well-targeted modeling strategy. Geological conditions provide external complementary information to early deformation observations, while parameter distributions are employed to characterize uncertainty in predictions.

  2. The ablation settings are aligned with the core design. Comparisons between numerical and textual representations help disentangle the roles of information sources and representation modalities, whereas comparisons among deterministic regression, posterior mean, and posterior sampling support the assessment of probabilistic learning and uncertainty propagation.

  3. The manuscript is clearly structured, with the framework diagram, spatial prediction comparisons, and temporal curves mutually complementing each other. The evaluation covers parameter estimation, deformation prediction, and the quality of prediction intervals, offering a relatively comprehensive presentation of the results.

Critical weaknesses
  1. The posterior learning objective is not described with sufficient specificity. Section 4.1 mentions parameter regression and distribution learning, but the loss formulation and the way it is combined with the physical fitting objective are not explicitly provided. These definitions should be supplemented to facilitate the understanding of the gains from probabilistic learning in Table 3.
  2. Certain definitions in the physics module require clarification. If Eq. (21) adopts the Euclidean norm and the integration domain does not vary with the direction angle, the rotation term will not alter the kernel function; likewise, under the definition of the predicted quantities provided in the paper, Eq. (25) and Eq. (27) are mathematically equivalent. It is recommended to clarify how the direction parameter acts and to explain the rationale for the separate design of the two loss terms.

Other issues: The temporal split of the data, the length of historical observations, and the prediction horizon should be explicitly specified to facilitate understanding of the task setup. In addition, the publication years and reference information for some methods in Table 4 need to be verified.

Questions for authors

See Weaknesses

Reviewer_P9QdOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes MPPI whichs infers a posterior over subsidence parameters using early InSAR deformation maps combined with text (statistical and geological), and uses a physics-based decoder to forecast deformation.

Critical strengths

A detailed assessment was not carried out; see Weaknesses.

Critical weaknesses

The main text runs to page 13 which is well beyond the 9-page limit. Under the ICLR 2027 author guidelines, papers exceeding the page limit will be desk-rejected. I flagged this to the AC/SAC/PCs. I have therefore not provided a full technical review.

Questions for authors

None, given the expected desk rejection.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.