acceptodds
Under review as a conference paper at ICLR 2027

PhyDiM: Physics-Guided Diffusion Model for Spatio-Temporal Satellite Image Reconstruction under Sparse Observations

Abstract

Reconstructing continuous spatio-temporal images from sparse and irregular visual observations is a challenging machine learning problem due to complex spatial dependencies and strong temporal dynamics that must be inferred from a limited number of temporal observations. In this setting, incorporating physical priors is particularly beneficial, as it enables the reconstructed images to respect underlying physical laws and produce more realistic and temporally consistent dynamics. Recent generative approaches, particularly diffusion models, have demonstrated good performance in different reconstruction tasks, but they remain largely data-driven and lack explicit physical structure to guide the generation process. Existing attempts treat physical constraints as external regularization terms in the loss function, which can lead to unstable optimization. In this paper, we introduce PhyDiM, a physics-guided diffusion model that incorporates partial differential equation (PDE) priors directly into the diffusion process. Specifically, we embed the physical constraints directly into both the forward diffusion process and the reverse sampling dynamics. PhyDiM is applied to land surface temperature reconstruction from sparsely sampled satellite observations, where the PDE is instantiated as a heat equation. Experimental results on large-scale and climatically diverse regions covering France, Spain, Germany, Belgium, and the Netherlands demonstrate that PhyDiM consistently outperforms both state-of-the-art reconstruction and diffusion-based methods, particularly at extreme diurnal time steps where existing approaches tend to degrade. Moreover, qualitative analysis further demonstrates that PhyDiM preserves sharper spatial structures and more realistic thermal patterns across diverse conditions.

Primary area:
generative models
Keywords:
spatio-temporal image reconstruction, diffusion models, physics-guided vision, satellite imagery

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_uYEbOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose PhyDiM, a physics-guided diffusion model for continuous spatio-temporal LST reconstruction from sparse satellite observations. Physical priors are integrated into both the forward diffusion and reverse sampling processes. The method supports both temporal interpolation and extrapolation, enabling reconstruction at arbitrary target times. Experiments on MODIS LST datasets across three European regions demonstrate improved reconstruction performance over existing methods.

Critical strengths

1.Physical priors are integrated into both the forward diffusion and reverse sampling processes, improving reconstruction performance.

2.The conditional denoising architecture enable flexible reconstruction at arbitrary target times, supporting both temporal interpolation and extrapolation.

Critical weaknesses

1.The justification for adopting the heat equation to represent LST dynamics is insufficient. Moreover, the PDE parameters are empirically selected, raising concerns about the physical validity of the proposed PDE-based formulation.

2.The reason behind the PDE-based forward process in Eq. (5) is unclear. If the heat equation represents the actual LST dynamics, applying it to the temperature field at the target timestamp in fact advances the field by an additional tΔtt\Delta t, causing it to deviate from the target physical state.

3.The proposed method achieves modest improvements over standard DDPM, raising concerns about whether the additional complexity of the physics-guided components is justified.

4.The methodological novelty is limited, as the proposed framework largely combines existing diffusion modeling techniques with PDE-based smoothing and sampling guidance.

Questions for authors

The PDE-based operations in both the forward and reverse processes appear to primarily introduce spatial smoothing. Can the authors demonstrate that these operations provide physically meaningful benefits beyond conventional smoothing? Could alternative smoothing methods achieve similar improvements?

9 comments
AuthorsOpenReview ↗

Response to Reviewer uYEb (1/4)

We sincerely thank the reviewer for the careful assessment of our manuscript and the constructive feedback.

We would like to highlight that Weakness 2, which was also raised as Weakness 2 by Reviewer GmWZ and Question 3 by Reviewer W6p8, identified an important methodological concern regarding the reverse sampling process. Addressing this issue strengthened the theoretical validity of the reverse transition and improved PhyDiM's performance over DDPM in both RMSE and SSIM (see our response to Weakness 3). Accordingly, we report the updated results throughout our responses. For ease of navigation, all text added to the revised manuscript has been highlighted in red.

Weakness 1: Physical Relevance of the Heat Equation to LST Dynamics

We clarify that the heat equation is not intended to capture the full dynamics of LST. A complete surface energy balance involves processes such as radiation, soil moisture, thermal inertia, and turbulent heat fluxes, which cannot be fully characterized using the MODIS observations available at inference time. Incorporating these processes would require additional parameters and input variables that may not be observable. We therefore adopt spatial heat diffusion as a simplified physical prior, since it operates directly on the observed temperature field and requires only a single effective diffusion coefficient.

This choice is motivated by three considerations. First, heat diffusion is a physical mechanism that smooths spatial temperature gradients at the surface and has been used in LST modeling [1]. Second, LST fields show spatial structures, including coherent warm and cold regions with smooth transitions, that the heat operator helps preserve while attenuating high-frequency variations [2]. Third, MODIS LST retrievals have errors of approximately 2∘C2^\circ\mathrm{C} [3]. A prior that suppresses potentially non-physical high-frequency fluctuations while preserving large-scale spatial patterns is therefore well suited to these data. The aspects of LST dynamics not captured by the heat equation, particularly diurnal forcing, are not imposed by the physical prior but are learned by the denoiser network. We have revised the text accordingly in Sec. 3.2.

Regarding the PDE parameters (we changed the notation from tt to τ\tau following the remark of Reviewer W6p8 to avoid confusion with the diffusion timestep tt), κ\kappa and Δτ\Delta\tau cannot be independently determined for each pixel from satellite imagery, as they depend on ground thermal properties that are not directly observed. We therefore treat their product, κΔτ\kappa\Delta\tau, as a single effective diffusion strength and select its value using the validation set (Sec. 4.4). The case κΔτ=0\kappa\Delta\tau=0 corresponds to Config. 2, which obtains an RMSE of 1.82 and an SSIM of 0.56 on ROI$_1$, compared with 1.78 and 0.61 for PhyDiM at κΔτ=2×10−4\kappa \Delta \tau = 2 \times 10^{-4}, showing that heat evolution brings a measurable gain over the no-physics case.

The table below presents a comparison between Config. 2 and PhyDiM for ROI$_1$. For the full ablation comparison, including the other configurations, please refer to Table 2 in Sec. 4.4:

ConfigurationForward DiffusionSampling PhaseInverse Heat OperatorRMSE↓SSIM↑
Config. 2✗✓✗1.820.56
PhyDiM✓✓✓1.780.61

[1] Saeid Asadzadeh and Carlos Roberto Souza Filho. Numerical modeling of land surface temperature over complex geologic terrains: A remote sensing approach. Remote Sensing, 15(19):4877, 2023.

[2] Song, J., Du, S., Feng, X., & Guo, L. (2014). The relationships between landscape compositions and land surface temperature: Quantifying their resolution sensitivity with spatial regression models. Landscape and urban planning, 123, 145-157.

[3] Si-Bo Duan, Zhao-Liang Li, Hua Li, Frank-M Gottsche, Hua Wu, Wei Zhao, Pei Leng, Xia Zhang, and Cesar Coll. Validation of collection 6 modis land surface temperature product using in situ measurements. Remote sensing of environment, 225:16–29, 2019.

AuthorsOpenReview ↗

Response to Reviewer uYEb (2/4)

Weakness 2: Clarification of the Physical-Time Advancement Introduced by the PDE-Based Forward Process

The evolution in Eq. (5) does not advance the target field in physical time. Rather, it acts along the diffusion (corruption) axis, indexed by tt, with pseudo-time τ=tΔτ\tau=t\Delta\tau. It is part of the corruption operator, in the same way as the Gaussian noise, and is not intended to model how LST evolves between satellite acquisition hours. Physical time is handled separately through conditioning on the acquisition hours hih_i and the target hour h∗h^*. To avoid confusion between these two notions of time, we now use tt exclusively for the diffusion timestep, hh for physical hours, and τ\tau for the heat-equation pseudo-time.

The target physical state is not altered by this construction. The evolved field satisfies F0(x0)=x0F_0(\mathbf{x}_0)=\mathbf{x}_0, meaning that the clean target corresponds to t=0t=0, whereas Ft(x0)F_t(\mathbf{x}_0) for t>0t>0 represents an intermediate corrupted state, just as xt\mathbf{x}_t represents a noisy version of the target in a standard DDPM.

We acknowledge that the previous version of Algorithm 2 did not invert the forward heat evolution, resulting in an output that was one Euler step away from x0\mathbf{x}_0. This inconsistency was also raised by Reviewers GmWZ (Weakness 2) and W6p8 (Question 3). In the revised sampler, we apply the exact inverse heat operator at each reverse step, thereby preserving the forward marginals (see the revised Sec. 3.4 and the derivation in App. B). The final iterate therefore estimates F0(x0)=x0F_0(\mathbf{x}_0)=\mathbf{x}_0, corresponding to the target physical state without any additional tΔτt\Delta\tau drift.

The motivation for this corruption process is that, unlike isotropic Gaussian noise, it degrades spatial structure in a coupled manner through the smoothing behavior of the heat operator. This encourages the denoiser to learn to reverse a structured degradation. Its benefit is supported empirically by the ablation results in Table 2.

Revised sampling procedure (Algorithm 2):

  1. Initialize the noisy field: xT∼N(0,I)\mathbf{x}_T \sim \mathcal{N}(0,\mathbf{I}).
  2. Compute the Fourier-domain heat operator: G^(kx,ky)=1+kd[2cos⁡(2πkx)+2cos⁡(2πky)−4]\widehat{G}(k_x,k_y) = 1+k_d[2\cos(2\pi k_x)+2\cos(2\pi k_y)-4].
  3. For t=T,…,1t=T,\ldots,1:
    • Predict noise: ϵθ=ϵθ(xt,C,h∗,t)\boldsymbol{\epsilon}_\theta = \epsilon_\theta(\mathbf{x}_t,\mathcal{C},h^*,t).
    • Estimate the physically evolved field: F^t(x0)=1αˉt(xt−1−αˉtϵθ)\hat{F}_t(\mathbf{x}_0) = \frac{1}{\sqrt{\bar{\alpha}_t}}\left(\mathbf{x}_t-\sqrt{1-\bar{\alpha}_t}\boldsymbol{\epsilon}_\theta\right).
    • Apply inverse heat evolution: F^t−1(x0)=F−1[F(F^t(x0))G^]\hat{F}_{t-1}(\mathbf{x}_0) = \mathcal{F}^{-1}\left[\frac{\mathcal{F}(\hat{F}_t(\mathbf{x}_0))}{\widehat{G}}\right].
    • Compute the DDPM posterior variance: σt2=βt(1−αˉt−1)1−αˉt\sigma_t^2 = \frac{\beta_t(1-\bar{\alpha}_{t-1})}{1-\bar{\alpha}_t}.
    • Sample Gaussian noise: z∼N(0,I)\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}).
    • Perform the reverse diffusion update: xt−1DDPM=αˉt−1F^t−1(x0)+1−αˉt−1−σt2ϵθ+σtz\mathbf{x}_{t-1}^{\mathrm{DDPM}} = \sqrt{\bar{\alpha}_{t-1}}\hat{F}_{t-1}(\mathbf{x}_0)+\sqrt{1-\bar{\alpha}_{t-1}-\sigma_t^2}\boldsymbol{\epsilon}_\theta+\sigma_t\mathbf{z}.
    • Compute the physical constraint: Lphys=∥∇2F^t(x0)∥22\mathcal{L}_{\mathrm{phys}} = \left\|\nabla^2\hat{F}_t(\mathbf{x}_0)\right\|_2^2.
    • Apply physics guidance: xt−1=xt−1DDPM−λ∇xtLphys\mathbf{x}_{t-1} = \mathbf{x}_{t-1}^{\mathrm{DDPM}}-\lambda\nabla_{\mathbf{x}_t}\mathcal{L}_{\mathrm{phys}}.
  4. Return the reconstructed field: x0\mathbf{x}_0.

All PhyDiM results have been regenerated using the consistent sampler, including Table 1, the qualitative figures, the diurnal curves, and the appendix results. In the ablation study (Table 2), the previous sampler is retained as Config 3 (without the inverse heat operator), while the new sampler is reported in the last row. The gains over DDPM have increased. Averaged across regions, PhyDiM now improves RMSE by 4.53% and SSIM by 20.47%, compared with the previous improvements of 2.1% and 12%, respectively. The abstract and conclusion have also been updated accordingly. We will also update the associated code to reflect these corrections in the camera-ready version, if accepted.

AuthorsOpenReview ↗

Response to Reviewer uYEb (3/4)

Weakness 3: Performance Improvements and Computational Overhead of the Physics-Guided Components

We believe this weakness is partially addressed by the revised physics-guided sampling procedure, as suggested by the reviewers. With the revised sampler, PhyDiM improves over DDPM by an average of 4.53% in RMSE and 20.47% in SSIM, compared to 2.1% and 12% with the previous sampler. The gains increase in all three ROIs. For example, the SSIM gains over DDPM at h1h_1-h4h_4 rise from 7.32%, 14.55%, 11.67%, 13.33% to 26.32%, 25.45%, 18.33%, 26.67% in ROI$_1$, from 8.51%, 5.80%, 7.25%, 12.77% to 19.15%, 13.04%, 11.59%, 23.40% in ROI$_2$, and from 15.15%, 12.50%, 11.67%, 12.90% to 30.30%, 21.43%, 20.00%, 19.35% in ROI$_3$. RMSE improves in every setting except h4h_4 in ROI$_1$ and ROI$_2$, where DDPM remains slightly better (the gap narrows from 0.07 to 0.06 in ROI$_1$ and from 0.07 to 0.02 in ROI$_2$).

Regarding computational complexity, PhyDiM adds an average of only 1.19 s to standard DDPM inference (3.93s vs. 2.74s, see the ablation study in Sec. 4.4). To quantify this trade-off, we define the gain efficiency as the relative improvement over DDPM (%) divided by the additional inference time (s). For PhyDiM, the 4.53% RMSE reduction and 20.47% SSIM improvement correspond to gain efficiencies of approximately 3.83.8 and 17.217.2% per second, respectively. In other words, each additional second of inference yields, on average, an approximately 3.8% reduction in RMSE and a 17.2% improvement in SSIM relative to DDPM. This analysis has been added to the ablation study in Sec. 4.4.

Weakness 4: Methodological Novelty

PhyDiM does use known building blocks (DDPM, the heat equation). Our contribution lies in how physical information is incorporated into the diffusion process, which distinguishes PhyDiM from existing physics-informed diffusion models in three main aspects.

  1. First, existing physics-informed diffusion models, such as [4], incorporate PDE residuals as training losses, while DiffusionPDE [5] uses PDE-residual guidance during sampling. In both approaches, the forward corruption process remains standard (We have expanded the comparison with these methods and clarified the methodological novelty of PhyDiM in the third paragraph of the Introduction). In contrast, PhyDiM modifies the corruption operator itself, xt=αˉtFt(x0)+1−αˉt ϵ\mathbf{x}_t=\sqrt{\bar\alpha_t}F_t(\mathbf{x}_0)+\sqrt{1-\bar\alpha_t}\,\epsilon, so that the denoiser learns to reverse a spatially coupled and structured degradation. This modification leaves the standard noise-prediction objective unchanged, as shown in Appendix A, and avoids introducing additional physics-based training losses and their associated loss-balancing requirements. Moreover, this formulation is independent of the denoiser architecture and can be used with different denoising networks.

  2. Second, modifying the forward process requires a mathematically consistent reverse process, which cannot be obtained simply by combining existing components. As highlighted by the reviewers' comments on the reverse transition, a standard DDPM reverse step is not consistent with the time-dependent heat evolution introduced in the forward process. We therefore derive a reverse update that applies the exact inverse of the discretized heat operator, implemented as element-wise division in the Fourier domain (Appendix B), while preserving the forward marginals. We further provide a theoretical analysis of the physics-guidance step in Appendix C. The ablation study in Table 2 isolates the contributions of the different components and shows that the full model improves over the variants.

  3. Third, we design a conditional multi-temporal architecture that maps three acquisitions and a continuous target hour to the field at that hour. It handles both interpolation and extrapolation, which goes beyond standard video frame interpolation and is important for real-world monitoring of the spatio-temporal evolution of image sequences. We evaluate it on LST, and the framework can be adapted to other systems by replacing the heat equation with a PDE whose discretized step is invertible (Algorithms 1 and 2).

[4] Jan-Hendrik Bastek, WaiChing Sun, and Dennis Kochmann. Physics-informed diffusion models. In International Conference on Learning Representations, volume 2025, pp. 3360–3385, 2025.

[5] Huang, Jiahe, et al. Diffusionpde: Generative pde-solving under partial observation. Advances in Neural Information Processing Systems 37 (2024): 130291-130323.

Reviewer_W6p8OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PhyDiM, a physics-guided diffusion model for land surface temperature (LST) reconstruction from sparse satellite observations. The method has three main components: a deterministic heat-equation-based evolution followed by Gaussian perturbation in the forward diffusion process; a conditional U-Net that integrates multi-temporal observations and their timestamps through context encoding and cross-attention; and a Laplacian-energy-based gradient correction during reverse sampling. Experiments on three European regions compare the proposed method with classical interpolation techniques, deep learning models, neural operators, and diffusion baselines.

Critical strengths
  1. The Introduction provides a well-organized overview of prior work, covering both traditional methods and recent research approaches. This review helps contextualize the proposed method within the existing literature.
  2. Appendix A provides a theoretical justification for retaining the original noise-prediction training objective when noise is added to the physically evolved state , rather than directly to the original sample . This analysis strengthens the methodological foundation of the proposed approach.
  3. The research problem is meaningful and relevant, and the proposed direction has promising potential for future development and extension to other applications.
Critical weaknesses
  1. Figure 5 may be potentially misleading. The method described in the text appears to add noise to each physically evolved image x~t\tilde{x}_t to obtain xtx_t, whereas Figure 5 seems to depict a process in which noise is progressively added to x~T\tilde{x}_T to obtain the noisy image.

  2. The notation is potentially confusing, as tt is used to represent both the diffusion timestep and physical time.

  3. The abstract states, we introduce PhyDiM, a physics-guided diffusion model that incorporates partial differential equation (PDE) priors directly into the diffusion process. The abstract may overstate the extent of PDE integration by stating that PDE priors are incorporated directly into the diffusion process. The reverse sampling process uses a Laplacian-based smoothness penalty rather than explicitly enforcing the full PDE dynamics. Clarifying this distinction would provide a more precise description of the proposed method.

  4. The average RMSE reduction over DDPM is only 2.1%, despite a more substantial SSIM improvement. Since the main contribution is a physics-guided diffusion model, comparisons with more recent diffusion-based methods would better demonstrate its effectiveness. Given the additional inference cost, the authors should also clarify whether the performance gains justify the overhead.

  5. The related-work review is incomplete and does not sufficiently cover closely related diffusion-based methods for physics-informed reconstruction. In particular, DiffusionPDE [1] and SDIFT [2] are highly relevant to the problem setting considered in this work, but are not adequately discussed. Authors are encouraged to clarify the differences between them.

[1] Huang, Jiahe, et al. "Diffusionpde: Generative pde-solving under partial observation." Advances in Neural Information Processing Systems 37 (2024): 130291-130323. [2] Chen, Panqi, et al. "Generating full-field evolution of physical dynamics from irregular sparse observations." Advances in Neural Information Processing Systems 38 (2026): 133720-133756.

Questions for authors
  1. The paper refers to “fixed and irregularly spaced timestamps.” Could this setting limit the model's generalization ability, particularly its ability to reconstruct images at timestamps that are not included in the training set?

  2. In Eq. (7), could directly subtracting the gradient of the regularization term from xtx_t cause the updated sample xt−1x_{t-1} to deviate from the data manifold? Refer to [1].

  3. In Algorithm 2, should x^0\hat{x}_0 in line 6 instead be interpreted as an estimate of Ft(x0)‾\overline{F_t(x_0)}?

If so, should the expected value of xt−1x_{t-1} in line 8 involve αˉt−1Ft(x0)\sqrt{\bar{\alpha}_{t-1}} F_t(x_0) rather than $ \sqrt{\bar{\alpha}{t-1}} F{t-1}(x_0)$? Could this potential inconsistency affect the validity of the reverse transition or the convergence of the sampling process?

  1. The paper claims arbitrary-time LST reconstruction, but direct LST ground truth is only available at four MODIS acquisition times per day. How does the current evaluation validate reconstruction accuracy at truly unobserved intermediate times, especially since the Appendix C hourly validation is indirect and relies on an LST-to-NSAT conversion model?

[1] Hamidi S M, Yang E H. \textit{Score-Based Manifold Projection for Diffusion-Based Inverse Problems} [C]//2025 IEEE International Symposium on Information Theory (ISIT). IEEE, 2025: 1--6.

8 comments
AuthorsOpenReview ↗

Response to Reviewer W6p8 (1/5)

We sincerely thank the reviewer for the careful reading of our manuscript, the positive assessment of our methodology, and the constructive feedback. We would particularly like to highlight Question 3, which raised an important methodological concern also noted by Reviewer GmWZ. Addressing this issue strengthened the theoretical validity of the reverse transition and improved PhyDiM's performance over DDPM across RMSE and SSIM.

We therefore address Question 3 first, as the responses to the remaining comments depend on the corrections made in response to this concern. For ease of navigation, all additions to the revised manuscript are highlighted in red.

Question 3: Theoretical Derivation of the Reverse Diffusion Process

We agree with the reviewer on this point. In the original Algorithm 2, x^0\hat{\mathbf{x}}_0 estimated $F_t(\mathbf{x}_0)$ rather than x0\mathbf{x}_0. Consequently, the standard DDPM reverse update targeted αˉt−1Ft(x0)\sqrt{\bar{\alpha}_{t-1}}F_t(\mathbf{x}_0) instead of the required αˉt−1Ft−1(x0)\sqrt{\bar{\alpha}_{t-1}}F_{t-1}(\mathbf{x}_0). Thus, the original sampler was not theoretically consistent with the time-dependent forward process, and we acknowledge that Laplacian guidance alone does not resolve this inconsistency.

Revised sampler (Sec 3.4, Alg. 2). We now interpret the network output as an estimate of the physically evolved field, F^t(x0)\hat F_t(\mathbf{x}_0), and apply the inverse heat evolution to step back to the previous physical state, F^t−1(x0)\hat F_{t-1}(\mathbf{x}_0), before performing the reverse update. The explicit Euler step is linear and shift-invariant, so under periodic boundary conditions, it becomes a per-frequency multiplication, F(Ft)=G^ F(Ft−1)\mathcal{F}(F_t) = \widehat G\ \mathcal{F}(F_{t-1}), where G^=1+κΔτ[2cos⁡(2πkx)+2cos⁡(2πky)−4]>0\widehat G = 1 + \kappa\Delta\tau[2\cos(2\pi k_x)+2\cos(2\pi k_y)-4] > 0.

The inverse is therefore an exact element-wise division in the Fourier domain, requiring no matrix inversion. (A detailed derivation has been added in Appendix B) The resulting reverse update is therefore:

xt−1=αˉt−1F^t−1(x0)+1−αˉt−1−σt2ϵθ+σtz,\mathbf{x}_{t-1} = \sqrt{\bar{\alpha}_{t-1}}\hat F_{t-1}(x_0) + \sqrt{1-\bar{\alpha}_{t-1}-\sigma_t^2}\epsilon_\theta + \sigma_t\mathbf{z},

with F^t−1(x0)=F−1[F(F^t(x0))/G^],\hat F_{t-1}(x_0) = \mathcal{F}^{-1}\left[\mathcal{F}(\hat F_t(x_0))/\widehat G\right], where σt2\sigma_t^2 denotes the DDPM posterior variance. The division is well-defined since G^>0\widehat{G}>0 at all frequencies for kd<1/8k_d<1/8. Therfore, this update preserves the forward marginals: qσ(xt−1∣x0)=N(αˉt−1Ft−1(x0),(1−αˉt−1)I)q_\sigma(\mathbf{x}_{t-1}\mid\mathbf{x}_0) = \mathcal{N}\big(\sqrt{\bar{\alpha}_{t-1}}F_{t-1}(\mathbf{x}_0),(1-\bar{\alpha}_{t-1})\mathbf{I}\big).

The training objective remains unchanged, and the final iterate estimates x0=F0(x0)\mathbf{x}_0 = F_0(\mathbf{x}_0). Laplacian guidance is retained as a separate soft prior, applied every 50 steps.

Revised sampling procedure (Algorithm 2):

  1. Initialize the noisy field: xT∼N(0,I)\mathbf{x}_T \sim \mathcal{N}(0,\mathbf{I}).
  2. Compute the Fourier-domain heat operator: G^(kx,ky)=1+kd[2cos⁡(2πkx)+2cos⁡(2πky)−4]\widehat{G}(k_x,k_y) = 1+k_d[2\cos(2\pi k_x)+2\cos(2\pi k_y)-4].
  3. For t=T,…,1t=T,\ldots,1:
    • Predict noise: ϵθ=ϵθ(xt,C,h∗,t)\boldsymbol{\epsilon}_\theta = \epsilon_\theta(\mathbf{x}_t,\mathcal{C},h^*,t).
    • Estimate the physically evolved field: F^t(x0)=1αˉt(xt−1−αˉtϵθ)\hat{F}_t(\mathbf{x}_0) = \frac{1}{\sqrt{\bar{\alpha}_t}}\left(\mathbf{x}_t-\sqrt{1-\bar{\alpha}_t}\boldsymbol{\epsilon}_\theta\right).
    • Apply inverse heat evolution: F^t−1(x0)=F−1[F(F^t(x0))G^]\hat{F}_{t-1}(\mathbf{x}_0) = \mathcal{F}^{-1}\left[\frac{\mathcal{F}(\hat{F}_t(\mathbf{x}_0))}{\widehat{G}}\right].
    • Compute the DDPM posterior variance: σt2=βt(1−αˉt−1)1−αˉt\sigma_t^2 = \frac{\beta_t(1-\bar{\alpha}_{t-1})}{1-\bar{\alpha}_t}.
    • Sample Gaussian noise: z∼N(0,I)\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}).
    • Perform the reverse diffusion update: xt−1DDPM=αˉt−1F^t−1(x0)+1−αˉt−1−σt2ϵθ+σtz\mathbf{x}_{t-1}^{\mathrm{DDPM}} = \sqrt{\bar{\alpha}_{t-1}}\hat{F}_{t-1}(\mathbf{x}_0)+\sqrt{1-\bar{\alpha}_{t-1}-\sigma_t^2}\boldsymbol{\epsilon}_\theta+\sigma_t\mathbf{z}.
    • Compute the physical constraint: Lphys=∥∇2F^t(x0)∥22\mathcal{L}_{\mathrm{phys}} = \left\|\nabla^2\hat{F}_t(\mathbf{x}_0)\right\|_2^2.
    • Apply physics guidance: xt−1=xt−1DDPM−λ∇xtLphys\mathbf{x}_{t-1} = \mathbf{x}_{t-1}^{\mathrm{DDPM}}-\lambda\nabla_{\mathbf{x}_t}\mathcal{L}_{\mathrm{phys}}.
  4. Return the reconstructed field: x0\mathbf{x}_0.

Updated results. All PhyDiM results have been regenerated using the consistent sampler, including quantitative and qualitative results, diurnal curves, and appendix experiments. Table 2 retains the previous sampler as Config 3 and reports the revised sampler in the final row. The gains over DDPM have increased. Averaged across regions, PhyDiM now improves RMSE by 4.53% and SSIM by 20.47%, compared with the previous improvements of 2.1% and 12%, respectively. The abstract and conclusion have also been updated accordingly. We will also update the associated code to reflect these corrections in the camera-ready version, if accepted.

AuthorsOpenReview ↗

Response to Reviewer W6p8 (2/5)

Weakness 1: Revision of Figure 2

We believe the comment refers to Fig. 2(a), which depicts the forward process, rather than Fig. 5 (diurnal reconstructions).

We agree that Fig. 2(a) was potentially misleading. It followed the conventional DDPM illustration of progressively increasing corruption, which does not accurately represent the process implemented in PhyDiM. In PhyDiM, xt\mathbf{x}_t is obtained by applying tt heat-equation steps to x0\mathbf{x}_0, followed by a single Gaussian perturbation (Eq. 5). To address this issue, we have made the following revisions:

  • Redrawn Fig. 2(a) to show that the physically evolved field Ft(x0)F_t(\mathbf{x}_0) is perturbed by a single Gaussian noise draw.
  • Updated Fig. 2(b) to show the revised sampler, in which the inverse heat evolution is computed before the DDPM update (Alg. 2).
  • Aligned the notation in both panels with the main text, using Ft(x0)F_t(\mathbf{x}_0) (instead of x~t\tilde{\mathbf{x}}_t), F^t(x0)\hat{F}_t(\mathbf{x}_0), and F^t−1(x0)\hat{F}_{t-1}(\mathbf{x}_0).

We also corrected Sec. 3.2, where this panel was incorrectly cited as Fig. 1(a).

Weakness 2: Clarification of the Notation and Separation of Time Variables

We agree with the reviewer that using tt to represent both the diffusion timestep and physical time could lead to confusion. We have therefore revised the notation throughout the manuscript to avoid this overlap. In the revised manuscript, tt denotes only the diffusion timestep, while the other quantities use distinct symbols:

  • hih_i and h∗h^{*} denote the satellite acquisition hours and target hour, respectively (Secs. 3.1 and 3.3). Accordingly, the target groups are now denoted by h1h_1-h4h_4 in Table~1 and the figures.
  • τ\tau and Δτ\Delta\tau denote the heat-equation pseudo-time and Euler step, respectively (Eqs. 3 and 5, Alg. 1, and the hyperparameter descriptions), with κΔτ\kappa\Delta\tau representing the effective diffusion strength.
  • g\mathbf{g} denotes the U-Net feature maps in Sec.3.3, avoiding a notational clash with hh.

Weakness 3: Clarification of the PDE Integration in the Abstract

We have revised the abstract to clarify the specific role of the heat-equation prior at each stage of the diffusion process:

  • Forward process: the clean field is evolved by the heat equation before Gaussian noise is added.
  • Reverse process: with the revised sampler, each step applies the inverse heat evolution, keeping the reverse update consistent with the forward process, followed by a Laplacian regularization.

The revised abstract reads:

"In this paper, we introduce PhyDiM, a physics-guided diffusion model that integrates a heat-equation prior at two stages of the diffusion process. In the forward process, the clean field is first evolved by a heat partial differential equation before Gaussian noise is added. In the reverse process, the sampler inverts this evolution at each step, with a lightweight Laplacian regularization promoting spatially coherent solutions. PhyDiM is applied to land surface temperature reconstruction from sparsely sampled satellite observations."

AuthorsOpenReview ↗

Response to Reviewer W6p8 (3/5)

Weakness 4: Performance Gains and Computational Overhead

We believe this weakness is partially addressed by the revised physics-guided sampling procedure, as suggested by the reviewer. With the revised sampler, PhyDiM improves over DDPM by an average of 4.53% in RMSE and 20.47% in SSIM, compared to 2.1% and 12%, respectively, with the previous sampler.

Regarding inference cost, PhyDiM adds an average of only 1.19 s to standard DDPM inference (3.93 s vs. 2.74 s, see the ablation study in Sec. 4.4). To quantify this trade-off, we define gain efficiency as the relative improvement over DDPM (%) divided by the additional inference time (s). For PhyDiM, the 4.53% RMSE reduction and 20.47% SSIM improvement correspond to gain efficiencies of approximately 3.83.8% and 17.217.2% per second, respectively. In other words, each additional second of inference yields an approximately 3.8% reduction in RMSE and a 17.2% improvement in SSIM relative to DDPM. This gain-efficiency analysis has also been added to the ablation study in Sec.4.4.

Regarding comparisons with more recent diffusion-based methods, existing physics-informed diffusion models incorporate physical constraints only through additional losses optimized during training. In contrast, our physics-guided formulation is independent of the denoiser. It modifies both the forward corruption (Alg. 1) and the sampling procedure (Alg. 2). It can therefore be applied to any diffusion model, including those with more complex denoising networks, without changing their architecture or training objective.

Our evaluation already includes classical interpolation methods, a transformer-based video frame interpolation model (VFIT), the Fourier Neural Operator (FNO), the Physics-Informed Neural Operator (PINO), a standard DDPM, and a physics-informed DDPM (ICLR 2025). The two additional works suggested by the reviewer in the following weakness are difficult to adapt to our setting. DiffusionPDE is designed for forward and inverse problems governed by a known PDE, where the governing equation and its associated residual can be incorporated directly into the sampling process. Our task, however, is LST reconstruction from sparse satellite observations, for which the heat equation serves as a guiding physical model rather than the exact governing equation, making a direct comparison unsuitable. SDIFT, on the other hand, is purely data-driven and incorporates no physical constraints. Its latent representation is designed for irregular point observations rather than our image-based setting. We therefore believe that our selected baselines provide a relevant evaluation of the proposed methodology against existing approaches.

Weakness 5: Expanded Related Work and Comparison with DiffusionPDE and SDIFT

We have expanded the related-work discussion in the Introduction (paragraph 3) to include both DiffusionPDE and SDIFT, with an explicit comparison to PhyDiM.

  • DiffusionPDE models the joint distribution of PDE coefficients and solutions and guides sampling using sparse observations and a PDE-residual term. It addresses forward and inverse problems governed by a known PDE. Its physical constraints are incorporated during sampling, while the forward corruption process remains standard. In contrast, PhyDiM targets the reconstruction of real-world LST at unobserved hours from timestamped satellite observations, and incorporates the heat-equation operator directly into the forward corruption process.
  • SDIFT reconstructs full-field evolution from irregular sparse observations using sequential diffusion in a functional Tucker latent space with Gaussian-process noise. Although its reconstruction setting is closely related to ours, SDIFT is purely data-driven and does not explicitly incorporate governing physical equations. Its inductive bias instead comes from its latent representation. We therefore discuss it as a data-driven diffusion approach for physical field reconstruction, rather than as a physics-informed method.
Reviewer_GmWZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PhyDiM reconstructs land surface temperature at requested times from three satellite observations. It modifies forward diffusion by applying heat-equation smoothing before Gaussian corruption and adds Laplacian-based guidance during sampling. A conditional U-Net integrates observation and timestamp information. Experiments cover three European regions, with additional indirect validation against ground-based air-temperature measurements.

Critical strengths
  1. Reconstructing temperature between sparse satellite overpasses is practically relevant. The dataset spans multiple regions and years, with explicitly separated temporal splits.
  2. Table 1 reports higher SSIM than DDPM across all listed region–time combinations, although temperature-error improvements are less uniform.
  3. The paper separates forward modification from sampling guidance, reports runtime, and attempts independent sensor-based validation while acknowledging that LST and air temperature differ.
Critical weaknesses
  1. The heat evolution operates along artificial diffusion time, while the sampling energy constrains an individual image’s spatial smoothness. Neither explicitly enforces a relationship between reconstructions at different physical times. The model also does not include forcing terms for diurnal heating and cooling. Consequently, smoother outputs should not be interpreted as demonstrating compliance with the governing temporal physics. A matched spatial-smoothing baseline is needed to distinguish physical modeling benefits from regularization benefits.
  2. The algorithm nevertheless labels this estimate as the clean field and applies a standard DDPM update. A reverse-process derivation consistent with the time-dependent heat transformation is needed; Laplacian guidance does not itself supply that derivation.
  3. Direct LST evaluation is restricted to satellite acquisition times. Appendix C adds useful evidence, but evaluates a composite pipeline: reconstructed LST is converted to air temperature by a retrained SPyCer model using temporal and environmental inputs. This model may explain part of the observed diurnal agreement, and no corresponding comparison using DDPM, interpolated LST, or reduced-input controls is reported. Held-out sensor locations validate the composite predictor’s spatial generalization, but do not isolate the accuracy of PhyDiM’s unobserved hourly LST.
Questions for authors

see weaknesses

5 comments
AuthorsOpenReview ↗

Response to Reviewer GmWZ (1/3)

We sincerely thank the reviewer for the careful reading of our manuscript, the positive assessment of our methodology, and the constructive feedback. We would particularly like to highlight Weakness 2, which raised an important methodological concern also noted by Reviewer W6p8. Addressing this issue strengthened the theoretical validity of the reverse transition and improved PhyDiM's performance over DDPM across RMSE and SSIM.

We therefore address Comment 2 first, as the responses to the remaining comments depend on the corrections made in response to this concern. For ease of navigation, all additions to the revised manuscript are highlighted in red.

Weakness 2: Theoretical Derivation of the Reverse Diffusion Process

We agree with the reviewer on this point. In the original Algorithm 2, x^0\hat{\mathbf{x}}_0 estimated $F_t(\mathbf{x}_0)$ rather than x0\mathbf{x}_0. Consequently, the standard DDPM reverse update targeted αˉt−1Ft(x0)\sqrt{\bar{\alpha}_{t-1}}F_t(\mathbf{x}_0) instead of the required αˉt−1Ft−1(x0)\sqrt{\bar{\alpha}_{t-1}}F_{t-1}(\mathbf{x}_0). Thus, the original sampler was not theoretically consistent with the time-dependent forward process, and we acknowledge that Laplacian guidance alone does not resolve this inconsistency.

Revised sampler (Sec 3.4, Alg. 2). We now interpret the network output as an estimate of the physically evolved field, F^t(x0)\hat F_t(\mathbf{x}_0), and apply the inverse heat evolution to step back to the previous physical state, F^t−1(x0)\hat F_{t-1}(\mathbf{x}_0), before performing the reverse update. The explicit Euler step is linear and shift-invariant, so under periodic boundary conditions, it becomes a per-frequency multiplication, F(Ft)=G^ F(Ft−1)\mathcal{F}(F_t) = \widehat G\ \mathcal{F}(F_{t-1}), where G^=1+κΔτ[2cos⁡(2πkx)+2cos⁡(2πky)−4]>0\widehat G = 1 + \kappa\Delta\tau[2\cos(2\pi k_x)+2\cos(2\pi k_y)-4] > 0.

The inverse is therefore an exact element-wise division in the Fourier domain, requiring no matrix inversion. (A detailed derivation has been added in Appendix B) The resulting reverse update is therefore:

xt−1=αˉt−1F^t−1(x0)+1−αˉt−1−σt2ϵθ+σtz,\mathbf{x}_{t-1} = \sqrt{\bar{\alpha}_{t-1}}\hat F_{t-1}(x_0) + \sqrt{1-\bar{\alpha}_{t-1}-\sigma_t^2}\epsilon_\theta + \sigma_t\mathbf{z},

with F^t−1(x0)=F−1[F(F^t(x0))/G^],\hat F_{t-1}(x_0) = \mathcal{F}^{-1}\left[\mathcal{F}(\hat F_t(x_0))/\widehat G\right], where σt2\sigma_t^2 denotes the DDPM posterior variance. The division is well-defined since G^>0\widehat{G}>0 at all frequencies for kd<1/8k_d<1/8. Therfore, this update preserves the forward marginals: qσ(xt−1∣x0)=N(αˉt−1Ft−1(x0),(1−αˉt−1)I)q_\sigma(\mathbf{x}_{t-1}\mid\mathbf{x}_0) = \mathcal{N}\big(\sqrt{\bar{\alpha}_{t-1}}F_{t-1}(\mathbf{x}_0),(1-\bar{\alpha}_{t-1})\mathbf{I}\big).

The training objective remains unchanged, and the final iterate estimates x0=F0(x0)\mathbf{x}_0 = F_0(\mathbf{x}_0). Laplacian guidance is retained as a separate soft prior, applied every 50 steps.

Revised sampling procedure (Algorithm 2):

  1. Initialize the noisy field: xT∼N(0,I)\mathbf{x}_T \sim \mathcal{N}(0,\mathbf{I}).
  2. Compute the Fourier-domain heat operator: G^(kx,ky)=1+kd[2cos⁡(2πkx)+2cos⁡(2πky)−4]\widehat{G}(k_x,k_y) = 1+k_d[2\cos(2\pi k_x)+2\cos(2\pi k_y)-4].
  3. For t=T,…,1t=T,\ldots,1:
    • Predict noise: ϵθ=ϵθ(xt,C,h∗,t)\boldsymbol{\epsilon}_\theta = \epsilon_\theta(\mathbf{x}_t,\mathcal{C},h^*,t).
    • Estimate the physically evolved field: F^t(x0)=1αˉt(xt−1−αˉtϵθ)\hat{F}_t(\mathbf{x}_0) = \frac{1}{\sqrt{\bar{\alpha}_t}}\left(\mathbf{x}_t-\sqrt{1-\bar{\alpha}_t}\boldsymbol{\epsilon}_\theta\right).
    • Apply inverse heat evolution: F^t−1(x0)=F−1[F(F^t(x0))G^]\hat{F}_{t-1}(\mathbf{x}_0) = \mathcal{F}^{-1}\left[\frac{\mathcal{F}(\hat{F}_t(\mathbf{x}_0))}{\widehat{G}}\right].
    • Compute the DDPM posterior variance: σt2=βt(1−αˉt−1)1−αˉt\sigma_t^2 = \frac{\beta_t(1-\bar{\alpha}_{t-1})}{1-\bar{\alpha}_t}.
    • Sample Gaussian noise: z∼N(0,I)\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}).
    • Perform the reverse diffusion update: xt−1DDPM=αˉt−1F^t−1(x0)+1−αˉt−1−σt2ϵθ+σtz\mathbf{x}_{t-1}^{\mathrm{DDPM}} = \sqrt{\bar{\alpha}_{t-1}}\hat{F}_{t-1}(\mathbf{x}_0)+\sqrt{1-\bar{\alpha}_{t-1}-\sigma_t^2}\boldsymbol{\epsilon}_\theta+\sigma_t\mathbf{z}.
    • Compute the physical constraint: Lphys=∥∇2F^t(x0)∥22\mathcal{L}_{\mathrm{phys}} = \left\|\nabla^2\hat{F}_t(\mathbf{x}_0)\right\|_2^2.
    • Apply physics guidance: xt−1=xt−1DDPM−λ∇xtLphys\mathbf{x}_{t-1} = \mathbf{x}_{t-1}^{\mathrm{DDPM}}-\lambda\nabla_{\mathbf{x}_t}\mathcal{L}_{\mathrm{phys}}.
  4. Return the reconstructed field: x0\mathbf{x}_0.

Updated results. All PhyDiM results have been regenerated using the consistent sampler, including quantitative and qualitative results, diurnal curves, and appendix experiments. Table 2 retains the previous sampler as Config 3 and reports the revised sampler in the final row. The gains over DDPM have increased. Averaged across regions, PhyDiM now improves RMSE by 4.53% and SSIM by 20.47%, compared with the previous improvements of 2.1% and 12%, respectively. The abstract and conclusion have also been updated accordingly. We will also update the associated code to reflect these corrections in the camera-ready version, if accepted.

AuthorsOpenReview ↗

Response to Reviewer GmWZ (2/3)

Weakness 1: Distinguishing Physics-Guided Modeling from Spatial Smoothing

We agree with the reviewer’s characterization of the physical components. Rather than modeling the full surface energy balance, which requires unobserved quantities such as net radiation, soil moisture, and turbulent heat fluxes, PhyDiM incorporates physical priors suited to remote sensing imagery. Heat evolution governs spatial propagation during forward corruption, while the sampling energy promotes spatial coherence in reconstructed images. The diurnal cycle is learned by the conditional multi-temporal cross-attention architecture of the denoiser network.

To distinguish the benefits of physical modeling from those of spatial regularization, as suggested, we added a matched spatial-smoothing baseline. Specifically, we apply Gaussian filters with σ∈{0.25,0.5,1,1.5,2,2.5,3}\sigma\in\{0.25,0.5,1,1.5,2,2.5,3\} pixels to the DDPM reconstructions and compare the resulting outputs with PhyDiM (Appendix D).

The table is provided below. For clarity and ease of reading, we have divided it into three separate tables, one for each region. We report only RMSE and SSIM here (Following another reviewer's suggestion, we now use tt exclusively for the diffusion timestep, hh for physical hours, and τ\tau for heat-equation pseudo-time to avoid ambiguity). Please refer to Appendix D for the full comparison across all evaluation metrics.

  1. ROI1:
Methodh1h2h3h4
RMSESSIMRMSESSIMRMSESSIMRMSESSIM
DDPM (σ=0.0\sigma=0.0)1.970.381.540.551.850.602.180.45
DDPM (σ=0.25\sigma=0.25)1.930.421.540.551.850.602.180.45
DDPM (σ=0.5\sigma=0.5)1.930.421.470.601.800.642.130.50
DDPM (σ=1.0\sigma=1.0)1.900.431.430.621.790.642.090.52
DDPM (σ=1.5\sigma=1.5)1.900.431.450.571.820.602.080.48
DDPM (σ=2.0\sigma=2.0)1.910.351.470.521.850.552.080.43
DDPM (σ=2.5\sigma=2.5)1.910.321.490.471.870.502.090.38
DDPM (σ=3.0\sigma=3.0)1.920.291.510.431.900.472.090.35
PhyDiM1.820.481.410.691.670.712.240.57
  1. ROI2:
Methodh1h2h3h4
RMSESSIMRMSESSIMRMSESSIMRMSESSIM
DDPM (σ=0.0\sigma=0.0)2.140.472.280.692.970.692.400.47
DDPM (σ=0.25\sigma=0.25)2.140.472.280.692.970.702.400.47
DDPM (σ=0.5\sigma=0.5)2.100.512.250.722.940.722.360.51
DDPM (σ=1.0\sigma=1.0)2.070.522.250.712.940.702.340.52
DDPM (σ=1.5\sigma=1.5)2.080.482.310.642.990.642.340.49
DDPM (σ=2.0\sigma=2.0)2.090.432.370.573.040.582.340.44
DDPM (σ=2.5\sigma=2.5)2.100.392.420.513.090.522.350.40
DDPM (σ=3.0\sigma=3.0)2.110.362.470.463.140.472.360.37
PhyDiM2.060.562.160.782.840.772.420.58
  1. ROI3:
Methodh1h2h3h4
RMSESSIMRMSESSIMRMSESSIMRMSESSIM
DDPM (σ=0.0\sigma=0.0)1.980.331.810.562.150.602.180.31
DDPM (σ=0.25\sigma=0.25)1.980.331.810.562.150.602.180.31
DDPM (σ=0.5\sigma=0.5)1.940.361.750.602.100.642.130.34
DDPM (σ=1.0\sigma=1.0)1.920.371.740.592.100.632.100.34
DDPM (σ=1.5\sigma=1.5)1.930.331.770.532.140.562.100.31
DDPM (σ=2.0\sigma=2.0)1.940.291.800.462.180.502.120.28
DDPM (σ=2.5\sigma=2.5)1.960.261.830.412.210.442.130.24
DDPM (σ=3.0\sigma=3.0)1.970.231.850.372.240.402.140.22
PhyDiM1.850.431.680.682.050.722.100.37

Post-hoc Gaussian smoothing yields only modest improvements over DDPM, with the best results around σ≈1\sigma\approx 1, beyond which SSIM declines. PhyDiM generally outperforms the best-smoothed DDPM in both RMSE and SSIM, particularly for the first three target groups across all ROIs. At h4h_4, DDPM retains a slightly lower RMSE in ROI$_1$ and ROI$_2$, although this advantage already exists without smoothing. PhyDiM achieves higher SSIM in both cases. Overall, these results suggest that PhyDiM's gains cannot be attributed to output smoothing alone, but rather to incorporating heat evolution directly into the generative process.

AuthorsOpenReview ↗

Response to Reviewer GmWZ (3/3)

Weakness 3: Evaluating Generalization to Unseen Hours and Hourly Reconstructions

We agree that our original evaluation did not explicitly assess generalization to unseen hours beyond the satellite acquisition times used for training. To address this limitation, we conducted two complementary evaluations of hourly reconstruction reliability.

First, we retrained PhyDiM from scratch while excluding one target hour (3h, 10h, 14h, or 21h) from the training data across all three regions. This ensures that the model never sees the omitted hour during training. We then evaluated each retrained model at the corresponding omitted hour and compared its performance with that of PhyDiM trained on all available hours. This evaluation has been added to the ablation study in Sec. 4.4.

The following table reports the results of this evaluation for the three regions, comparing PhyDiM trained with each target hour held out against PhyDiM trained on all available hours. The symbol # denotes the number of test samples at the corresponding hour.

Omitted hourTraining settingROI₁ROI₂ROI₃
#RMSE ↓SSIM ↑#RMSE ↓SSIM ↑#RMSE ↓SSIM ↑
3hPhyDiM (held-out hour)752.520.50992.100.56252.190.24
PhyDiM (full training)751.970.46991.960.56252.550.19
10hPhyDiM (held-out hour)1502.300.681252.000.681502.000.68
PhyDiM (full training)1501.450.691252.250.771501.680.68
14hPhyDiM (held-out hour)1502.550.681254.750.74753.140.72
PhyDiM (full training)1501.730.711252.660.78751.810.76
21hPhyDiM (held-out hour)1502.010.581252.560.541501.970.47
PhyDiM (full training)1502.010.581252.560.571502.100.37

SSIM is largely preserved at unseen hours, remaining within 0.04 of the full-training results in 10 of the 12 region-hour settings. This suggests that the model can transfer spatial structure to timestamps not seen during training. In terms of RMSE, performance is unchanged at 21h in ROI$_1$ and ROI$_2$, and improves in ROI$_3$. At 10h, the degradation is moderate in ROI$_3$ (from 1.68 to 2.00), while ROI$_2$ actually shows an improvement (from 2.25 to 2.00). The largest increase occurs in ROI$_1$ (from 1.45 to 2.30). The held-out model also outperforms the full-training model in some cases, notably at 3h and 21h in ROI$_3$, and in RMSE at 10h in ROI$_2$. The clearest degradation occurs at 14h, when temperatures peak and diurnal dynamics are fastest. Furthermore, on average, PhyDiM trained with a held-out hour still outperforms several baselines trained on all hours (Table 1). We acknowledge the increased errors around the temperature peak as a limitation and have stated this explicitly in the revised manuscript.

Second, as correctly pointed out by the reviewer, we use a separate model to infer NSAT from the reconstructed LST, since meteorological stations measure NSAT rather than LST. In the revised manuscript, we additionally report RMSE and MAE for each hour of the day, as shown in the following table:

Metric00:0001:0002:0003:0004:0005:0006:0007:0008:0009:0010:0011:00
#412410414410412414416420418422416420
RMSE ↓1.671.711.691.761.532.263.173.002.682.262.342.46
MAE ↓1.311.331.351.441.251.692.522.412.051.791.831.94
Metric12:0013:0014:0015:0016:0017:0018:0019:0020:0021:0022:0023:00
#418412416422420418414416416418418416
RMSE ↓2.702.812.972.882.742.642.211.721.601.751.711.74
MAE ↓2.142.242.412.382.272.151.751.331.231.281.341.36

PhyDiM generalizes well to unseen target hours, achieving RMSEs of 1.74, 1.67, and 1.71$^\circ$C at 23:00, 00:00, and 01:00, respectively, comparable to observed nighttime hours. Errors are similarly low at 19:00 and 22:00 (1.72 and 1.71$^\circ$C). Larger errors occur during rapid post-sunrise warming (06:00-08:00) and in the afternoon, peaking at 14:00. This is consistent with the held-out-hour experiment.

Taken together, the two evaluations show that PhyDiM can generalize to unseen hours.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

We sincerely thank all reviewers for their careful evaluation of our manuscript and their constructive comments and suggestions. We have carefully considered all feedback and made substantial revisions to improve the clarity, theoretical justification, and experimental validation of our work.

A summary of the main revisions is provided below:

  • Revised the reverse sampling algorithm to consistently invert the heat-based forward operator.
  • Added ablation studies, Gaussian-filtering comparisons, and further quantitative analyses to assess the contributions of individual components.
  • Updated the quantitative results, clarified the improvements over standard DDPM, and reported the additional inference time.
  • Expanded the evaluation on unseen target acquisition times.
  • Clarified the physical motivation and assumptions underlying the heat-equation prior, the distinction between diffusion pseudo-time and physical time, and the role of physics in the forward and reverse processes.
  • Strengthened the discussion of PhyDiM's novelty and its distinction from existing physics-informed diffusion methods, supported by additional theoretical analysis.

Below, we address each reviewer's comments individually and detail the corresponding revisions made to the manuscript. All changes in the revised manuscript are highlighted in red.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.