This manuscript introduces a deep learning-based data restoration method, ESTD-Net, aimed at recovering surface temperature from high-resolution observations. This method is built upon advanced models and restores temperature data with pixel-level accuracy, enhancing the quality of the reconstructed images. The manuscript writing motivation is clear, and the experimental results show improvements. However, there are still many issues that need to be further revised.

Major Comments:

1. The ESTD-Net proposed in the manuscript effectively solves the time complexity of traditional Transformers in image quality restoration tasks, but there is no experimental comparison with the training time of Transformer models presented in the article.
2. The manuscript presents the superiority of ESTD-Net in the task of recovering high-resolution temperature data, but lacks experimental comparisons with more recent extreme models for data recovery tasks, which cannot fully demonstrate the effectiveness of the proposed method. Additional evidence can be provided by comparing experimental results with more advanced models.
3. The overall model diagram and description are not clear enough. It is unclear how the Conv-U-Net in the second stage further improves the reconstruction accuracy. Additionally, the structure of the discriminator is unclear. Is it composed only of fully connected layers?
4. The overall workflow of training and inference is not very intuitive.
5. The manuscript directly reconstructs the processed  brightness temperature data. Is the  brightness temperature first processed into surface temperature and then verified with the surface temperature of ERA5? The description and rationality of the dataset need to be further explained.

Minor Comments:

1. Regarding the mask based contextual attention module proposed in section 3.3.1 of this manuscript, although a detailed calculation process is introduced, no corresponding figure is attached for explanation, which will reduce the readability of the paper.
2. This manuscript validated the impact of various loss factors in the loss function on model performance in the ablation experiment section, but it seems that there were no confirmatory experiments on the impact of contextual attention and other modules on the model, which may lead to ambiguity in the study of model architecture.

Overall, this article contributes a new method architecture in the field of surface temperature restoration. By introducing improved modules such as context attention mechanism, the overall edge feature extraction ability of the model is improved. I suggest that the authors resolve the aforementioned issues before proceeding with publication.

A. Summary

This manuscript presents ESTD-Net, a two-stage hybrid architecture for inpainting missing Earth Surface Temperature (EST) fields derived from MWRI/FY-3D satellite data. The first stage employs a modified Transformer (“boundary-aware” multi-head context attention, dynamic masks, no LayerNorm, concatenated residuals) to perform global reconstruction. The second stage refines outputs via a convolutional U-Net. Losses include a Weighted Reconstruction Loss, Gradient Consistency Regularization, and an adversarial GAN loss. Experiments on ERA5-simulated gaps compare ESTD-Net against inverse-distance weighting and a partial-conv U-Net, reporting improvements in MAE, RMSE, PSNR, and SSIM.

B. General Comments

1. High-Frequency Refinement by U-Net

Why do you claim that the convolutional U-Net in ESTD-Net can "refine" high-frequency details? What theoretical basis supports this? Could other network architectures achieve similar refinement? Which component specifically drives the refinement? Moreover, the U-Net design here is very crude—how should it be improved?

2. Missing Diffusion-Model Comparison

The absence of any diffusion-based baseline is a serious gap. By 2025, diffusion models aren’t just "nice to have," they’ve become the gold standard for high-fidelity inpainting. Showing that your hybrid Transformer-GAN outperforms a 2018 CNN is only a sanity check; it tells us nothing about where ESTD-Net sits relative to the true state of the art. I would strongly recommend benchmarking against at least one modern diffusion inpainting model 

e.g., RePaint in CVPR 2022 (see https://arxiv.org/abs/2201.09865) or Palette in SIGGRAPH 2022 (see https://arxiv.org/abs/2111.05826). 

You could also consult the state-of-the-art leaderboard for image inpainting (see https://paperswithcode.com/task/image-inpainting).

3. Limited Baselines and Ablation

The set of baselines is very limited, which makes it hard to highlight the proposed method’s advantages. Additionally, there is no comprehensive ablation study or stability analysis, and the choice of model hyperparameters is not discussed.

4. Lack of Domain-Specific Adaptation

The proposed method appears entirely generic—usable for standard image inpainting or sea-surface-temperature fields alike—without any targeted adaptation. It seems transplanted wholesale from computer vision, with no domain-specific modifications. Are surface-temperature gaps really analogous to arbitrary image holes? Authors should justify their design choices from a physical/meteorological standpoint.

C. Minor Comments

1. “Key Innovations” should be stated more objectively

The paper claims, "The model is augmented by two key innovations," yet both the "Weighted Reconstruction Loss" and the "Gradient Consistency Regularization" are commonplace in the image-inpainting literature. It is unclear whether these truly qualify as "innovations."

2. Definition of PSNR’s MAX Value

Why is the PSNR’s MAX defined as it is? How does this differ from standard computer-vision images? Does MAX relate to temperature units? Is it a constant or variable? Please explain the phrase on line 350: "the maximum possible pixel value of an image.

3. Edge-Case Temperature Variations

How does the proposed method perform in scenarios of rapid or high-amplitude temperature variation? These edge cases may reveal significant weaknesses.

4. Gradient-Consistency Implementation and Hyperparameter

Eq. (5)’s gradient-consistency term $L_{gp}$ may require second-order derivatives—how is this implemented in code and what is the computational cost? Why is $\alpha$ set to 0.001? Similar hyperparameter questions apply to the Gradient Consistency Regularization.

5. Equation Numbering and Punctuation

The definition of $M_{ij}^\prime$ is missing an equation number—it should be Eq. (2). Furthermore, none of the equations ends with a comma or period, which is non-standard.

6. MTB Module Design Motivation

The motivation for the MTB module in Figure 3 is unclear. Why is the concatenation placed as shown? Why the specific sequence MCA–C–FC–MLP? Why eliminate Layer Normalization? It is not explained how these choices realize the authors’ stated goal at line 245: "To address these challenges…"

7. Definition of FC vs. MLP

In Eq. (2), what exactly is "FC"? A fully connected layer (e.g. PyTorch’s nn.Linear)? How does FC differ from MLP? If FC plus an activation is an MLP, why distinguish them? The authors provide no discussion.

8. Adversarial Loss Stability

Eqs. (3) and (4) define the adversarial loss—yet is this formulation stable? Did the authors assess training convergence and robustness? In image tasks, these losses are notoriously unstable; more analysis is needed beyond Table 2 to support claims of "robust and reliable" performance.

D. Recommendation

Major Revision: The paper addresses an important interdisciplinary problem but currently lacks sufficient AI/ML rigor and contemporary benchmarking. Addressing the general comments—particularly theory behind U-Net refinement, domain-specific adaptations, comprehensive ablations, clarity on loss implementations, and inclusion of modern diffusion-model baselines—will be essential before the manuscript can be considered for acceptance.