Review Reports - Cross-Modality Spectral Expansion Combined with Physical–Semantic Dual Priors for Cloud Detection in GF-1 Imagery

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors are encouraged to elaborate on the pseudo-SWIR reconstruction process and to further explore the physical plausibility of the restored bands.
The semantic labels created by the LLM should be more elaborated covering the prompt design, model version, label verification method.
It would be helpful to clarify the dataset split, training and testing samples, and scene distribution for reproducibility.
Ablation experiments should be strengthened to better show the individual contribution of each suggested module.
The English and presentation may be enhanced for clarity.

Author Response

Comment 1

Reviewer comment: The authors are encouraged to elaborate on the pseudo-SWIR reconstruction process and to further explore the physical plausibility of the restored bands.

Response: Thank you for this important suggestion. We have further expanded the explanation of the pseudo-SWIR reconstruction process and clarified the role of the generated bands in this work. In the revised manuscript, the spectral expansion network is explicitly described as a task-oriented cross-modality transfer module. It maps four-band GF-1 inputs to a 13-channel pseudo-Sentinel-2A representation while preserving the GF-1 spatial grid. We also clarify that the reconstructed SWIR-like bands are not claimed to be complete physical replacements for real Sentinel-2 SWIR radiance/reflectance measurements; rather, they are used to preserve SWIR-sensitive cues that are useful for downstream cloud-snow discrimination. To support their physical plausibility, we added a Sentinel-2 band-holding experiment. In this experiment, Sentinel-2 B2, B3, B4, and B8 were used as four-band inputs to reconstruct B11 and B12, and the reconstructed SWIR-like bands were compared with real Sentinel-2 SWIR observations. The evaluation metrics include RMSE, Pearson correlation coefficient, SAM, and NDSI consistency. The RMSE values of reconstructed B11/B12 are 0.086228 and 0.064325, respectively; the corresponding correlation coefficients are 0.812641 and 0.852589; the SAM is 7.159606 degrees; and the NDSI correlation coefficient is 0.830303. These results indicate that the reconstructed bands preserve meaningful spatial and spectral variation trends, while also clarifying their limitation as proxy bands.

Location of changes in the revised manuscript: Section 3.1, lines 291-322; Section 4.2, lines 666-697; Discussion, lines 814-818; Table 2.

Comment 2

Reviewer comment: The semantic labels created by the LLM should be more elaborated covering the prompt design, model version, label verification method.

Response: We agree with the reviewer that the construction of LLM-derived semantic labels should be made more transparent. Therefore, we added a dedicated subsection entitled “LLM-Based Semantic Label Generation and Validation”. In the revised manuscript, we specify that DeepSeek-R1 is used to infer scene-level semantic attributes from the available metadata, including latitude-longitude coordinates, acquisition time, imaging angle, and sensor-related information. Instead of relying on unconstrained free-text generation, we use fixed binary prompts for each semantic attribute. For example, the prompt asks whether snow cover exists in the satellite image and requires the model to output only “1” or “0”. The same procedure is used for four attributes: snow cover, dense vegetation, desert surface, and large water bodies. We also added an automatic output-checking process: if the LLM output is not a valid binary result, the same question is re-issued until a valid output is obtained. Finally, the accepted labels are checked against the corresponding metadata background and representative image samples, and then concatenated into a four-dimensional semantic label vector.

Location of changes in the revised manuscript: Section 3.4.1, lines 517-564; Figure 8 caption.

Comment 3

Reviewer comment: It would be helpful to clarify the dataset split, training and testing samples, and scene distribution for reproducibility.

Response: Thank you for pointing out the need for more reproducibility details. We have expanded the dataset and experimental settings subsections. The revised manuscript states that the GF-1 data are from the GF1_WHU dataset, the original scenes were cropped into 512 x 512 image patches, patches containing invalid NoValue pixels were removed, and the labels were remapped for the cloud detection task. We also added details on Sentinel-2A preprocessing: before cross-modality training, the 13-channel Sentinel-2A target patches were spatially normalized to the same tensor size as the GF-1 patches. In addition, we explicitly state that the generated pseudo-Sentinel-2A representation preserves the GF-1 WFV spatial grid, patch size, and original 16 m spatial resolution. The two-stage training protocol is also described in more detail, including batch size, learning rate, number of training epochs, optimizer, hardware, and software environment. All reported ablation and comparison experiments are conducted on the same 22-scene test set to ensure fair evaluation.

Location of changes in the revised manuscript: Section 3.5.1, lines 566-599; Section 3.5.2, lines 600-612; Section 4.3, lines 700-750.

Comment 4

Reviewer comment: Ablation experiments should be strengthened to better show the individual contribution of each suggested module.

Response: We have strengthened the ablation experiments and revised the related analysis to make the contribution of each module clearer. The revised Table 1 reports more controlled configurations, including the baseline, baseline+SEG, baseline+SEG+SSFM, baseline+SEG/SSFM/SAFAGM, a leave-one-out variant that removes SSFM under the complete dual-prior setting, and the full CMSE-DPNet. These results separately reflect the effects of spectral expansion, spatial-spectral fusion, physical-prior attention, and semantic-prior attention. In addition, for SAFAGM, we conducted a dedicated comparison under snow-interference scenarios. This experiment shows that enabling SAFAGM increases Precision from 96.73% to 98.39%, Recall from 91.97% to 93.73%, and reduces FAR from 2.54% to 1.26%, while improving F1-score and IoU by 1.71 and 3.12 percentage points, respectively. For LG-CAM, the full model achieves 96.94% Precision, 94.41% F1-score, and 89.40% IoU, indicating that semantic priors provide additional gains beyond spectral expansion and physical guidance. The SEG module is the prerequisite for all subsequent modules and therefore cannot be removed independently in the full pipeline.

Location of changes in the revised manuscript: Table 1; Section 4.3.1, lines 699-721; Section 4.3.2, lines 722-746; Section 4.3.3, lines 747-767; Table 3.

Comment 5

Reviewer comment: The English and presentation may be enhanced for clarity.

Response: Thank you for the suggestion. We have carefully revised the English expression, logical organization, and overall presentation of the manuscript. The Introduction has been rewritten in the order of research background, research status, research gap, research idea, and main contributions. The original literature-review subsections have been moved into the new Related Work section. We also revised figure captions and operator legends, added explanatory text for the main information flow, and corrected references and formatting throughout the manuscript.

Location of changes in the revised manuscript: Introduction, lines 25-98; Related Work, lines 99-252; captions of Figures 1 and 3-8; revised reference list.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript addresses an important problem in remote sensing, namely cloud detection under limited spectral conditions, and proposes a unified framework that combines cross-modality spectral expansion with physical and semantic priors. The overall design is coherent, and the experimental results demonstrate consistent performance improvements. The integration of spectral reconstruction with both physics-based and semantic guidance is a noteworthy and timely contribution.

The method reconstructs 13-band pseudo-spectral data from 4-band inputs, yet no quantitative evaluation of spectral fidelity is provided. In particular, the reconstructed SWIR bands are used to compute NDSI, but their physical reliability is not verified. Without demonstrating that the reconstructed spectra preserve meaningful physical properties, the interpretation of “physical priors” becomes questionable. It is therefore necessary to include quantitative metrics (e.g., SAM, RMSE, correlation) and analysis comparing reconstructed and real multispectral data.

the use of LLM-based semantic priors is an interesting idea, but its added value could be better demonstrated. A comparison with simpler metadata-based or rule-based approaches would help clarify whether the LLM provides a measurable advantage. Additionally, a brief discussion of potential uncertainties in LLM outputs would improve transparency.

Although ablation studies are provided, the model includes multiple components (SEG, SSFM, SAFAGM, LG-CAM), making it difficult to clearly attribute performance gains. Consider adding more controlled experiments, This would help clarify the relative importance of each component.

Author Response

Comment 1

Reviewer comment: The method reconstructs 13-band pseudo-spectral data from 4-band inputs, yet no quantitative evaluation of spectral fidelity is provided. In particular, the reconstructed SWIR bands are used to compute NDSI, but their physical reliability is not verified. Without demonstrating that the reconstructed spectra preserve meaningful physical properties, the interpretation of “physical priors” becomes questionable. It is therefore necessary to include quantitative metrics (e.g., SAM, RMSE, correlation) and analysis comparing reconstructed and real multispectral data.

Response: We fully agree with this comment. Accordingly, we added a subsection to evaluate the spectral fidelity of the reconstructed SWIR-like bands. Specifically, we designed a Sentinel-2 band-holding experiment: B2, B3, B4, and B8 were used as four-band inputs, while B11 and B12 were used as reference SWIR bands. The reconstructed SWIR-like bands were then compared with real Sentinel-2 SWIR observations using RMSE, Pearson correlation coefficient, and SAM. Since NDSI is the core of the physical-prior branch, we also compared NDSI values computed from real B11 and reconstructed B11. The results show that the correlation coefficients of B11 and B12 reach 0.812641 and 0.852589, respectively, with RMSE values of 0.086228 and 0.064325. The SAM is 7.159606 degrees, and the NDSI correlation coefficient is 0.830303. These results indicate that the reconstructed bands preserve meaningful task-related SWIR-like cues. At the same time, we explicitly state in the manuscript that these bands should not be interpreted as direct physical SWIR measurements. This modification makes the term “physical prior” more precise: the NDSI branch uses task-oriented SWIR-sensitive proxy cues rather than claiming sensor-level physical equivalence.

Location of changes in the revised manuscript: Section 4.2, lines 666-697; Table 2; Discussion, lines 814-818.

Comment 2

Reviewer comment: the use of LLM-based semantic priors is an interesting idea, but its added value could be better demonstrated. A comparison with simpler metadata-based or rule-based approaches would help clarify whether the LLM provides a measurable advantage. Additionally, a brief discussion of potential uncertainties in LLM outputs would improve transparency.

Response: Thank you for this constructive comment. We revised the manuscript from two aspects. First, we made the LLM label-generation process more constrained and transparent. The revised method no longer relies on open-ended language generation. Instead, for each semantic attribute, a fixed binary prompt is used, and the required output is restricted to either “0” or “1”. This design reduces uncontrolled linguistic variation and makes the LLM output functionally similar to a structured metadata-to-label reasoning module. We also added format checking and metadata/sample-level verification to reduce label noise. Second, we further quantified the contribution of LLM-based semantic priors through the LG-CAM ablation experiment. On the same 22-scene test set, adding LG-CAM increases Precision to 96.94%, F1-score to 94.41%, and IoU to 89.40%, while reducing FAR to 2.02%. Compared with the variant using SEG, SSFM, and SAFAGM but without LG-CAM, the full model improves F1-score by 0.62 percentage points and IoU by 1.24 percentage points, and reduces FAR by 0.78 percentage points. We also added a limitations discussion, noting that the semantic-prior branch depends on the availability and reliability of geographic metadata. We agree that an independently designed manual rule-based geographic baseline is a valuable future direction; however, such a baseline would require additional external GIS/seasonal threshold rules, and the rules may vary across regions. In this revision, we mainly evaluate the proposed LLM semantic-prior mechanism through controlled branch ablation and leave the manual-rule baseline as future work.

Location of changes in the revised manuscript: Section 3.4.1, lines 517-564; Section 4.3.3, lines 747-767; Discussion, lines 826-834.

Comment 3

Reviewer comment: Although ablation studies are provided, the model includes multiple components (SEG, SSFM, SAFAGM, LG-CAM), making it difficult to clearly attribute performance gains. Consider adding more controlled experiments, This would help clarify the relative importance of each component.

Location of changes in the revised manuscript: Table 1; Section 4.3, lines 698-767; Table 3.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper addresses the issue that current cloud detection methods ignore physical constraints, and proposes a method that combines spectral expansion and dual priors of physics and semantics. The paper has novelty in the improvements of CycleGAN, the design of SAFAGM and LG-CAM modules. However, the paper still has deficiencies in terms of structural organization and research motivation. The specific comments are as follows:

Why don't the authors directly use Sentinel-2 data and instead use CycleGAN to perform spectral reconstruction on GF-1 data? If the factor of high spatial resolution is taken into account, has the super-resolution reconstruction of Sentinel-2 been considered?
The introduction section of the paper needs to be reorganized according to the research background and significance, the research status, the research gap, the research idea of this study, and the main contributions.
The paper needs to add a "Related Work" section and move Sections 1.1, 1.2 and 1.3 into this section.
The operators in Figures 1, 3-8 require supplementary legends for clarification.
The academic rigor of the paper needs to be enhanced. Many of the cited methods, models, formulas, modules, etc. have no reference citations marked. The authors need to conduct a thorough review and make corrections throughout the paper.
GF-1 and Sentinel-2 have significant differences in terms of spatial resolution and spectral band composition. How to balance the requirements of band and spatial resolution in image translation needs to be further explained; in addition, the spatial resolution of the generated images should be described in the text.
In the comparison method of the paper, the SOTA methods for the years 2024-2026 are relatively few. It is suggested to include the methods for the past two years for comparison.

Author Response

Comment 1

Reviewer comment: Why don't the authors directly use Sentinel-2 data and instead use CycleGAN to perform spectral reconstruction on GF-1 data? If the factor of high spatial resolution is taken into account, has the super-resolution reconstruction of Sentinel-2 been considered?

Response: Thank you for raising this important methodological question. The objective of this study is to achieve more reliable cloud detection on GF-1 WFV imagery. GF-1 WFV imagery offers relatively high spatial resolution and good practical availability, but its hardware sensor contains only visible and near-infrared bands and lacks SWIR bands, which are critical for cloud-snow discrimination. Therefore, the core idea of this paper is not to replace GF-1 data with Sentinel-2 data, but to use spectral expansion at the software algorithm level to compensate, as much as possible, for the insufficient spectral configuration of the hardware, while preserving the original spatial geometry and application scenario of GF-1.

Directly using Sentinel-2 can provide richer spectral information, but it would change the target sensor and practical application conditions and therefore cannot answer the question of how to improve cloud detection when only GF-1 imagery is available. In addition, Sentinel-2 bands have inconsistent spatial resolutions, including 10 m, 20 m, and 60 m, whereas GF-1 WFV multispectral imagery has a unified 16 m spatial grid. Direct replacement or fusion with Sentinel-2 data would introduce cross-sensor spatial registration and scale inconsistency issues and would also weaken the practical significance of this work for GF-1 data itself.

Accordingly, we adopt a spectral expansion strategy, using the improved CycleGAN to reconstruct 13-band pseudo-Sentinel-2 spectral representations from 4-band GF-1 imagery. Essentially, this is a software-based enhancement of hardware capability: under the constraint that the GF-1 spatial grid, patch size, and ground sampling distance are unchanged, the model learns to supplement SWIR-related discriminative spectral information, enabling physical priors such as NDSI to be introduced into the GF-1 cloud detection task. The revised manuscript now explicitly states that this stage performs spectral expansion rather than spatial super-resolution; the output has 13 spectral channels but still preserves the 16 m spatial resolution and original geometric structure of GF-1 WFV.

We have also explained why Sentinel-2 super-resolution is not the focus of this study. Super-resolution of Sentinel-2 mainly addresses the spatial refinement of Sentinel-2 data itself and may introduce additional resampling errors or artifacts. In contrast, this study focuses on enhancing the spectral representation of GF-1 through software modeling under the condition of limited hardware spectral capability, and on using the enhanced representation for downstream cloud detection. Therefore, the proposed method is more consistent with real GF-1 application scenarios and better reflects the research significance of using algorithms to optimize and compensate for sensor hardware limitations.

Location of changes in the revised manuscript: Introduction, lines 25-63; Section 3, lines 253-264; Section 3.1, lines 305-316; Section 3.5.1, lines 589-599.

Comment 2

Reviewer comment: The introduction section of the paper needs to be reorganized according to the research background and significance, the research status, the research gap, the research idea of this study, and the main contributions.

Response: We agree with this suggestion. We have rewritten and reorganized the Introduction following the logical order suggested by the reviewer. The revised Introduction first presents the research background and significance of cloud detection in optical remote sensing, then summarizes the research status of deep cloud detection, image translation/spectral reconstruction, and LLM-based semantic context. Next, it identifies the key research gaps: GF-1 lacks SWIR-dependent indices, the connection between reconstructed bands and downstream cloud-snow discrimination is insufficiently addressed, and scene-level semantic information is still rarely used in a structured manner for dense cloud segmentation. Finally, the Introduction presents the core idea of CMSE-DPNet and summarizes the main contributions.

Location of changes in the revised manuscript: Introduction, lines 25-98.

Comment 3

Reviewer comment: The paper needs to add a "Related Work" section and move Sections 1.1, 1.2 and 1.3 into this section.

Response: Thank you for this helpful structural suggestion. We have added Section 2, “Related Work”, and moved the original literature-review subsections into this section. The revised Related Work contains three subsections: Image Translation, Cloud Detection, and LLM-Driven Semantic Priors in Remote Sensing. This adjustment makes the Introduction more concise and the literature review more systematic.

Location of changes in the revised manuscript: Section 2, lines 99-252.

Comment 4

Reviewer comment: The operators in Figures 1, 3-8 require supplementary legends for clarification.

Response: We have revised the relevant figure captions and added explanatory legends for the main symbols and operators. For Figure 1, we explain the arrows and prior-information flow. For Figure 3, we describe 1 x 1 Conv, SSA, Res, and the soft/hard mask branch. For Figure 4, we explain the downsampling/upsampling operators, skip connections, and dual-output branch. For Figure 5, we define Q, K, V, softmax, and the residual connection. For Figure 6, we explain the operators such as addition, subtraction, and cr. For Figure 7, we define ms, cs, and the multiplication operator. For Figure 8, we explain the metadata-derived label vector, linear projection, Sigmoid activation, and channel reweighting process. These additions help readers better understand the network structure diagrams.

Location of changes in the revised manuscript: Captions of Figures 1 and 3-8, especially lines 613-620, 830-833, 879-884, 888-891, 1018-1023, 1103-1108, and 1335-1340.

Comment 5

Reviewer comment: The academic rigor of the paper needs to be enhanced. Many of the cited methods, models, formulas, modules, etc. have no reference citations marked. The authors need to conduct a thorough review and make corrections throughout the paper.

Response: Thank you for the reminder. We have carefully checked the references throughout the manuscript and added or corrected citations for key methods, models, formulas, and modules, including CycleGAN, U-Net/RS-Net, NDSI, self-attention and spectral reconstruction, cross-entropy loss, the Adam optimizer, attention mechanisms, FiLM-style conditional modulation, LLM/LVLM studies in remote sensing, and recent cloud detection baselines. We also revised the relevant descriptions in the Methods section to clarify which components are newly designed in this work and which components are inspired by existing mechanisms. These changes improve the academic rigor of the manuscript and make the technical lineage of the proposed method clearer.

Location of changes in the revised manuscript: Throughout the revised manuscript; for example, Section 2, lines 99-252; Section 3.1, lines 291-322; Section 3.4.1, lines 517-564; Section 3.5.2, lines 600-612; revised reference list.

Comment 6

Reviewer comment: GF-1 and Sentinel-2 have significant differences in terms of spatial resolution and spectral band composition. How to balance the requirements of band and spatial resolution in image translation needs to be further explained; in addition, the spatial resolution of the generated images should be described in the text.

Response: We agree with this comment and have added a clear explanation. In the revised manuscript, we explain that the spatial-resolution difference is handled at the preprocessing and network-input levels. Sentinel-2A samples are spatially normalized to the same patch size as GF-1 samples, and the generator is formulated as a mapping from X in R^(H x W x 4) to Y in R^(H x W x 13), where H and W remain unchanged. Thus, the model learns a Sentinel-2A-like spectral distribution without changing the GF-1 spatial geometry. The spectral-band difference is handled by the channel-alignment layer and the Spectral Expansion Generator, which map four GF-1 bands to a 13-channel pseudo-Sentinel-2A representation. To avoid spatial distortion, the model uses a residual branch, a structural-style separation loss, and an edge-guided cloud-mask constraint to preserve GF-1 spatial structures and cloud boundaries. We also explicitly state that each generated image patch remains 512 x 512 pixels on the GF-1 WFV grid, corresponding to the original 16 m GF-1 WFV spatial resolution.

Location of changes in the revised manuscript: Section 3, lines 260-264; Section 3.1, lines 305-316; Section 3.5.1, lines 589-599.

Comment 7

Reviewer comment: In the comparison method of the paper, the SOTA methods for the years 2024-2026 are relatively few. It is suggested to include the methods for the past two years for comparison.

Response: Thank you for the suggestion. We revised the comparison-experiment section and added recent public GF1_WHU results of CloudMamba to improve completeness. Because these methods do not report all five metrics under the same evaluation protocol, unavailable entries are marked with “-”, and we explicitly distinguish these published results from the baselines re-evaluated under our unified experimental setting. We also updated the related references by adding recent LLM/LVLM and remote-sensing studies to improve the timeliness of the manuscript.

Location of changes in the revised manuscript: Table 4; Section 4.4, lines 768-793; updated reference list.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript has been sufficiently improved, and I recommend acceptance for publication.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors addressed all my comments and the manuscript can be accepted.