Review Reports - An End-to-End Foundation Model-Based Framework for Robust LAI Retrieval Under Cloud Cover

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript presents an end-to-end deep learning framework based on a pretrained foundation model (Prithvi) for retrieving Leaf Area Index (LAI) from cloud-contaminated HLS data. The study compares this approach with a cascaded pipeline (cloud reconstruction followed by LAI retrieval) and demonstrates that direct mapping yields improved performance, particularly under moderate to high cloud cover. The topic is timely and relevant, especially given the increasing interest in foundation models and gap-free Earth observation products.

Overall, the manuscript is well structured, clearly written, and presents a coherent experimental framework. The comparison between end-to-end and cascaded approaches is valuable, and the reported U-shaped performance response to cloud fraction is an interesting finding. However, several important conceptual and methodological issues need to be addressed before the manuscript can be considered for publication.

First, the novelty is somewhat overstated. While the use of a pretrained foundation model is of interest, the concept of direct regression from reflectance to LAI is well established in the literature. The main contribution lies in the application of foundation models to cloud-contaminated inputs, rather than a fundamentally new retrieval paradigm. The authors should moderate their claims and better position their work with respect to existing machine learning and hybrid retrieval approaches.

Second, the use of Hi-GLASS LAI as ground truth represents a significant limitation. As Hi-GLASS is itself a model-derived product (based on PROSAIL simulations and machine learning), the study effectively evaluates consistency with an existing retrieval rather than true LAI accuracy. This raises concerns about potential bias propagation and limits the interpretability of the reported performance gains. At minimum, this limitation should be clearly acknowledged and discussed. Ideally, validation against independent datasets (e.g., field measurements or GBOV products) should be included.

Third, the manuscript lacks any form of uncertainty quantification. This is particularly problematic given the focus on cloud-contaminated observations, where uncertainty is expected to be substantial. The absence of uncertainty estimates (either epistemic or aleatoric) weakens the robustness of the conclusions. Even a qualitative discussion of uncertainty limitations would improve the manuscript.

Fourth, the interpretation of the observed U-shaped performance pattern with increasing cloud cover is not sufficiently supported. The explanation based on implicit regularization is speculative, and alternative explanations (e.g., sampling effects or masking artefacts) are not explored. The authors should either provide stronger evidence or adopt a more cautious interpretation.

Fifth, the manuscript would benefit from a more explicit discussion of physical consistency. The approach is entirely data-driven, yet issues such as spectral sensitivity, LAI saturation, and consistency with radiative transfer principles are only briefly mentioned. A deeper discussion would strengthen the scientific grounding of the work.

Sixth, the manuscript repeatedly refers to the proposed approach as an “end-to-end” framework. However, the meaning of this term is not sufficiently clarified and may be ambiguous for the remote sensing community. In the current context, “end-to-end” appears to refer to a direct statistical mapping from cloud-contaminated reflectance inputs to LAI outputs, without intermediate reconstruction or physically interpretable variables. This should be explicitly stated. I recommend that the authors:

Clearly define what is meant by “end-to-end” early in the manuscript;
Distinguish it from cascaded pipelines (e.g., reconstruction + retrieval);
Avoid implying that the approach is physically consistent or fundamentally different from other direct retrieval models.

Finally, the experimental design has limitations in terms of generalization. The analysis is restricted to China and a relatively short time period, and the sample size is modest for fine-tuning a foundation model. Claims about general applicability should therefore be moderated.

In summary, the manuscript addresses an important problem and presents promising results, but requires substantial revision to improve scientific rigor and positioning.

Author Response

General Comment: This manuscript presents an end-to-end deep learning framework based on a pretrained foundation model (Prithvi) for retrieving Leaf Area Index (LAI) from cloud-contaminated HLS data. The study compares this approach with a cascaded pipeline (cloud reconstruction followed by LAI retrieval) and demonstrates that direct mapping yields improved performance, particularly under moderate to high cloud cover. The topic is timely and relevant, especially given the increasing interest in foundation models and gap-free Earth observation products.

Response: We thank the reviewer for their encouraging assessment of the manuscript’s structure and relevance. We have carefully addressed the conceptual and methodological issues, particularly regarding novelty, ground truth limitations, and physical consistency, as detailed below.

Comment 1: First, the novelty is somewhat overstated. While the use of a pretrained foundation model is of interest, the concept of direct regression from reflectance to LAI is well established in the literature. The main contribution lies in the application of foundation models to cloud-contaminated inputs, rather than a fundamentally new retrieval paradigm. The authors should moderate their claims and better position their work with respect to existing machine learning and hybrid retrieval approaches.

Response: We thank the reviewer for this thoughtful and constructive comment. We fully agree that the concept of direct regression from reflectance to LAI is not new, and that our manuscript overstated the novelty of the end-to-end paradigm itself. The reviewer is correct that the main contribution of our work lies in the application of a pretrained foundation model to cloud-contaminated LAI retrieval , not in inventing a fundamentally new retrieval approach.

We have therefore revised the manuscript to better reflect this understanding. Specifically, we have toned down the language throughout the Abstract, Introduction, and Conclusions, removing claims such as “a novel paradigm” and replacing them with more measured statements , explicitly clarifying that our contribution is to extend such approaches to the challenging scenario of cloud‑affected observations. Moderated the conclusions to state that, under the research conditions, the end-to-end framework consistently outperforms the cascaded baselines, rather than claiming universal superiority.

We believe these revisions make the manuscript better positioned, and more useful to the community. We are grateful to the reviewer for helping us sharpen the focus of our work.

Comment 2: Second, the use of Hi-GLASS LAI as ground truth represents a significant limitation. As Hi-GLASS is itself a model-derived product (based on PROSAIL simulations and machine learning), the study effectively evaluates consistency with an existing retrieval rather than true LAI accuracy. This raises concerns about potential bias propagation and limits the interpretability of the reported performance gains. At minimum, this limitation should be clearly acknowledged and discussed. Ideally, validation against independent datasets (e.g., field measurements or GBOV products) should be included.

Response: We thank the reviewer for the valuable suggestion regarding independent validation. We carefully considered using GBOV or in‑situ LAI measurements, but identified a potential spatial scale mismatch that prevents a direct and meaningful comparison. GBOV provides point‑scale measurements, whereas our model predictions are made for 224 × 224 pixel patches (approximately 6.7 km × 6.7 km at 30 m resolution). This discrepancy implies that a single point measurement cannot adequately represent the patch‑level LAI features learned by the model. Moreover, the only GBOV sites in Asia are located in Japan, which does not cover our study area in China, making direct cross‑validation geographically infeasible. Given this limitation, we explicitly acknowledge the scale issue and the lack of independent validation as a key limitation of the current study. We have added a paragraph in the Discussion (Section X.X) stating that our training and evaluation rely on the Hi‑GLASS LAI product, and that independent validation against in‑situ or GBOV data is desirable but currently hindered by the spatial scale mismatch and geographic coverage gap. We also propose that future work should design dedicated upscaling experiments to bridge the scale gap and assess the absolute accuracy of the model. We believe this transparent acknowledgment of the limitation, together with a concrete future direction, adequately addresses the reviewer’s concern. We thank the reviewer for prompting us to think more carefully about validation scales.

Comment 3: Third, the manuscript lacks any form of uncertainty quantification. This is particularly problematic given the focus on cloud-contaminated observations, where uncertainty is expected to be substantial. The absence of uncertainty estimates (either epistemic or aleatoric) weakens the robustness of the conclusions. Even a qualitative discussion of uncertainty limitations would improve the manuscript.

Response: We sincerely thank the reviewer for identifying this important methodological shortcoming. We fully agree that addressing uncertainty under severe cloud contamination is essential to confirm the reliability and validity of our conclusions. In response to this constructive comment, we have adopted a twopronged approach, including quantitative bootstrap confidence intervals for the statistical results of the three LAI retrieval models and a qualitative discussion of uncertainty.

To directly address the reviewer’s concern about the robustness of our comparative conclusions, we introduced a quantitative uncertainty estimate for all performance metrics using the bootstrap method. Specifically, we performed 1,000 iterations of random resampling (with replacement) on the validation set to compute the 95% confidence intervals (CIs) for RMSE and R². These CIs have been added to the results in Section 3.1 to robustly quantify model performance differences. The newly added 95% CIs demonstrate that the performance gaps between the endtoend model and the cascaded baselines are statistically significant and stable under data variability. Furthermore, while the bootstrap analysis ensures the statistical robustness of the evaluation metrics, the current deterministic deep learning framework still lacks pixelwise predictive uncertainty maps. Following the reviewer’s practical suggestion, we have added a qualitative discussion in the Discussion section. We now explicitly articulate the main sources of predictive uncertainty in our task: aleatoric uncertainty arising from cloud contamination, atmospheric correction errors, sensor noise, and the inherent variability of vegetation reflectance; and epistemic uncertainty stemming from the use of a modelderived product (HiGLASS) as labels, as well as the geographical restriction to China. We note that future iterations of this foundation model must incorporate more comprehensive uncertainty analyses.

Comment 4: Fourth, the interpretation of the observed U-shaped performance pattern with increasing cloud cover is not sufficiently supported. The explanation based on implicit regularization is speculative, and alternative explanations (e.g., sampling effects or masking artefacts) are not explored. The authors should either provide stronger evidence or adopt a more cautious interpretation.

Response: We thank the reviewer for this critical and constructive comment. We agree that our initial explanation based solely on implicit regularization (MAE analogy) was speculative and that alternative explanations were not adequately explored. To address this, we performed an additional analysis of the land cover composition across different cloud cover strata, computing the percentage distribution of major vegetation types (forest, shrubland, grassland, cropland, and others) for each cloud cover interval in both the training and validation sets. The results show that under extreme cloud cover (70‑80%), the remaining valid pixels originate from a variety of land cover types, including shrubland, grassland, and cropland, indicating that the prediction task still involves considerable complexity and is not merely a matter of predicting a constant background baseline.

Based on this evidence, we have substantially revised the interpretation in the Discussion. The revised text now acknowledges the speculative nature of the implicit regularization hypothesis, presents sampling bias and spatial homogeneity as partial contributing factors while noting that the presence of diverse land cover types under extreme cloud cover suggests other mechanisms may also be involved, and adopts a more cautious conclusion that the U‑shaped pattern is an empirical observation with multiple plausible contributing factors rather than a definitive mechanistic conclusion. We also explicitly note that dedicated experiments (e.g., controlled subsampling or synthetic data) are required to fully disentangle the underlying mechanisms. We believe these revisions transform the earlier speculative claim into a more balanced, evidence‑informed discussion that properly acknowledges alternative explanations. The manuscript now presents the U‑shaped pattern as an interesting observation requiring further investigation, rather than as a proven regularization effect.

Comment 5: Fifth, the manuscript would benefit from a more explicit discussion of physical consistency. The approach is entirely data-driven, yet issues such as spectral sensitivity, LAI saturation, and consistency with radiative transfer principles are only briefly mentioned. A deeper discussion would strengthen the scientific grounding of the work.

Response: We thank the reviewer for this constructive suggestion. In the revised manuscript, we have added a dedicated paragraph in the Discussion (Section X.X) to explicitly address physical consistency. We now state that the systematic decline in estimation accuracy for LAI values above approximately 4 is not a modelspecific deficiency but a direct manifestation of spectral saturation, a fundamental physical constraint of optical remote sensing arising from multiple scattering and canopy closure. This saturation ceiling applies to any opticalonly retrieval method, including our endtoend framework, and therefore the observed advantage of the endtoend model in the highLAI regime is statistical rather than physical. We also acknowledge that our purely datadriven approach does not explicitly enforce radiative transfer principles such as energy conservation or directionalhemispherical consistency, which may limit its extrapolation capability under extreme viewing geometries or atmospheric conditions. Future work could incorporate lightweight physical priors to improve consistency without sacrificing representational power. We believe these additions provide the more explicit discussion of physical consistency requested by the reviewer and strengthen the scientific grounding of the work.

Comment 6: Sixth, the manuscript repeatedly refers to the proposed approach as an “end-to-end” framework. However, the meaning of this term is not sufficiently clarified and may be ambiguous for the remote sensing community. In the current context, “end-to-end” appears to refer to a direct statistical mapping from cloud-contaminated reflectance inputs to LAI outputs, without intermediate reconstruction or physically interpretable variables. This should be explicitly stated. I recommend that the authors:·Clearly define what is meant by “end-to-end” early in the manuscript;·Distinguish it from cascaded pipelines (e.g., reconstruction + retrieval);·Avoid implying that the approach is physically consistent or fundamentally different from other direct retrieval models.

Response: We thank the reviewer for this important clarification and agree that the term “endtoend” can be ambiguous and should be explicitly defined for the remote sensing community. To address this, we have added a clear definition early in the manuscript (Introduction, paragraph 3), stating that in this study “endtoend” refers to a direct statistical mapping from cloudcontaminated, multitemporal reflectance inputs to LAI outputs without any explicit intermediate steps such as cloud detection, reflectance reconstruction, or retrieval of physically interpretable intermediate variables. Within the same paragraph, we explicitly distinguish this framework from the cascaded pipeline by noting that the latter separates the problem into two sequential stages: first reconstructing cloudfree reflectance, then inverting LAI from the reconstructed reflectance. We have also removed any language that could imply physical consistency or fundamental novelty of the endtoend paradigm; for instance, we have replaced phrases such as “a novel retrieval paradigm” with “an endtoend framework based on a foundation model” and added a sentence clarifying that our approach is purely datadriven and does not claim physical interpretability. These revisions directly address the comment, and all changes have been highlighted in the revised manuscript.

Comment 7: Finally, the experimental design has limitations in terms of generalization. The analysis is restricted to China and a relatively short time period, and the sample size is modest for fine-tuning a foundation model. Claims about general applicability should therefore be moderated.

Response:

We thank the reviewer for this honest assessment. We agree that our experimental design has clear limitations in terms of geographical scope, temporal coverage , and sample size. These constraints indeed limit the general applicability of the quantitative results.

To address this, we have made the following revisions. First, we added a dedicated paragraph on limitations in the Discussion, explicitly acknowledging the geographical restriction to China and the lack of validation across other continents or climate zones, as well as the limited temporal span and the absence of interannual variability assessment.

Regarding the risk associated with the limited size of the cloudaffected training set, we note that our finetuning experiments have demonstrated that approximately 2000 samples are sufficient for the downstream task, as discussed in the methods section.

Second, we have moderated claims about general applicability throughout the manuscript, particularly in the Abstract, Conclusions, and the final summary of the Discussion. We now state that the applicability of the proposed framework to other regions and longer time series requires further validation.

Third, we have reframed the contribution as methodological (demonstrating the superiority of the endtoend paradigm over cascaded architectures under controlled conditions) rather than claiming universal numerical accuracy. The primary value of this study is now presented as a proofofconcept and a reproducible framework, with the explicit note that future work should extend the evaluation to broader geographical regions, longer time spans, and larger datasets.

We believe these revisions appropriately temper our claims while preserving the significance of the methodological findings. The revised manuscript now clearly separates what has been demonstrated (relative performance under the tested conditions) from what requires future investigation (generalizability beyond the current experimental design).

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript explores the application of a remote sensing foundation model (Prithvi) to address the challenging issue of LAI retrieval under cloud-occluded conditions. The chosen topic is highly relevant and holds significant practical value for the remote sensing community. However, the manuscript exhibits several flaws and logical loopholes regarding its data scale, validation benchmarks, physical mechanism explanations, and experimental design. These issues should be thoroughly addressed before the manuscript can be considered for publication. “Major Revision” is needed. My detailed critical comments are outlined below.

The authors fine-tuned the Prithvi foundation model, which possesses hundreds of millions of parameters. However, the magnitude of the data used for this process casts doubt on the model's generalization capabilities. The manuscript states that only 9347 clear-sky samples and 2000 cloud-affected sample groups were used for training and validation. For a foundation model, utilizing merely 1000 time-series cloud-occluded training samples is prone to severe overfitting (a phenomenon the authors explicitly admit regarding the end-to-end model).

Although the authors mention stratifying the data based on China's annual average cloud cover, it remains unclear which specific ecological zones (Biomes) and vegetation types (e.g., evergreen broadleaf forest, deciduous needleleaf forest, croplands) are covered by these 2000 sample groups. Without extensive sampling across regions with distinct phenological differences and surface heterogeneity, the model likely merely memorizes local spatial textures rather than learning the actual radiative transfer laws of the canopy under cloud interference.

This study relies entirely on the Hi-GLASS LAI product as the supervised label and the ultimate ground truth for validation. However, Hi-GLASS itself is a Level-2 product retrieved based on the PROSAIL physical model and Random Forest algorithms.

Training a deep learning model using products generated by another model is essentially knowledge distillation. This implies that the upper accuracy bound of the proposed end-to-end model is strictly limited by the inherent errors of the Hi-GLASS product itself. Without independent validation from global/regional in-situ LAI networks, it is impossible to prove the model's absolute accuracy in the real physical world.

The paper presents a novel but highly counter-intuitive phenomenon that model accuracy exhibits a U-shaped curve relative to cloud cover, performing worst at 50-60% cloud cover, but rebounding at extreme cloud fractions of 70-80%. The authors attribute this rebound to the Masked Autoencoder pre-training mechanism being activated under extreme occlusion, producing implicit regularization. This is purely speculative algorithmic reasoning, lacking validation.

At 70-80% high cloud cover, available optical information is minimal. The accuracy rebound is highly likely due to the background LAI values in these regions being extremely homogeneous (e.g., large areas of dense forests or deserts). In this scenario, the model is likely just predicting the climatology (background baseline) rather than genuinely retrieving dynamic details. The authors should rule out this statistical survivorship bias.

The cascaded models constructed as comparative baselines merely apply basic Transformers or GANs for reflectance reconstruction, followed by a simple regressor. In the field of ecological remote sensing time-series reconstruction, there is a wealth of mature models based on spatiotemporal fusion or data assimilation algorithms. The performance of these dedicated cloud-removal/gap-filling algorithms typically far exceeds that of naive GANs. Failing to compare the proposed method with mainstream remote sensing time-series interpolation and cloud-removal retrieval algorithms severely weakens the claim that the end-to-end model comprehensively outperforms the alternatives.

Optical remote sensing inherently suffers from spectral saturation when LAI exceeds 3 or 4, driven by multiple scattering and occlusion effects within the canopy. The manuscript acknowledges that all models exhibit systemic underestimation in high LAI regions. While the end-to-end model shows smaller variance in the high LAI range, both the scatter plots (Fig. 3) and trend lines (Fig. 5) clearly indicate that the foundation model fails to break the optical saturation bottleneck physically. The end-to-end framework appears to merely smooth the errors mathematically, rather than extracting deeper and unsaturated features based on physical mechanisms, such as exploiting deep correlations in SWIR bands.

The paper introduces the SSIM into its composite loss function. SSIM was originally designed as an image quality assessment metric based on human visual perception. For a highly quantitative physical parameter like LAI, forcefully imposing visual smoothness on spatial structures risks erasing genuine high-frequency spatial abrupt changes (local gradient differences) caused by topography, soil moisture variations, or human disturbances (e.g., crop harvesting). This logically explains why the end-to-end model's results appear overly smoothed and blurry upon visual inspection. The authors need to justify the ecological rationale behind penalizing these natural physical gradients.

Author Response

General Comment:This manuscript explores the application of a remote sensing foundation model (Prithvi) to address the challenging issue of LAI retrieval under cloud-occluded conditions. The chosen topic is highly relevant and holds significant practical value for the remote sensing community. However, the manuscript exhibits several flaws and logical loopholes regarding its data scale, validation benchmarks, physical mechanism explanations, and experimental design. These issues should be thoroughly addressed before the manuscript can be considered for publication. “Major Revision” is needed. My detailed critical comments are outlined below.

Response:We sincerely thank the reviewer for acknowledging the relevance and practical value of our work. We also appreciate the critical assessment of the manuscript’s shortcomings and logical gaps concerning data scale, validation benchmarks, physical mechanism explanations, and experimental design. We fully agree that these issues must be thoroughly addressed, and we have undertaken major revisions to resolve them. In the following, we provide a pointbypoint response to each of the reviewer’s detailed comments.

Comment 1: The authors fine-tuned the Prithvi foundation model, which possesses hundreds of millions of parameters. However, the magnitude of the data used for this process casts doubt on the model's generalization capabilities. The manuscript states that only 9347 clear-sky samples and 2000 cloud-affected sample groups were used for training and validation. For a foundation model, utilizing merely 1000 time-series cloud-occluded training samples is prone to severe overfitting (a phenomenon the authors explicitly admit regarding the end-to-end model).

Response: We thank the reviewer for raising the concern about sample size. To address this, we have added an experiment on the amount of data required for downstream fine‑tuning (described in Section 2.1.2 ), which systematically evaluates the effect of training set size on model performance. Specifically, we randomly subsampled the cloud‑affected training set at increasing sizes and fine‑tuned the Prithvi model under identical settings. Statistical results were recorded for each subsample. The results indicate that the current dataset size (approximately 2000 patches) is sufficient for fine‑tuning the Prithvi model on this specific task, and that further increasing the sample size would not meaningfully improve generalization.

Overfitting was ruled out through repeated training and multiple epoch attempts. The description of overfitting in the manuscript is intended to emphasize that, despite overfitting during training, an upper bound of estimation remains unbreakable.

This analysis has been added to the Data section to provide empirical justification for the sample size. The new figure and associated discussion are highlighted in the revised manuscript.

Comment 2: Although the authors mention stratifying the data based on China's annual average cloud cover, it remains unclear which specific ecological zones (Biomes) and vegetation types (e.g., evergreen broadleaf forest, deciduous needleleaf forest, croplands) are covered by these 2000 sample groups. Without extensive sampling across regions with distinct phenological differences and surface heterogeneity, the model likely merely memorizes local spatial textures rather than learning the actual radiative transfer laws of the canopy under cloud interference.

Response:We thank the reviewer for this important insight. We agree that without explicit information on the vegetation types covered by the cloudaffected samples, it is unclear whether the model has learned generalizable features or merely memorized local spatial patterns. To address this, we performed a detailed analysis of the vegetation type composition of the 2,000 cloudaffected samples.

In the revised manuscript (Section 2.1.2), we compiled statistics on the vegetation types covered by the samples according to the MODIS IGBP land cover classification system. The results show that the samples cover a wide range of vegetation types as well as nonvegetated land cover types. Furthermore, we calculated the distribution of each vegetation type across different cloud cover intervals, the valid pixels originate from multiple land cover types rather than being confined to a single homogeneous vegetation class. This diversity strongly suggests that the model was exposed to land surface information with varying phenological characteristics and spatial heterogeneity during training, thereby reducing the likelihood that it merely memorized local textures.

It should be noted that the primary objective of this study is not to enable the model to learn the complete physical laws of canopy radiative transfer, but to verify whether the endtoend paradigm can achieve more robust LAI estimation than the cascaded paradigm under cloudcontaminated conditions. Nevertheless, we agree with the reviewer that broader sampling across ecological zones would enhance the model’s generalization capability. In the limitations subsection of the Discussion, we have explicitly noted the current shortfall in ecological zone coverage and identified this as an important direction for future research. We believe these additional analyses and clarifications adequately address the reviewer’s concern.

Comment 3: This study relies entirely on the Hi-GLASS LAI product as the supervised label and the ultimate ground truth for validation. However, Hi-GLASS itself is a Level-2 product retrieved based on the PROSAIL physical model and Random Forest algorithms.

Response: We thank the reviewer for this insightful observation. We fully agree that using Hi‑GLASS, a model‑derived product, as the reference label effectively constitutes knowledge distillation from the Hi‑GLASS retrieval system. Consequently, the upper bound of our model’s absolute accuracy is indeed limited by the inherent errors of Hi‑GLASS (e.g., saturation biases, cloud masking artifacts). Without independent validation against in‑situ measurements, we cannot claim absolute physical accuracy. Nevertheless, the primary contribution of this study is not to assert absolute accuracy, but to demonstrate the relative superiority of the end‑to‑end paradigm over cascaded architectures under identical reference conditions. Because all models were trained and evaluated on the same Hi‑GLASS labels, the comparative conclusion remains valid regardless of any systematic biases in the labels.

Regarding independent validation, we carefully considered using GBOV LAI data. However, a spatial scale mismatch prevents a direct and meaningful comparison. GBOV provides point‑scale measurements, while our model predictions are made for 224×224 pixel patches. Upscaling or downscaling would introduce substantial uncertainty.

To response these limitations, we have added a paragraph in the Discussion explicitly acknowledging that: Our evaluation measures consistency with Hi‑GLASS, not absolute ground truth；The absolute accuracy metrics should be interpreted with caution；Independent validation is desirable but currently hindered by scale mismatch and geographic coverage；Future work should design dedicated upscaling experiments or use high‑resolution imagery to bridge the scale gap.

We believe this transparent acknowledgment and the proposed future directions can address the reviewer’s concern. We thank the reviewer for prompting us to clarify these critical limitations.

Comment 4: The paper presents a novel but highly counter-intuitive phenomenon that model accuracy exhibits a U-shaped curve relative to cloud cover, performing worst at 50-60% cloud cover, but rebounding at extreme cloud fractions of 70-80%. The authors attribute this rebound to the Masked Autoencoder pre-training mechanism being activated under extreme occlusion, producing implicit regularization. This is purely speculative algorithmic reasoning, lacking validation.

Response:We thank the reviewer for this critical and constructive comment. We agree that our initial explanation based solely on implicit regularization (MAE analogy) was speculative and that the alternative explanation of survivorship bias (i.e., valid pixels under extreme cloud cover being confined to homogeneous areas) needed to be tested.

To address this, we performed an additional analysis of the land cover composition across different cloud cover strata using the MODIS IGBP classification system. The results, now included in the revised manuscript (Section 2.1.2), show that under extreme cloud cover (7080%), the remaining valid pixels originate from a variety of vegetation types—including forest, shrubland, grassland, and cropland—rather than being exclusively from homogeneous forest or desert areas. This diversity indicates that the prediction task under extreme cloud cover still involves considerable complexity and is not merely a matter of predicting a constant background baseline.

Based on this evidence, we have revised the interpretation in the Discussion. The revised text now acknowledges the speculative nature of the implicit regularization hypothesis, presents the survivorship bias as a potential contributing factor that is largely mitigated by the observed land cover diversity, and adopts a more cautious conclusion: the Ushaped pattern is an empirical observation with multiple plausible contributing factors rather than a definitive mechanistic conclusion. We also explicitly note that dedicated experiments (e.g., controlled subsampling or synthetic data) would be required to fully disentangle the underlying mechanisms. We believe these revisions transform the earlier speculative claim into a more balanced, evidenceinformed discussion that properly addresses the reviewer’s concern about survivorship bias.

Comment 5: The cascaded models constructed as comparative baselines merely apply basic Transformers or GANs for reflectance reconstruction, followed by a simple regressor. In the field of ecological remote sensing time-series reconstruction, there is a wealth of mature models based on spatiotemporal fusion or data assimilation algorithms. The performance of these dedicated cloud-removal/gap-filling algorithms typically far exceeds that of naive GANs. Failing to compare the proposed method with mainstream remote sensing time-series interpolation and cloud-removal retrieval algorithms severely weakens the claim that the end-to-end model comprehensively outperforms the alternatives.

Response: We thank the reviewer for this important observation. We acknowledge that the cascaded baselines in our study (basic Transformer and GAN) are not state‑of‑the‑art in the field of ecological remote sensing time‑series reconstruction. More advanced methods such as spatiotemporal fusion (e.g., STARFM, ESTARFM) or data assimilation algorithms may achieve higher reconstruction accuracy, and we agree that our claim of “comprehensive superiority” over all alternatives was overstated.

However, we would like to clarify the primary goal of this study. Our objective is not to benchmark the proposed end-to-end framework against every possible cloud reconstruction algorithm, but to compare two task paradigms under identical conditions: direct mapping (end-to-end) versus reconstruction‑then‑retrieval (cascaded). To isolate the effect of task design, we intentionally kept the reconstruction modules simple (basic Transformer and GAN), so that the only difference between the two paradigms is the presence or absence of the intermediate reconstruction step. This controlled comparison allows us to conclude that, under the same data and simple reconstruction backbone, the end‑to‑end paradigm yields more accurate and robust LAI retrieval.

We agree that a more comprehensive comparison with advanced gap‑filling methods (e.g., spatiotemporal fusion or data assimilation) would be valuable. However, such methods often require additional input data (e.g., MODIS daily observations) or complex parameter tuning, which would introduce confounding factors beyond the scope of this study. We have therefore moderated our claims throughout the manuscript, replacing phrases such as “comprehensively outperforms the alternatives” with “outperforms the cascaded baselines constructed in this study” or “demonstrates superiority over the evaluated cascaded approach.”

We believe these response can appropriately address the reviewer’s concern without overclaiming, while preserving the validity of our core finding: under the tested conditions and with comparable task designs, the end-to-end paradigm outperforms the cascaded paradigm.

Comment 6: Optical remote sensing inherently suffers from spectral saturation when LAI exceeds 3 or 4, driven by multiple scattering and occlusion effects within the canopy. The manuscript acknowledges that all models exhibit systemic underestimation in high LAI regions. While the end-to-end model shows smaller variance in the high LAI range, both the scatter plots (Fig. 3) and trend lines (Fig. 5) clearly indicate that the foundation model fails to break the optical saturation bottleneck physically. The end-to-end framework appears to merely smooth the errors mathematically, rather than extracting deeper and unsaturated features based on physical mechanisms, such as exploiting deep correlations in SWIR bands.

Response:We thank the reviewer for this technically accurate observation. We fully agree that spectral saturation is an inherent physical constraint of optical remote sensing when LAI exceeds approximately 3 to 4, driven by multiple scattering and canopy closure. No optical‑only method including our end‑to‑end framework can fundamentally break this saturation bottleneck. The reviewer is correct that the end-to-end model’s smaller variance in the high LAI range does not represent a physical breakthrough; it is a statistical property indicating more stable and less variable underestimation.

We also acknowledge the reviewer’s point regarding SWIR bands. Although the model input includes SWIR1 and SWIR2 bands as stated in Section 2.1, these bands do not eliminate saturation. The fundamental limitation arises from canopy structure including multiple scattering and occlusion, and SWIR can only partially mitigate rather than remove the saturation effect. The model has not learned unsaturated features from SWIR that bypass physics, and we have revised the manuscript to avoid any such implication.

To address the reviewer’s concern, we have made the following revisions in the manuscript. In the Discussion section, we now explicitly state that spectral saturation is a physical ceiling of optical remote sensing. None of the models evaluated including the end‑to‑end framework can retrieve LAI accurately beyond approximately 4. The advantage of the end-to-end model in this regime is limited to lower prediction variance and more stable error behavior, that is statistical robustness, not higher absolute accuracy or saturation free retrieval. We have removed any language that could be misinterpreted as claiming physical learning; for example, the phrase “captures canopy structure” has been replaced with “learns empirical spectral patterns correlated with vegetation structure.”

We believe these revisions fully align with the reviewer’s correct physical insight and honestly position our contribution as empirical and statistical, not as a physical breakthrough. We thank the reviewer for pushing us to clarify this critical limitation.

Comment 7: The paper introduces the SSIM into its composite loss function. SSIM was originally designed as an image quality assessment metric based on human visual perception. For a highly quantitative physical parameter like LAI, forcefully imposing visual smoothness on spatial structures risks erasing genuine high-frequency spatial abrupt changes (local gradient differences) caused by topography, soil moisture variations, or human disturbances (e.g., crop harvesting). This logically explains why the end-to-end model's results appear overly smoothed and blurry upon visual inspection. The authors need to justify the ecological rationale behind penalizing these natural physical gradients.

Response: We thank the reviewer for this insightful and technically accurate observation. We agree that SSIM, originally designed for human visual perception, may penalize genuine high‑frequency spatial variations in a quantitative parameter such as LAI, including those caused by topography, soil moisture heterogeneity, or anthropogenic disturbances like crop harvesting. This could indeed contribute to the visually smoother appearance of our end‑to‑end model predictions.

However, the primary rationale for including SSIM in the composite loss function is to address the substantial noise and missing information inherent in cloud-contaminated inputs. In the cascaded models, this often leads to unrealistic artifacts in the reconstructed LAI maps, such as speckle-like noise or boundary discontinuities. The SSIM term helps the end-to-end model produce spatially coherent outputs that are more physically plausible and easier to interpret, albeit at the potential cost of sacrificing some fine-scale detail. We acknowledge this trade-off: smoothing reduces both noise and genuine local gradients. And adding a section in 2.2.3 to explain the function.

Crucially, all three models (end-to-end, cascaded, and GAN based cascaded models) were trained with the identical composite loss function, which includes the SSIM term. Therefore, if SSIM were the primary cause of over-smoothing, the effect should be similar across all models rather than being unique to the end-to-end framework. The fact that the end-to-end model appears visually smoother suggests that the smoothing is not attributable to SSIM alone, but rather to the inherent characteristic of the end-to-end paradigm itself. Specifically, direct mapping from cloud-contaminated inputs to LAI tends to suppress high-frequency noise and produce more continuous outputs, which is a feature of the paradigm rather than a side effect of the loss function.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thanks for the revised version. While the manuscript is mature now, I have a few more comments.

1) Please carefully revise the GBOV dataset. That is actually upscaled data, designed for validation of satellite data (e.g., S2/S3). So your remark that it would require upscaled data is not correct regarding GBOV.

2) While appreciating the new parts in Discussion, it reads all too hastily (perhaps AI-generated) prepared. I miss references and reflections on fundamental papers and similar works. For instance, the authors seem to be ignorant about uncertainty quantification methods/results presented in similar studies. As a general remark, this work cites only a few references. I strongly advise the authors to critically reflect on the Discussion and make it more in-depth and scientifically sound.

3) At the same time, the Conclusions are overly long. Please focus on presenting the main conclusions and keep all side information for discussion. Keep in mind that in Conclusions the reader wants to read the main conclusions of the work.

Author Response

Response to Reviewer 1 ( round 2)

Thanks for the revised version. While the manuscript is mature now, I have a few more comments.

1)Please carefully revise the GBOV dataset. That is actually upscaled data, designed for validation of satellite data (e.g., S2/S3). So your remark that it would require upscaled data is not correct regarding GBOV.

Response:
We thank the reviewer for this important clarification. We acknowledge that our previous remark about “requiring dedicated upscaling experiments” was inaccurate with respect to GBOV. GBOV indeed provides upscaled products (e.g., LP3) specifically designed for validating satellite data, derived from pointscale measurements (RM7) using transfer functions calibrated over homogeneous sites. Our earlier focus was on the RM7 point measurements, which led to the mistaken impression that upscaling would be necessary.

However, even when considering the LP3 product, two fundamental issues prevent its use for independent validation in our study. First, the geographical coverage is severely limited in Asia. The only GBOV sites are located in Japan, which does not cover our study area in China. Second, the transfer functions used to generate LP3 products were developed and optimized primarily over sites in North America and Europe (with a few in Australia). Their applicability to the diverse ecological and atmospheric conditions across China has not been validated. Applying these transfer functions directly to our region would introduce unknown and potentially large uncertainties, undermining the goal of independent validation.

Given these constraints, the feasibility of independent validation using GBOV under the current experimental design remains questionable. We have revised the Discussion to clarify this point, explicitly distinguishing between the RM7 and LP3 products and stating the geographical and methodological limitations. We also note that establishing a similar reference dataset for China would be a valuable future direction. We thank the reviewer again for correcting our understanding of GBOV.

2)While appreciating the new parts in Discussion, it reads all too hastily (perhaps AI-generated) prepared. I miss references and reflections on fundamental papers and similar works. For instance, the authors seem to be ignorant about uncertainty quantification methods/results presented in similar studies. As a general remark, this work cites only a few references. I strongly advise the authors to critically reflect on the Discussion and make it more in-depth and scientifically sound.

Response:
We sincerely apologize for the lack of depth, the insufficient literature integration, and the somewhat superficial tone in the newly added Discussion section from the previous revision. We deeply appreciate the reviewer’s candid and constructive criticism. You are completely correct that our previous discussion failed to critically reflect on fundamental Uncertainty Quantification (UQ) methodologies and their applications in similar remote sensing studies.

To address this critical shortcoming, we have extensively rewritten and expanded the Discussion section to ensure it is scientifically sound and thoroughly grounded in the existing literature. Specifically, we have implemented the following major improvements:

Clarification of Performed vs. Missing UQ: We clarified that while our non-parametric bootstrapping provides robust statistical uncertainty quantification for the evaluation metrics (confirming the macroscopic stability of our method), it does not substitute for pixel-wise predictive uncertainty (aleatoric and epistemic).
Critical Reflection on Similar Works: We have integrated citations to fundamental UQ studies in remote sensing, explicitly discussing how research have utilized Monte Carlo (MC) Dropout and Deep Ensembles for biophysical parameter retrieval.
Addressing the Foundation Model Bottleneck: We critically analyzed why these conventional, sample-intensive UQ methods, which require multiple forward passes or multiple models,are computationally prohibitive when applied to massive foundation models like Prithvi for large-scale operational tasks.
Outlining FM-Specific Future Directions: We outlined mathematically yet computationally efficient future pathways for foundation models (e.g., parameter-efficient probabilistic adaptation and evidential deep learning), providing a much more in-depth and forward-looking scientific perspective.

We have refined the academic language throughout this section to ensure a rigorous, objective tone. We believe these revisions, supported by newly added high-quality references, cab elevate the theoretical depth of the manuscript.

Response:

We thank the reviewer for this observation. We agree that the original Conclusions section was overly long and contained side information that belongs in the Discussion. In the revised manuscript, we have substantially shortened the Conclusions by removing detailed analyses, intermediate findings, and methodological justifications. The Conclusions now focus exclusively on the three main contributions of this work: (1) the empirical superiority of the endtoend paradigm over cascaded architectures, supported by quantitative metrics and bootstrap confidence intervals; (2) the Ushaped performance pattern and the model’s robustness under extreme cloud cover; and (3) the stability of the endtoend model near the spectral saturation limit. All side information, including the tradeoff between quantitative accuracy and perceptual fidelity, detailed comparisons of model behaviors under different conditions, and future research directions, has been moved to the Discussion. The revised Conclusions are now concise and present only the essential findings that the reader expects. We believe this revision addresses the reviewer’s concern.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

Thanks for the added experiments. The manuscript looks much better in this round. I recommend publish it in its current form.

Author Response

Thanks for the added experiments. The manuscript looks much better in this round. I recommend publish it in its current form.

Response: We sincerely thank the reviewer for the positive evaluation and for acknowledging the added experiments. We are very grateful for the recommendation to publish the manuscript in its current form. The reviewer’s insightful comments throughout the review process have significantly improved the quality of our work. Thank you again for your time and expertise.