We thank the reviewer for pointing us to PANGAEA/GeoBench (also suggested by reviewer HaH9). To facilitate reproducibility and fair comparison, we integrated our PhySwin encoder into the PANGAEA evaluation framework and re‑ran experiments under its unified training/evaluation recipes. Due to time constraints and a few checkpoints/datasets that required extensive manual fixes to download or run, we temporarily skipped them. We believe the results below already provide a comprehensive and representative picture of PhySwin’s performance.

We are continuously running the remaining experiments, including integrating our original baselines (SeCo, CaCo, and SatMAE) into PANGAEA to unify the benchmarks and baselines (an extra benefit beyond our core contribution). We will add the remaining items in the final paper.

TABLE 1: New PANGAEA results (mean ± std from three random seeds (42, 177, 892) as suggested by reviewer qjEV).
We will denote the best results in bold and the second-best results as underlined in the paper. Here, since OpenReview does not support underlining, we used bold+italic for the best results and bold for the second-best results.

Takeaways. PhySwin‑B is consistently ranking among the top 2 performers across all benchmarks with S2-exclusive data. PhySwin‑T is competitive given its smaller capacity, supporting our efficiency‑first design.

We will release our PANGAEA configs and scripts for these runs to ensure full reproducibility. We also fixed our table styling as suggested (bold best, underlined second‑best) and will maintain this convention throughout the paper and supplement.

Point 1: F1 scores is missing

Thanks for the suggestion. We now report macro‑F1 (mean ± std) alongside accuracy/mAP under the unified PANGAEA recipe.

• EuroSAT: single‑label and balanced, so macro‑F1 is equivalent to accuracy.
• BigEarthNet: multi‑label and long‑tailed.

Under the PANGAEA settings, PhySwin‑B ranks Top‑2 in Acc and mAP (see TABLE 2), and shows only slightly lower performance compared to SOTA baselines. As we noted to reviewers qjEV and HaH9, Skysense is not open‑sourced, so the ~4% gap in our paper may stem from setting differences. These results, produced using the unified and open‑sourced PANGAEA framework, demonstrate that our method incurs only a minor compromise in classification performance but still top-tier, which is consistent with the findings reported in Table 3 of our paper.

We will explicitly label which results are from our experiments and which are taken from the literature in the revised version (as suggested by reviewer FjVd). This will give a clearer picture and may further justify our method.

TABLE 2: Current classification results (mean ± std).

Note, GeoBench uses only a subset of the BigEarthNet dataset (20,000 patches for training and 1,000 for testing), so results may differ from those obtained using the full dataset in our original training‑testing setup.

We will update the full classification table in the revised version.

Point 2: Additional benchmarks & comparisons

Please see TABLE 1 for the new results on benchmarks against recent EO models.
These results align with our paper’s conclusions: PhySwin consistently ranks top‑1/top‑2 while remaining efficient.

All results are reproducible with PANGAEA’s default recipe using S2 bands and the command below (3 random seeds: 42, 177, 892):

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
  --config-name=train \
  decoder=seg_upernet preprocessing=seg_resize criterion=cross_entropy \
  task=segmentation task.trainer.n_epochs=80 batch_size=16 \
  encoder=ENCODER dataset=DATASET

Replace ENCODER/DATASET with the model and dataset names in configs/encoder and configs/dataset used in TABLE 1. The experiments are run on RTX 5090 GPU 24GB.

Point 3: Claim on pretraining data volume/quality/token count (line 288)

We appreciate the reviewer’s attention to this statement. Since none of our results or conclusions depend on this statement, we will remove the claim from the paper (main text and supplement) and tighten the wording to avoid any over‑statement. We will treat scaling effects as future work and only report such correlations once supported by dedicated experiments.

Point 4: Evaluation heads clarity

We agree and will clarify the evaluation heads in the main text. We will also add the exact configs used (encoder/head/loss) to ensure reproducibility in the Supplementary.

Point 5: Multimodality scope clarity

Thank you for raising this. To avoid over‑reach, we will clarify the scope in the title/abstract and main text, positioning PhySwin as an efficient multispectral encoder pretraining method rather than a full EO foundation model.

Since our initial focus was on developing an on‑satellite, deployment‑level efficient solution, multimodality was not included in this first stage. With the progress we have made, we are now actively working on enhancing performance with multimodal strategies and exploring efficient solutions in this direction. Furthermore, we believe our current settings are valuable for future multimodal training, as stronger S2 representations could benefit the overall multimodal system. We will continue exploring this angle in future work.

Point 6: Physical regularizations

Thank you for the thoughtful feedback. We agree our wording can better convey the intent and limits of the two regularizations. Our goal is to inject physics‑motivated priors into pretraining, not to claim full radiative‑transfer modeling.

• Spectral smoothness: leverages the well‑established property that surface reflectance varies smoothly with wavelength (Fig. 1b, §3.1), promoting stable inter‑band relationships and more physically consistent features. Our implementation operates on adjacent S2 bands in spectral order; while it does not explicitly weight by wavelength spacing Δλ, we will state this modeling choice and note Δλ‑aware weighting as future work.

• Energy bounds: enforce the physical principle that surface reflectance is bounded (Fig. 1a, §3.1), guiding the model toward realistic ranges and avoiding unphysical reconstructions. We will clarify that this is a soft, physics‑motivated barrier (upper bound relaxed to tolerate noise).

Table 5a shows that each prior improves performance and their combination yields the strongest gains, confirming complementarity. We will make this physical motivation and the above clarifications explicit in the revised manuscript.

Point 1: EuroSAT F1 scores

We thank the reviewer for raising this. We agree that our earlier phrasing was imprecise, where we did not intend to claim that macro-F1 and accuracy are generally equal. What we meant is that in our current runs (using the PANGAEA evaluation protocol), these metrics were numerically very close.

To clarify, we now report per-class F1 scores and macro-F1 for three models: SatlasNet, PhySwin-T, and PhySwin-B. This table makes the classwise variation and summary metrics explicit:

EuroSAT Per-class F1 and summary metrics (PANGAEA setting, test split)

Notes.

• mF1 is the unweighted mean of the ten per-class F1 scores.
• OA is the standard overall accuracy.

Why the metrics coincide here.
EuroSAT has near-balanced classes, and all three models achieve uniformly high per-class F1. Under these conditions, macro-F1 and OA tend to be numerically close, which is what we observe in our runs. This is, however, not a general identity; if the dataset were more imbalanced or class performance varied more, the two metrics would diverge.

Reproducibility.
All results above come from the PANGAEA evaluation protocol and can be reproduced with:

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
  --config-name=train decoder=cls_linear preprocessing=cls_resize \
  criterion=cross_entropy task=linear_classification \
  task.trainer.n_epochs=50 batch_size=64 \
  encoder=YOURENCODER dataset=meurosat

This command will generate the OA.

We have also updated pangaea/engine/evaluator.py to explicitly log macro-averaged metrics using:

precision_macro, recall_macro, f1_macro, _ = sklearn.metrics.precision_recall_fscore_support(
    all_targets, all_preds, average="macro", zero_division=0
)

This will generate the macro-F1.

Point 2: Additional benchmarks & comparisons

Setup clarity. All new results were produced with the PANGAEA evaluation protocol and are fully reproducible. For PASTIS and CropMap, we used the PANGAEA recipes but, to keep S2 setting, we ran with:

• multi_temporal = 1 (single timestamp)
• multi_modal = False (no additional modalities beyond S2)

This design choice could be the main reason some PASTIS numbers are lower than time-series or multi-modal results reported elsewhere. But our results show that CropMap is less sensitive to temporal stacking. In many cases, our scores even exceed numbers commonly reported in the literature, for example, in the pangaea paper.

On comparisons to AnySat / Galileo / TerraMind

Due to time constraints, we prioritized SOTA baselines included with PANGAEA to avoid manual, method-specific fixes. For the reviewer mentioned models, we met the following constraints:

• TerraMind: Public checkpoint in PANGAEA is incompatible with PANGAEA encoder code (dimensionality/layer mismatch); we are developing an adapter and will include results once aligned.
• AnySat and Galileo: Not included in PANGAEA; we will try our best to integrate via their official repos.

We hope this clarifies the observed gaps and the work we have done.

Point 6: Physical regularizations

We appreciate this feedback and agree our priors are physics-motivated rather than full physical models. We will adjust the wording accordingly and clarify our design choices.

Energy upper bound sensitivity

We chose 1.20 to allow for illumination variance, calibration noise, and bright surfaces. To show sensitivity, we will add an ablation table (Appendix) sweeping {1.0, 1.05, 1.1, 1.2} in the final version. Our expectation is that tighter bounds may reduce over-bright predictions but could also over-penalize genuine high-albedo or low-shadow cases, potentially hurting downstream accuracy.

Non-uniform wavelength handling for spectral smoothness

Our current pretraining data has fairly regular band spacing, so we did not explicitly handle non-uniform wavelength distributions. We agree this is important for broader settings. One improvement could be to weight the smoothness penalty by spectral distance:
L_smooth = sum_b (1 / Δλ_b) * (r_{b+1} - r_b)^2,
where Δλ_b is the difference in central wavelength. We leave this as future work, especially for datasets with highly non-uniform spacing.

We welcome further discussion. Thanks.

We sincerely thank the reviewer's active engagement and constructive feedback.

Point 1. Our analysis aligns with yours, and the per-class results confirm the explanation. Thank you for highlighting this interesting point.

Point 2. We will explicitly state the evaluation setting for each dataset to ensure reproducibility. Upon acceptance, we will open-source our pretraining and benchmarking code.

We are actively working to resolve the terramind_large compatibility issue in PANGAEA and will report back as soon as it is fixed. For AnySat, we are integrating it into the unified framework based on the gastruc/AnySat repo. But if time runs out during the discussion, we will continue these efforts post-discussion and, in the meantime, include reported-from-paper results to ensure modern baselines are represented.

Point 3. Thank you for encouraging the upper-bound ablation. We now start to run the energy bound sweep and include the results in the final version/supplementary. Since our current value was chosen empirically, we expect the ablation to confirm the choice and may reveal useful trade-offs.

I thank the authors for their prompt response.

Point 1:

I find it surprising that the F1 score and accuracy are exactly equal up to three significant digits. This can likely be explained by the near-perfect performance of most models on this dataset, combined with the dataset being approximately balanced.

Point 2:

The inclusion of selected datasets from the PANGEA benchmark represents a meaningful step toward improving the evaluation of the proposed method. However, all baseline methods used for comparison are at least two years old, which represents a significant gap considering the rapid development in this field. While I do not expect nor require the proposed method to achieve state-of-the-art performance on all datasets, I believe a proper comparison with modern encoders is essential for a comprehensive evaluation.

For datasets that are typically multitemporal or multimodal, I encourage the authors to clearly specify that their evaluation is conducted using monotemporal S2 data only.

Point 3:

An ablation study on the upper bound relaxation would indeed be valuable to justify this design choice. While proposing an improved formulation for spectral smoothness is interesting (and could potentially be evaluated using a subset of S2 wavelengths), I agree that this falls outside the scope of the current paper.

We sincerely thank you for your valuable comments and constructive engagement. We are continually working on integrating the latest models into our comparison. We would appreciate it if you could let us know whether our response sufficiently addresses your concern.

Thank you for your suggestion regarding the inclusion of SOTA models for comparison. After fixing Terramind, we have taken steps to add AnySat (Base) into Pangaea, to enable fair and consistent comparisons.

Since AnySat is not originally supported by Pangaea, we manually integrated it based on the publicly available repo gastruc/AnySat. Below, we detail the setup and configuration used to ensure transparency and reproducibility:

• Modality: We used S2 data.
• Input Configuration: According to AnySat's official setup, the S2 configuration is: s2 BxTx10xHxW B2, B3, B4, B5, B6, B7, B8, B8a, B11, B12 10m

We followed this specification, and used the single timestep in Pangaea:

• We manually extracted intermediate features from layers 1, 3, 5, and 7 of the Transformer blocks after the S2 projector.
• This was done to align with Pangaea’s feature extraction convention, enabling compatibility with segmentation downstream heads.

We have currently obtained preliminary results on three datasets:

• EuroSat: 92.30%
• Sen1Floods11: 87.57%
• Mados: 62.94%

We note that AnySat does not outperform PhySwin under the current setting. However, we acknowledge that our setup may not fully leverage AnySat’s capacity. According to the original AnySat paper, the model can achieve 91.1% mIoU on the Sen1Floods11 dataset.

To provide a comprehensive picture, we are continuing experiments with other datasets and also plan to include both:

• Our re-implemented results under Pangaea setting, and
• Reported results from the literature

in the Supplementary Material.

We hope this addition addresses the reviewer’s comment and reinforces the fairness and comprehensiveness of our comparisons.

We appreciate the reviewer’s positive assessment of our comparison with EO foundation models and are pleased that the performance of the proposed PhySwin encoder was found compelling.

As suggested, we will revise the manuscript to state explicitly that all evaluations are conducted under an S2-only and monotemporal setting. We sincerely thank the reviewer for the constructive feedback and valuable discussion throughout this process.

Point 1: Ablations isolating the effect of physics‑informed losses

Thank you for the suggestion. To isolate the contribution of the physics priors, we ran physics‑only pretraining (no Spectral Group Masking, no refined MixMAE), alongside group‑only, refined‑MixMAE‑only, and the full with‑all configuration. As suggested by reviewer qjEV, all results are reported as mean ± std over three random seeds (42, 177, 892).

Implementation details.

• Due to time constraints, we ran 15 epochs for PhySwin‑T on each variant to demonstrate relative performance. Full training results will be presented in the final paper.
• Base (SimMIM): native patch embedding and SimMIM pretraining.
• Physics‑only: same embedding as the base, but pretraining loss includes physical constraints.
• Group‑only: uses our grouped spectral embedding with SimMIM on the spatial dimension.
• Refined‑MixMAE‑only: uses our refined MixMAE pretraining with native embedding and no physics constraints in the loss.
• With‑all: combines refined MixMAE, physics priors, and grouped embedding.
• We also included a non‑trained SwinV2 (scratch) for comparison.

Takeaways.

• Physics‑only improves over the SimMIM base consistently across all tasks: +2.92 F1 (OSCD), +2.07 mIoU (SegMunich), +1.22 F1 (EuroSAT), confirming that physics‑informed objectives contribute independently.
• With‑all maintains gains over the base while enabling the efficiency benefits of grouped embedding and refined MixMAE.
• Group‑only and refined‑MixMAE‑only do not explain the observed accuracy gains, underscoring that the physics priors are a major driver of performance, while other components mainly deliver efficiency.

We will add a short note in §4.4 to clearly attribute accuracy gains to the physics‑informed losses and efficiency to the architectural choices.

Point 2: Physics‑informed losses on MAE SOTA backbones

Thank you for this suggestion. To test generality beyond PhySwin, we added our two physics losses to a ViT-B model pretrained with standard MAE (no grouped embedding, no refined MixMAE).

We can observe that the gains arise again. This is align with our statements in the paper that the physics regularizers encourage inter‑band consistency and suppress out‑of‑range reconstructions, guiding features toward a physically plausible direction and faster adaptation to valid ranges. We will add a brief note in §4.4 as a baseline for refined MixMAE and Spectral Group Masking.

(Compute note: these runs follow the same reduced‑epoch style, 15 epochs, used in our ablations; we did not retune hyperparameters for the ViT + physics.)

Point 3: Clarifying importance, physics link and architectural emphasis

Thank you for this constructive critique. We will rebalance the presentation to make the motivation and contributions clearer.

Why it matters
We observe a growing number of foundation models emerging in EO, with a strong focus on scaling parameters and pretraining datasets to push performance. While this trend has yielded impressive accuracy, there is also a rapidly growing interest in deployment‑level ML for EO. This interest is motivated by the fact that large amounts of EO data can be discarded or delayed because of the high cost and latency of communicating with ground stations. Our work is designed with this balance in mind. We aim not only for strong accuracy but also for efficiency at both training/finetuning and inference. Specifically, we introduce a grouped spectral embedding to reduce token counts without sacrificing spectral structure and refine MixMAE to maintain Swin’s window‑shifted context while being more efficient during the pretraining. Together with our physics‑guided priors, these components address the dual challenges of learning meaningful multispectral representations and achieving the throughput/memory profile required for practical deployment.

Physics priors are not generic "smoothing" or "clipping"

• Spectral smoothness encodes the well‑established observation that surface reflectance varies smoothly with wavelength; we implement this as a first‑order difference on adjacent bands in spectral order and apply it to the reconstructed reflectance during pretraining.
• Energy bounds encode the physical boundedness of reflectance, serving as a soft barrier that keeps reconstructions in a realistic range while tolerating sensor noise.

These losses operate on reconstructed reflectance vectors, tying them directly to spectral physics rather than generic pixel heuristics. Our ablations (Point 1) and MAE‑backbone test (Point 2) show consistent gains even without architectural changes, indicating the priors improve representation quality in a plug‑in manner.

Architectural choices

• Backbone: SwinV2 for hierarchical, windowed attention → near‑linear scaling and strong dense‑task transfer.
• Refined MixMAE: we retain SW‑MSA and coordinate masking with window shifts, preserving cross‑window context while avoiding [MASK] tokens.
• Token‑efficient grouped spectral embedding (+MaskSpec): preserves rich band structure without increasing token count, delivering high throughput with robust features.

Our efficiency results in the paper highlight that these choices approach ConvNet‑level throughput while remaining competitive on accuracy.

We will (i) move a concise design‑rationale paragraph to the end of §3, explicitly mapping each component to accuracy vs efficiency goals; (ii) add a short physics‑motivation paragraph in §3.1 linking the losses to spectral smoothness and bounded reflectance. We hope these changes make the importance and novelty clearer.

Point 4: Additional suggestions and clarity improvements

We thank the reviewer for these practical suggestions and will address each point:

1. Marking replicated results: We will clearly label in all tables which results are independently reproduced by us and which are reported directly from the original papers.

2. Alternative masking strategies: While orthogonal to our current design, we agree that exploring single‑component masking (e.g., uniformly sampled masks across spectral groups) could offer additional insights into feature robustness and inter‑group relationships. We will explicitly note this as an interesting direction for future work.

3. Figure 1 captions: We will expand the captions for Figure 1a–c to better explain:

   • Figure 1a: The bounded reflectance distribution in Sentinel‑2 multispectral data and how this motivates the energy‑bound loss.
   • Figure 1b: The smooth transitions across adjacent MS bands, showing empirical spectral differences and motivating the spectral‑smoothness loss.
   • Figure 1c: The trade‑off between accuracy and inference throughput for EO foundation models, positioning PhySwin’s improvements.

   These expanded captions will explicitly link the visuals to the physics‑informed objectives (§3.1) and the overall efficiency‑accuracy motivation of PhySwin.

Point 1: "Physically‑grounded" claim clarification

Thank you for raising this. We would like to clarify that our physical component comprises two priors, not only a bound:

• Spectral smoothness: enforces continuity across adjacent spectral bands, reflecting the well‑documented smooth variation of surface reflectance with wavelength (Fig. 1b, §3.1).
• Energy bounds: applies a soft reflectance‑range prior so reconstructions remain physically plausible (Fig. 1a, §3.1).

As suggested by reviewer FjVd, we also ran the new ablations as the evidence that the priors matter (beyond architecture). More information can be found in our answers of reviewer FjVd.

A) Isolating the physics priors within our setup (PhySwin‑T, 15‑epoch for all variants due to compute limits). (As suggested by reviewer qjEV, we report mean ± std over 3 seeds 42/177/892)

This isolates the contribution of the priors: they improve accuracy without grouped embedding or refined MixMAE.

B) Adding the priors to a separate backbone (ViT‑B + MAE, 15 epoch).

These arise that the physics regularizers encourage inter‑band consistency and suppress out‑of‑range reconstructions, guiding features toward a physically plausible direction and faster adaptation to valid value ranges.

Furthermore, to avoid overstatement, we will adjust the wording to “physics‑informed priors” and add explicit pointers to Fig. 1a–b and §3.1 where these losses are defined and motivated.

Point 2: "Incremental changes"

We appreciate the concern and will make the design rationale more explicit. Our contributions target the dual goal of accuracy + deployment‑level efficiency for multispectral EO, and they work together rather than as small tweaks in isolation:

• Token‑efficient grouped spectral embedding.
  Unlike standard patch embeddings whose token count grows with channels, our grouped spectral embedding keeps the token count fixed while preserving band structure via lightweight per‑group mappings and concatenation. This directly reduces memory/FLOPs without discarding spectral information. Meanwhile the randomly masking spectral groups during the pretraining ecourage the model to learn from different groups, avoiding the bias in features.

• Refined MixMAE that preserves SW‑MSA.
  We coordinate MixMAE’s masking with Swin’s shifted windows, retaining cross‑window context and eliminating [MASK] tokens that otherwise slow hierarchical transformers. This lets us pretrain Swin‑v2 faithfully and then fine‑tune without modifying the architecture.

• Physics‑informed priors (two losses).
  Beyond architecture, we introduce spectral smoothness (adjacent‑band continuity) and energy bounds (plausible reflectance range). These priors improve accuracy independently of architectural changes (see Point 1 ablations and the ViT‑MAE + Physics result).

Action we will take: we will refine design‑rationale paragraph for mapping each component to its role (grouped embedding → token/compute efficiency; refined MixMAE → Swin‑compatible pretraining with high throughput; physics priors → accuracy/robustness), and we will put these connections earlier to avoid the impression of small, incremental tweaks. Thank you for raising this point.

Point 3: Classification results and "near‑SOTA" concern

Thank you for flagging this. The cited +4 pts comes from literature‑reported Skysense numbers; Skysense is not open‑sourced, so we cannot evaluate it under the same protocol/settings.

To ensure scope and fairness, and as suggested by reviewer HaH9 and HFfE, we moved classification to the open‑sourced PANGAEA/GeoBench framework and now report macro‑F1 in addition to accuracy/mAP. Under this unified setup, we get TABLE 1 in our response for reviewer qjEV.

These standardized, reproducible runs show that PhySwin‑B is still top‑tier compared with the most recent baselines while maintaining strong efficiency. We believe this can address the "near‑SOTA" concern under a unified evaluation protocol.

Our actions: We will label ours (reproduced) vs from literature in all tables for transparency. Some strong baselines (e.g., non‑open‑sourced checkpoints) may reflect recipe differences.

Point 4: Baseline coverage concern

Thanks for pointing this out. To address it fairly and reproducibly, we moved our evaluation to the PANGAEA/GeoBench framework and added recent EO FMs to the comparison, following the unified recipe.

TABLE 1: PANGAEA results (S2 benchmarks).

Takeaway. Under a standardized pipeline, PhySwin‑B is Top‑1/Top‑2 across all four tasks, while preserving efficiency.

Reproducibility. We use PANGAEA defaults with S2 bands and hyperparameter settings; runs can be reproduced with the following command:

torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
  --config-name=train \
  decoder=seg_upernet preprocessing=seg_resize criterion=cross_entropy \
  task=segmentation task.trainer.n_epochs=80 batch_size=16 \
  encoder=ENCODER dataset=DATASET

Note: Some models (e.g., terramind_large) require manual fixes to download and run experiments, so we skip them for now. We are continuously adding those that can be executed in PANGAEA and will include any newly completed entries in the later discussion and revised version if required. More experimental settings can be found in our answers for reviewer HFfE.

Point 5: Temporal aspect

Thank you for raising this. We totally agree that Temporal modeling is important for EO (e.g., crop phenology, flooding dynamics).

Why we did not include it in this release.
Our first objective was to balance accuracy and deployment‑level efficiency under a fixed token/compute budget (on‑satellite/edge constraints). Multi‑temporal pretraining multiplies the token count and memory footprint; we therefore prioritized an efficient multispectral encoder and validated it thoroughly under a standardized pipeline. Note that while pretraining is single‑timestamp, our change‑detection downstream head already consumes bi‑temporal inputs in fine‑tuning/evaluation following the benchmark recipe. Meanwhile, this performance ranks among the top in all SOTA benchmark tests.

For question Is it important / can the model handle multiple images? The answer is Yes. PhySwin is readily extensible to multi‑temporal inputs, and this is under active development. Concretely, we are trying to implement (i) temporal patching/merging to keep the token budget roughly fixed, and (ii) spatio‑temporal masking aligned with SW‑MSA. These are directly compatible with our PhySwin design.

We will explicitly state in the paper that this version focuses on efficient pretraining, and we will list temporal extensions (temporal patching/merging + spatio‑temporal masking) as future work. This roadmap preserves the efficiency goals while adding the temporal dimension.

Point 6: Accuracy–efficiency trade‑offs

Yes, PhySwin is able to trade accuracy for speed/memory.

Practical aspects to tune the trade‑off.

• Model size: choose T for edge/real‑time or B for higher accuracy with moderate extra cost.
• Token budget: adjust spectral grouping granularity (fewer groups → fewer params/tokens; more groups → richer per‑band specialization).
• Swin window/shift settings: larger windows capture more context at extra compute.

For example, across our new PANGAEA/GeoBench results (Tables 1–2), PhySwin‑B consistently outperforms PhySwin‑T on segmentation, change detection, and classification, while PhySwin‑T delivers higher throughput and lower memory (Table 4 in the paper). This shows scaling model size helps accuracy, but we intentionally keep both variants compact to meet our deployment constraints.

We will add a brief paragraph in the paper summarizing these tunes' effects to help readers select a configuration aligned with their compute budget.

We thank the reviewer for the valuable comments. We would appreciate it if you could let us know whether our response sufficiently addresses the concerns.

We thank the reviewer for the valuable comments. We would appreciate it if you could let us know whether our response sufficiently addresses the concerns.

Point 1: Classification performance clarification

Thank you for raising this. The cited relatively large +4 pts gap comes from literature‑reported Skysense numbers. Skysense is not open‑sourced, so we cannot evaluate it under the same protocol. Following other reviewers’ suggestions, we now use the unified, open‑sourced PANGAEA/GeoBench framework and report macro‑F1 (in addition to Acc/mAP), with mean ± std over 3 seeds (42/177/892).

TABLE 1: Classification under P/GEO (EuroSAT & BigEarthNet‑subset).

Notes: (i) GeoBench uses a subset of BigEarthNet (20k train / 1k test), so numbers differ from full‑dataset training; (ii) Provenance will be labeled in all tables (ours (reproduced) vs from literature).

As shown, PhySwin‑B remains top‑tier and aligns with our paper’s comparisons (excluding Skysense), while maintaining the efficiency needed for deployment‑level use. Similar discussions can be also seen in point 3 for reviewer HaH9.

Point 2: Uncertainty estimates & significance

Thank you for noting this. In our updated experiments we now report mean ± std over 3 random seeds (42, 177, 892) for all tasks and tables (including newly added experiments based on PANGAEA/GeoBench). Across seeds, results are stable, normally less than 1% (small std), and we will add seed info and mean ± std to table captions in the paper and supplement. Additional results can also be found in the tables included in our answers for other reviewers.

These additions address the reviewer’s request for uncertainty reporting and statistical support.

Point 3: Spectral smoothness vs. band misregistration

Thank you for the thoughtful question. We agree misregistration can bias a spectral‑smoothness prior if adjacent bands are not perfectly aligned.

Currently, in our pipeline, the smoothness loss is implemented as a soft penalty (with moderate weight) applied to the reconstructed reflectance vector per pixel. This means it serves as a gentle prior rather than a strict constraint. However, we acknowledge that residual misregistration remains a limitation of the current formulation. As future work, we plan to explore misregistration‑robust smoothness variants, or alternatively, more straightforward approaches such as data preprocessing or filtering methods to avoid this happens.

Point 4: Cross‑source attention constraint (L141–142) & dual reconstruction

Thank you for the question. We will clarify how we ensure "tokens attend only to others from the same source image, as dictated by M."

During pretraining with mixing, we build an attention mask from the mixing map M and apply it to every self‑attention layer (both W‑MSA and SW‑MSA). Concretely, if tokens i and j come from different source images according to M, the attention score entry is set to the very large nagative value before softmax (i.e., a block‑diagonal attention mask over sources). In addition, window partitioning packs tokens per source so that windows never span tokens from two sources; the mask remains active after the shift in SW‑MSA to prevent any cross‑source leakage.

For Dual reconstruction (Eq. 3), we will add one sentence noting that the dual reconstruction loss reconstructs the masked regions of each source image from its own unmasked context, i.e., the model is trained to recover missing content from real within‑source evidence, not from cross‑source tokens.

We will include more details in Section 3 and new pseudo‑code in the Supplementary.

Point 5: Why Skysense is absent from Table 4

As noted in Point 1, Skysense is not open‑sourced, so we cannot reproduce it under our unified pipeline. Table 4 includes only models we can run end‑to‑end.

Point 6: Figure 3 color legend

Thanks for catching this. We will add a color legend to Figure 3 and clarify the caption. Quick details:

• Top: OSCD change‑detection masks (foreground/background).
• Bottom: SegMunich (13 LULC classes) and Dyna.-S2 semantic segmentation—colors will follow each dataset’s official palette and be listed in the legend.

We will update the caption accordingly.

We thank the reviewer for providing the final score and for the valuable comments on our work. We would appreciate it if you could let us know whether our response sufficiently addresses your concern.