Spiking Meets Attention: Efficient Remote Sensing Image Super-Resolution with Attention Spiking Neural Networks

Response to Reviewer WHSi

Dear Reviewer WHSi:

We sincerely thank the reviewer for the encouraging and insightful feedback. We are pleased that you consider our first attempt as a valuable and highly original contribution to the field, and our observation of active learning states provides a new perspective in large-scale Earth observation scenarios. We are especially encouraged by your recognition that our performance gain against ANN-based models is non-trivial, demonstrating a significant technical superiority. We hope our responses below can address your concerns.

Weaknesses

W1: Clarify the application level of deformable convolution.

A1: To save computational cost, deformable convolution is applied at the image level. This enables our model to flexibly mitigate geometric transformations across entire images while maintaining efficiency. We will clarify this detail in the revised manuscript!

W2: Please provide more ablation on different selections of T.

A2: To assess its impact, we conducted additional ablation studies with T=1,2,4,8. The results are presented below. It can be observed that as T increases, more spatiotemporal information is incorporated, leading to progressively improved performance. However, this improvement saturates when T=4. Considering that model complexity also grows with increasing T, we choose T=4 as the final setting to strike a favorable balance between performance and efficiency. This ablation will be included in the Appendix!

W3: More clarification of pixel-level intensity observed in Fig. 1a should be discussed.

A3: Yes, the smoothed pixel-level intensity indeed reflects the tendency of ANN-based architecture (e.g., convolution kernel) to over-smooth local textures, which may hinder their ability to capture fine-grained spatial details. This phenomenon highlights the importance of temporal dynamics and sparse activation in enhancing spatial sensitivity. To clarify this, we will revise the caption of Fig. 1 to explicitly point out that the smoothing effect serves as an intuitive motivation for adopting SNN architectures in SR tasks.

W4: The details of AID-tiny are unclear.

A4: Thanks for your suggestion. The AID-tiny is a lightweight dataset derived from the AID dataset, where we randomly select one image from each of the 30 categories, resulting in a total of 30 samples. These samples are non-overlapping with the training and testing sets. To ensure reproducibility, we will release the AID-tiny dataset publicly on our GitHub repository.

Questions

Q1: Some typos should be corrected.

A1: Thank you for pointing this out. We have carefully reviewed the manuscript and corrected the noted typos, including those following the values 90, 180, and 270. In addition, we have added the missing section title for Section 1. These revisions will be reflected in the updated manuscript!

Dear reviewer,

Thank you for the comments on our paper.

We have submitted the responses to your comments. Please let us know if you have additional questions so that we can address them during the discussion period.

Thank you!

Best,

Authors

Response to Reviewer Ecnp

Dear Reviewer Ecnp:

Thank you for your encouraging comments and insightful questions. We are glad that you found our method technically sound, and recognized the application of SNN to SR as relatively novel and conceptually promising. We also appreciate your recognition of the well-articulated motivation behind our work, and your acknowledgment that our research direction is timely, meaningful, and worth exploring, which highlights the originality and potential impact of our contribution. We hope that the responses below will address your concerns, and we would be truly grateful if you could consider updating the score!

Weaknesses and Questions

W1 & Q3: The paper would benefit from more ablation of time steps and connection to the underlying spiking dynamics.

A1: To better understand the influence of time steps: we conducted additional ablation experiments by setting different numbers of simulation steps T={1, 2, 4, 8}. As shown in the table below, increasing T allows our model to integrate more spatiotemporal cues, which leads to increased performance improvements. However, this gain saturates around T=4, suggesting that further increasing T may introduce redundant or noisy information. Moreover, a larger T also leads to additional computational overhead. Therefore, we selected T=4 as a practical trade-off between efficiency and reconstruction quality. These ablations will be included in the Appendix to strengthen the argument for the effectiveness of spike-based representations in the SR context.

To further clarify the connection between HDA and DSA modules and the spiking temporal dynamics: we conducted an in-depth visualization of feature and spiking representations. First, HDA effectively filters out redundant spike activations along both temporal and channel dimensions, thereby enhancing the quality of feature representations, as illustrated in Fig. 7(b). Furthermore, we additionally visualized spiking maps at different time steps and observed that, with the guidance of DSA, the temporal signals exhibit notable variation across different time steps. This indicates that DSA significantly enriches the temporal dynamics and mitigates the homogenization issue commonly observed in SNNs over the temporal dimension. These visualization results and analyses will be included in the supplementary materials!

W2 & Q1: More explanation of HDA could significantly strengthen the clarity.

A2: In the revised main body of our paper, we will describe the working mechanism of the HDA module and briefly introduce its structure. Specifically, HDA first applies a squeeze operation to compress membrane potentials into a compact matrix. It then performs 1D convolutions independently along the temporal and channel dimensions to extract attention weights that reflect the dynamic importance of features across both axes. To model their interdependence, we perform element-wise multiplication to fuse the two attention weights into a unified temporal-channel attention map, which is ultimately used to modulate the membrane potential representations. Instead of operating directly on raw spike trains, HDA enhances feature representation by adaptively adjusting membrane potentials, enabling it to capture rich spatiotemporal information while preserving the discrete nature of SNNs.

Moreover, we will release the source code to ensure technical completeness and reproducibility of HDA. Also, we will ensure the appendix is available and sufficiently descriptive.

W3: More investigation of spiking dynamics.

A3: We have added more visual analyses to illustrate how spike-based temporal information flow facilitates texture restoration and detail propagation. Specifically, we present visualizations of spiking maps across different time steps, which reveal that the proposed DSA module enhances the variation of spiking dynamics at each temporal stage by introducing self-similarity priors.

Moreover, features across different time steps exhibit distinct levels of detail saliency, which aligns with the biological inspiration of SNNs—where the human brain selectively attends to different information at different moments. After removing the HDA module, we observe a noticeable loss of high-frequency textures and more homogeneous patterns across time steps. These visualizations further validate the effectiveness of our model in addressing the key challenges of super-resolution.

Q2: The connection between spiking responses and actual SR reconstruction quality in Fig.1 should be clarified.

We sincerely appreciate your recognition of Figure 1 as an interesting perspective.

As clarified, the high-voltage regions do not directly correspond to high-frequency areas in the final SR outputs. Rather, the visualization in Fig. 1 is intended to illustrate our motivation, not to serve as direct empirical evidence. Specifically, it demonstrates that spiking neurons can sustain an active state even when high-frequency details are reduced. In contrast, intuitively, convolutional networks tend to exhibit reduced activation in response to degraded high-frequency inputs, which may limit their representational capacity. We will revise the caption of Fig. 1 to better clarify this point and avoid potential misinterpretation.

Dear reviewer,

Thank you for the comments on our paper.

We have submitted the responses to your comments. Please let us know if you have additional questions so that we can address them during the discussion period.

Thank you!

Best,

Authors

Dear Reviewer Ecnp:

Many thanks for your time!

We noticed there has been no response to our feedback, and would like to kindly confirm whether our updates have addressed your concerns.

As the discussion phase is approaching its end, please be assured that we are ready to clarify any remaining questions you might have!

Best regards,

All authors

Response to Reviewer DFR5

Dear Reviewer DFR5:

Thank you for your insightful comments. We are pleased that you think our SpikeSR is a novel framework, recognize its practical applications, and consider our manuscript well-written and well-organized. We appreciate your recognition of its achievement of state-of-the-art performance across multiple remote sensing benchmarks, while maintaining high computational efficiency. We hope our responses below address your concerns, and we kindly ask you to consider updating the score.

Weakness

W1: What are the benefits over lightweight SR networks?

A1: Thank you for your valuable feedback. Compared to lightweight SR networks, SNNs offer unique advantages due to their biological plausibility and sparse spatiotemporal dynamics. Thus:

1. These characteristics enable SNNs to achieve significantly lower energy consumption (as listed in the table below), which motivates us to apply SNN to reconstruct large-scale and high-resolution remote sensing images.

2. We observe spiking neurons maintain an active learning state across low-resolution images, even in severely damaged textures, which motivates us to leverage the inherent properties of SNNs to handle image degradation for efficient yet high-quality SR.

3. With the growing interest and demand for in-orbit processing of remote sensing data, our method enjoys highly energy-efficient benefits and favorable performance, making it a promising solution for practical scenarios.

We will polish the motivation in the revised manuscript to explicitly highlight these benefits!

W2: Concerns about the novelty

A2: Thank you for your thoughtful comment. We appreciate the opportunity to clarify the novelty and significance of our work：

1. The first attempt to bridge research gap for SNN-based SR

While our framework indeed integrates these components, we emphasize that it is the first attempt to successfully adapt and customize attention mechanisms specifically to the SNNs for SR tasks.

Although SNNs are known for their energy efficiency, they have rarely been explored in remote sensing SR tasks due to their limited representation power compared to ANN-based models. This is primarily due to the inherent difficulty of adapting continuous, dense attention operations to the discrete, sparse spiking signals and their temporal dynamics, which pose unique architectural and training challenges.

Our work is the first to mitigate this gap, achieving ANN-comparable or superior performance while maintaining significantly low energy consumption. This novelty is recognized by Reviewer WHSi, who states our approach makes a valuable and highly original contribution to the field and provides a new perspective on developing efficient models, and Reviewer bD3C, who highlights for the first time and filling the gap of SNN in the field of remote sensing SR, further affirming the impact and novelty of our contribution.

2. Concise is not simple: Our method is concise yet highly effective

Although our method is concise, it is far from simplistic. Our design achieves significant performance improvements with modest architectural modifications, demonstrating that intuitive and efficient solutions can be highly effective. As reported in the table below, the performance improvement demonstrates that simple yet spike-compatible attention mechanisms can improve the limited capacity of SNN by a large margin, making it competitive to ANN-based SR methods. We believe this aligns with the broader research ethos that: a well-justified, efficient improvement is often more impactful than a complex, over-parameterized solution. Reviewer WHSi also acknowledges that “achieving performance comparable to or even surpassing that of ANNs using SNNs is non-trivial.”

3. Customized technical novelty for attention-based SNNs

While the design of combining SNN backbones with attention might seem straightforward, our contribution lies in the careful adaptation and customization of attention modules (HDA and DSA) to the unique spike-based characteristics. In particular:

1. Our HDA and DSA operate on membrane potentials rather than dense activations, aligning with the discrete and temporal nature of spiking dynamics;

2. By exploring joint temporal and channel-wise dependencies within the sparse spiking sequences, we achieve a performance improvement of +0.07dB, compared to naive temporal and channel attention mechanisms.

3. Developing a pyramid structure to exploit rich multi-scale self-similarity priors. This design significantly enhances the performance against conventional spatial attention, achieving an improvement of +0.27 dB.

As reviewer WHSi notes: our technical superiority is significant.

We hope that the above clarifications can address your concerns, and we sincerely look forward to your updated score.

W3: Concerns about the performance

A3: Thank you for raising this valuable concern. We would like to clarify that our SpikeSR does not exhibit a performance disadvantage compared to HiT-SR. While our model has significantly fewer parameters (472K vs. 792K), it achieves a slightly lower performance (31.88 vs. 31.90), which is reasonable given the smaller model size. When comparing models with similar parameters (763K vs. 792K), our approach delivers comparable results. Furthermore, upon scaling up the model, SpikeSR outperforms HiT-SR (31.95 vs. 31.90), demonstrating the superiority and scalability of our framework.

To further highlight the favorable performance of our SpikeSR, we have conducted additional comparisons with two SOTA efficient SR methods, SeemoRe (ICML 2024) and CATANet (CVPR 2025). As shown below, SpikeSR remains highly competitive against these cutting-edge approaches. We will add these results to our manuscript!

W4 & W5: Missing energy consumption and memory overhead

A4 & A5: We have additionally evaluated these metrics to address this concern. As shown in the table below, our SpikeSR demonstrates significantly lower energy consumption compared to ANN-based methods, highlighting the inherent energy efficiency advantage of SNNs. We will include a comprehensive discussion on model efficiency in the revised manuscript!

Q1: Provide a more detailed theoretical analysis of the proposed SAB

A1: Thank you for raising this insightful question. We provide a more in-depth theoretical analysis of the proposed SAB. Specifically, our SAB is composed of two key modules: Hybrid Dimension Attention (HDA), which operates in temporal-channel joint dimensions, and Deformable Similarity Attention (DSA), which adjusts membrane potentials across spatial dimensions. Together, they serve to modulate the dynamics of membrane potential accumulation in a spike-compatible manner, thus improving information flow and representation quality.

Mathematically, in standard Leaky Integrate-and-Fire (LIF) neuron models, the membrane potential $u_t$ of neuron $i$ at time $t$ evolves as: $u_t^i=\tau u_{t-1}^i+x_t^i-V_{\mathrm{th}}s_{t-1}^i$, where $\tau$ is the decay constant, $x_t^i$ is the input current, $s_{t-1}^i$ is the spike signal from the previous time step, and $V_{\mathrm{th}}$ is the firing threshold. Without attention guidance, membrane potential integration is uniform, lacking selective representation in time or importance across channels and positions, which significantly limits the representation power of SNN.

To address this, we introduce attention-guided modulation over the membrane potential update: $u_t=\tau u_{t-1}+\alpha_t\odot x_t+\beta_t$. Here, $\alpha_t\in\mathbb{R}^C$ is a joint temporal-channel attention map from HDA, $\beta_t\in\mathbb{R}^{H\times W}$ is a spatial modulation map from DSA. In this form, HDA re-weights input activations across channels and time, focusing accumulation on informative temporal-channel combinations. DSA refines the spatial saliency, emphasizing regions that contribute more meaningfully to super-resolution. The refined update thus becomes: $u_t^{i,j}=\tau u_{t-1}^{i,j}+\alpha_t^c\cdot x_t^{i,j,c}+\beta_t^{i,j}$.

In the backward pass, the pseudo-gradient of loss L w.r.t. the membrane potential is: $\frac{\partial L}{\partial u_t}\approx\frac{\partial L}{\partial s_t}\cdot\sigma^{\prime}(u_t-V_\mathrm{th})$. Due to attention-induced focus in $u_t$, more informative gradients are passed to salient positions and time steps, yielding stronger feature representation power.

We will include this theoretical analysis in the Appendix!

Dear reviewer,

Thank you for the comments on our paper.

We have submitted the responses to your comments. Please let us know if you have additional questions so that we can address them during the discussion period.

Thank you!

Best,

Authors

Dear Reviewer DFR5:

Thank you again for your thoughtful and constructive comments. We have made our best effort to address your concerns in this round of discussion. We sincerely hope our responses resolve your doubts, and we would be truly grateful if you would consider raising your score — your recognition means a great deal to us.

A1: The differences: We agree that the differences between natural and remote sensing images need clearer elaboration. Specifically, two key differences make SNNs suitable for remote sensing SR tasks:

1. Remote sensing images often cover significantly larger areas and have much higher resolutions. These characteristics pose more critical demands on efficient reconstruction of large-scale images, especially for on-orbit processing. By employing SNNs, which consume extremely low energy for computation, our method offers a new perspective on developing efficient models in large-scale Earth observation scenarios.
2. Remote sensing images present unique data properties such as scale diversity of objects and weaker texture details than natural images. In this case, we reveal a novel finding that SNNs maintain active neuro state across different scales of degradation, even when high-frequency textures are severely damaged. This insight suggests that SNNs are well-suited for remote sensing image SR tasks.

A1: To address these unique issues: We have technically adapted the SNN architecture, making it more capable for remote sensing applications. In particular, our DSA exploits the rich self-similarity priors inherent in large-scale remote sensing scenarios to modulate the spatial-wise membrane potential, thus enabling faithful and high-fidelity reconstruction.

A1: The reason for comparing with natural image SR methods: Currently, existing works of remote sensing SR were mainly adopted from or built upon natural image SR models. Following the common practice in the field, we chose widely recognized and representative natural SR for comparison. To address this concern, we further added ESTNet (TIP 2024) [R1], a lightweight model specifically developed for remote sensing images. We will try our best to complete the evaluation before the end of the discussion phase and ensure the results are included in the revised manuscript!

[R1] Kang X, Duan P, Li J, Li S. Efficient Swin Transformer for Remote Sensing Image Super-Resolution. IEEE Transactions on Image Processing, 2024, 33: 6367-6379.

We will incorporate these responses into the manuscript to better establish the connection between SNNs and remote sensing!

A2: We appreciate the reviewer pointing out these important works. While our method is inspired by attention spiking neural networks, we would like to clarify that it exhibits significant differences in both motivation, architectural design, and experimental focus:

1. Task-Specific Design for Super-Resolution: Prior works [1–3] focus mainly on object classification or general vision tasks, whereas our method is specifically designed for the regression-based SR tasks, which pose specific challenges in recovering pixel-level details. This fundamental difference leads to distinct design choices in spike encoding, attention mechanism adaptation, and decoding structure.

2. First to Explore SNN in Remote Sensing SR: To our best knowledge, we are the first to explore the application of SNNs for remote sensing image SR, a scenario distinguished by unique challenges such as large-scale imaging, scale diversity, weak textures, etc.

3. New Observations on Spiking Behavior under Degradation: We reveal a novel empirical observation that SNN model maintains active neural states across diverse degradations, even when the high-frequency details are severely damaged. This behavior has not been reported in prior studies, and we believe it provides a new perspective on the potential of SNNs in handling degraded or low-quality visual inputs, particularly in large-scale Earth observation areas.

We will discuss these instructive works carefully and add a detailed discussion in the revised manuscript!

A3: We appreciate the valuable feedback. In response, we have added MambaIR v2 for comparison. We will do our best to report the results before the end of the discussion phase; otherwise, we assure that the updated comparisons will be included in the revised manuscript!

Once again, thank you for your valuable time and thoughtful response!

Response to Reviewer bD3C

Dear Reviewer bD3C:

We sincerely appreciate your time and effort in reviewing our paper. Thank you for your positive feedback on the significance of our work, especially recognizing our first attempt to combine spiking neural network (SNN) with attention mechanism for remote sensing image super-resolution. We are glad to hear that you find our SpikeSR fills the gap of SNN in the field of remote sensing SR, and that our experimental design and ablation are sufficient. We hope our responses below will address your concerns. We would greatly appreciate it if you would consider a score update.

Weaknesses

W1: Can the author provide more information to reflect efficiency?

A1: Thank you for this valuable comment. We have additionally evaluated three metrics to address this concern: (1) energy consumption (Joule/inference), (2) GPU memory usage, and (3) inference time. The results are presented in the table below. It can be found that our SpikeSR exhibits significantly lower energy consumption compared to ANN-based methods, which further demonstrates the high energy efficiency of SNN. We will add a comprehensive discussion of model efficiency in our revised manuscript!

It should be noted that the energy consumption is a theoretical approximation, calculated based on the following principle: the energy cost of ANNs is derived from multiply-and-accumulate (MAC) operations, while that of SNNs comes from AC operations (i.e., an addition). In a typical CMOS system [1], one MAC operation consumes approximately 3.2 pJ, while one accumulate AC consumes about 0.2 pJ. The energy consumption of an ANN model can be estimated by $E_{\text{ANN}} = N_{\text{MAC}} \times E_{\text{MAC}}$, where $N_{\text{MAC}}$ is the number of MAC and $E_{\text{MAC}}$ is set to 3.2. For an SNN model, the energy consumption depends on the number of AC operations, the synaptic activity sparsity, and the number of time steps. The estimated energy is computed as $E_{\text{SNN}} = s \times T \times N_{\text{AC}} \times E_{\text{AC}}$, where $s$ denote the mean sparsity, which is set to 16.42% [2], $T$ is the number of time setp, $N_{\text{AC}}$ means the additon mumber, and $E_{\text{AC}}$ is set to 0.2.

The GPU consumption and inference time were evaluated on a single NVIDIA RTX 4090 GPU. The reported inference time is the total time required to super-resolve 100 LR images of size 160×160.

Reference

[1] Sze V, Chen Y H, Yang T J, et al. Efficient processing of deep neural networks: A tutorial and survey. Proceedings of the IEEE, 2017, 105(12): 2295-2329.

[2] Guo Y, Chen Y, Liu X, et al. Ternary spike: Learning ternary spikes for spiking neural networks. Proceedings of the AAAI conference on artificial intelligence. 2024, 38(11): 12244-12252.

W2: Explain the adaptation and technical details of the proposed HDA.

Thank you for your insightful comment. To effectively adapt the HDA module to the spiking dynamics of SNNs, we first apply a squeeze operation to compress membrane potentials into an average matrix. We then perform 1D convolutions independently along the temporal and channel dimensions to extract attention weights that reflect the dynamic importance of features across both scopes. To further model their interdependence, we employ element-wise multiplication to fuse two weights into a unified temporal-channel attention map, which is ultimately used to modulate the membrane potential representations. Instead of operating directly on raw spike trains, HDA enhances feature representation by dynamically adjusting membrane potentials, enabling it to leverage rich spatiotemporal information while respecting the discrete nature of SNNs. This explanation will be included in the revised manuscript!

Due to space limitations, we have provided more technical details of HDA in the supplementary materials, which is ensured to be publicly accessible. Besides, we will release the source code to ensure technical completeness and reproducibility.

Dear reviewer,

Thank you for the comments on our paper.

We have submitted the responses to your comments. Please let us know if you have additional questions so that we can address them during the discussion period.

Thank you!

Best,

Authors

Dear Reviewers,

Yes, we sincerely welcome any feedback you may have on our rebuttal and look forward to discussing it further with you!

Authors