MambaOutRS: A Hybrid CNN-Fourier Architecture for Remote Sensing Image Classification

Changbae Mun; Minjong Cheon

REVIEW 5 major objections 4 minor 16 references

MambaOutRS: A Hybrid CNN-Fourier Architecture for Remote Sensing Image Classification

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MambaOutRS claims that a hybrid of gated convolutions and a learned Fourier-domain gate reaches state-of-the-art remote sensing classification with fewer parameters than transformer and Mamba baselines, and that dropping the recurrent SSM…

desk verdict The architecture is a reasonable engineering combination, but the main tables carry exact duplicated numbers across different datasets that make the headline SOTA claims untrustworthy as printed. read the letter →

arxiv 2506.19561 v1 pith:6PF4SSYR submitted 2025-06-24 cs.CV cs.AI

Minjong Cheon , Changbae Mun This is my paper

classification cs.CVcs.AI

keywords remotesensingimageclassificationFourierFilterGategatedconvolutionstatespacemodelsMambafrequency-domainfilteringaerialscenehybridCNNarchitecture

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that recurrent state-space models are not needed for remote sensing scene classification. It builds MambaOutRS from stacked Gated CNN blocks for local features and a new Fourier Filter Gate (FFG) that multiplies the 2D Fourier transform of the feature map by a learned frequency mask, then transforms back. The authors report that the tiny variant (24.0M parameters) reaches 98.41% F1 on UC Merced and 95.99% on AID, above larger transformers and Mamba-based models, and that an ablation removing the FFG drops the UC Merced F1 to 94.58%. The intended upshot is that gated convolutions plus frequency-domain global mixing can substitute for recurrent SSMs while using fewer parameters.

What carries the argument

The load-bearing mechanism is the Fourier Filter Gate (FFG), wrapped in a residual Fourier Gate Block. For an input tensor the FFG computes the real 2D Fourier transform $\hat{x}=\mathrm{rfft2}(x)$, applies a learned sigmoid gate $g=\sigma(w)\in[0,1]$ element-wise in the frequency domain, and returns $\mathrm{irfft2}(\hat{x}\odot g)$. Because the gate is learned and shared over all spatial positions, the block can suppress or amplify selected frequencies to create a global receptive field without attention or recurrence. The Gated CNN blocks provide local token mixing via depthwise convolution on one branch, gated by another linear branch; the four-stage hierarchy places gated convolutions early and late, with Fourier Gate Blocks in the middle stages.

What would settle it

Rerun Table 3 under one shared protocol with explicitly fixed split ratios for UC Merced, AID, and RESISC45, the same 224x224 input size, 25 epochs, augmentations, and best-validation checkpoint selection. If MambaOutRS-t no longer exceeds RSMamba-H and VMamba-T on UC Merced and AID, or if the removal of the Fourier Gate Block no longer causes a large accuracy drop, the paper's central claims are falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a purely convolutional backbone with a frequency-domain gate outperforms the compared transformer and Mamba models on two remote sensing benchmarks. MambaOutRS-t (24.0M parameters) reports F1 of 98.41% on UC Merced and 95.99% on AID, surpassing RSMamba-H (33.1M) and ViT-L (303.0M); on RESISC45 it reports competitive but slightly lower F1 than RSMamba-H. The ablation shows the Fourier Gate Block is the decisive component: removing it from MambaOutRS-t reduces UC Merced F1 from 98.41% to 94.58%. The paper concludes that the complexities of recurrent SSMs can be superseded by combining gated convolutions for spatial mixing and frequency-based gates for spectral global context.

Load-bearing premise

The state-of-the-art comparison depends on every baseline in Table 3 being evaluated under exactly the same protocol as MambaOutRS, including the train/test split ratios and training schedule; the paper does not report those split ratios or how the baseline numbers were reproduced, so the comparison is unverifiable.

Editorial extensions

If this is right

Recurrent SSMs are not required for state-of-the-art remote sensing scene classification under the authors' protocol; gated convolutions plus frequency gating suffice.
A 24.0M-parameter model can beat much larger transformer and SSM baselines on UC Merced and AID.
The Fourier Gate Block is the main source of global-context gains, so frequency-domain filtering is the component worth retaining and refining.
On EuroSAT, adding the Fourier Gate Block raises the tiny variant from 94.58% to 98.30% F1, extending the benefit beyond the two headline datasets.
On RESISC45 the advantage narrows, which means the method's benefit is not uniform across remote sensing datasets.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial extension: the paper's ablation removes the whole FFG block, so it does not isolate whether the gain comes from the learnable frequency mask itself or simply from adding another residual nonlinear module; a freeze-gate comparison would separate these.
Editorial extension: because frequency masking is global and parameter-light, the same block is a plausible drop-in replacement for attention in other dense prediction tasks, but the paper does not test that transfer.
Editorial extension: remote sensing scenes contain repeated structures such as fields and grids, so a natural test of the mechanism is to probe sensitivity to particular frequency bands, or to evaluate on datasets with more periodic textures.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The architecture idea is fine and worth a quick look: stacking MambaOut's gated CNN blocks with a learnable frequency-domain gate and testing on remote sensing benchmarks is a sensible engineering exercise. The FFG is close to GFNet's learnable frequency filter and should cite it, but the real-valued rfft twist is a minor variant, not a scandal. The writing is clear and the method section is readable. That is where the credit ends.

The results section has a load-bearing problem that I would not send to a referee without fixing. In Table 3, MambaOutRS-f reports P/R/F1 = 91.20/90.58/90.60 for both AID and RESISC45, which are different datasets with different class counts and input resolutions. The same triple reappears in Table 4 for the same variant with FGB on both datasets. The f-variant without FGB also reports 97.63/97.46/97.44 on both UC Merced and EuroSAT. Those coincidences are not credible; they indicate row copying or mislabeling. If the f-variant's AID and RESISC45 numbers are wrong, then the claimed SOTA on those datasets and the FGB ablation conclusions on them are unsupported as presented. The paper also never specifies the train/test split ratios, even though the comparison against RSMamba baselines depends on identical protocol. The baseline numbers are imported from Chen et al. without re-evaluation, and MambaOut itself is absent from the comparison table, which weakens the positioning claim. The ablation shows large FGB gains that, if real, would be interesting, but given the duplicated numbers I cannot trust the table without the code or corrected tables.

The internal inconsistencies are the load-bearing flaw. The architecture is not deeply novel, but it could be a useful empirical data point if the experiments were redone carefully with full protocol disclosure and corrected tables. As it stands, the paper should not be published in its current form. If the authors provide corrected tables, code, and exact split ratios, a serious referee could then judge the actual gains. My recommendation: do not desk-reject outright, but send back for major revision with a request for raw results and protocol details. The idea deserves a fair hearing, but the presented evidence does not support the current SOTA claims.

Referee Report

5 major / 4 minor

Summary. The paper proposes MambaOutRS, a hybrid CNN-Fourier architecture for remote sensing scene classification. It replaces the recurrent SSM in Mamba-based models with Gated CNN blocks for local feature extraction and a learnable Fourier Filter Gate (FFG) for global frequency-domain mixing, arranged in a four-stage hierarchical backbone. Experiments on UC Merced, AID, NWPU-RESISC45, and EuroSAT report state-of-the-art F1 scores for the tiny variant (98.41% on UC Merced, 95.99% on AID), and an ablation study attributes the gains to the FFG. The paper argues that the complexity of recurrent SSMs is unnecessary for this task.

Significance. The architectural idea is plausible and timely: testing whether a convolutional backbone with a frequency-domain gate can match SSM-based models is a useful question, and the FFG is simple and easy to implement. The paper also compares across four standard datasets and reports parameter counts, which is helpful. However, the empirical evidence, which is the entire basis for the claims, is not internally consistent: several metric triples are exactly duplicated across datasets, the ablation contains a direct contradiction, the comparison protocol is unspecified, and the "consistent SOTA" claim is contradicted by the authors' own table. No code or trained models are provided. As submitted, the results cannot be verified and the central claims are unsupported.

major comments (5)

[Section 3.3, Table 3] MambaOutRS-f is reported with identical Precision/Recall/F1 values (91.20/90.58/90.60) on AID and NWPU-RESISC45, despite these datasets having different numbers of classes (30 vs. 45) and native resolutions (600x600 vs. 256x256). The same triple is repeated for MambaOutRS-f with FGB on both datasets in Table 4, and Table 4 also contains other exact or near-exact coincidences across different datasets (e.g., MambaOutRS-f without FGB on UC Merced and EuroSAT both 97.63/97.46/97.44; MambaOutRS-t without FGB on UC Merced and EuroSAT both P=95.00 and F1=94.58). These exact matches are not credible for independent evaluations and indicate that numbers were copied or assigned to the wrong cells. Since these rows underpin the reported state-of-the-art scores and the FGB ablation, the core empirical claims are corrupted.
[Section 3.4, Table 4] The ablation results contradict the accompanying text. For MambaOutRS-f on AID, adding the FFG/FGB changes P/R/F1 from 91.82/91.50/91.47 (without FGB) to 91.20/90.58/90.60 (with FGB), i.e., all three metrics decrease. The text states that "integrating the FGB substantially improves model performance, particularly on UC Merced, AID, and EuroSAT" and that "FGB contributes consistently to gains in precision, recall, and F1-score." This is false for the AID column of the smallest variant. Moreover, the "without FGB" architecture is not defined: it is not stated which modules are used in place of the Fourier Gate Blocks in stages 2 and 3, so the ablation is not interpretable even where the numbers are internally consistent.
[Section 3.2, Section 3.3] The evaluation protocol is underspecified. The paper reports 25 epochs, 224x224 input, Adam, and best-validation checkpoint selection, but it does not state the train/test split ratios for any dataset (common conventions are 80/20 for UC Merced and EuroSAT and 50/50 for AID and RESISC45), the random seed or number of runs, or how the 64x64 EuroSAT images are resized and preprocessed to 224x224. The baseline numbers in Table 3 are imported from Chen et al. [2024] rather than reproduced under the authors' protocol; no evidence is given that those baselines used the same splits, preprocessing, or training settings. Without this information, the SOTA comparisons in Table 3 are not verifiable, and the claim of a "fair comparison" is unsupported. No code or trained models are provided to check the numbers.
[Abstract, Section 3.3] The "consistently achieved state-of-the-art (SOTA) performance across these benchmarks" claim is contradicted by the authors' own Table 3. On NWPU-RESISC45, RSMamba-H reports F1=95.18, while the best MambaOutRS variant (MambaOutRS-t) reports 94.09. The text in Section 3.3 acknowledges that "RSMamba-H achieves the top score on RESISC45," so the abstract's claim of consistent SOTA is internally inconsistent. The claim should be restricted to UC Merced and AID, or the table and abstract must be reconciled.
[Section 3.3, Table 3] The metric definitions are not given, and several F1 entries are not consistent with the standard relation F1 = 2PR/(P+R). For example, MambaOutRS-f on UC Merced has P=98.27, R=98.10, and F1=98.10, but the harmonic mean is about 98.18; MambaOutRS-t on the same dataset has P=98.53, R=98.41, and F1=98.41, while the harmonic mean is about 98.47. The same discrepancy also appears in many baseline rows. If macro-F1 is intended, the paper should say so and provide the per-class F1 values; as written, the reported F1 scores cannot be independently checked.

minor comments (4)

[Section 2.1.1, Algorithm 1] Please proofread for typographical issues: "recurrentState Space Model(SSM)" and "GatedCNNBlocks" are missing spaces, and "Gated CNN blocks?" contains a stray question mark.
[Section 2.2, Algorithm 2] The text describes a "real-valued 2D Fourier transform," but the algorithm uses rfft2, which produces a complex half-spectrum; please clarify the distinction and define the exact shape of the learned gate.
[Section 3.1, Table 2] The EuroSAT entry lists "13-band MS / RGB," but Section 3.2 treats the input as a fixed 224x224 RGB image; please specify which bands are used and how the native 64x64 EuroSAT images are resized or cropped.
[References] The reference formatting is inconsistent (e.g., "EuroSA THelber et al." appears in the dataset list, and several citations in the introduction lack separating spaces); please apply a consistent citation style throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the SOTA and FGB-ablation claims rest on empirical benchmark measurements against external baselines, not on derivations that reduce to their own inputs.

full rationale

The paper has no derivation chain whose output is defined in terms of an input quantity; it proposes an architecture (Gated CNN blocks plus a Fourier Filter Gate, FFG) and supports its claims empirically. The FFG's contribution is measured by a with/without-FGB ablation in Table 4, which is a controlled comparison rather than a restatement of the module's definition, so the 'FGB is critical' conclusion is not circular. The SOTA claim is benchmarked against external baselines (ResNet, ViT, Swin, VMamba, RSMamba) whose numbers are imported from Chen et al. 2024; importing external results raises reproducibility and protocol-matching concerns (split ratios and baseline training details are unreported), but that is a correctness risk, not circularity. The only self-citations (Cheon and Mun 2024; Cheon 2024) appear in the introduction as background for general deep-learning progress and carry no load in any conclusion. One integrity flag: Table 4 contains exact internal duplicates (MambaOutRS-f reports identical P/R/F1 of 91.20/90.58/90.60 for AID and RESISC45, and MambaOutRS-f without FGB reports 97.63/97.46/97.44 for both UC Merced and EuroSAT), which suggests data-copying errors and would undermine the ablation and SOTA claims if confirmed; this is a verification and integrity issue, not a circularity issue. Accordingly the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

This is an empirical engineering paper. The 'free parameters' are design choices and training settings rather than fitted constants in a derivation. The main axioms are assumptions about the benchmarking protocol and the correctness of borrowed baseline numbers.

free parameters (3)

Variant hyperparameters (depths, dims) per model = f: [3,3,9,3]/[48,96,192,288]; k: [3,3,15,3]/[48,96,192,288]; t: [3,3,9,3]/[96,192,384,576]
Hand-picked configuration; no search or sensitivity analysis is provided.
Fourier gate mask (fully learned) = Learned during training, initialization unstated
The gate g is a learned tensor; its initialization and training dynamics are not specified, so the ablation result is not reproducible.
Training hyperparameters = 25 epochs, lr 1e-3, Adam, Mixup/CutMix, 224x224
Chosen without justification; comparing to baselines trained for different lengths or optimizers may confound architecture gains.

assumptions (4)

standard math rfft2/irfft2 with ortho normalization is a differentiable, invertible transform
Used implicitly in FFG (Section 2.2); no proof needed.
domain assumption The benchmark labels and the reported baseline numbers from Chen et al. are accurate
Table 3 reproduces baselines from RSMamba without re-running; if those numbers are wrong or were obtained under different protocols, the comparison fails.
ad hoc to paper 25 training epochs are sufficient to reach converged performance for fair comparison
No convergence analysis is provided; many RS baselines are trained longer.
ad hoc to paper The real-valued mask on complex FFT coefficients is an effective mechanism for global feature mixing
This is the core design hypothesis, untested against alternative global mixing modules.

invented entities (2)

Fourier Filter Gate (FFG)
purpose: Learns a frequency-wise real mask to gate the rfft2 spectrum before inverse transform, capturing global context
No code, no theoretical guarantees, and the concept closely resembles prior global filter networks (GFNet/FCC) not cited; only internal benchmark results support it.
MambaOutRS variants (f/k/t)
purpose: Hybrid backbones with Gated CNN and FGB blocks
No pre-trained weights or code released; benchmark numbers are internally inconsistent.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MambaOutRS: A Hybrid CNN-Fourier Architecture for Remote Sensing Image Classification." pith.science (2026). https://pith.science/paper/6PF4SSYR

@misc{pith2026250619561,
  author       = {Pith},
  title        = {Pith review of: MambaOutRS: A Hybrid CNN-Fourier Architecture for Remote Sensing Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6PF4SSYR}},
  note         = {Machine review of arXiv:2506.19561}
}

read the original abstract

Recent advances in deep learning for vision tasks have seen the rise of State Space Models (SSMs) like Mamba, celebrated for their linear scalability. However, their adaptation to 2D visual data often necessitates complex modifications that may diminish efficiency. In this paper, we introduce MambaOutRS, a novel hybrid convolutional architecture for remote sensing image classification that re-evaluates the necessity of recurrent SSMs. MambaOutRS builds upon stacked Gated CNN blocks for local feature extraction and introduces a novel Fourier Filter Gate (FFG) module that operates in the frequency domain to capture global contextual information efficiently. Our architecture employs a four-stage hierarchical design and was extensively evaluated on challenging remote sensing datasets: UC Merced, AID, NWPU-RESISC45, and EuroSAT. MambaOutRS consistently achieved state-of-the-art (SOTA) performance across these benchmarks. Notably, our MambaOutRS-t variant (24.0M parameters) attained the highest F1-scores of 98.41\% on UC Merced and 95.99\% on AID, significantly outperforming existing baselines, including larger transformer models and Mamba-based architectures, despite using considerably fewer parameters. An ablation study conclusively demonstrates the critical role of the Fourier Filter Gate in enhancing the model's ability to capture global spatial patterns, leading to robust and accurate classification. These results strongly suggest that the complexities of recurrent SSMs can be effectively superseded by a judicious combination of gated convolutions for spatial mixing and frequency-based gates for spectral global context. Thus, MambaOutRS provides a compelling and efficient paradigm for developing high-performance deep learning models in remote sensing and other vision domains, particularly where computational efficiency is paramount.

Figures

Figures reproduced from arXiv: 2506.19561 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 12 canonical work pages

[1]

Machine learning in modelling land-use and land cover-change (lulcc): Current status, challenges and prospects

Junye Wang, Michael Bretz, M Ali Akber Dewan, and Mojtaba Aghajani Delavar. Machine learning in modelling land-use and land cover-change (lulcc): Current status, challenges and prospects. Science of the Total Environment, 822: 0 153559, 2022

work page 2022
[2]

Pure data correction enhancing remote sensing image classification with a lightweight ensemble model

Huaxiang Song, Hanglu Xie, Yingying Duan, Xinyi Xie, Fang Gan, Wei Wang, and Jinling Liu. Pure data correction enhancing remote sensing image classification with a lightweight ensemble model. Scientific Reports, 15 0 (1): 0 5507, 2025

work page 2025
[3]

Combining kan with cnn: Konvnext’s performance in remote sensing and patent insights

Minjong Cheon and Changbae Mun. Combining kan with cnn: Konvnext’s performance in remote sensing and patent insights. Remote Sensing, 16 0 (18): 0 3417, 2024

work page 2024
[4]

Rsmamba: Remote sensing image classification with state space model

Keyan Chen, Bowen Chen, Chenyang Liu, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsmamba: Remote sensing image classification with state space model. IEEE Geoscience and Remote Sensing Letters, 2024

work page 2024
[5]

Remote sensing image classification using deep learning

Shreya Gupta, Rakesh K Dwivedi, Vivek Kumar, Riya Jain, Shruti Jain, and Mayank Singh. Remote sensing image classification using deep learning. In 2021 10th International Conference on System Modeling & Advancement in Research Trends (SMART), pages 274--279. IEEE, 2021

work page 2021
[6]

Review of vision transformer models for remote sensing image scene classification

Pengyuan Lv, Wenjun Wu, Yanfei Zhong, and Liangpei Zhang. Review of vision transformer models for remote sensing image scene classification. In IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium, pages 2231--2234. IEEE, 2022

work page 2022
[7]

Kolmogorov-arnold network for satellite image classification in remote sensing

Minjong Cheon. Kolmogorov-arnold network for satellite image classification in remote sensing. arXiv preprint arXiv:2406.00600, 2024

arXiv 2024
[8]

Vivit-prob: A radar echo extrapolation model based on video vision transformer and spatiotemporal sparse attention

Yunan Qiu, Bingjian Lu, Wenrui Xiong, Zhenyu Lu, Le Sun, and Yingjie Cui. Vivit-prob: A radar echo extrapolation model based on video vision transformer and spatiotemporal sparse attention. Remote Sensing, 17 0 (12): 0 1966, 2025

work page 1966

Show all 16 references

[9]

Mamba: Linear-time sequence modeling with selective state spaces

Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arxiv 2023. arXiv preprint arXiv:2312.00752, 2023

2023 arXiv
[10]

Vision mamba: A comprehensive survey and taxonomy

Xiao Liu, Chenxu Zhang, and Lei Zhang. Vision mamba: A comprehensive survey and taxonomy. arXiv preprint arXiv:2405.04404, 2024

2024 arXiv
[11]

Mambavision: A hybrid mamba-transformer vision backbone

Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 25261--25270, 2025

2025
[12]

Mambaout: Do we really need mamba for vision? In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4484--4496, 2025

Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4484--4496, 2025

2025
[13]

Aid: A benchmark data set for performance evaluation of aerial scene classification

Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55 0 (7): 0 3965--3981, 2017

2017
[14]

Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification

Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12 0 (7): 0 2217--2226, 2019

2019
[15]

Remote sensing image scene classification: Benchmark and state of the art

Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105 0 (10): 0 1865--1883, 2017

2017
[16]

Bag-of-visual-words and spatial extensions for land-use classification

Yi Yang and Shawn Newsam. Bag-of-visual-words and spatial extensions for land-use classification. In Proceedings of the 18th SIGSPATIAL international conference on advances in geographic information systems, pages 270--279, 2010

2010

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Machine learning in modelling land-use and land cover-change (lulcc): Current status, challenges and prospects

Junye Wang, Michael Bretz, M Ali Akber Dewan, and Mojtaba Aghajani Delavar. Machine learning in modelling land-use and land cover-change (lulcc): Current status, challenges and prospects. Science of the Total Environment, 822: 0 153559, 2022

work page 2022

[2] [2]

Pure data correction enhancing remote sensing image classification with a lightweight ensemble model

Huaxiang Song, Hanglu Xie, Yingying Duan, Xinyi Xie, Fang Gan, Wei Wang, and Jinling Liu. Pure data correction enhancing remote sensing image classification with a lightweight ensemble model. Scientific Reports, 15 0 (1): 0 5507, 2025

work page 2025

[3] [3]

Combining kan with cnn: Konvnext’s performance in remote sensing and patent insights

Minjong Cheon and Changbae Mun. Combining kan with cnn: Konvnext’s performance in remote sensing and patent insights. Remote Sensing, 16 0 (18): 0 3417, 2024

work page 2024

[4] [4]

Rsmamba: Remote sensing image classification with state space model

Keyan Chen, Bowen Chen, Chenyang Liu, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsmamba: Remote sensing image classification with state space model. IEEE Geoscience and Remote Sensing Letters, 2024

work page 2024

[5] [5]

Remote sensing image classification using deep learning

Shreya Gupta, Rakesh K Dwivedi, Vivek Kumar, Riya Jain, Shruti Jain, and Mayank Singh. Remote sensing image classification using deep learning. In 2021 10th International Conference on System Modeling & Advancement in Research Trends (SMART), pages 274--279. IEEE, 2021

work page 2021

[6] [6]

Review of vision transformer models for remote sensing image scene classification

Pengyuan Lv, Wenjun Wu, Yanfei Zhong, and Liangpei Zhang. Review of vision transformer models for remote sensing image scene classification. In IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium, pages 2231--2234. IEEE, 2022

work page 2022

[7] [7]

Kolmogorov-arnold network for satellite image classification in remote sensing

Minjong Cheon. Kolmogorov-arnold network for satellite image classification in remote sensing. arXiv preprint arXiv:2406.00600, 2024

arXiv 2024

[8] [8]

Vivit-prob: A radar echo extrapolation model based on video vision transformer and spatiotemporal sparse attention

Yunan Qiu, Bingjian Lu, Wenrui Xiong, Zhenyu Lu, Le Sun, and Yingjie Cui. Vivit-prob: A radar echo extrapolation model based on video vision transformer and spatiotemporal sparse attention. Remote Sensing, 17 0 (12): 0 1966, 2025

work page 1966