Pith. sign in

REVIEW 4 major objections 4 minor 76 references

STeInFormer: Spatial-Temporal Interaction Transformer Architecture for Remote Sensing Change Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper claims a spatial-temporal interaction Transformer with a parameter-free DCT token mixer reaches state-of-the-art change-detection accuracy while using only 1.26M parameters.

desk verdict A well-ablated lightweight architecture for change detection with a genuinely new DCT frequency mixer; the frequency set is under-specified and the prior-transfer advantage is statistically thin, but the core contribution holds. read the letter →

arxiv 2412.17247 v1 pith:WIPKDKVH submitted 2024-12-23 cs.CV

Xiaowen Ma , Zhenkai Wu , Mengting Ma , Mengjiao Zhao , Fan Yang , Zhenhong Du , Wei Zhang This is my paper · ORCID
classification cs.CV
keywords remotesensingchangedetectionspatial-temporalinteractionTransformerbackbonemulti-frequencytokenmixerdiscretecosinetransformcross-temporalgatingefficiency-accuracytrade-offSiamesenetwork
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

Remote sensing change detection compares two images of the same place taken at different times and labels which pixels changed. Most deep learning systems extract features from the two images independently with a Siamese network and only fuse them later in a change-detection head. STeInFormer argues that this misses the task-specific need for spatial and temporal interaction during feature extraction itself. The paper proposes a backbone with cross-temporal interactors that gate each image's features by the difference between the two, cross-spatial interactors built on a U-shaped per-stage design that preserves spatial detail, and a parameter-free multi-frequency token mixer that uses a small set of pre-selected discrete cosine transform frequencies. The reported result is the highest F1 among the compared methods on three public benchmarks (WHU-CD 89.61, LEVIR-CD 91.47, CLCD 73.83) with 1.26M parameters and 9.42G FLOPs, which would make accurate change detection substantially cheaper to run.

What carries the argument

The central machinery is the pairing of three components: a cross-temporal interactor (CTI), which gates each temporal feature stream by a learned weight map derived from the difference between the two streams, a cross-spatial interactor (CSI), which reshapes each encoder stage into a U-shaped module so that deep semantic features and shallow spatial-detail features are fused, and a multi-frequency token mixer, which splits the projected feature map into channel groups and applies fixed 2D discrete cosine transform basis functions, selected using ImageNet frequency-importance priors, to each group before concatenating and projecting the result. The DCT bases carry the frequency-domain token mixing, the gating mechanism carries the cross-temporal interaction, and the U-shaped CSI carries the cross-spatial interaction, together giving the model its reported accuracy at low parameter and FLOP counts.

What would settle it

Train the same architecture on WHU-CD, LEVIR-CD, and CLCD with the fixed DCT bases re-selected by per-dataset frequency importance, measured by the F1 change when each frequency is added or removed, instead of by ImageNet priors. If per-dataset selection clearly improves over the ImageNet-prior selection, the prior-transfer assumption is not the source of the mixer's gain; if it matches or underperforms, the ImageNet priors are doing the work claimed.

Watch

Extended reading notes

Core claim

STeInFormer is presented as the first general backbone network designed specifically for remote sensing change detection, replacing the non-interactive Siamese feature extractor with a spatial-temporal interaction Transformer. The cross-temporal interactor computes the element-wise difference of bi-temporal features and learns per-pixel gates from depthwise-separable-convolution-processed concatenations, amplifying changes of interest and suppressing non-interest changes. The cross-spatial interactor reorganizes each feature-extraction stage into a U-shape so semantic and spatial-detail information interact within the encoder. In addition, a multi-frequency token mixer based on discrete cosine transform with fixed frequency priors mixes tokens in the frequency domain with linear complexity. The paper reports that this combination outperforms the compared state-of-the-art methods on WHU-CD, LEVIR-CD, and CLCD while using far fewer parameters and FLOPs, and that swapping STeInFormer into existing methods as a backbone improves their accuracy.

Load-bearing premise

The load-bearing premise is that the frequency bands judged most important by ImageNet classification remain the informative ones for distinguishing changed from unchanged pixels in remote sensing imagery after the feature map is split into channel groups; if that transfer fails, the multi-frequency mixer loses its advantage over ordinary convolution.

Editorial extensions

If this is right

  • Replacing standard CNN backbones with STeInFormer raises reported F1 for IFNet, BIT, and SNUNet on all three datasets, so the architecture can serve as a generic change-detection backbone.
  • A fixed set of eight DCT frequencies outperforms both plain 3x3 convolution and a single-frequency mixer, indicating that multiple frequency components contribute to the change signal.
  • The encoder alone carries the spatio-temporal interaction, so a simple MLP decoder is enough to reach the top results, removing the need for heavy interaction modules in the decoder.
  • At 1.26M parameters and 9.42G FLOPs, the model reports a better efficiency-accuracy trade-off than compared lightweight methods such as BIT, which uses 3.55M parameters and 10.60G FLOPs.

Reading between the lines

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

  • The paper does not test whether its ImageNet-derived frequency priors are optimal for RSCD; re-ranking the DCT bases on each dataset's validation split would reveal whether the mixer's gain comes from frequency diversity or from the specific priors.
  • Because the DCT mixer has linear complexity and no learnable weights in its bases, a natural extension is to transplant it into other dense prediction or bi-temporal tasks where self-attention's quadratic cost is prohibitive.
  • The CTI gating ablation suggests a broader design lesson: cross-attention can be replaced by a much cheaper learned gate without losing accuracy, which could inform other multi-temporal fusion modules.
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.

Referee Report

4 major / 4 minor

Summary. The paper presents STeInFormer, a change-detection architecture that replaces the standard non-interactive Siamese design with per-stage cross-temporal interactors (CTIs) based on a gating mechanism and cross-spatial interactors (CSIs) built on a U-shaped structure. It also proposes a 'multi-frequency token mixer' that applies a fixed set of 2D-DCT basis functions to split feature channels and mixes tokens in the frequency domain. The method is evaluated on WHU-CD, LEVIR-CD, and CLCD with five metrics, compared with nine baselines, and tested as a drop-in backbone for IFNet, BIT, and SNUNet. Tables report the best F1 on all three datasets (89.61/91.47/73.83) with 1.26M parameters and 9.42G FLOPs, and ablations analyze frequency selection, component design, number of frequencies, CTI design, and loss weighting.

Significance. If its claims are confirmed, the paper contributes a lightweight backbone with explicit spatio-temporal interactions, which are shown to give large gains over a Siamese baseline (Table IV, design-ablation rows), and demonstrates that frequency-domain token mixing can be competitive at very low parameter cost. The evaluation uses public datasets, multiple metrics, a reasonable set of recent baselines, and additional backbone-transfer experiments, and a code link is provided. The main weaknesses are reproducibility and statistical support for the frequency-prior component rather than the overall architecture.

major comments (4)
  1. [§III-D-c and §IV-F-c] The selected frequency indices (u_i, v_i) for M=8 are never enumerated, so the central 'pre-trained priors' strategy cannot be reproduced from the text. Section III-D-c says only that the frequencies were chosen 'according to the experimental results' of [40]; Figure 3 shows B_{u_i,v_i} without giving the indices. Please list the chosen indices and their ordering, and state whether the p×p convolution implementing the DCT basis has fixed weights; otherwise the 'parameter-free' claim and the reported gains of PP over RS/DA cannot be verified.
  2. [§IV-F-a and Table IV] The statement that PP 'reaches the best results on the three datasets' is contradicted by Table IV on WHU-CD, where DA achieves F1=89.64 versus PP's 89.61; on LEVIR-CD the PP advantage over DA is only 0.54 F1. No standard deviation or number of seeds is reported for any ablation row, so the PP-vs-DA difference may be within run-to-run noise. Please correct the claim and add multi-seed statistics (mean±std or a significance test) for at least the frequency-selection comparison.
  3. [§IV-F-c and §IV-F-f] The choices M=8 and λ_focal:λ_dice=1:1 are justified by F1 values in Tables IV and VI, but the paper does not state whether those values were obtained on a held-out validation split or on the test set. If the test set was used for model selection, the reported final F1 values are optimistically biased. Please either perform ablations on the validation splits (WHU-CD and LEVIR-CD have explicit validation sets) and then report the selected configuration on the test set, or explicitly disclose and justify the test-set selection.
  4. [§III-D-b and Figure 3] The batch implementation is under-specified: it is not defined how the input feature map R^p is split into A_i, what p is (patch size or feature-map size), how the M frequency heads partition the channel dimension, and whether the p×p convolution uses fixed DCT weights or learnable filters. These details are necessary for exact reimplementation and for the linear-complexity and parameter-free claims.
minor comments (4)
  1. [Table II] Table II labels ChangeFormer with reference [68], whereas Table I and the text cite ChangeFormer as [35]; the reference list entry [68] is ChangeStar, so the citation should be corrected.
  2. [§III-E, Eq. (12)] Eq. (12) is not a well-formed dice loss as written: E and softmax(E′) are not scalar quantities, and the formula lacks the sum over pixels; please rewrite using standard dice notation.
  3. [§IV-C] Section IV-C states an 'average F1 improvement of 4' over DMINet on the three datasets, but Table I contains no CLCD result for DMINet; the per-dataset margins (0.92 on WHU-CD and 0.76 on LEVIR-CD) do not support an average of 4 over three datasets.
  4. [§III-D-a, Eqs. (4)-(8)] The notation in Eqs. (4)-(8) switches the roles of spatial indices (h,w) and frequency indices (x,y) relative to the standard 2D-DCT definition; please define the axes consistently to avoid ambiguity.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: STeInFormer is validated against external baselines and ablations, with only a minor under-specification of the ImageNet-derived frequency prior.

full rationale

The central claims of the paper are empirical. STeInFormer's F1 scores on WHU-CD, LEVIR-CD, and CLCD are obtained by running the proposed architecture against published methods with default settings, and the proposed modules (CSI, CTI, multi-frequency mixer) are evaluated in ablation Table IV, where removing or replacing them changes results. The multi-frequency mixer's fixed frequency bases are imported from FCANet's ImageNet experiments (Sec. III-D-c, ref. [40]) as a prior, not fitted to the RSCD test sets, so the reported gains are not a re-statement of the prior. Loss weights and M are tuned on validation/test sets, but the paper reports these tuning choices explicitly in Tables IV and VI and the overall comparison is against external baselines, so the headline results are not forced by construction. The frequency selection is a correctness/reproducibility weakness: the paper never lists the selected (u_i, v_i) frequency indices or their ordering, and the PP-vs-DA differences on WHU-CD (89.61 vs 89.64) and LEVIR-CD (91.47 vs 90.93) are small with no standard deviations per ablation row, making the advantage of ImageNet priors over dynamic assignment statistically fragile. However, this is lack of independent support or experimental rigor, not circularity under the criteria: there is no equation making the prediction equal to its input, no fitted parameter renamed as a prediction, and no load-bearing self-citation chain. The 'first general backbone for RSCD' claim is a novelty assertion rather than a derivation. Therefore the paper merits a low circularity score of 1, reflecting one minor self-referential weakness (reliance on an unenumerated prior from FCANet) that does not make the central derivation circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method relies on standard DCT mathematics, standard binary segmentation framing, and an external ImageNet frequency prior. The only target-data-dependent quantities are architectural hyperparameters such as M, loss weights, and channel widths, chosen by hand or ablation. No new physical entities are introduced.

free parameters (4)
  • Number of DCT frequency values M = 8
    Selected from ablation on the three target datasets (Table IV, Section IV-F-c). It directly trades off expressiveness and cost.
  • Loss weight ratio (lambda_focal:lambda_dice) = 1:1
    Chosen from ablation on WHU-CD test set (Table VI, Section IV-F-f).
  • Pre-trained DCT frequency index set = Top-8 frequencies from ImageNet experiments
    Taken from FCANet-style ImageNet evaluation (Section III-D-c). This is an external prior rather than fitted to RSCD data, but it is a hand-selected design input.
  • Stage channel widths = 32, 48, 64, 96
    Hand-specified in Section IV; these determine model capacity and affect the reported efficiency.
assumptions (4)
  • standard math 2D DCT basis functions form an orthonormal basis, so selected frequency coefficients are a valid invertible transform of the input feature map.
    Section III-D-a, Eq. 4-8; standard mathematical result, not in question.
  • domain assumption Change detection can be treated as binary semantic segmentation.
    Section I states this; it is the standard framing in the field but ignores some multi-class change nuances.
  • domain assumption ImageNet-derived frequency importance priors transfer to remote sensing change detection.
    Section III-D-c adopts the pre-trained priors strategy based on ImageNet experiments [40]; transferability is assumed, not directly validated on RSCD data.
  • domain assumption Random cropping and splitting of the large WHU-CD image yields independent train and test patches.
    Section IV-A; potential spatial autocorrelation could inflate scores, following standard practice in the field.

how reviews work

0 comments
Cite this review

Pith. "Pith review of STeInFormer: Spatial-Temporal Interaction Transformer Architecture for Remote Sensing Change Detection." pith.science (2026). https://pith.science/paper/WIPKDKVH

@misc{pith2026241217247,
  author       = {Pith},
  title        = {Pith review of: STeInFormer: Spatial-Temporal Interaction Transformer Architecture for Remote Sensing Change Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WIPKDKVH}},
  note         = {Machine review of arXiv:2412.17247}
}
read the original abstract

Convolutional neural networks and attention mechanisms have greatly benefited remote sensing change detection (RSCD) because of their outstanding discriminative ability. Existent RSCD methods often follow a paradigm of using a non-interactive Siamese neural network for multi-temporal feature extraction and change detection heads for feature fusion and change representation. However, this paradigm lacks the contemplation of the characteristics of RSCD in temporal and spatial dimensions, and causes the drawback on spatial-temporal interaction that hinders high-quality feature extraction. To address this problem, we present STeInFormer, a spatial-temporal interaction Transformer architecture for multi-temporal feature extraction, which is the first general backbone network specifically designed for RSCD. In addition, we propose a parameter-free multi-frequency token mixer to integrate frequency-domain features that provide spectral information for RSCD. Experimental results on three datasets validate the effectiveness of the proposed method, which can outperform the state-of-the-art methods and achieve the most satisfactory efficiency-accuracy trade-off. Code is available at https://github.com/xwmaxwma/rschange.

Figures

Figures reproduced from arXiv: 2412.17247 by the authors.

Figure 1
Figure 1. Visualization of two challenges in RSCD: frequent non-interest [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the STeInFormer. Given as input bi-temporal images, multi-scale features are extracted by each CSI, which is U-shaped and relies [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Structure of the multi-frequency mixer. The input feature map [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Example outputs from our STeInFormer and other methods for comparison on WHU-CD (first and second rows), LEVIR-CD (third and fourth rows), [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparison of performance statistics on three datasets. Bars represent F1-scores, with standard deviations at the top. The bar corresponds to the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Class activation maps for features of Cross-Spatial Interactor (CSI) [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 42 canonical work pages

  1. [40]

    Fcanet: Frequency channel attention networks,

    Z. Qin, P. Zhang, F. Wu, and X. Li, “Fcanet: Frequency channel attention networks,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 783–792

    2021

  2. [1]

    Unsupervised spatial-spectral hyperspectral image reconstruction and clustering with diffusion geometry,

    K. Cui, R. Li, S. L. Polk, J. M. Murphy, R. J. Plemmons, and R. H. Chan, “Unsupervised spatial-spectral hyperspectral image reconstruction and clustering with diffusion geometry,” in 2022 12th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS). IEEE, 2022, pp. 1–5

    work page 2022

  3. [2]

    Real-time localization and bimodal point pattern analysis of palms using uav imagery,

    K. Cui, W. Tang, R. Zhu, M. Wang, G. D. Larsen, V . P. Pauca, S. Alqahtani, F. Yang, D. Segurado, P. Fineet al., “Real-time localization and bimodal point pattern analysis of palms using uav imagery,” arXiv preprint arXiv:2410.11124, 2024

    arXiv

  4. [3]

    Superpixel-based and spatially-regularized diffusion learning for unsupervised hyperspectral image clustering,

    K. Cui, R. Li, S. L. Polk, Y . Lin, H. Zhang, J. M. Murphy, R. J. Plemmons, and R. H. Chan, “Superpixel-based and spatially-regularized diffusion learning for unsupervised hyperspectral image clustering,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    work page 2024

  5. [4]

    Deep learning for change detection in remote sensing images: Comprehensive review and meta-analysis,

    L. Khelifi and M. Mignotte, “Deep learning for change detection in remote sensing images: Comprehensive review and meta-analysis,” Ieee Access, vol. 8, pp. 126 385–126 400, 2020

    work page 2020

  6. [5]

    Contrastive multiview subspace clustering of hyperspectral images based on graph convolutional networks,

    R. Guan, Z. Li, W. Tu, J. Wang, Y . Liu, X. Li, C. Tang, and R. Feng, “Contrastive multiview subspace clustering of hyperspectral images based on graph convolutional networks,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–14, 2024

    2024

  7. [6]

    Spatial-spectral graph contrastive clustering with hard sample mining for hyperspectral images,

    R. Guan, W. Tu, Z. Li, H. Yu, D. Hu, Y . Chen, C. Tang, Q. Yuan, and X. Liu, “Spatial-spectral graph contrastive clustering with hard sample mining for hyperspectral images,” IEEE Transactions on Geoscience and Remote Sensing, pp. 1–16, 2024

    work page 2024

  8. [7]

    Multi- decadal mangrove forest change detection and prediction in honduras, central america, with landsat imagery and a markov chain model,

    C.-F. Chen, N.-T. Son, N.-B. Chang, C.-R. Chen, L.-Y . Chang, M. Valdez, G. Centeno, C. A. Thompson, and J. L. Aceituno, “Multi- decadal mangrove forest change detection and prediction in honduras, central america, with landsat imagery and a markov chain model,” Remote Sensing, vol. 5, no. 12, pp. 6408–6426, 2013

    work page 2013

Show all 76 references
  1. [8]

    Building change detection in multitemporal very high resolution sar images,

    C. Marin, F. Bovolo, and L. Bruzzone, “Building change detection in multitemporal very high resolution sar images,” IEEE transactions on geoscience and remote sensing , vol. 53, no. 5, pp. 2664–2682, 2014

    2014

  2. [9]

    A polsar change detection index based on neighborhood information for flood mapping,

    S. Mahdavi, B. Salehi, W. Huang, M. Amani, and B. Brisco, “A polsar change detection index based on neighborhood information for flood mapping,” Remote Sensing, vol. 11, no. 16, p. 1854, 2019

    2019

  3. [10]

    A superresolution land- cover change detection method using remotely sensed images with different spatial resolutions,

    X. Li, F. Ling, G. M. Foody, and Y . Du, “A superresolution land- cover change detection method using remotely sensed images with different spatial resolutions,” IEEE Transactions on Geoscience and Remote Sensing, vol. 54, no. 7, pp. 3822–3841, 2016

    2016

  4. [11]

    An automatic approach for land-change detection and land updates based on integrated ndvi timing analysis and the cvaps method with gee support,

    Y . Hu, Y . Donget al., “An automatic approach for land-change detection and land updates based on integrated ndvi timing analysis and the cvaps method with gee support,” ISPRS journal of photogrammetry and remote sensing, vol. 146, pp. 347–359, 2018

    2018

  5. [12]

    Cross- match: Cross-view matching for semi-supervised remote sensing image segmentation,

    R. Liu, T. Luo, S. Huang, Y . Wu, Z. Jiang, and H. Zhang, “Cross- match: Cross-view matching for semi-supervised remote sensing image segmentation,” IEEE Transactions on Geoscience and Remote Sensing , pp. 1–1, 2024

    2024

  6. [13]

    Self-supervised neuron segmentation with multi-agent reinforcement learning,

    Y . Chen, W. Huang, S. Zhou, Q. Chen, and Z. Xiong, “Self-supervised neuron segmentation with multi-agent reinforcement learning,” in IJCAI, 2023, pp. 609–617

    2023

  7. [14]

    Learning multiscale consistency for self-supervised electron microscopy instance segmentation,

    Y . Chen, W. Huang, X. Liu, S. Deng, Q. Chen, and Z. Xiong, “Learning multiscale consistency for self-supervised electron microscopy instance segmentation,” in ICASSP. IEEE, 2024, pp. 1566–1570

    2024

  8. [15]

    Maskfactory: Towards high-quality synthetic data generation for dichotomous image segmentation,

    H. Qian, Y . Chen, S. Lou, F. Khan, X. Jin, and D.-P. Fan, “Maskfactory: Towards high-quality synthetic data generation for dichotomous image segmentation,” in NeurIPS, 2024

    2024

  9. [16]

    Change detection in the tropical forest environment of northeastern india using landsat,

    A. Singh, “Change detection in the tropical forest environment of northeastern india using landsat,” Remote sensing and tropical land management, vol. 44, pp. 273–254, 1986

    1986

  10. [17]

    Urban and regional land use change detected by using landsat data,

    W. J. Todd, “Urban and regional land use change detected by using landsat data,” Journal of Research of the US Geological Survey , vol. 5, no. 5, pp. 529–534, 1977

    1977

  11. [18]

    Unsupervised change detection in satellite images using principal component analysis and k-means clustering,

    T. Celik, “Unsupervised change detection in satellite images using principal component analysis and k-means clustering,” IEEE geoscience and remote sensing letters , vol. 6, no. 4, pp. 772–776, 2009

    2009

  12. [19]

    Unsupervised deep change vector analysis for multiple-change detection in vhr images,

    S. Saha, F. Bovolo, and L. Bruzzone, “Unsupervised deep change vector analysis for multiple-change detection in vhr images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 57, no. 6, pp. 3677–3693, 2019

    2019

  13. [20]

    Tokenunify: Scalable autoregressive visual pre-training with mixture token prediction,

    Y . Chen, H. Shi, X. Liu, T. Shi, R. Zhang, D. Liu, Z. Xiong, and F. Wu, “Tokenunify: Scalable autoregressive visual pre-training with mixture token prediction,” arXiv preprint arXiv:2405.16847 , 2024

    2024 arXiv

  14. [21]

    Bimcv-r: A landmark dataset for 3d ct text-image retrieval,

    Y . Chen, C. Liu, X. Liu, R. Arcucci, and Z. Xiong, “Bimcv-r: A landmark dataset for 3d ct text-image retrieval,” in MICCAI. Springer, 2024, pp. 124–134

    2024

  15. [22]

    Program: Prototype graph model based pseudo-label learning for test-time adaptation,

    H. Sun, L. Xu, S. Jin, P. Luo, C. Qian, and W. Liu, “Program: Prototype graph model based pseudo-label learning for test-time adaptation,” in The Twelfth International Conference on Learning Representations

  16. [23]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Communications of the ACM, vol. 60, no. 6, pp. 84–90, 2017

    2017

  17. [24]

    Fully convolutional siamese networks for change detection,

    R. C. Daudt, B. Le Saux, and A. Boulch, “Fully convolutional siamese networks for change detection,” in 2018 25th IEEE International Con- ference on Image Processing (ICIP) . IEEE, 2018, pp. 4063–4067

    2018

  18. [25]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556 , 2014

    2014 arXiv

  19. [26]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  20. [29]

    Building change detection for remote sensing images using a dual-task constrained deep siamese convolutional network model,

    Y . Liu, C. Pang, Z. Zhan, X. Zhang, and X. Yang, “Building change detection for remote sensing images using a dual-task constrained deep siamese convolutional network model,” IEEE Geoscience and Remote Sensing Letters, vol. 18, no. 5, pp. 811–815, 2020

    2020

  21. [30]

    Remote sensing change detection via temporal feature interaction and guided refinement,

    Z. Li, C. Tang, L. Wang, and A. Y . Zomaya, “Remote sensing change detection via temporal feature interaction and guided refinement,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–11, 2022

    2022

  22. [31]

    Snunet-cd: A densely connected siamese network for change detection of vhr images,

    S. Fang, K. Li, J. Shao, and Z. Li, “Snunet-cd: A densely connected siamese network for change detection of vhr images,” IEEE Geoscience and Remote Sensing Letters , vol. 19, pp. 1–5, 2021

    2021

  23. [32]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

    2017

  24. [33]

    Swinsunet: Pure transformer network for remote sensing image change detection,

    C. Zhang, L. Wang, S. Cheng, and Y . Li, “Swinsunet: Pure transformer network for remote sensing image change detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–13, 2022

    2022

  25. [34]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022

    2021

  26. [35]

    A transformer-based siamese net- work for change detection,

    W. G. C. Bandara and V . M. Patel, “A transformer-based siamese net- work for change detection,” in IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2022, pp. 207–210

    2022

  27. [36]

    Remote sensing image change detection with transformers,

    H. Chen, Z. Qi, and Z. Shi, “Remote sensing image change detection with transformers,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–14, 2021

    2021

  28. [37]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929 , 2020

    2010 arXiv

  29. [38]

    Fourier features let networks learn high frequency functions in low dimensional domains,

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, and R. Ng, “Fourier features let networks learn high frequency functions in low dimensional domains,” SUBMIT TO IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSE...

    2020

  30. [39]

    Learning frequency domain priors for image demoireing,

    B. Zheng, S. Yuan, C. Yan, X. Tian, J. Zhang, Y . Sun, L. Liu, A. Leonardis, and G. Slabaugh, “Learning frequency domain priors for image demoireing,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 7705–7717, 2021

    2021

  31. [41]

    Structured graph based image regression for unsupervised multimodal change detection,

    Y . Sun, L. Lei, X. Tan, D. Guan, J. Wu, and G. Kuang, “Structured graph based image regression for unsupervised multimodal change detection,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 185, pp. 16–31, 2022

    2022

  32. [42]

    A tm tasseled cap equivalent transformation for reflectance factor data,

    E. P. Crist, “A tm tasseled cap equivalent transformation for reflectance factor data,” Remote sensing of Environment, vol. 17, no. 3, pp. 301–306, 1985

    1985

  33. [43]

    Spectral–spatial-aware unsupervised change detection with stochastic distances and support vector machines,

    R. G. Negri, A. C. Frery, W. Casaca, S. Azevedo, M. A. Dias, E. A. Silva, and E. H. Alcˆantara, “Spectral–spatial-aware unsupervised change detection with stochastic distances and support vector machines,” IEEE Transactions on Geoscience and Remote Sensing , vol. 59, no. 4, pp...

    2020

  34. [44]

    Fusion of sar and multispectral images using random forest regression for change detection,

    D. K. Seo, Y . H. Kim, Y . D. Eo, M. H. Lee, and W. Y . Park, “Fusion of sar and multispectral images using random forest regression for change detection,” ISPRS International Journal of Geo-Information , vol. 7, no. 10, p. 401, 2018

    2018

  35. [45]

    Triplet-based semantic relation learning for aerial remote sensing image change detection,

    M. Zhang, G. Xu, K. Chen, M. Yan, and X. Sun, “Triplet-based semantic relation learning for aerial remote sensing image change detection,” IEEE Geoscience and Remote Sensing Letters , vol. 16, no. 2, pp. 266– 270, 2018

    2018

  36. [46]

    Change detection on remote sensing images using dual-branch multilevel intertemporal network,

    Y . Feng, J. Jiang, H. Xu, and J. Zheng, “Change detection on remote sensing images using dual-branch multilevel intertemporal network,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–15, 2023

    2023

  37. [47]

    A deeply supervised image fusion network for change detection in high resolution bi-temporal remote sensing images,

    C. Zhang, P. Yue, D. Tapete, L. Jiang, B. Shangguan, L. Huang, and G. Liu, “A deeply supervised image fusion network for change detection in high resolution bi-temporal remote sensing images,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 166, pp. 183–200, 2020

    2020

  38. [48]

    A deeply supervised attention metric-based network and an open aerial image dataset for remote sensing change detection,

    Q. Shi, M. Liu, S. Li, X. Liu, F. Wang, and L. Zhang, “A deeply supervised attention metric-based network and an open aerial image dataset for remote sensing change detection,” IEEE transactions on geoscience and remote sensing , vol. 60, pp. 1–16, 2021

    2021

  39. [49]

    Mstdsnet-cd: Multiscale swin transformer and deeply supervised network for change detection of the fast-growing urban regions,

    F. Song, S. Zhang, T. Lei, Y . Song, and Z. Peng, “Mstdsnet-cd: Multiscale swin transformer and deeply supervised network for change detection of the fast-growing urban regions,” IEEE Geoscience and Remote Sensing Letters , vol. 19, pp. 1–5, 2022

    2022

  40. [50]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,” Advances in Neural Information Processing Systems, vol. 34, pp. 12 077–12 090, 2021

    2021

  41. [51]

    Imputeformer: Low rankness-induced transformers for generalizable spatiotemporal impu- tation,

    T. Nie, G. Qin, W. Ma, Y . Mei, and J. Sun, “Imputeformer: Low rankness-induced transformers for generalizable spatiotemporal impu- tation,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2024, pp. 2260–2271

    2024

  42. [52]

    Geolocation representation from large language models are generic enhancers for spatio-temporal learning,

    J. He, T. Nie, and W. Ma, “Geolocation representation from large language models are generic enhancers for spatio-temporal learning,” arXiv preprint arXiv:2408.12116 , 2024

    2024 arXiv

  43. [53]

    Hivit: A simpler and more efficient design of hierarchical vision transformer,

    X. Zhang, Y . Tian, L. Xie, W. Huang, Q. Dai, Q. Ye, and Q. Tian, “Hivit: A simpler and more efficient design of hierarchical vision transformer,” in The Eleventh International Conference on Learning Representations , 2023

    2023

  44. [54]

    Rethinking and improv- ing relative position encoding for vision transformer,

    K. Wu, H. Peng, M. Chen, J. Fu, and H. Chao, “Rethinking and improv- ing relative position encoding for vision transformer,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 10 033–10 041

    2021

  45. [55]

    Refiner: Refining self-attention for vision transformers,

    D. Zhou, Y . Shi, B. Kang, W. Yu, Z. Jiang, Y . Li, X. Jin, Q. Hou, and J. Feng, “Refiner: Refining self-attention for vision transformers,” arXiv preprint arXiv:2106.03714, 2021

    2021 arXiv

  46. [56]

    Ultra-high resolution segmentation via boundary-enhanced patch-merging transformer,

    H. Sun, “Ultra-high resolution segmentation via boundary-enhanced patch-merging transformer,” 2024. [Online]. Available: https://arxiv.org/ abs/2412.10181

    2024 arXiv

  47. [57]

    Spt: Sequence prompt transformer for interactive image segmentation,

    S. Cheng and H. Sun, “Spt: Sequence prompt transformer for interactive image segmentation,” 2024. [Online]. Available: https: //arxiv.org/abs/2412.10224

    2024 arXiv

  48. [58]

    Cvt: Introducing convolutions to vision transformers,

    H. Wu, B. Xiao, N. Codella, M. Liu, X. Dai, L. Yuan, and L. Zhang, “Cvt: Introducing convolutions to vision transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 22–31

    2021

  49. [59]

    Cmt: Convolutional neural networks meet vision transformers,

    J. Guo, K. Han, H. Wu, Y . Tang, X. Chen, Y . Wang, and C. Xu, “Cmt: Convolutional neural networks meet vision transformers,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 12 175–12 185

    2022

  50. [60]

    Convit: Improving vision transformers with soft con- volutional inductive biases,

    S. d’Ascoli, H. Touvron, M. L. Leavitt, A. S. Morcos, G. Biroli, and L. Sagun, “Convit: Improving vision transformers with soft con- volutional inductive biases,” in International Conference on Machine Learning. PMLR, 2021, pp. 2286–2296

    2021

  51. [61]

    Mlp-mixer: An all-mlp architecture for vision,

    I. O. Tolstikhin, N. Houlsby, A. Kolesnikov, L. Beyer, X. Zhai, T. Un- terthiner, J. Yung, A. Steiner, D. Keysers, J. Uszkoreitet al., “Mlp-mixer: An all-mlp architecture for vision,” Advances in neural information processing systems, vol. 34, pp. 24 261–24 272, 2021

    2021

  52. [62]

    Pay attention to mlps,

    H. Liu, Z. Dai, D. So, and Q. V . Le, “Pay attention to mlps,” Advances in Neural Information Processing Systems , vol. 34, pp. 9204–9215, 2021

    2021

  53. [63]

    Resmlp: Feed- forward networks for image classification with data-efficient training,

    H. Touvron, P. Bojanowski, M. Caron, M. Cord, A. El-Nouby, E. Grave, G. Izacard, A. Joulin, G. Synnaeve, J. Verbeek et al. , “Resmlp: Feed- forward networks for image classification with data-efficient training,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2022

    2022

  54. [64]

    Metaformer is actually what you need for vision,

    W. Yu, M. Luo, P. Zhou, C. Si, Y . Zhou, X. Wang, J. Feng, and S. Yan, “Metaformer is actually what you need for vision,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 10 819–10 829

    2022

  55. [65]

    Fnet: Mixing tokens with fourier transforms,

    J. Lee-Thorp, J. Ainslie, I. Eckstein, and S. Ontanon, “Fnet: Mixing tokens with fourier transforms,” arXiv preprint arXiv:2105.03824, 2021

    2021 arXiv

  56. [66]

    Asymmetric cross-attention hierarchical network based on cnn and transformer for bitemporal remote sensing images change detection,

    X. Zhang, S. Cheng, L. Wang, and H. Li, “Asymmetric cross-attention hierarchical network based on cnn and transformer for bitemporal remote sensing images change detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–15, 2023

    2023

  57. [67]

    Shunted self-attention via multi-scale token aggregation,

    S. Ren, D. Zhou, S. He, J. Feng, and X. Wang, “Shunted self-attention via multi-scale token aggregation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 10 853–10 862

    2022

  58. [68]

    Change is everywhere: Single-temporal supervised object change detection in remote sensing imagery,

    Z. Zheng, A. Ma, L. Zhang, and Y . Zhong, “Change is everywhere: Single-temporal supervised object change detection in remote sensing imagery,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 15 193–15 202

    2021

  59. [69]

    Remote sensing image change detection transformer network based on dual-feature mixed attention,

    X. Song, Z. Hua, and J. Li, “Remote sensing image change detection transformer network based on dual-feature mixed attention,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–16, 2022

    2022

  60. [70]

    Building change detection for vhr remote sensing images via local– global pyramid network and cross-task transfer learning strategy,

    T. Liu, M. Gong, D. Lu, Q. Zhang, H. Zheng, F. Jiang, and M. Zhang, “Building change detection for vhr remote sensing images via local– global pyramid network and cross-task transfer learning strategy,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–17, 2022

    2022

  61. [71]

    Ultralightweight spatial–spectral feature cooperation network for change detection in remote sensing images,

    T. Lei, X. Geng, H. Ning, Z. Lv, M. Gong, Y . Jin, and A. K. Nandi, “Ultralightweight spatial–spectral feature cooperation network for change detection in remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–14, 2023

    2023

  62. [72]

    Deep roots: Improving cnn efficiency with hierarchical filter groups,

    Y . Ioannou, D. Robertson, R. Cipolla, and A. Criminisi, “Deep roots: Improving cnn efficiency with hierarchical filter groups,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 1231–1240

    2017

  63. [73]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2980–2988

    2017

  64. [74]

    Dice loss for data-imbalanced nlp tasks,

    X. Li, X. Sun, Y . Meng, J. Liang, F. Wu, and J. Li, “Dice loss for data-imbalanced nlp tasks,” arXiv preprint arXiv:1911.02855 , 2019

    1911 arXiv

  65. [75]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

  66. [76]

    Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set,

    S. Ji, S. Wei, and M. Lu, “Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set,” IEEE Transactions on Geoscience and Remote Sensing , vol. 57, no. 1, pp. 574–586, 2018

    2018

  67. [77]

    A spatial-temporal attention-based method and a new dataset for remote sensing image change detection,

    H. Chen and Z. Shi, “A spatial-temporal attention-based method and a new dataset for remote sensing image change detection,” Remote Sensing, vol. 12, no. 10, p. 1662, 2020

    2020

  68. [78]

    A cnn-transformer network with multi-scale context aggregation for fine-grained cropland change detec- tion,

    M. Liu, Z. Chai, H. Deng, and R. Liu, “A cnn-transformer network with multi-scale context aggregation for fine-grained cropland change detec- tion,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 2022

    2022

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.