Pith. sign in

REVIEW 3 major objections 4 minor 76 references

Be the Change You Want to See: Revisiting Remote Sensing Change Detection Practices

T0 review · 3 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 Fundamental design choices, not new architectures, drive most recent change detection gains.

desk verdict A genuinely useful empirical study of design choices in change detection, whose strongest claim is undercut by test-set selection but whose baseline and guidelines are solid contributions. read the letter →

arxiv 2507.03367 v1 pith:GEU7CVL5 submitted 2025-07-04 cs.CV cs.AI

Blaž Rolih , Matic Fučka , Filip Wolf , Luka Čehovin Zajc This is my paper · ORCID
classification cs.CVcs.AI
keywords changedetectionremotesensingdesignchoicesbackbonepretrainingdataaugmentationDicelosslearningrateschedulerSiamesenetwork
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

The paper claims that much of the recent progress in remote sensing change detection comes not from new architectural components but from basic design choices: which pretrained backbone to use, what data augmentation to apply, which loss function to optimize, and how to schedule the learning rate. By systematically varying these choices on a deliberately simple model, the authors raise average binary F1 from 71.5% to 80.9% across six datasets, and they show the same choices improve existing methods when applied to them. A sympathetic reader would care because the result redirects effort: standardize and optimize the basics before adding architectural complexity.

What carries the argument

The machinery is a deliberately simple change detection model used as an ablation harness: a Siamese encoder with shared weights applied to both images, element-wise subtraction of the two feature maps at each level, and a UPerNet decoder that turns the fused features into a change map. This setup isolates the effect of each design choice, and the winning recipe—a Swin-B backbone pretrained on Cityscapes semantic segmentation, flip augmentation, Dice loss, and a cosine scheduler—carries the empirical result.

What would settle it

Train the exact BTC recipe on a held-out change detection benchmark from a different sensor family, such as SAR or multispectral imagery, or on datasets with strongly different geography; if the 9.4 percentage point gain over the base configuration shrinks or reverses, the claim that these choices are general best practices is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that fundamental design choices—backbone selection, backbone size, pretraining data and task, augmentation, loss function, and learning-rate scheduler—account for a large share of the performance attributed to architectural novelty in change detection. Starting from a randomly initialized Swin-T with no augmentation, no scheduler, and cross-entropy loss, the authors incrementally apply ImageNet1k pretraining, flip augmentation, Cityscapes semantic-segmentation pretraining, a cosine scheduler, a larger Swin-B backbone, and Dice loss. These additions raise the average binary F1 by 9.4 percentage points, from 71.5% to 80.9% across six datasets, and the resulting simple model, BTC, matches or surpasses state-of-the-art methods, particularly on smaller datasets.

Load-bearing premise

The conclusion rests on the six RGB datasets and the binary-F1 protocol being a fair sample of change detection as a field; if other sensors or regions behave differently, the recommended defaults may not transfer.

Editorial extensions

If this is right

  • Applying the identified recipe raises average F1 by 9.4 percentage points over a randomly initialized baseline, with the largest single gain coming from pretraining.
  • An architecturally simple Siamese Swin encoder with subtraction fusion and a UPerNet decoder can match or beat published state-of-the-art change detection methods on six datasets.
  • The design choices transfer: adding flip augmentation, Dice loss, and a cosine scheduler improves remote sensing foundation models by 1.9 to 3.8 percentage points and improves the FC-Siam-Diff baseline by 16.9 percentage points.
  • Segmentation pretraining on Cityscapes outperforms ImageNet1k classification pretraining and also outperforms the tested remote sensing pretraining, suggesting that task similarity matters more than domain match.
  • Past methods may have underperformed because of suboptimal defaults rather than architectural limitations, so future comparisons should use well-optimized baselines.

Reading between the lines

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

  • The paper leaves implicit that a cheap diagnostic for the field is to re-run recently published change detection architectures inside the BTC training recipe; margins that vanish under that recipe are better attributed to suboptimal defaults than to architectural ideas.
  • The pretraining result points to an untested extension: pretraining on larger, task-aligned natural-image segmentation data may keep closing the gap to specialized remote sensing foundation models.
  • Because the six datasets are all RGB, the recipe's transfer to SAR, multispectral, and hyperspectral change detection is untested; re-running the same ablations on multi-modal benchmarks is a direct next experiment.
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

3 major / 4 minor

Summary. The paper investigates the contribution of fundamental design choices (backbone architecture and scale, pre-training dataset/task, data augmentation, loss function, learning-rate scheduler) to remote sensing change detection, using a simple Siamese encoder with subtraction fusion and an UPerNet decoder as a test bed. On six RGB datasets it reports that these choices improve average binary F1 from 71.5% to 80.9% (Table VII), and that the resulting BTC model matches or surpasses state-of-the-art methods on several datasets (Table IX). The authors also show that applying the identified insights to existing foundation models and change-detection architectures yields consistent gains (Table VIII), and they propose a standardized evaluation protocol (Section III-B3).

Significance. The paper is a potentially valuable systematic study: it is among the few to isolate the contribution of training and backbone choices in change detection, it uses three seeds and reports standard deviations, it releases code, and the transfer experiments are a useful addition. If the conclusions are robust, they support a practical message that the community should optimize fundamental components rather than only architectural novelty. However, the main claim is currently supported by test-set selection and a narrow dataset scope, so the headline result should be interpreted with care.

major comments (3)
  1. [Section III-B, Tables I-VII] The design choices are selected by average F1 on the same six predefined test sets that are later used for the state-of-the-art comparison, with no held-out validation split. Because per-dataset F1 is noisy (e.g., OSCD standard deviation up to 12.0 p.p. in Table S9), the argmax over roughly 30 configurations is expected to overestimate the true performance of the selected configuration, and the 9.4 p.p. gain in Table VII may partly reflect selection bias rather than a generalizable property of the choices. The authors should either introduce a validation split for model selection, use nested cross-validation, or clearly relabel the findings as "best on these test sets" rather than "best practices".
  2. [Section VI-A and Table IX] The comparison to state-of-the-art is asymmetric: BTC is the product of test-set selection, whereas competing methods are run with their default training recipes (Section VI-B). The transfer experiments in Section VI-A apply insights that were themselves selected on these same six test sets, so they do not break the selection-on-test-set circularity. To support the claim that fundamental design choices rather than architecture drive gains, the authors should tune the competing methods under the same protocol on a validation split, or at minimum report results for a pre-registered configuration fixed before test-set evaluation.
  3. [Section III-B1 and Abstract] The general claim that fundamental design choices matter more than architectural novelty is established only on six RGB datasets, all at relatively high resolution and dominated by building/urban/cropland changes. Section VII acknowledges that multi-modal data (SAR, multispectral) is future work, but the Abstract and Introduction state the conclusion without this scope restriction. The authors should add an explicit caveat or extend the evaluation to at least one dataset with a different modality or substantially different acquisition geometry before claiming a field-wide generalization.
minor comments (4)
  1. [Table I] In Table I, the "None" row appears to be column-shifted relative to the values in Table S9: as printed, the EGYBCD, GVLM, CLCD, and OSCD entries are each one column to the left of their correct positions, which is confusing for readers comparing the two tables.
  2. [Section S2] In Section S2, "we urge researches to correctly use binary F1 metric" should read "we urge researchers to correctly use the binary F1 metric."
  3. [Table VIII header] The naming is inconsistent: "SwinSUN" in the Table VIII header should be "SwinSUNet", and "GaSSL" appears in places where "GeSSL" is used elsewhere (e.g., Table IX and the text).
  4. [Figure 1 caption] The Figure 1 caption states that "The first two rows contain the input image pair," which is hard to parse because the figure appears to show one image pair and then two prediction rows; please rephrase for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the 9.4 p.p. gain is an empirical ablation result, not a derived or fitted quantity.

full rationale

This paper is an empirical design-space study; there is no formal derivation whose output is equivalent to its input. The 9.4 p.p. improvement is a measured difference between trained baseline models and trained optimized models on fixed test splits, not a fitted parameter renamed as a prediction. Design choices such as Cityscapes pre-training, flip augmentation, cosine scheduler, Dice loss, and Swin-B are selected by comparing average F1 on the same six test sets later used for the SOTA comparison (Section III-B and Tables VII/IX), which is a benchmark-selection caveat that can inflate apparent gains and weaken generalization claims, but it is not definitional or self-referential circularity. The paper reports three-seed standard deviations and conducts transfer experiments to other methods, providing independent checks. The only self-citation ([20], a prior paper by two of the authors on training recipes) appears in a related-work list and is not load-bearing. No structural claim reduces to its own premise.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

This paper introduces no invented physical or mathematical entities. However, the final model configuration depends on a set of design choices selected by test-set performance across the six datasets; these are treated here as fitted hyperparameters. The axioms include evaluation and domain-representativeness assumptions that underpin the empirical claims.

free parameters (6)
  • Pre-training dataset and task = Cityscapes semantic segmentation (Swin-B)
    Selected as the best among tested options based on average F1 across six test datasets (Tables I, VII, S3).
  • Augmentation policy = Flip (horizontal, vertical, rotation in [-90,90]) at 30% probability
    Chosen based on average F1 across datasets; the 30% probability and specific flip family are hand-picked (Section V-A, S3-D).
  • Loss function = Dice loss
    Dice loss had the best average F1, especially on OSCD, and was chosen for the final model (Section V-C, Table VI).
  • Learning rate scheduler = Cosine decay (used with augmentations)
    No scheduler was best in isolation; cosine was chosen because it improves when combined with augmentations (Section V-B, Table VII).
  • Backbone scale = Swin-B
    Swin-B gives the best average F1 (+0.4 p.p. over Swin-T) but with near-double compute; selected for final BTC-B (Section IV-C, Table III).
  • Training hyperparameters = LR 1e-4, weight decay 1e-4, batch size 32, 100 epochs (50 for OSCD); ViT uses LR 6e-5 and WD 0.05
    Hand-chosen defaults held constant across experiments; these choices affect all comparisons (Section III-C, S3-B).
assumptions (5)
  • domain assumption Binary F1 computed on the change class is the appropriate evaluation metric for change detection; class-averaged mF1 inflates scores.
    Section III-B2 and Supplementary S2 present this as a recommendation for the field. It is a reasonable but debatable evaluation choice, not an externally validated truth.
  • domain assumption The six selected RGB datasets are representative of remote sensing change detection.
    Section III-B1 claims the datasets cover diverse sensors, resolutions, and change types, but they are all RGB; no SAR or multispectral data, which limits generalizability.
  • domain assumption A Siamese encoder with element-wise subtraction and UPerNet decoder is a strong, representative base architecture.
    Section III-A states the authors preliminarily evaluated other fusion and decoder designs and adopted well-established choices. The validity of the design-choice conclusions depends on this base being a fair representative.
  • domain assumption Pre-trained weights are necessary for good performance, and standard ImageNet pre-training is a valid starting point.
    Section III-C uses ImageNet pre-training for the baseline; the paper never questions whether training from scratch with more epochs could change conclusions.
  • domain assumption Three random seeds and final-epoch test evaluation are sufficient to compare methods.
    Section III-B reports mean F1 over three seeds and evaluation on the final epoch; this assumes low seed variance and no need for early stopping or validation-based model selection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Be the Change You Want to See: Revisiting Remote Sensing Change Detection Practices." pith.science (2026). https://pith.science/paper/GEU7CVL5

@misc{pith2026250703367,
  author       = {Pith},
  title        = {Pith review of: Be the Change You Want to See: Revisiting Remote Sensing Change Detection Practices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GEU7CVL5}},
  note         = {Machine review of arXiv:2507.03367}
}
read the original abstract

Remote sensing change detection aims to localize semantic changes between images of the same location captured at different times. In the past few years, newer methods have attributed enhanced performance to the additions of new and complex components to existing architectures. Most fail to measure the performance contribution of fundamental design choices such as backbone selection, pre-training strategies, and training configurations. We claim that such fundamental design choices often improve performance even more significantly than the addition of new architectural components. Due to that, we systematically revisit the design space of change detection models and analyse the full potential of a well-optimised baseline. We identify a set of fundamental design choices that benefit both new and existing architectures. Leveraging this insight, we demonstrate that when carefully designed, even an architecturally simple model can match or surpass state-of-the-art performance on six challenging change detection datasets. Our best practices generalise beyond our architecture and also offer performance improvements when applied to related methods, indicating that the space of fundamental design choices has been underexplored. Our guidelines and architecture provide a strong foundation for future methods, emphasizing that optimizing core components is just as important as architectural novelty in advancing change detection performance. Code: https://github.com/blaz-r/BTC-change-detection

Figures

Figures reproduced from arXiv: 2507.03367 by the authors.

Figure 1
Figure 1. Comparison of change detection performance. The first two rows con [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Outline of our change detection method. Each image in the given [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of different augmentations used during the analysis. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Summary of important fundamental building blocks analysed in our [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of the predictions made by models with different combined configurations. The pair of considered images is shown in the first [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of the predictions made by our method and [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 70 canonical work pages

  1. [1]

    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

    work page 2021

  2. [2]

    Urban Change Detection for Multispectral Earth Observation Using Convolutional Neu- ral Networks,

    R. C. Daudt, B. Le Saux, A. Boulch, and Y . Gousseau, “Urban Change Detection for Multispectral Earth Observation Using Convolutional Neu- ral Networks,” in IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2018, pp. 2115–2118

    work page 2018

  3. [3]

    Segment Any Change,

    Z. Zheng, Y . Zhong, L. Zhang, and S. Ermon, “Segment Any Change,” in Advances in Neural Information Processing Systems , vol. 37, 2024, pp. 81 204–81 224

    work page 2024

  4. [4]

    Remote Sensing of Land Change: A Multifaceted Perspective,

    Z. Zhu, S. Qiu, and S. Ye, “Remote Sensing of Land Change: A Multifaceted Perspective,” Remote Sensing of Environment , vol. 282, p. 113266, 2022

    work page 2022

  5. [5]

    Earth Observation and Machine Learning for Climate Change,

    R. H ¨ansch and M. A. Chaurasia, “Earth Observation and Machine Learning for Climate Change,” in IEEE International Geoscience and Remote Sensing Symposium . IEEE, 2024, pp. 1676–1682

    work page 2024

  6. [6]

    Bifa: Remote Sensing Image Change Detection With Bitemporal Feature Alignment,

    H. Zhang, H. Chen, C. Zhou, K. Chen, C. Liu, Z. Zou, and Z. Shi, “Bifa: Remote Sensing Image Change Detection With Bitemporal Feature Alignment,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    work page 2024

  7. [7]

    DDPM-CD: Denoising Diffusion Probabilistic Models as Feature Extractors for Remote Sensing Change Detection,

    W. G. C. Bandara, N. G. Nair, and V . Patel, “DDPM-CD: Denoising Diffusion Probabilistic Models as Feature Extractors for Remote Sensing Change Detection,” in IEEE/CVF Winter Conference on Applications of Computer Vision, 2025, pp. 5250–5262

    work page 2025

  8. [8]

    A Transformer-Based Siamese Network for Change Detection,

    W. G. C. Bandara and V . M. Patel, “A Transformer-Based Siamese Network for Change Detection,” in IEEE International Geoscience and Remote Sensing Symposium . IEEE, 2022, pp. 207–210

    work page 2022

Show all 76 references
  1. [9]

    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

  2. [10]

    Lightweight Remote Sensing Change Detection With Progressive Fea- ture Aggregation and Supervised Attention,

    Z. Li, C. Tang, X. Liu, W. Zhang, J. Dou, L. Wang, and A. Y . Zomaya, “Lightweight Remote Sensing Change Detection With Progressive Fea- ture Aggregation and Supervised Attention,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–12, 2023

    2023

  3. [11]

    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 IEEE/CVF International Conference on Computer Vision , 2021, pp. 10 012–10 022

    2021

  4. [12]

    Swin Transformer V2: Scaling up Capacity and Resolution,

    Z. Liu, H. Hu, Y . Lin, Z. Yao, Z. Xie, Y . Wei, J. Ning, Y . Cao, Z. Zhang, L. Dong et al., “Swin Transformer V2: Scaling up Capacity and Resolution,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 12 009–12 019

    2022

  5. [13]

    A Convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A Convnet for the 2020s,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 11 976–11 986

    2022

  6. [14]

    Change is Everywhere: Single-Temporal Supervised Object Change Detection in Remote Sens- ing Imagery,

    Z. Zheng, A. Ma, L. Zhang, and Y . Zhong, “Change is Everywhere: Single-Temporal Supervised Object Change Detection in Remote Sens- ing Imagery,” in IEEE/CVF International Conference on Computer Vision, 2021, pp. 15 193–15 202

    2021

  7. [15]

    SwinSUNet: Pure Trans- former Network for Remote Sensing Image Change Detection,

    C. Zhang, L. Wang, S. Cheng, and Y . Li, “SwinSUNet: Pure Trans- former Network for Remote Sensing Image Change Detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–13, 2022

    2022

  8. [16]

    ChangeMamba: Remote Sensing Change Detection With Spatiotemporal State Space Model,

    H. Chen, J. Song, C. Han, J. Xia, and N. Yokoya, “ChangeMamba: Remote Sensing Change Detection With Spatiotemporal State Space Model,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–20, 2024. PREPRINT 10

    2024

  9. [17]

    Exploring the Importance of Pretrained Feature Extractors for Unsupervised Anomaly Detection and Localization,

    L. Heckler, R. K ¨onig, and P. Bergmann, “Exploring the Importance of Pretrained Feature Extractors for Unsupervised Anomaly Detection and Localization,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2023, pp. 2916–2925

    2023

  10. [18]

    In Defense of Pre- Trained Imagenet Architectures for Real-Time Semantic Segmentation of Road-Driving Images,

    M. Orsic, I. Kreso, P. Bevandic, and S. Segvic, “In Defense of Pre- Trained Imagenet Architectures for Real-Time Semantic Segmentation of Road-Driving Images,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019

    2019

  11. [19]

    DINO: DETR With Improved Denoising Anchor Boxes for End-to-End Object Detection,

    H. Zhang, F. Li, S. Liu, L. Zhang, H. Su, J. Zhu, L. Ni, and H.-Y . Shum, “DINO: DETR With Improved Denoising Anchor Boxes for End-to-End Object Detection,” in International Conference on Learning Representations, 2023

    2023

  12. [20]

    SuperSimpleNet: Unifying Unsu- pervised and Supervised Learning for Fast and Reliable Surface Defect Detection,

    B. Rolih, M. Fu ˇcka, and D. Sko ˇcaj, “SuperSimpleNet: Unifying Unsu- pervised and Supervised Learning for Fast and Reliable Surface Defect Detection,” in International Conference on Pattern Recognition , 2024

    2024

  13. [21]

    Review Article Digital Change Detection Techniques Using Remotely-Sensed Data,

    A. Singh, “Review Article Digital Change Detection Techniques Using Remotely-Sensed Data,” International journal of remote sensing, vol. 10, no. 6, pp. 989–1003, 1989

    1989

  14. [22]

    Urban Change Detection in SAR Images by Interactive Learning,

    B. Le Saux and H. Randrianarivo, “Urban Change Detection in SAR Images by Interactive Learning,” in IEEE International Geoscience and Remote Sensing Symposium . IEEE, 2013, pp. 3990–3993

    2013

  15. [23]

    Urban Change Forecasting From Satellite Images,

    N. Metzger, M. ¨O. T ¨urkoglu, R. C. Daudt, J. D. Wegner, and K. Schindler, “Urban Change Forecasting From Satellite Images,” PFG– Journal of Photogrammetry, Remote Sensing and Geoinformation Sci- ence, vol. 91, no. 6, pp. 443–452, 2023

    2023

  16. [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 IEEE International Conference on Image Processing. IEEE, 2018, pp. 4063–4067

    2018

  17. [25]

    An Image is Worth 16x16 Words: Trans- formers 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, J. Uszkoreit, and N. Houlsby, “An Image is Worth 16x16 Words: Trans- formers for Image Recognition at Scale,” in International Conference on Learni...

    2021

  18. [26]

    Mamba: Linear-Time Sequence Modeling With Selective State Spaces,

    A. Gu and T. Dao, “Mamba: Linear-Time Sequence Modeling With Selective State Spaces,” in First Conference on Language Modeling ,

  19. [27]

    VMamba: Visual State Space Model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, and Y . Liu, “VMamba: Visual State Space Model,” in Advances in Neural Information Processing Systems , vol. 37, 2024, pp. 103 031–103 063. [Online]. Available: https://openreview.net/forum?id=ZgtLQQR1K7

    2024

  20. [28]

    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, p. 1662, 2020

    2020

  21. [29]

    Remote Sensing Change Detection With Transformers Trained From Scratch,

    M. Noman, M. Fiaz, H. Cholakkal, S. Narayan, R. M. Anwer, S. Khan, and F. S. Khan, “Remote Sensing Change Detection With Transformers Trained From Scratch,” IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  22. [30]

    Imagenet: a Large-Scale Hierarchical Image Database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: a Large-Scale Hierarchical Image Database,” in IEEE Conference on Computer Vision and Pattern Recognition . IEEE, 2009, pp. 248–255

    2009

  23. [31]

    A New Learning Paradigm for Foundation Model-Based Remote-Sensing Change Detection,

    K. Li, X. Cao, and D. Meng, “A New Learning Paradigm for Foundation Model-Based Remote-Sensing Change Detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–12, 2024

    2024

  24. [32]

    AnySat: An Earth Observation Model for Any Resolutions, Scales, and Modalities,

    G. Astruc, N. Gonthier, C. Mallet, and L. Landrieu, “AnySat: An Earth Observation Model for Any Resolutions, Scales, and Modalities,” arXiv preprint arXiv:2412.14123, 2024

    2024 arXiv

  25. [33]

    OmniSat: Self-Supervised Modality Fusion for Earth Observa- tion,

    ——, “OmniSat: Self-Supervised Modality Fusion for Earth Observa- tion,” in European Conference on Computer Vision, 2024, pp. 409–427

    2024

  26. [34]

    Geography-Aware Self-Supervised Learning,

    K. Ayush, B. Uzkent, C. Meng, K. Tanmay, M. Burke, D. Lobell, and S. Ermon, “Geography-Aware Self-Supervised Learning,” in IEEE/CVF International Conference on Computer Vision, 2021, pp. 10 181–10 190

    2021

  27. [35]

    Seasonal Contrast: Unsupervised Pre-Training From Uncurated Remote Sensing Data,

    O. Manas, A. Lacoste, X. Gir ´o-i Nieto, D. Vazquez, and P. Rodriguez, “Seasonal Contrast: Unsupervised Pre-Training From Uncurated Remote Sensing Data,” in IEEE/CVF International Conference on Computer Vision, 2021, pp. 9414–9423

    2021

  28. [36]

    Change-Aware Sampling and Contrastive Learning for Satellite Images,

    U. Mall, B. Hariharan, and K. Bala, “Change-Aware Sampling and Contrastive Learning for Satellite Images,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 5261–5270

    2023

  29. [37]

    SatMAE: Pre-Training Transformers for Temporal and Multi-Spectral Satellite Imagery,

    Y . Cong, S. Khanna, C. Meng, P. Liu, E. Rozi, Y . He, M. Burke, D. Lobell, and S. Ermon, “SatMAE: Pre-Training Transformers for Temporal and Multi-Spectral Satellite Imagery,” Advances in Neural Information Processing Systems , vol. 35, pp. 197–211, 2022

    2022

  30. [38]

    Cross-Scale Mae: A Tale of Multiscale Exploitation in Remote Sensing,

    M. Tang, A. Cozma, K. Georgiou, and H. Qi, “Cross-Scale Mae: A Tale of Multiscale Exploitation in Remote Sensing,” Advances in Neural Information Processing Systems , vol. 36, pp. 20 054–20 066, 2024

    2024

  31. [39]

    Towards Geospatial Foundation Models via Continual Pretraining,

    M. Mendieta, B. Han, X. Shi, Y . Zhu, and C. Chen, “Towards Geospatial Foundation Models via Continual Pretraining,” in IEEE/CVF Interna- tional Conference on Computer Vision , 2023, pp. 16 806–16 816

    2023

  32. [40]

    MTP: Advancing Remote Sensing Foundation Model via Multi-Task Pretraining,

    D. Wang, J. Zhang, M. Xu, L. Liu, D. Wang, E. Gao, C. Han, H. Guo, B. Du, D. Tao et al. , “MTP: Advancing Remote Sensing Foundation Model via Multi-Task Pretraining,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 2024

    2024

  33. [41]

    An Empirical Study of Remote Sensing Pretraining,

    D. Wang, J. Zhang, B. Du, G.-S. Xia, and D. Tao, “An Empirical Study of Remote Sensing Pretraining,” IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–20, 2022

    2022

  34. [42]

    Self-Supervised Pretraining via Multimodality Images With Transformer for Change Detection,

    Y . Zhang, Y . Zhao, Y . Dong, and B. Du, “Self-Supervised Pretraining via Multimodality Images With Transformer for Change Detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–11, 2023

    2023

  35. [43]

    Self-Supervised Pre-Training Enhances Change Detection in Sentinel-2 Imagery,

    M. Leenstra, D. Marcos, F. Bovolo, and D. Tuia, “Self-Supervised Pre-Training Enhances Change Detection in Sentinel-2 Imagery,” in Pattern Recognition. ICPR International Workshops and Challenges, Proceedings, Part VII. Springer, 2021, pp. 578–590

    2021

  36. [44]

    Unified Building Change Detection Pre-Training Method With Masked Semantic Annotations,

    Y . Quan, A. Yu, W. Guo, X. Lu, B. Jiang, S. Zheng, and P. He, “Unified Building Change Detection Pre-Training Method With Masked Semantic Annotations,” International Journal of Applied Earth Observation and Geoinformation, vol. 120, p. 103346, 2023

    2023

  37. [45]

    Scalable Multi- Temporal Remote Sensing Change Data Generation via Simulating Stochastic Change Process,

    Z. Zheng, S. Tian, A. Ma, L. Zhang, and Y . Zhong, “Scalable Multi- Temporal Remote Sensing Change Data Generation via Simulating Stochastic Change Process,” in IEEE/CVF International Conference on Computer Vision, 2023, pp. 21 818–21 827

    2023

  38. [46]

    Bag of Tricks for Image Classification With Convolutional Neural Networks,

    T. He, Z. Zhang, H. Zhang, Z. Zhang, J. Xie, and M. Li, “Bag of Tricks for Image Classification With Convolutional Neural Networks,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2019, pp. 558–567

    2019

  39. [47]

    Bag of Tricks for Training Data Extraction From Language Models,

    W. Yu, T. Pang, Q. Liu, C. Du, B. Kang, Y . Huang, M. Lin, and S. Yan, “Bag of Tricks for Training Data Extraction From Language Models,” in International Conference on Machine Learning , 2023, pp. 40 306– 40 320

    2023

  40. [48]

    How Effective is Pre-Training of Large Masked Autoencoders for Downstream Earth Observation Tasks?

    J. Sosa, M. Aloulou, D. Rukhovich, R. Sleimi, B. Changaival, A. Kacem, and D. Aouada, “How Effective is Pre-Training of Large Masked Autoencoders for Downstream Earth Observation Tasks?” Workshop on Machine Vision for Earth Observation and Environment Monitoring , 2024

    2024

  41. [49]

    A Change Detection Reality Check,

    I. Corley, C. Robinson, and A. Ortiz, “A Change Detection Reality Check,” arXiv preprint arXiv:2402.06994 , 2024

    2024 arXiv

  42. [50]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778

    2016

  43. [51]

    Unified Perceptual Parsing for Scene Understanding,

    T. Xiao, Y . Liu, B. Zhou, Y . Jiang, and J. Sun, “Unified Perceptual Parsing for Scene Understanding,” inEuropean Conference on Computer Vision, 2018, pp. 418–434

    2018

  44. [52]

    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, 2022

    2022

  45. [53]

    AFDE-net: Building Change Detection Using Attention-Based Feature Differential Enhancement for Satellite Imagery,

    S. Holail, T. Saleh, X. Xiao, and D. Li, “AFDE-net: Building Change Detection Using Attention-Based Feature Differential Enhancement for Satellite Imagery,” IEEE Geoscience and Remote Sensing Letters , vol. 20, pp. 1–5, 2023

    2023

  46. [54]

    Cross-Domain Landslide Mapping From Large-Scale Remote Sensing Images Using Prototype- Guided Domain-Aware Progressive Representation Learning,

    X. Zhang, W. Yu, M.-O. Pun, and W. Shi, “Cross-Domain Landslide Mapping From Large-Scale Remote Sensing Images Using Prototype- Guided Domain-Aware Progressive Representation Learning,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 197, pp. 1–17, 2023

    2023

  47. [55]

    A CNN-Transformer Net- work With Multiscale Context Aggregation for Fine-Grained Cropland Change Detection,

    M. Liu, Z. Chai, H. Deng, and R. Liu, “A CNN-Transformer Net- work With Multiscale Context Aggregation for Fine-Grained Cropland Change Detection,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 15, pp. 4297–4306, 2022

    2022

  48. [56]

    Decoupled Weight Decay Regularization,

    I. Loshchilov and F. Hutter, “Decoupled Weight Decay Regularization,” in International Conference on Learning Representations , 2017

    2017

  49. [57]

    Scene Parsing Through ADE20k Dataset,

    B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba, “Scene Parsing Through ADE20k Dataset,” in IEEE Conference on Computer Vision and Pattern Recognition , 2017, pp. 633–641

    2017

  50. [58]

    The Cityscapes Dataset for Semantic Urban Scene Understanding,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Be- nenson, U. Franke, S. Roth, and B. Schiele, “The Cityscapes Dataset for Semantic Urban Scene Understanding,” in IEEE Conference on Computer Vision and Pattern Recognition , 2016

    2016

  51. [59]

    Microsoft Coco: Common Objects in Context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft Coco: Common Objects in Context,” in European Conference on Computer Vision , 2014, pp. 740– 755. PREPRINT 11

    2014

  52. [60]

    EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification,

    P. Helber, B. Bischke, A. Dengel, and D. 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 , 2019

    2019

  53. [61]

    Introducing EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification,

    ——, “Introducing EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification,” in IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2018, pp. 204–207

    2018

  54. [62]

    Position: Mission critical – satellite data is a distinct modality in machine learning,

    E. Rolf, K. Klemmer, C. Robinson, and H. Kerner, “Position: Mission critical – satellite data is a distinct modality in machine learning,” in International Conference on Machine Learning , 2024. [Online]. Available: https://openreview.net/forum?id=PQ0ERKKYJu

    2024

  55. [63]

    SGDR: Stochastic Gradient Descent With Warm Restarts,

    I. Loshchilov and F. Hutter, “SGDR: Stochastic Gradient Descent With Warm Restarts,” in International Conference on Learning Representations, 2017. [Online]. Available: https://openreview.net/for um?id=Skq89Scxx

    2017

  56. [64]

    V-Net: Fully Convolutional Neural Networks for V olumetric Medical Image Segmentation,

    F. Milletari, N. Navab, and S.-A. Ahmadi, “V-Net: Fully Convolutional Neural Networks for V olumetric Medical Image Segmentation,” in International Conference on 3D Vision , 2016, pp. 565–571

    2016

  57. [65]

    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 IEEE International Conference on Computer Vision, 2017, pp. 2980–2988

    2017

  58. [66]

    Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations,

    C. H. Sudre, W. Li, T. Vercauteren, S. Ourselin, and M. Jorge Cardoso, “Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations,” in Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third Inter...

    2017

  59. [67]

    U-Net: Convolutional Net- works for Biomedical Image Segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Net- works for Biomedical Image Segmentation,” in Medical Image Comput- ing and Computer-assisted Intervention–MICCAI, proceedings, part III

  60. [68]

    Springer, 2017, pp. 240–248

    2017

  61. [70]

    Springer, 2015, pp. 234–241. PREPRINT 1 Be the Change You Want to See: Revisiting Remote Sensing Change Detection Practices Supplementary Material S1. A DDITIONAL DATASET INFORMATION In this section, we provide additional details about the datasets used in our experiments. Tec...

    2015

  62. [71]

    Implementation Details: We use pre-trained weights from Huggingface 2 in our experiments. The corresponding Swin-T weights for each dataset-task pair are as follows: • ImageNet1k - classification: microsoft/swin-tiny-patch4-window7-224 • EuroSat - RS classification: nielsr/swi...

  63. [72]

    Here, we fur- ther analyize different segmentation tasks – semantic, instance, and panoptic – available for CityScapes [58] and COCO [59]

    Pre-training Task: We report only the best performing dataset-task pairs in Section IV-A of main paper. Here, we fur- ther analyize different segmentation tasks – semantic, instance, and panoptic – available for CityScapes [58] and COCO [59]. The results are summarized in Tabl...

  64. [73]

    Remote Sensing Foundation Models: We use the official code and pre-trained weights provided by authors for all remote sensing foundation models. The specific versions of the code used in our experiments can be accessed through the following GitHub links: • SeCo [35]: https://g...

  65. [74]

    Change Detection Specific Methods: We use official code, provided by authors, for all methods and just integrate our datasets. The following are links to used versions of their code on GitHub: • FCS-Diff [24]: https://github.com/rcdaudt/fully convolu tional change detection/co...

  66. [75]

    GFLOPs are measured using the official PyTorch profiler 5

    Implementation Details: We measure 3 different com- putational efficiency metrics: parameter count, inference time (also expressed as frames per second - FPS) and GFLOPs. GFLOPs are measured using the official PyTorch profiler 5. For inference time measurement, we use a pair o...

  67. [76]

    The FPS is calculated from inference time with the equation: F P S= 1second/inference time

    Additional Results: We report the extended results in Table S7. The FPS is calculated from inference time with the equation: F P S= 1second/inference time. The results reveal that GFLOPs (billion floating-point op- erations per second) do not correlate well with inference time...

  68. [2024]

    Available: https://openreview.net/forum?id=tEYskw1V Y2

    [Online]. Available: https://openreview.net/forum?id=tEYskw1V Y2

Pith tools

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