Pith. sign in

REVIEW 3 major objections 4 minor 59 references

Compress-Align-Detect: onboard change detection from unregistered images

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 One end-to-end network can compress, align, and detect changes onboard a satellite, keeping F1 near the uncompressed bound at roughly 3 bits per pixel and running at about 685K pixels per second on a 15-watt embedded board.

desk verdict A genuinely new invariant change-detection architecture with credible onboard efficiency, but the central onboard claim rests on synthetic homography warps rather than real non-orthorectified revisits. read the letter →

arxiv 2507.15578 v1 pith:FC4LWAAP submitted 2025-07-21 cs.CV eess.IV

Gabriele Inzerillo , Diego Valsesia , Aniello Fiengo , Enrico Magli This is my paper
classification cs.CVeess.IV
keywords onboardchangedetectionimagecompressionco-registrationtemporalinvariancehomographyestimationlightweightneuralnetworksremotesensingformachines
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 change detection on satellites need not wait for ground processing: a single end-to-end neural network can compress, align, and compare image pairs entirely onboard, cutting latency from hours or days to near real time. The framework, CAD, couples a learned compressor with a lightweight homography regressor (Light2Reg) and a change detector (TieCD) that is mathematically invariant to the order in which the two images are presented. On the LevirCD benchmark with synthetically misaligned pairs, CAD reaches an F1 score near 70%, close to the uncompressed reference, at about 3 bits per pixel, and runs at roughly 685K pixels per second on a 15W Jetson Orin Nano. The authors position this as the first combined compression-registration-detection pipeline designed under onboard constraints.

What carries the argument

The load-bearing objects are three: the Scale Hyperprior compression module, chosen for low FLOPs rather than peak rate–distortion; Light2Reg, a coarse-to-fine cascade of three homography regressors that estimates a projective transform at quarter, half, and full resolution; and TieCD, a U-Net-style detector built from Temporally-Equivariant Downsample and Upsample Blocks with a Temporally-Equivariant Self-Attention (TESA) that mixes the two images as a length-2 sequence, followed by a Temporal Fusion Gate (TFG) that combines average and absolute-difference features so the output is invariant to input order. The framework operates in learned feature space after the first homography is computed, warping compressed-decoder features rather than pixel images, which is what keeps the inference cost low.

What would settle it

A concrete test is to run the trained CAD pipeline on real raw, non-orthorectified bitemporal satellite image pairs with ground-truth change labels and compare F1 against the ~70% reported under simulated homographies; if real-data F1 falls well below that, or if any revisit pair exhibits local misalignment that no single global warp can remove, the central claim would be falsified.

Watch

Extended reading notes

Core claim

The central discovery is that the three traditionally separate tasks—storage-efficient compression, geometric alignment, and change detection—can be optimized jointly in one small network, and that optimizing them together is what makes the pipeline usable. The paper shows that independently trained modules fail (F1 around 56 in the solo registration configuration) while joint training brings unregistered change detection to about 70 F1, nearly matching a much heavier SuperPoint+LightGlue matcher. It also establishes a design principle for temporal symmetry: instead of training a change detector to tolerate swapped inputs, TieCD's architecture guarantees identical outputs for either ordering by construction, using permutation-equivariant layers and an invariant fusion gate.

Load-bearing premise

The load-bearing premise is that the geometric relationship between two revisits is a projective homography, which the paper states explicitly in Section III-C and approximates in all training and evaluation by applying random affine and perspective transforms to one image of each pair; should real misregistration involve parallax, terrain-dependent distortion, or sensor artifacts beyond a single global homography, the registration module and the entire CAD pipeline may fail on actual onboard imagery.

Editorial extensions

If this is right

  • Satellites could downlink change maps instead of image volumes, reducing bandwidth and latency for disaster monitoring and near-real-time surveillance.
  • Joint end-to-end optimization matters more than any single module's sophistication: the paper's ablation shows that independently trained modules drop F1 far below the jointly trained pipeline, so future onboard designs should train the whole chain together.
  • Temporal invariance is achievable by architecture rather than data augmentation, so reversing the input order produces identical output and removes a failure mode that degrades many change detection models to unusable levels.
  • Running change detection in compressed feature space, rather than full-resolution pixel space, keeps onboard memory and compute low enough for a 15-watt accelerator.

Reading between the lines

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

  • Because the paper simulates unregistered revisits with random affine and perspective transforms, the immediate next test is to acquire or generate real non-orthorectified revisit pairs with terrain-dependent distortion; the authors themselves call for such datasets, and the homography assumption's validity on real imagery is the main open risk.
  • TieCD's pattern of permutation-equivariant encoders followed by an invariant fusion gate could transfer to other pairwise vision tasks, such as stereo matching or video frame differencing, wherever order symmetry is a desirable prior.
  • The rate–F1 curve degrades smoothly down to about 0.1 bpp, so a storage-limited mission could plausibly operate below 1 bpp with only a modest F1 loss; this is an extrapolation beyond the paper's reported operating points and deserves direct measurement before being relied upon.
  • The chosen compression backbone underperforms ELIC on rate–distortion, so the practical advantage of CAD comes from task-aware end-to-end training; a lighter or better compressor that is also trainable end-to-end could shift the whole rate–F1 curve further.
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 proposes CAD (Compress-Align-Detect), an end-to-end neural network for onboard satellite change detection that integrates image compression, a lightweight co-registration module (Light2Reg), and a temporally invariant change detection network (TieCD). The compression module stores compact latent representations; Light2Reg regresses a coarse-to-fine homography to warp one image onto the other; TieCD produces a change map from the aligned features and is architecturally guaranteed to be invariant to the order of the input pair. The method is evaluated on LevirCD with synthetic affine and perspective distortions applied to one image of each pair. The main results show that CAD achieves an F1 score close to the uncompressed upper bound at approximately 3 bpp (about 70% on LevirCD) and sustains about 685K pixels/s on a 15W Jetson Orin Nano. Additional experiments compare the standalone compression, registration, and change detection modules with state-of-the-art baselines and include ablations demonstrating the necessity of joint training.

Significance. If the results transfer to real onboard data, this is a meaningful contribution: it is the first framework to jointly address compression, co-registration, and change detection in a single end-to-end model with a strict low-power constraint. The paper provides a mathematically explicit construction of temporal invariance for TieCD, which is a stronger property than the soft invariance obtained by data augmentation in prior work. The experimental methodology is generally sound, with error bars over 100 random distortion runs, comparisons against external baselines (LightGlue, OpenCD models, ELIC, Factorized Prior), and a clear ablation showing that independent module training fails (Fig. 10). The main weakness is external validity: all unregistered-image experiments use synthetic homography-distorted pairs, so the applicability to real non-orthorectified satellite imagery with parallax, terrain-induced local displacement, or sensor-specific distortion remains unproven.

major comments (3)
  1. [III-C, IV-B, V] The central claim of the paper is that CAD performs change detection from unregistered images, but the only evidence for the unregistered condition is synthetic random affine and perspective warps applied to one image of each pair. Real satellite revisits can exhibit parallax, terrain-dependent local displacements, and sensor-specific distortions that a single global homography cannot represent. Because both training and evaluation (Fig. 7, Table III) are confined to homography-correctable warps, the reported F1 scores may not transfer to actual non-orthorectified onboard imagery. The authors acknowledge this limitation and call for public datasets, but the abstract and title should be qualified accordingly, and a sensitivity analysis with non-homography deformations (e.g., random thin-plate spline or local warps) would substantially strengthen the external validity claim.
  2. [I, III-D, IV-E] Temporal invariance is demonstrated only for the standalone TieCD module (Table II). In the full CAD pipeline, Light2Reg warps the second image onto the first via an asymmetric homography, so swapping the input order changes the warping direction and the actual feature maps that TieCD receives. Consequently, the end-to-end CAD system is not architecturally invariant to temporal order, despite TieCD's mathematical property. The paper should either evaluate the full CAD framework under swapped inputs or clearly state that the onboard pipeline assumes a fixed chronological ordering, which is a limitation of the system-level claim.
  3. [III-E, Eq. (10)] Equation (10) defines w(s)_t2 = Warp(x(s/2)_t2; H(s)), but the registration losses in Eqs. (11)-(13) compare w(s)_t2 with x(s)_t1 at the same scale s. For s = 1, this would warp a half-resolution image and compare it to a full-resolution reference, which is dimensionally inconsistent. This appears to be a typographical error; it should likely be x(s)_t2. As written, the loss definition is not self-consistent and would prevent exact reproduction of the training procedure.
minor comments (4)
  1. [IV-F] The comparison with LightGlue uses SuperPoint for keypoint extraction, which was trained on different data and is not optimized for remote sensing; the paper should acknowledge this potential source of bias in the registration comparison.
  2. [IV-E, Table II] The statement that TieCD is the only method whose predictions are exactly the same under swapped inputs is correct by construction, but ChangerEx and ChangeStar achieve near-invariance with differences of 0.02 and 0.23 F1 points, respectively; a brief discussion of the practical benefit of hard versus soft invariance would be useful.
  3. [IV-B, IV-H] The paper reports 121.47 GFLOPs for CAD with TieCD-L in Table I and Table V, while standalone TieCD-L is 58.98 GFLOPs in Table II; it would be helpful to explain how the feature-space CD FLOPs differ from the standalone pixel-space CD FLOPs.
  4. [III-C] The coarse-to-fine registration description is somewhat ambiguous about the scales of the feature maps used at each cascade stage; clarifying the resolution of the tensors that receive H(1/4), H(1/2), and H(1) would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework's claims rest on held-out evaluation, external baselines, and explicit architectural construction rather than on fitted inputs or load-bearing self-citation.

full rationale

The paper's central claims are (1) CAD achieves change detection on compressed, unregistered image pairs with F1 close to the uncompressed upper bound, (2) TieCD is temporally invariant by construction, and (3) the whole pipeline runs efficiently on low-power hardware. None of these reduce to its inputs by definition. The temporal-invariance claim is a mathematical construction: TieCD uses permutation-equivariant layers (TEDB, TEUB, TESA) followed by the Temporal Fusion Gate, whose equations use only symmetric statistics—average and absolute difference—so swapped inputs provably give identical outputs. This is a construction, not a prediction fitted from data. The end-to-end F1-versus-rate results are evaluated on a held-out LevirCD test set with 100 random distortion runs and are compared against external baselines (LightGlue, OpenCD models, ELIC, Factorized Prior), so the performance numbers are not fitted into existence. The registration module is trained to regress homographies and tested on the same family of synthetic projective distortions, but the paper explicitly acknowledges this limitation: 'our experiments rely on simulations based on random projective transformations ... this synthetic approach can never fully reproduce real onboard data.' That is an external-validity gap, not circular reasoning. The only self-citation, reference [17], is a preliminary IGARSS version of the same work and is used purely as a prior-version pointer, not as evidence for any claim. The authors' previous works cited for context (e.g., onboard lossless compression) are not load-bearing for the CAD results. Overall, the derivation chain is self-contained and empirically grounded, so no circular step is present.

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

The central rate-F1 trade-off depends on four hand-chosen training parameters (lambda, alpha, latent widths, registration weights), and the real-world claim depends on the unverified assumption that homography-plus-synthetic-warp models real unregistered revisits. No new physical entities are introduced.

free parameters (4)
  • Rate-distortion trade-off lambda (compression) = Multiple values; produces rates from about 0.05 to about 3 bpp in Fig. 7
    Each operating point of the claimed rate-F1 curve is a different model trained with a hand-chosen lambda; the curve is the central result.
  • Loss-balancing weight alpha = 0.3 for Reg-CD pretraining, 0.5 for end-to-end finetuning (Sec. IV-B)
    Hand-chosen to weight CD loss versus registration loss in Eqs. (15)-(16); affects the trained models.
  • Compression latent channels = 150 main, 225 hyperprior
    Custom Scale Hyperprior widths chosen by hand to balance complexity and rate-distortion (Sec. IV-B).
  • Registration loss scale weights alpha_1, alpha_2, alpha_3 = Not reported
    Weights in Eqs. (11)-(13); the paper does not give their numerical values, yet they define the registration pretraining objective.
assumptions (4)
  • domain assumption Geometric disparities between revisits are a projective (homography) transformation
    Stated in Sec. III-C: 'We assume that the geometric relationships between images acquired at multiple revisits can be approximated by a projective transformation.' Real non-orthorectified imagery can have parallax and local terrain distortions that a single homography cannot capture.
  • ad hoc to paper Random affine and perspective warps simulate real unregistered satellite revisits
    Sec. III-C and IV-B: no public unregistered CD dataset exists, so all training and evaluation uses synthetic torchvision transformations; the paper acknowledges this cannot fully reproduce real data.
  • domain assumption The entropy-model cross-entropy R equals the storage bitrate actually needed
    Eqs. (7)-(8) use negative log-likelihood as rate; no arithmetic coder or actual bitstream is implemented, so the bpp values in Fig. 7 are estimates.
  • domain assumption Compression decoder penultimate-layer features are sufficient for registration warping and change detection
    Sec. III-A and III-D: the framework warps feature maps z_t2 and feeds them to TieCD; this design choice assumes the learned features preserve the geometric and semantic information needed for CD.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Compress-Align-Detect: onboard change detection from unregistered images." pith.science (2026). https://pith.science/paper/FC4LWAAP

@misc{pith2026250715578,
  author       = {Pith},
  title        = {Pith review of: Compress-Align-Detect: onboard change detection from unregistered images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FC4LWAAP}},
  note         = {Machine review of arXiv:2507.15578}
}
read the original abstract

Change detection from satellite images typically incurs a delay ranging from several hours up to days because of latency in downlinking the acquired images and generating orthorectified image products at the ground stations; this may preclude real- or near real-time applications. To overcome this limitation, we propose shifting the entire change detection workflow onboard satellites. This requires to simultaneously solve challenges in data storage, image registration and change detection with a strict complexity constraint. In this paper, we present a novel and efficient framework for onboard change detection that addresses the aforementioned challenges in an end-to-end fashion with a deep neural network composed of three interlinked submodules: (1) image compression, tailored to minimize onboard data storage resources; (2) lightweight co-registration of non-orthorectified multi-temporal image pairs; and (3) a novel temporally-invariant and computationally efficient change detection model. This is the first approach in the literature combining all these tasks in a single end-to-end framework with the constraints dictated by onboard processing. Experimental results compare each submodule with the current state-of-the-art, and evaluate the performance of the overall integrated system in realistic setting on low-power hardware. Compelling change detection results are obtained in terms of F1 score as a function of compression rate, sustaining a throughput of 0.7 Mpixel/s on a 15W accelerator.

Figures

Figures reproduced from arXiv: 2507.15578 by the authors.

Figure 1
Figure 1. The proposed CAD framework takes as input a non-co-registered image pair, 1 Sof Dro [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Design of the homography regressor in Light2Reg. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Design of the efficient encoder block employed in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Unregistered transformed image pair. In addition to the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Left: designs of the Temporally-Equivariant Downsample Block (TEDB) and Temporally-Equivariant Upsample Block (TEUB), used as main building blocks of TieCD. Right: design of Temporally-Invariant and Efficient Change Detector (TieCD). By concatenating the input images o…
Figure 6
Figure 6. Figure 6: Design of the temporally-equivariant self-attention [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Trade-off between compression rate and F1-Score. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Qualitative co-registration results on two misaligned LevirCD image pairs. For each pair, the [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Rate-distortion curves of compression models trained [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 11
Figure 11. Figure 11: Rate–F1 curves comparing CAD with initial module [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 47 canonical work pages

  1. [1]

    The phi-sat-1 mis- sion: The first on-board deep neural network demonstrator for satellite earth observation,

    G. Giuffrida, L. Fanucci, G. Meoni, M. Bati ˇc, L. Buckley, A. Dunne, C. van Dijk, M. Esposito, J. Hefele, N. Vercruyssen, G. Furano, M. Pastena, and J. Aschbacher, “The phi-sat-1 mis- sion: The first on-board deep neural network demonstrator for satellite earth observation,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–14, 2022

    work page 2022

  2. [2]

    Unlocking the use of raw multispectral earth observation imagery for onboard artificial intelligence,

    G. Meoni, R. Del Prete, F. Serva, A. De Beusscher, O. Colin, and N. Long ´ep´e, “Unlocking the use of raw multispectral earth observation imagery for onboard artificial intelligence,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2024

    work page 2024

  3. [3]

    Phi-sat-2: Onboard ai apps for earth observation,

    A. Marin, C. Coelho, F. Deconinck, I. Babkina, N. Longepe, and M. Pastena, “Phi-sat-2: Onboard ai apps for earth observation,” Proc. Space Artif. Intell , vol. 6, 2021

    work page 2021

  4. [4]

    Benchmarking deep learning for on-board space applications,

    M. Ziaja, P. Bosowski, M. Myller, G. Gajoch, M. Gumiela, J. Protich, K. Borda, D. Jayaraman, R. Dividino, and J. Nalepa, “Benchmarking deep learning for on-board space applications,” Remote Sensing, vol. 13, no. 19, p. 3981, 2021

    work page 2021

  5. [5]

    On-board ship detection in micro-nano satellite based on deep learning and cots component,

    Y . Yao, Z. Jiang, H. Zhang, and Y . Zhou, “On-board ship detection in micro-nano satellite based on deep learning and cots component,” Remote Sensing, vol. 11, no. 7, p. 762, 2019

    work page 2019

  6. [6]

    Efficient onboard multi- task ai architecture based on self-supervised learning,

    G. Inzerillo, D. Valsesia, and E. Magli, “Efficient onboard multi- task ai architecture based on self-supervised learning,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2024

    work page 2024

  7. [7]

    Ravaen: unsupervised change detection of extreme events using ml on-board satellites,

    V . Ruvzivcka, A. Vaughan, D. De Martini, J. Fulton, V . Sal- vatelli, C. Bridges, G. Mateo-Garcia, and V . Zantedeschi, “Ravaen: unsupervised change detection of extreme events using ml on-board satellites,” Scientific reports, vol. 12, no. 1, p. 16939, 2022

    work page 2022

  8. [8]

    Lightcdnet: Lightweight change detection network based on vhr images,

    Y . Xing, J. Jiang, J. Xiang, E. Yan, Y . Song, and D. Mo, “Lightcdnet: Lightweight change detection network based on vhr images,” IEEE Geoscience and Remote Sensing Letters , vol. 20, pp. 1–5, 2023

    work page 2023

Show all 59 references
  1. [9]

    Lightweight remote sensing change detection with progressive feature 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 feature aggregation and supervised attention,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–12, 2023

    2023

  2. [10]

    Tinycd: A (not so) deep learning model for change detection,

    A. Codegoni, G. Lombardi, and A. Ferrari, “Tinycd: A (not so) deep learning model for change detection,” Neural Computing and Applications, vol. 35, no. 11, pp. 8471–8486, 2023

    2023

  3. [11]

    End-to-end remote sensing change detection of unregistered bi-temporal images for natural disasters,

    G. Zhao, L. Shan, and W. Wang, “End-to-end remote sensing change detection of unregistered bi-temporal images for natural disasters,” in International Conference on Artificial Neural Networks. Springer, 2023, pp. 259–270

    2023

  4. [12]

    Changerd: A registration- integrated change detection framework for unaligned remote sensing images,

    W. Jing, K. Chi, Q. Li, and Q. Wang, “Changerd: A registration- integrated change detection framework for unaligned remote sensing images,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 220, pp. 64–74, 2025

    2025

  5. [13]

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

    Z. Zheng, A. Ma, L. Zhang, and Y . Zhong, “Change is ev- erywhere: 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

  6. [14]

    Change- mask: Deep multi-task encoder-transformer-decoder architec- ture for semantic change detection,

    Z. Zheng, Y . Zhong, S. Tian, A. Ma, and L. Zhang, “Change- mask: Deep multi-task encoder-transformer-decoder architec- ture for semantic change detection,” ISPRS Journal of Pho- togrammetry and Remote Sensing, vol. 183, pp. 228–239, 2022

    2022

  7. [15]

    Segment any change,

    Z. Zheng, Y . Zhong, L. Zhang, and S. Ermon, “Segment any change,” arXiv preprint arXiv:2402.01188 , 2024

    2024 arXiv

  8. [16]

    Changer: Feature interaction is what you need for change detection,

    S. Fang, K. Li, and Z. Li, “Changer: Feature interaction is what you need for change detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–11, 2023

    2023

  9. [17]

    Towards storage-aware onboard change detection,

    G. Inzerillo, D. Valsesia, E. Magli, and A. Fiengo, “Towards storage-aware onboard change detection,” inIGARSS 2025-2025 IEEE International Geoscience and Remote Sensing Sympo- sium. IEEE, 2025

    2025

  10. [18]

    A tutorial on image compression for optical space imaging systems,

    I. Blanes, E. Magli, and J. Serra-Sagrista, “A tutorial on image compression for optical space imaging systems,” IEEE Geoscience and Remote Sensing Magazine , vol. 2, no. 3, pp. 8–26, 2014

    2014

  11. [19]

    The ccsds 123.0- b-2 “low-complexity lossless and near-lossless multispectral and hyperspectral image compression

    M. Hern ´andez-Cabronero, A. B. Kiely, M. Klimesh, I. Blanes, J. Ligo, E. Magli, and J. Serra-Sagrist `a, “The ccsds 123.0- b-2 “low-complexity lossless and near-lossless multispectral and hyperspectral image compression” standard: A comprehen- sive review,” IEEE Geoscience an...

    2021

  12. [20]

    End-to-end opti- mized image compression,

    J. Ball ´e, V . Laparra, and E. P. Simoncelli, “End-to-end opti- mized image compression,” arXiv preprint arXiv:1611.01704 , 2016

    2016 arXiv

  13. [21]

    Variational image compression with a scale hyperprior,

    J. Ball ´e, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Variational image compression with a scale hyperprior,” arXiv preprint arXiv:1802.01436, 2018

    2018 arXiv

  14. [22]

    Learned image compression with discretized gaussian mixture likelihoods and attention modules,

    Z. Cheng, H. Sun, M. Takeuchi, and J. Katto, “Learned image compression with discretized gaussian mixture likelihoods and attention modules,” in Proceedings of the IEEE/CVF conference 14 on computer vision and pattern recognition , 2020, pp. 7939– 7948

    2020

  15. [23]

    Overview of the versatile video coding (vvc) standard and its applications,

    B. Bross, Y .-K. Wang, Y . Ye, S. Liu, J. Chen, G. J. Sullivan, and J.-R. Ohm, “Overview of the versatile video coding (vvc) standard and its applications,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 31, no. 10, pp. 3736– 3764, 2021

    2021

  16. [24]

    Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding,

    D. He, Z. Yang, W. Peng, R. Ma, H. Qin, and Y . Wang, “Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5718–5727

    2022

  17. [25]

    Raw sar data compression with deep learning,

    G. Pilikos, M. Azcueta, R. Camarero, and N. Floury, “Raw sar data compression with deep learning,” in IGARSS 2024- 2024 IEEE International Geoscience and Remote Sensing Sym- posium. IEEE, 2024, pp. 2546–2549

    2024

  18. [26]

    Onboard deep lossless and near-lossless predictive coding of hyperspectral images with line-based attention,

    D. Valsesia, T. Bianchi, and E. Magli, “Onboard deep lossless and near-lossless predictive coding of hyperspectral images with line-based attention,” IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  19. [27]

    Object recognition from local scale-invariant fea- tures,

    D. Lowe, “Object recognition from local scale-invariant fea- tures,” in Proceedings of the Seventh IEEE International Con- ference on Computer Vision, vol. 2, 1999, pp. 1150–1157 vol.2

    1999

  20. [28]

    Superglue: Learning feature matching with graph neural net- works,

    P.-E. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superglue: Learning feature matching with graph neural net- works,” in Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , 2020, pp. 4938–4947

    2020

  21. [29]

    Lightglue: Local feature matching at light speed,

    P. Lindenberger, P.-E. Sarlin, and M. Pollefeys, “Lightglue: Local feature matching at light speed,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 17 627–17 638

    2023

  22. [30]

    A multiscale framework with unsupervised learning for remote sensing image registration,

    Y . Ye, T. Tang, B. Zhu, C. Yang, B. Li, and S. Hao, “A multiscale framework with unsupervised learning for remote sensing image registration,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–15, 2022

    2022

  23. [31]

    A unified deep learning network for remote sensing image registration and change detection,

    R. Zhou, D. Quan, S. Wang, C. Lv, X. Cao, J. Chanussot, Y . Li, and L. Jiao, “A unified deep learning network for remote sensing image registration and change detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–16, 2024

    2024

  24. [32]

    Digital change detection in forest ecosystems with remote sensing imagery,

    P. R. Coppin and M. E. Bauer, “Digital change detection in forest ecosystems with remote sensing imagery,” Remote sensing reviews, vol. 13, no. 3-4, pp. 207–234, 1996

    1996

  25. [33]

    Change vector analysis: A technique for the multispectral monitoring of land cover and condition,

    R. D. Johnson and E. Kasischke, “Change vector analysis: A technique for the multispectral monitoring of land cover and condition,” International journal of remote sensing , vol. 19, no. 3, pp. 411–426, 1998

    1998

  26. [34]

    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

  27. [35]

    Object-based classification of remote sensing data for change detection,

    V . Walter, “Object-based classification of remote sensing data for change detection,” ISPRS Journal of photogrammetry and remote sensing, vol. 58, no. 3-4, pp. 225–238, 2004

    2004

  28. [36]

    Forest change de- tection by statistical object-based method,

    B. Descl ´ee, P. Bogaert, and P. Defourny, “Forest change de- tection by statistical object-based method,” Remote sensing of environment, vol. 102, no. 1-2, pp. 1–11, 2006

    2006

  29. [37]

    Use of a dark object concept and support vector machines to automate forest cover change analysis,

    C. Huang, K. Song, S. Kim, J. R. Townshend, P. Davis, J. G. Masek, and S. N. Goward, “Use of a dark object concept and support vector machines to automate forest cover change analysis,” Remote sensing of environment , vol. 112, no. 3, pp. 970–985, 2008

    2008

  30. [38]

    A change detection model based on neighborhood correlation image analysis and decision tree classification,

    J. Im and J. R. Jensen, “A change detection model based on neighborhood correlation image analysis and decision tree classification,” Remote Sensing of Environment , vol. 99, no. 3, pp. 326–340, 2005

    2005

  31. [39]

    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 conference on image processing (ICIP) . IEEE, 2018, pp. 4063–4067

    2018

  32. [40]

    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

  33. [41]

    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 IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Sympo- sium. IEEE, 2022, pp. 207–210

    2022

  34. [42]

    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

  35. [43]

    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

  36. [44]

    Superpoint: Self-supervised interest point detection and description,

    D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superpoint: Self-supervised interest point detection and description,” in Proceedings of the IEEE conference on computer vision and pattern recognition workshops , 2018, pp. 224–236

    2018

  37. [45]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7132–7141

    2018

  38. [46]

    Advances and challenges in multimodal remote sensing image registration,

    B. Zhu, L. Zhou, S. Pu, J. Fan, and Y . Ye, “Advances and challenges in multimodal remote sensing image registration,” IEEE Journal on Miniaturization for Air and Space Systems , vol. 4, no. 2, pp. 165–174, 2023

    2023

  39. [47]

    Fast and robust matching for multimodal remote sensing image registration,

    Y . Ye, L. Bruzzone, J. Shan, F. Bovolo, and Q. Zhu, “Fast and robust matching for multimodal remote sensing image registration,” IEEE Transactions on Geoscience and Remote Sensing, vol. 57, no. 11, pp. 9059–9070, 2019

    2019

  40. [48]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 . S...

    2015

  41. [49]

    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

  42. [50]

    Mish: A self regularized non-monotonic activation function,

    D. Misra, “Mish: A self regularized non-monotonic activation function,” arXiv preprint arXiv:1908.08681 , 2019

    1908 arXiv

  43. [51]

    Scalable image coding for humans and machines,

    H. Choi and I. V . Baji´c, “Scalable image coding for humans and machines,” IEEE Transactions on Image Processing, vol. 31, pp. 2739–2754, 2022

    2022

  44. [52]

    X. Ye. (2023) Calflops: a flops and params calculate tool for neural networks in pytorch framework. [Online]. Available: https://github.com/MrYxJ/calculate-flops.pytorch

    2023

  45. [53]

    Cloudsen12, a global dataset for semantic understanding of cloud and cloud shadow in sentinel-2,

    E. Aybar, L. Ysuhuaylas, J. Loja, K. Gonzales, F. Herrera, L. Bautista, R. Yali, A. Flores, L. Diaz, N. Cuenca et al. , “Cloudsen12, a global dataset for semantic understanding of cloud and cloud shadow in sentinel-2,” Scientific Data, vol. 9, no. 1, p. 782, 2022

    2022

  46. [54]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” CoRR, vol. abs/1412.6980, 2014. [Online]. Available: https://api.semanticscholar.org/CorpusID:6628106

    2014 arXiv

  47. [55]

    Com- pressai: a pytorch library and evaluation platform for end-to-end compression research,

    J. B ´egaint, F. Racap ´e, S. Feltman, and A. Pushparaja, “Com- pressai: a pytorch library and evaluation platform for end-to-end compression research,” arXiv preprint arXiv:2011.03029, 2020

    2011 arXiv

  48. [56]

    Open-CD: A comprehensive toolbox for change detection,

    K. Li, J. Jiang, A. Codegoni, C. Han, Y . Deng, K. Chen, Z. Zheng, H. Chen, Z. Liu, Y . Gu, Z. Zou, Z. Shi, S. Fang, D. Meng, Z. Wang, and X. Cao, “Open-CD: A comprehensive toolbox for change detection,”arXiv preprint arXiv:2407.15317, 2024

    2024 arXiv

  49. [57]

    Hanet: A hierarchical attention network for change detection with bitemporal very-high-resolution remote sensing images,

    C. Han, C. Wu, H. Guo, M. Hu, and H. Chen, “Hanet: A hierarchical attention network for change detection with bitemporal very-high-resolution remote sensing images,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 16, pp. 3867–3878, 2023. 15

    2023

  50. [58]

    Time travelling pixels: Bitemporal features integra- tion with foundation model for remote sensing image change detection,

    K. Chen, C. Liu, W. Li, Z. Liu, H. Chen, H. Zhang, Z. Zou, and Z. Shi, “Time travelling pixels: Bitemporal features integra- tion with foundation model for remote sensing image change detection,” arXiv preprint arXiv:2312.16202 , 2023

    2023 arXiv

  51. [59]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo et al. , “Segment anything,” in Proceedings of the IEEE/CVF international conference on computer vision , 2023, pp. 4015– 4026. Gabriele Inzerillo received the M.Sc....

    2023

Pith tools

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