Semantic-CD: Remote Sensing Image Semantic Change Detection towards Open-vocabulary Setting

Chenyang Liu; Fugen Zhou; Keyan Chen; Lu Li; Yongshuo Zhu; Zhenwei Shi

REVIEW 3 major objections 5 minor 2 cited by

Semantic-CD: Remote Sensing Image Semantic Change Detection towards Open-vocabulary Setting

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Semantic-CD injects CLIP category knowledge into bi-temporal satellite images and names the changes.

desk verdict Strong closed-set SCD numbers with a sensible CLIP-adapter architecture, but the open-vocabulary claim is asserted, not measured—all six classes are in training and test. read the letter →

arxiv 2501.06808 v1 pith:PDJPOCC4 submitted 2025-01-12 cs.CV

Yongshuo Zhu , Lu Li , Keyan Chen , Chenyang Liu , Fugen Zhou , Zhenwei Shi This is my paper · ORCID

classification cs.CV

keywords remotesensingsemanticchangedetectionopen-vocabularyCLIPvision-languagefoundationmodelbi-temporalimagesmulti-tasklearningSECONDdataset

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

Semantic-CD asks a vision-language model to do more than detect that a scene changed: it tries to name the change, telling an analyst whether a building appeared, water shifted, or vegetation regrew. The paper's core proposal is to inject CLIP text embeddings into a bi-temporal change-detection network through an instance-level open semantic prompter, and to train binary change detection and semantic change detection as two fully decoupled stages. On the SECOND benchmark the method reports the best scores across all four metrics (OA 91.31, F1 56.11, mIoU 75.10, SeK 23.85), with qualitative results showing sharper masks and fewer semantic mix-ups. The paper frames this as a degraded open-vocabulary setting, meaning all six categories are allowed in training, so what is demonstrated is the value of CLIP semantic priors on known classes, with generalization to unseen categories left as the promised capability.

What carries the argument

The load-bearing component is the open semantic prompter, an instance-level adapter that turns category names into CLIP text embeddings conditioned on the image pair. It combines a meta token derived from the visual features, $\Phi_{\mathrm{meta}}(F_i)$, with task-specific learnable tokens $p^t$, appends the result to each category word embedding $p^j_{\mathrm{text}}$, and encodes the whole string with the frozen CLIP text encoder; cosine similarity between the resulting embedding $t_{i,j}$ and the visual feature map produces the semantic cost volume $F^i_{\mathrm{cv}}$. Around this sit two supporting mechanisms: the BCSF adapter, inserted into the four global attention layers of the CLIP visual encoder to make frozen features sensitive to genuine bi-temporal change, and the fully decoupled two-stage training that first learns the BCD branch with the prompter frozen, then learns the SCD branch with the visual encoder and BCD decoder frozen.

What would settle it

Train Semantic-CD with one or more SECOND categories held out of the SCD stage, then evaluate whether the prompter assigns correct semantics to changes involving the held-out class; if accuracy on that class is near chance or the classification confusion rises sharply relative to the all-categories model, the open-vocabulary generalization claim is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a fully decoupled binary/semantic change-detection training scheme wrapped around a frozen CLIP backbone can outperform dedicated semantic change detection models that are trained end-to-end on the same data. The mechanism is the open semantic prompter: for each image, a meta network reads the visual features and emits instance-level meta tokens, which are summed with shared learnable context tokens, prepended to category words, and passed through the CLIP text encoder; the resulting text embeddings are compared with the visual feature map by cosine similarity to form a cost volume. That volume, fused with visual features in the SCD decoder, supplies categorical semantics to the change mask. The authors report best-in-class numbers on SECOND, and a baseline without the prompter converges far more slowly, which they read as evidence that the prompt-injected semantic priors carry the improvement.

Load-bearing premise

The open-vocabulary claim rests on the assumption that a text prompter trained with all six SECOND categories present will generalize to category words it never saw during training; the paper explicitly degrades the open-vocabulary setting into a closed-set training scheme and reports no held-out category test, so this generalization is assumed rather than measured.

Editorial extensions

If this is right

Applying vision-language semantic priors to semantic change detection yields more accurate change masks and fewer semantic classification errors than the previous best method on SECOND.
Because the prompter maps category words to cost volumes through the CLIP text encoder, the output vocabulary is set by text input at inference time rather than by a fixed classifier head, which is the key open-vocabulary property.
The fully decoupled training scheme lets the SCD branch be trained for only five epochs on top of a frozen BCD branch, sharply reducing training cost relative to the 100-epoch baseline that omits the prompter.
Removing the prompter slows convergence dramatically, indicating that the semantic priors injected through the text encoder are what make the fast and accurate training feasible.

Reading between the lines

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

The paper explicitly degrades the open-vocabulary setting by letting all six SECOND categories participate in training; a direct test it leaves implicit is to hold out one or more categories during SCD training and check whether the prompter still assigns correct semantics to changes involving the held-out class.
Because the cost volume is built by cosine similarity with CLIP text embeddings, the same machinery could be reused for change captioning or for user-specified queries at inference time, where the vocabulary need not be fixed at train time.
The decoupled training schedule suggests an asymmetric strategy for new datasets: invest most of the budget in the BCD branch and the prompter, then adapt the SCD head to a new category set with very few epochs.
If held-out category tests fail, the reported SOTA gains would still support the closed-set claim; the open-vocabulary claim and the benchmark improvement are separable and should be evaluated independently.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague, quick read of Semantic-CD. What you should know: it gets real gains on SECOND (OA 91.31, F1 56.11, mIoU 75.10, SeK 23.85), beating MambaSCD and others, using a frozen RemoteCLIP ViT-L-14 with BCSF adapters plus a learned prompt that fuses text embeddings with visual features. The architecture is coherent and the decoupled BCD/SCD training is a sensible way to keep the CLIP encoder frozen. That part is worth taking seriously—if the numbers reproduce, it is a useful closed-set SCD method.

The soft spot is the title claim. Open vocabulary means generalizing to categories not seen in training. This paper never tests that. Section III-C explicitly says they degraded the setting and let all six SECOND categories participate in training. The prompt learner gets learnable context tokens and instance-conditioned meta tokens, all trained on exactly those six class names. So the model could be memorizing the six categories rather than acquiring a compositional text-conditioned procedure. The qualitative figure is not evidence of open-vocabulary behavior. A held-out category test—train on five classes, evaluate on the sixth—is the missing experiment, and it is the experiment the method needs to carry its name.

There are two secondary issues. First, the ablation baseline is trained for 100 SCD epochs versus 5 for Semantic-CD; the paper explains the baseline needed more epochs to converge, but that confound plus no seed variance means the margin over the baseline is not robust. Second, the setup is under-specified: no optimizer, LR, train/val/test split, or number of runs. Given the paper compares against published numbers, that is sloppy and makes verification harder. Code would help.

The citation pattern is fine—BCSF comes from the authors' prior work, but that is a legitimate component, not circular, and they cite it.

Bottom line: this is a plausible closed-set SCD improvement with a misleading open-vocabulary framing. I would send it to a serious referee, because the architecture and numbers deserve scrutiny, but I would expect major revision: a proper held-out category evaluation, a controlled baseline, and full training details. If the open-vocabulary test fails, the paper should be repositioned as closed-set SCD with CLIP priors. I would not cite it in my own work until that test is done.

Referee Report

3 major / 5 minor

Summary. The paper proposes Semantic-CD, a semantic change detection method for bi-temporal remote sensing images that combines a frozen CLIP (RemoteCLIP ViT-L-14) vision encoder with a BCSF change-semantic adapter, a prompt learner that maps category words and learned context tokens to text embeddings, and two decoupled decoders for binary change detection (BCD) and semantic change detection (SCD). The BCD and SCD branches are trained in separate stages, and the SCD branch fuses CLIP visual features with text-derived cost volumes to produce per-class change masks. Experiments on the SECOND dataset report OA 91.31, F1 56.11, mIoU 75.10, and SeK 23.85, exceeding the compared methods, and qualitative results show cleaner masks. The paper frames the method as an open-vocabulary approach, while Section III-C concedes that the evaluation is a degraded setting in which all six dataset categories participate in training.

Significance. If the open-vocabulary claim were supported, the work would be significant: it would show that frozen CLIP text priors plus a lightweight adapter can outperform fully trained SCD models and generalize to categories not seen during training. The architecture is reasonable, the decoupled training idea is clearly presented, and the paper is honest about the degraded setting in Section III-C. However, the current evaluation cannot substantiate the central claim: all experiments use the same six SECOND categories in training and test, and the comparison to the baseline and to prior methods is confounded by training schedules and omitted hyperparameters. The reported closed-set numbers are useful, but they do not by themselves demonstrate open-vocabulary semantic change detection.

major comments (3)

[Section III-C and Eq. (2)] The central open-vocabulary claim is asserted but never measured. Section III-C explicitly degrades the setting by allowing all six SECOND categories to participate in training, and Eq. (2) learns the context tokens p_t and image-dependent meta tokens p_m on exactly those categories; no held-out category is evaluated. The gains in Table I are therefore compatible with a closed-set SCD model that uses CLIP features. Please either add an evaluation with category words and labels not seen during training (e.g., held-out SECOND classes or a cross-dataset/zero-shot transfer test), or revise the title, abstract, and conclusion to claim only closed-set SCD with open-vocabulary potential.
[Section III-A and Table I] The state-of-the-art comparison is confounded by the training protocol. The baseline without the prompter is trained for 100 SCD epochs, while Semantic-CD is trained for 5 epochs, and the paper does not report the optimizer, learning rate, schedule, data split, or number of runs; the margin over MambaSCD (0.95 OA, 0.79 F1, 1.42 mIoU, 0.93 SeK) may be within run-to-run variation. Please report a matched training budget for the baseline, full implementation details, and mean and standard deviation over at least three seeds, or otherwise justify that the observed margins are statistically meaningful.
[Section II-B and Table I] The contribution of the BCSF adapter is not isolated. The full model differs from the baseline in several simultaneous ways: presence of the prompt learner, context token length, and SCD training schedule. No ablation fixes all other factors and toggles only the BCSF adapter, so the claim that BCSF, originally designed for SAM in prior work [20], transfers to CLIP and contributes to the reported gains is not supported by the presented experiments.

minor comments (5)

[Section III-A.1] The word 'evaluatd' should be corrected to 'evaluated', and the method name is spelled inconsistently as both 'MamabaSCD' and 'MambaSCD'.
[Equation (5)] The output of the SCD decoder is written as M^{1/2}_{BCD}, but it should be M^{1/2}_{SCD}; the typo makes Eq. (6) harder to follow.
[Figure 3] The legend entries 'N.v.g. surface' and 'Buliding' are unclear or misspelled; please spell out the land-cover classes and fix the typo.
[Section III-A.2] The implementation details give batch size and epochs but omit the image size, optimizer, learning rate, weight decay, and number of runs; these are needed for reproducibility and for fair comparison with Table I.
[Section II-D] The BCD decoder is described as following the head in reference [20]; the paper should clearly state which components are reused from prior work and which are new, since the contribution statement in Section I does not mention this inheritance.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the SECOND benchmark results are self-contained, though the open-vocabulary generalization claim is asserted rather than measured.

full rationale

The central quantitative claim is an independent benchmark comparison on the SECOND dataset, not a derivation from the self-cited BCSF module or from CLIP. The BCSF adapter is reused from the authors' prior work [20], but it is an architectural component whose contribution is validated by the full-system experiments; no equation in the paper reduces the reported SOTA numbers to that citation. The open semantic prompter's context tokens are learned on the six target categories during ordinary supervised training, which is not a fitted parameter being relabeled as a prediction. The open-vocabulary claim is weakened by the paper's own admission in Sec. III-C that the setting is degraded and all categories participate in training, with no held-out category evaluation; however, this is a validity gap in the generalization claim, not circular reasoning. The self-citation of BCSF is present but not load-bearing, so the score is 2 rather than 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper's central result rests on a handful of domain assumptions about the dataset, the CLIP text encoder, and the transferability of the BCSF adapter, rather than on free constants. The main operational choices are training schedules and context token length, several of which are unreported or unjustified.

free parameters (5)

Context token length = 256
Set to 256 without ablation; controls the capacity of the open semantic prompter.
SCD training schedule = 5 epochs for Semantic-CD, 100 for baseline
The unequal schedules make the ablation comparison ambiguous; no sensitivity analysis is reported.
BCD training schedule = 300 epochs
Chosen schedule; no sensitivity analysis or convergence evidence.
Batch size = 1
Fixed to 1 for all training; no sensitivity analysis.
BCD/SCD loss weights = unspecified
Equation (6) defines both losses but no coefficients are given, so the relative weighting is an unreported modeling choice.

assumptions (4)

domain assumption SECOND dataset with its standard evaluation protocol is a valid benchmark for comparing SCD methods.
The paper uses only SECOND and does not describe the train/test split or verify that the cited methods used identical protocols.
domain assumption RemoteCLIP text embeddings carry usable semantics for the six land-cover categories.
The open semantic prompter depends on these embeddings; no zero-shot experiment or unseen-category test checks this.
ad hoc to paper BCSF, designed for SAM in the authors' prior work, transfers to CLIP ViT-L-14.
Section II-B inserts BCSF into four CLIP attention layers without ablation or justification beyond the prior paper.
ad hoc to paper Allowing all dataset categories in training still constitutes open-vocabulary evaluation.
Section III-C states the setting is degraded; the paper never holds out categories, so open-vocabulary generalization is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Semantic-CD: Remote Sensing Image Semantic Change Detection towards Open-vocabulary Setting." pith.science (2026). https://pith.science/paper/PDJPOCC4

@misc{pith2026250106808,
  author       = {Pith},
  title        = {Pith review of: Semantic-CD: Remote Sensing Image Semantic Change Detection towards Open-vocabulary Setting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PDJPOCC4}},
  note         = {Machine review of arXiv:2501.06808}
}

read the original abstract

Remote sensing image semantic change detection is a method used to analyze remote sensing images, aiming to identify areas of change as well as categorize these changes within images of the same location taken at different times. Traditional change detection methods often face challenges in generalizing across semantic categories in practical scenarios. To address this issue, we introduce a novel approach called Semantic-CD, specifically designed for semantic change detection in remote sensing images. This method incorporates the open vocabulary semantics from the vision-language foundation model, CLIP. By utilizing CLIP's extensive vocabulary knowledge, our model enhances its ability to generalize across categories and improves segmentation through fully decoupled multi-task learning, which includes both binary change detection and semantic change detection tasks. Semantic-CD consists of four main components: a bi-temporal CLIP visual encoder for extracting features from bi-temporal images, an open semantic prompter for creating semantic cost volume maps with open vocabulary, a binary change detection decoder for generating binary change detection masks, and a semantic change detection decoder for producing semantic labels. Experimental results on the SECOND dataset demonstrate that Semantic-CD achieves more accurate masks and reduces semantic classification errors, illustrating its effectiveness in applying semantic priors from vision-language foundation models to SCD tasks.

Figures

Figures reproduced from arXiv: 2501.06808 by the authors.

**Figure 1.** The architecture of the Semantic-CD consists of four main components: an adapted CLIP vision encoder for extracting robust features from bi-temporal [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** The structure of the Semantic Change Detection Decoder. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Qualitative comparison results with other methods on the SECOND [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Foundation Model-Driven Semantic Change Detection in Remote Sensing Imagery
cs.CV 2026-02 unverdicted novelty 6.0 of 10

PerASCD sets new state-of-the-art Sek scores on SECOND and LandsatSCD datasets by using a modular cascaded gated decoder on PerA foundation model features plus a new consistency loss.
DynamicEarth: How Far are We from Open-Vocabulary Change Detection?
cs.CV 2025-01 reject novelty 4.0 of 10

The paper shows that composing mask proposal, feature comparison, and open-vocabulary classification models can detect arbitrary-category changes in satellite images without training.

Reference graph

Works this paper leans on

23 extracted references · 16 canonical work pages · cited by 2 Pith papers

[20]

Semantic- cc: Boosting remote sensing image change captioning via foundational knowledge and semantic guidance,

Y . Zhu, L. Li, K. Chen, C. Liu, F. Zhou, and Z. Shi, “Semantic- cc: Boosting remote sensing image change captioning via foundational knowledge and semantic guidance,” arXiv preprint arXiv:2407.14032, 2024

arXiv 2024
[1]

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

work page 2022
[2]

Difference enhancement and spatial–spectral nonlocal network for change detection in vhr remote sensing images,

T. Lei, J. Wang, H. Ning, X. Wang, D. Xue, Q. Wang, and A. K. Nandi, “Difference enhancement and spatial–spectral nonlocal network for change detection in vhr remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–13, 2021

work page 2021
[3]

Siamese network with multi-level features for patch-based change detection in satellite imagery,

F. Rahman, B. Vasu, J. Van Cor, J. Kerekes, and A. Savakis, “Siamese network with multi-level features for patch-based change detection in satellite imagery,” in 2018 IEEE Global Conference on Signal and Information Processing (GlobalSIP) . IEEE, 2018, pp. 958–962

work page 2018
[4]

From w-net to cdgan: Bitemporal change detection via deep learning techniques,

B. Hou, Q. Liu, H. Wang, and Y . Wang, “From w-net to cdgan: Bitemporal change detection via deep learning techniques,” IEEE Transactions on Geoscience and Remote Sensing , vol. 58, no. 3, pp. 1790–1802, 2019

work page 2019
[5]

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
[6]

Dasnet: Dual attentive fully convolutional siamese networks for change detection in high-resolution satellite im- ages,

J. Chen, Z. Yuan, J. Peng, L. Chen, H. Huang, J. Zhu, Y . Liu, and H. Li, “Dasnet: Dual attentive fully convolutional siamese networks for change detection in high-resolution satellite im- ages,” IEEE Journal of Selected Topics in Applied Earth Ob- servations and Remote Sensing , vol. 14, pp. 1194–1206, 2020

work page 2020
[7]

Pga-siamnet: Pyramid feature-based attention-guided siamese network for remote sensing orthoimagery building change de- tection,

H. Jiang, X. Hu, K. Li, J. Zhang, J. Gong, and M. Zhang, “Pga-siamnet: Pyramid feature-based attention-guided siamese network for remote sensing orthoimagery building change de- tection,” Remote Sensing , vol. 12, no. 3, p. 484, 2020

work page 2020

Show all 23 references

[8]

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
[9]

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
[10]

Joint spatio-temporal modeling for semantic change detection in remote sensing images,

L. Ding, J. Zhang, H. Guo, K. Zhang, B. Liu, and L. Bruzzone, “Joint spatio-temporal modeling for semantic change detection in remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing , 2024

2024
[11]

Change- mamba: Remote sensing change detection with spatio-temporal state space model,

H. Chen, J. Song, C. Han, J. Xia, and N. Yokoya, “Change- mamba: Remote sensing change detection with spatio-temporal state space model,” arXiv preprint arXiv:2404.03425 , 2024

2024 arXiv
[12]

Smnet: symmetric multi-task network for semantic change detection in remote sensing images based on cnn and transformer,

Y . Niu, H. Guo, J. Lu, L. Ding, and D. Yu, “Smnet: symmetric multi-task network for semantic change detection in remote sensing images based on cnn and transformer,” Remote Sensing, vol. 15, no. 4, p. 949, 2023

2023
[13]

Scdnet: A novel convolutional network for semantic change detection in high resolution optical remote sensing imagery,

D. Peng, L. Bruzzone, Y . Zhang, H. Guan, and P. He, “Scdnet: A novel convolutional network for semantic change detection in high resolution optical remote sensing imagery,” International Journal of Applied Earth Observation and Geoinformation , vol. 103, p. 102465, 2021

2021
[14]

Multi- task learning for large-scale semantic change detection,

R. C. Daudt, B. Le Saux, A. Boulch, and Y . Gousseau, “Multi- task learning for large-scale semantic change detection,” Com- puter Vision and Image Understanding , vol. 187, p. 102783, 2019

2019
[15]

Bi-temporal semantic reasoning for the semantic change de- tection in hr remote sensing images,

L. Ding, H. Guo, S. Liu, L. Mou, J. Zhang, and L. Bruzzone, “Bi-temporal semantic reasoning for the semantic change de- tection in hr remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–14, 2022

2022
[16]

Asymmetric siamese networks for semantic change detection in aerial images,

K. Yang, G.-S. Xia, Z. Liu, B. Du, W. Yang, M. Pelillo, and L. Zhang, “Asymmetric siamese networks for semantic change detection in aerial images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–18, 2021

2021
[17]

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
[18]

Temporal-agnostic change region proposal for semantic change detection,

S. Tian, X. Tan, A. Ma, Z. Zheng, L. Zhang, and Y . Zhong, “Temporal-agnostic change region proposal for semantic change detection,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 204, pp. 306–320, 2023

2023
[19]

Learning transferable visual models from natural language supervision,

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al. , “Learning transferable visual models from natural language supervision,” PMLR, pp. 8748–8763, 2021

2021
[21]

Exploring plain vision transformer backbones for object detection,

Y . Li, H. Mao, R. Girshick, and K. He, “Exploring plain vision transformer backbones for object detection,” in European conference on computer vision . Springer, 2022, pp. 280–296

2022
[22]

Semantic change detection with asymmetric siamese networks,

K. Yang, G.-S. Xia, Z. Liu, B. Du, W. Yang, M. Pelillo, and L. Zhang, “Semantic change detection with asymmetric siamese networks,” arXiv preprint arXiv:2010.05687 , 2020

2010 arXiv
[23]

Remoteclip: A vision language foundation model for remote sensing,

F. Liu, D. Chen, Z. Guan, X. Zhou, J. Zhu, Q. Ye, L. Fu, and J. Zhou, “Remoteclip: A vision language foundation model for remote sensing,” IEEE Transactions on Geoscience and Remote Sensing, 2024

2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [20]

Semantic- cc: Boosting remote sensing image change captioning via foundational knowledge and semantic guidance,

Y . Zhu, L. Li, K. Chen, C. Liu, F. Zhou, and Z. Shi, “Semantic- cc: Boosting remote sensing image change captioning via foundational knowledge and semantic guidance,” arXiv preprint arXiv:2407.14032, 2024

arXiv 2024

[2] [1]

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

work page 2022

[3] [2]

Difference enhancement and spatial–spectral nonlocal network for change detection in vhr remote sensing images,

T. Lei, J. Wang, H. Ning, X. Wang, D. Xue, Q. Wang, and A. K. Nandi, “Difference enhancement and spatial–spectral nonlocal network for change detection in vhr remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–13, 2021

work page 2021

[4] [3]

Siamese network with multi-level features for patch-based change detection in satellite imagery,

F. Rahman, B. Vasu, J. Van Cor, J. Kerekes, and A. Savakis, “Siamese network with multi-level features for patch-based change detection in satellite imagery,” in 2018 IEEE Global Conference on Signal and Information Processing (GlobalSIP) . IEEE, 2018, pp. 958–962

work page 2018

[5] [4]

From w-net to cdgan: Bitemporal change detection via deep learning techniques,

B. Hou, Q. Liu, H. Wang, and Y . Wang, “From w-net to cdgan: Bitemporal change detection via deep learning techniques,” IEEE Transactions on Geoscience and Remote Sensing , vol. 58, no. 3, pp. 1790–1802, 2019

work page 2019

[6] [5]

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

[7] [6]

Dasnet: Dual attentive fully convolutional siamese networks for change detection in high-resolution satellite im- ages,

J. Chen, Z. Yuan, J. Peng, L. Chen, H. Huang, J. Zhu, Y . Liu, and H. Li, “Dasnet: Dual attentive fully convolutional siamese networks for change detection in high-resolution satellite im- ages,” IEEE Journal of Selected Topics in Applied Earth Ob- servations and Remote Sensing , vol. 14, pp. 1194–1206, 2020

work page 2020

[8] [7]

Pga-siamnet: Pyramid feature-based attention-guided siamese network for remote sensing orthoimagery building change de- tection,

H. Jiang, X. Hu, K. Li, J. Zhang, J. Gong, and M. Zhang, “Pga-siamnet: Pyramid feature-based attention-guided siamese network for remote sensing orthoimagery building change de- tection,” Remote Sensing , vol. 12, no. 3, p. 484, 2020

work page 2020