Pith. sign in

REVIEW 4 major objections 7 minor 63 references

Multimodal Feature Fusion Network with Text Difference Enhancement for Remote Sensing Change Detection

T0 review · 4 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adding a vision-language text branch to remote sensing change detection—caption each date, amplify the caption difference, then fuse it with image features—consistently improves IoU and F1 over image-only state-of-the

desk verdict The TDE module's equations are identical, making the paper's central novelty degenerate as written—worth a referee, but the authors must fix the formulas and stop tuning the prompt on the test set. read the letter →

arxiv 2509.03961 v1 pith:2UAW6SSN submitted 2025-09-04 cs.CV cs.AI

Yijun Zhou , Yikui Zhai , Zilu Ying , Tingfeng Xian , Wenlve Zhou , Zhiheng Zhou , Xiaolin Tian , Xudong Jia
show 2 more authors
Hongsheng Zhang C. L. Philip Chen
This is my paper · ORCID
classification cs.CVcs.AI
keywords remotesensingchangedetectionmultimodallearningvision-languagemodeltextualdifferenceenhancementimage-textfeaturefusionattentionmechanismLEVIR-CDWHU-CD/SYSU-CD
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 argues that change detection in satellite and aerial image pairs can be improved by adding a text modality: a vision-language model writes a caption for each of the two dates, a Textual Difference Enhancement module amplifies the semantic gap between captions, and an Image-Text Feature Fusion module injects that gap into image features before decoding a change mask. On three standard benchmarks, the resulting MMChange network is reported to beat twelve prior methods on IoU and F1, with the largest margin on WHU-CD. The paper also reports that the text branch confers robustness to added noise and illumination shifts, and that each of its three modules contributes in ablations. The claimed novelty centers on TDE, described as the first text-difference enhancement module for multimodal remote sensing change detection.

What carries the argument

The object doing the work is the textual difference feature T_diff, produced by TDE from the subtraction of two caption embeddings. TDE is the paper's claimed novelty: instead of feeding static labels or category names, it computes T_text1 − T_text2, applies scaled dot-product attention to that difference, and gates convolved caption features by the attention output, amplifying semantic shifts. Around it sit IFR, which refines image features by grouped residual convolutions and coordinate/channel integration, and ITFF, which fuses the two modalities through channel, spatial, and pixel attention so the text difference can modulate image features at four scales. Together the three modules inst

What would settle it

Run the TDE module exactly as written: Eq. (1) and Eq. (2) compute the same function of T_text1 − T_text2, so T1 = T2, making Eq. (4)'s input zero; if the released code instead produces non-degenerate T1 and T2, the equations are not the implementation. Separately, retrain MMChange with the caption source replaced by random captions drawn from the same vocabulary: if IoU and F1 on WHU-CD stay near 90.90 and 95.23, the reported gains come from the fusion architecture rather than from semantic text.

Watch

Extended reading notes

Core claim

The central claim is that bi-temporal remote-sensing change detection can be improved by treating the semantic difference between two image captions as a first-class feature. MMChange first runs a small vision-language model on each date's image to get free-text descriptions, encodes them with a CLIP text encoder, then subtracts the two encodings and passes the difference through a Textual Difference Enhancement module built around scaled dot-product attention and convolutional gating. In parallel, a ResNet image encoder produces multi-scale features refined by an Image Feature Refinement module. An Image-Text Feature Fusion module combines the refined image features with the enhanced text-d

Load-bearing premise

The load-bearing premise is that the reported comparison is a fair test and that the printed TDE equations specify the architecture that was actually evaluated—yet Eqs. (1) and (2) are identical operations on the same input, so taken literally the textual-difference branch collapses and the published ablation numbers cannot be reconstructed from the equations alone.

Editorial extensions

If this is right

  • Change detectors that already use Siamese image encoders could add a caption-difference branch without changing the decoder, gaining semantic guidance at the cost of one captioning pass.
  • Under the reported noise and illumination experiments, the text modality acts as a stabilizer: performance degrades less than the image-only baseline when inputs are corrupted.
  • The module ablations imply that each of IFR, TDE, and ITFF contributes, with the largest single-module drops on WHU-CD and SYSU-CD.
  • The method positions multimodal captioning as a general enhancement for remote sensing change detection rather than a replacement for image differencing.

Reading between the lines

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

  • Because Eqs. (1) and (2) are identical operations on the same input, the architecture as written makes T1 = T2, hence T4 = 0 and the difference pathway collapses; a reader should treat the figure and released code as the operative specification and expect the equations to need correction.
  • A decisive test of the text-branch claim would be to retrain MMChange with the vision-language captions replaced by random captions drawn from the same vocabulary: if IoU/F1 hold near the reported values, the gains come from the fusion structure rather than from semantic text.
  • The caption-difference machinery suggests a route to weakly supervised change detection: caption differences are available without pixel labels, so TDE/ITFF could be trained on image-caption pairs before pixel-level fine-tuning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes MMChange, a multimodal (image + text) remote sensing change detection network that integrates a VLM (TinyLLaVA) to generate textual descriptions of bitemporal images. Three modules are introduced: an Image Feature Refinement (IFR) module, a Textual Difference Enhancement (TDE) module, and an Image-Text Feature Fusion (ITFF) module. The authors report state-of-the-art results on LEVIR-CD, WHU-CD, and SYSU-CD, and they provide extensive comparisons, ablations, robustness experiments under noise/illumination, and prompt/VLM selection studies. The central claim is that MMChange consistently outperforms existing methods and that the TDE module is the key to capturing fine-grained semantic shifts between bitemporal textual descriptions.

Significance. If the architecture and experiments are reproducible as described, the paper would make a useful contribution to multimodal RSCD, particularly in showing that VLM-generated text descriptions can be integrated with image features for change detection. The proposed IFR/ITFF modules are reasonable and the experimental scope is broad, including three standard benchmarks, twelve baselines, and robustness ablations. The paper also provides a code link. However, the significance is conditional because the TDE module, which is a central claimed novelty, is described by internally inconsistent equations that make the module degenerate as written. The experimental protocol also has selection-on-test-set issues. Therefore the contribution is potentially valuable but not presently established.

major comments (4)
  1. [III-B, Eqs. (1)–(6)] Eqs. (1) and (2) are literally identical operations on the same input: T1 = ReLU(BN(Conv3x3(UP(Ttext1 - Ttext2)))) and T2 = ReLU(BN(Conv3x3(UP(Ttext1 - Ttext2)))). Hence T1=T2 for every input, making T1-T2 = 0. Eq. (4) then gives T4 = ReLU(BN(Conv3x3(0))), which is either identically zero (if the convolution has no bias) or a constant independent of the input (if bias is present). Eq. (5) collapses to Conv3x3(Cat(T1, T2)), and the only input-dependent text-difference path is T3 = SDPA(Ttext1 - Ttext2). The claimed 'fine-grained semantic shift' enhancement, and the large TDE ablation gains in Table II (e.g., WHU-CD IoU drops 2.45%, SYSU-CD IoU drops 3.72% without TDE), cannot be reproduced from the published equations. This is a load-bearing inconsistency: the central architectural novelty is not specified. The authors must correct Eqs. (1)–(2) (likely T1 should be computed from Ttext1 an
  2. [IV-B and IV-D] The experimental comparison rests on a fairness claim that is not adequately documented. Section IV-B states 'all training parameters were set to their default values,' and Section IV-D states that 'we used the open-source code to re-implement these methods,' but no commit hashes, training schedules, input resolutions, number of epochs, or hyperparameters are given for the baselines. Without this information, the reported gains (e.g., WHU-CD IoU 90.90% vs. BiFA 88.94%) cannot be independently reproduced or compared apples-to-apples. Please provide a complete protocol for both MMChange and every baseline, including any hyperparameter tuning that was performed.
  3. [IV-F, Table V] The prompt 'What are the components in this picture?' and the choice of TinyLLaVA are selected because they perform best on the same three test datasets (Table V). This is a form of model selection on the test set, which can inflate reported performance and complicates the claim that the method generalizes. The paper should either use a held-out validation split for prompt/VLM selection or report the selection procedure transparently and discuss the risk of overfitting. The same concern applies to the comparison of TinyLLaVA with CLIP Interrogator and mPLUG, which is also evaluated on the test datasets.
  4. [IV-E, Table II] The ablation experiments are presented as evidence for each module's contribution, but the TDE rows in Table II are not interpretable independently of the equation-level inconsistency in Section III-B. If the TDE module is implemented differently from Eqs. (1)–(6), the ablation table measures an undocumented architecture. The IFR and ITFF ablations are clearer, but the TDE ablation should be re-run and reported with a corrected, unambiguous module specification.
minor comments (7)
  1. [III-B] The text says features T4 and T5 are generated and then T3 and T5 are multiplied, but in Eq. (6) the multiplication is T5 · T3. Please make the notation consistent and clarify the channel dimension handling in Cat(·).
  2. [III-C] Eq. (10) uses a subscript 'n=4' but does not define how the feature map is split into groups, what operation is performed per group, or how the groups are recombined. Please provide the exact tensor operations for the groupwise refinement.
  3. [IV-E] The sentence 'We employs 1×1 convolution and element-wise addition for dimension alignment and feature fusion' is not reflected in Eq. (12) or Fig. 5, where only element-wise addition is shown. Please clarify where the 1×1 convolution is used.
  4. [IV-C] Implementation details give the optimizer and learning-rate schedule but omit input crop size, number of training epochs/iterations per dataset, batch size per dataset, and the train/validation split. These are needed for reproducibility.
  5. [IV-B] 'All training parameters were set to their default values' is ambiguous: which defaults, and for which models? Specify the exact configuration for each baseline or refer to a released configuration file.
  6. [Table IV] The term 'baseline' is used without definition. It appears to correspond to the image-only row in Table II, but this should be stated explicitly for the reader.
  7. [Throughout] There are numerous typos and formatting issues, e.g., 'Relate Works' in the Section II heading, 'Conν' instead of 'Conv' in Eqs. (1)–(10), 'Ourst+noise' in Table IV, and inconsistent capitalization of dataset names. A careful proofread is needed.

Circularity Check

2 steps flagged · score 6.0 of 10

TDE equations (1)-(2) are identical, nullifying the difference path; prompt/VLM are selected on the test benchmarks, making the reported SOTA results partly fitted rather than predicted.

  1. self definitional [Section III-B, Eqs. (1)-(2) and (4)-(6), Fig. 3]
    "T1 = ReLU (BN (Conν 3×3(U P(Ttext1 − Ttext2)))) (1) T2 = ReLU (BN (Conν 3×3(U P(Ttext1 − Ttext2)))) (2)"

    Equations (1) and (2) are identical functions of the same input, so T1 ≡ T2 for every input by construction. Substituting into Eq. (4) gives T4 = ReLU(BN(Conv3x3(T1−T2))) = ReLU(BN(Conv3x3(0))), which is zero or an input-independent constant. Eq. (5) then degenerates to Conv(Cat(T1, T2)), and the only input-dependent path is T3 = SDPA(Ttext1 − Ttext2). The paper claims the TDE module 'captures fine-grained semantic shifts' and attributes Table II ablation gains (e.g., WHU-CD IoU −2.45%, SYSU-CD IoU −3.72%) to this module, but those gains cannot be produced by the equations as written. The module's difference-enhancement path cancels by definition, so the reported effect is not a consequence of the specified architecture.

  2. fitted input called prediction [Section IV.F, Table V, and Section III.A (prompt choice)]
    "we designed four distinct prompts and employed them to produce bitemporal image descriptions using TinyLLaVA... As reported in Table V, although these prompts do not surpass the performance of the MMChange-specific prompt 'What are the components in this picture?'... we compared TinyLLaVA with two alternative VLM (CLIP Interrogator [62] and mPLUG [63]) under the same prompt setting. As presented in Table V, experimental results on the WHU-CD dataset show that TinyLLaVA yields higher F1 and IOU"

    The final prompt and text encoder were selected after observing their F1/IoU on the same three datasets (LEVIR-CD, WHU-CD, SYSU-CD) used for the headline comparisons in Table I. The prompt 'What are the components in this picture?' was chosen because it beat four alternatives on these datasets, and TinyLLaVA was chosen because it beat CLIP Interrogator and mPLUG on WHU-CD. Thus the reported SOTA numbers are the result of test-set selection among prompt/VLM options: the input is fitted to the benchmark, and then the benchmark result is presented as the performance of a fixed method. This is fitted-input-called-prediction: the advantage over baselines is partly an artifact of choosing the best-performing configuration on the evaluation set.

full rationale

The paper's central claim—that MMChange consistently surpasses SOTA on LEVIR-CD, WHU-CD, and SYSU-CD—is undermined by two specific reductions. First, the TDE module, which is a core contribution and is credited with notable ablation improvements, is specified by Eqs. (1) and (2) as two identical operations on the same input. This makes T1 = T2 by definition, so the subtraction in Eq. (4) is identically zero; the module's only input-dependent branch is the scaled dot-product attention in Eq. (3). The published equations therefore do not implement the claimed 'text difference enhancement,' and the ablation numbers attributed to TDE cannot be reproduced from the paper alone. Second, the choice of prompt and VLM is explicitly based on their performance on the same three benchmark datasets (Table V and Section IV.F). The prompt 'What are the components in this picture?' and TinyLLaVA were selected because they outperformed alternatives on WHU-CD (and the other datasets), meaning the reported results are a selected maximum over tested configurations rather than the prediction of a pre-specified method. This is a form of fitting the input to the evaluation set. The paper does not rely on load-bearing self-citations; its baseline comparisons are against external methods, and the reported numbers are empirical. However, the combination of a degenerate central module and test-set-selected inputs means the core 'predictions' are partly forced by construction and by selection. A score of 6 reflects this partial circularity. If Eqs. (1)-(2) are a typo, the paper must correct them; as published, the architecture is inconsistent with the claimed enhancements and ablation results.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new theoretical entities. The ledger shows that its main degrees of freedom are the VLM prompt/choice (tuned on the evaluation datasets) and a handful of architecture choices (n=4, attention layout). These are empirical design choices, not parameters derived from first principles.

free parameters (4)
  • TinyLLaVA prompt = "What are the components in this picture?"
    Selected by comparing 4 prompts on the three evaluation datasets (Table V); the prompt is tuned on the test benchmarks.
  • VLM choice = TinyLLaVA
    Chosen because it outperforms CLIP Interrogator and mPLUG on WHU-CD (Table V), i.e., selected on the evaluation set.
  • IFR group count n=4 = 4
    Stated as a design choice (Section III-C) with no ablation, so it is a hand-picked hyperparameter.
  • Training hyperparameters = lr=0.0005, power=0.9, max_iter=40000, batch=32, Adam beta2=0.99
    Reported as fixed values with no sensitivity analysis; these are standard but still fitted choices.
assumptions (3)
  • domain assumption ResNet50 provides useful image features for change detection
    Adopted from prior work (He et al. 2016); standard backbone.
  • domain assumption TinyLLaVA produces text descriptions that are 'semantically aligned and task-relevant'
    Paper assumes VLM output is useful for change localization, but shows redundant text examples in Fig. 2 and does not quantify VLM error.
  • domain assumption The evaluation protocol of three public datasets is a valid proxy for real-world change detection quality
    Standard for the field; no out-of-distribution validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Feature Fusion Network with Text Difference Enhancement for Remote Sensing Change Detection." pith.science (2026). https://pith.science/paper/2UAW6SSN

@misc{pith2026250903961,
  author       = {Pith},
  title        = {Pith review of: Multimodal Feature Fusion Network with Text Difference Enhancement for Remote Sensing Change Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2UAW6SSN}},
  note         = {Machine review of arXiv:2509.03961}
}
read the original abstract

Although deep learning has advanced remote sensing change detection (RSCD), most methods rely solely on image modality, limiting feature representation, change pattern modeling, and generalization especially under illumination and noise disturbances. To address this, we propose MMChange, a multimodal RSCD method that combines image and text modalities to enhance accuracy and robustness. An Image Feature Refinement (IFR) module is introduced to highlight key regions and suppress environmental noise. To overcome the semantic limitations of image features, we employ a vision language model (VLM) to generate semantic descriptions of bitemporal images. A Textual Difference Enhancement (TDE) module then captures fine grained semantic shifts, guiding the model toward meaningful changes. To bridge the heterogeneity between modalities, we design an Image Text Feature Fusion (ITFF) module that enables deep cross modal integration. Extensive experiments on LEVIRCD, WHUCD, and SYSUCD demonstrate that MMChange consistently surpasses state of the art methods across multiple metrics, validating its effectiveness for multimodal RSCD. Code is available at: https://github.com/yikuizhai/MMChange.

Figures

Figures reproduced from arXiv: 2509.03961 by the authors.

Figure 1
Figure 1. illustrates the architecture of the MMChange model, which is divided into two image and text stages. In the image stage, the image encoder generates [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 3
Figure 3. TDE Structure. T4 = ReLU(BN(Conν3×3(T1 − T2))) (4) T5 = Conv3×3(Cat(T1 + (T1 · T4), T2 + (T2 · T4))) (5) Tdif f = ReLU(BN(Conv3×3(T5 · T3))) (6) where SDPA(·) refers to the scaled dot-product attention operation. C. Image Feature Refinement Module (IFR) In multimodal RSCD, image features are refined to en￾hance their clarity and prominence, providing high-quality features for subsequent multimodal fusion. The IFR mo… view at source ↗
Figure 4
Figure 4. IFR Structure. change detection. Therefore, feature refinement establishes a solid foundation for multimodal fusion and markedly improves the performance of CD. The IFR module is depicted in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: ITFF Structure. spatial descriptors S1 and S2, which capture different types of activation responses. These are concatenated and passed through a convolutional layer followed by a sigmoid activation to generate the spatial attention map, thereby enhancing the model’s f…
Figure 6
Figure 6. Figure 6: Visual comparative results of MMChange and other state-of-the-art methods on LEVIR-CD, where white denotes true positives, black denotes true [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Visual comparative results of MMChange and other advanced methods on WHU-CD, where white indicates true positive, black indicates true negative, [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Visual comparative results of MMChange and other state-of-the-art methods on SYSU-CD, where white denotes true positives, black denotes true [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Visualisation of the results of different modular ablation experiments in MMChange, where white indicates true-positive, black indicates true-negative, [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Heatmap results of the ablation experiments on different modules in MMChange. Red indicates areas with higher attention, while blue represents [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Text Descriptions of Bitemporal Remote Sensing Images Generated by TinyLLaVA with Different Prompts [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 60 canonical work pages

  1. [1]

    An Empirical Review of Deep Learn- ing Frameworks for Change Detection: Model Design, Experimental Frameworks, Challenges and Research Needs,

    M. Mandal and S. K. Vipparthi, “An Empirical Review of Deep Learn- ing Frameworks for Change Detection: Model Design, Experimental Frameworks, Challenges and Research Needs,” IEEE Transactions on Intelligent Transportation Systems , vol. 23, no. 7, pp. 6101–6122, Jul. 2022

    work page 2022

  2. [2]

    A survey of sample-efficient deep learning for change detection in remote sensing: Tasks, strategies, and challenges,

    L. Ding, D. Hong, M. Zhao, H. Chen, C. Li, J. Deng, N. Yokoya, L. Bruzzone, and J. Chanussot, “A survey of sample-efficient deep learning for change detection in remote sensing: Tasks, strategies, and challenges,” IEEE Geoscience and Remote Sensing Magazine , 2025

    work page 2025

  3. [3]

    Competition between biogeochemical drivers and land-cover changes determines urban greening or browning,

    L. Li, W. Zhan, W. Ju, J. Pe ˜nuelas, Z. Zhu, S. Peng, X. Zhu, Z. Liu, Y . Zhou, J. Li, J. Lai, F. Huang, G. Yin, Y . Fu, M. Li, and C. Yu, “Competition between biogeochemical drivers and land-cover changes determines urban greening or browning,” Remote Sensing of Environ- ment, vol. 287, p. 113481, Mar. 2023

    work page 2023

  4. [4]

    Building damage assessment for rapid disaster response with a deep object-based semantic change detection framework: From natural disasters to man- made disasters,

    Z. Zheng, Y . Zhong, J. Wang, A. Ma, and L. Zhang, “Building damage assessment for rapid disaster response with a deep object-based semantic change detection framework: From natural disasters to man- made disasters,” Remote Sensing of Environment , vol. 265, p. 112636, Nov. 2021

    work page 2021

  5. [5]

    Improving Urban Change Detection From Multitemporal SAR Images Using PCA-NLM,

    O. Yousif and Y . Ban, “Improving Urban Change Detection From Multitemporal SAR Images Using PCA-NLM,” IEEE Transactions on Geoscience and Remote Sensing , vol. 51, no. 4, pp. 2032–2041, Apr. 2013

    work page

  6. [6]

    Timeliness in forest change monitoring: A new assessment framework demonstrated using Sentinel-1 and a continuous change detection algorithm,

    E. L. Bullock, S. P. Healey, Z. Yang, R. Houborg, N. Gorelick, X. Tang, and C. Andrianirina, “Timeliness in forest change monitoring: A new assessment framework demonstrated using Sentinel-1 and a continuous change detection algorithm,” Remote Sensing of Environment , vol. 276, p. 113043, Jul. 2022

    work page 2022

  7. [7]

    Land cover characterization and change detection for environmental monitoring of pan-Europe,

    C. A. Mucher, K. T. Steinnocher, F. P. Kressler, and C. Heunks, “Land cover characterization and change detection for environmental monitoring of pan-Europe,” International Journal of Remote Sensing , vol. 21, no. 6-7, pp. 1159–1181, Jan. 2000

    work page 2000

  8. [8]

    Landslide Recognition by Deep Convolutional Neural Network and Change Detec- tion,

    W. Shi, M. Zhang, H. Ke, X. Fang, Z. Zhan, and S. Chen, “Landslide Recognition by Deep Convolutional Neural Network and Change Detec- tion,” IEEE Transactions on Geoscience and Remote Sensing , vol. 59, no. 6, pp. 4654–4672, Jun. 2021. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14

    work page 2021

Show all 63 references
  1. [9]

    A Coarse-to-Fine Semi-Supervised Change Detection for Multispectral Images,

    W. Zhang, X. Lu, and X. Li, “A Coarse-to-Fine Semi-Supervised Change Detection for Multispectral Images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 56, no. 6, pp. 3587–3599, Jun. 2018

    2018

  2. [10]

    Quantification of the impact of misregistration on the accuracy of remotely sensed change detection,

    X. Dai and S. Khorram, “Quantification of the impact of misregistration on the accuracy of remotely sensed change detection,” in IGARSS’97. 1997 IEEE International Geoscience and Remote Sensing Symposium Proceedings. Remote Sensing - A Scientific Vision for Sustainable Develop...

    1997

  3. [11]

    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 Geo- science and Remote Sensing Letters , vol. 6, no. 4, pp. 772–776, Oct. 2009

    2009

  4. [12]

    Unsupervised Deep Change Vector Analysis for Multiple-Change Detection in VHR Images,

    S. Saha, F. Bovolo, and L. Bruzzone, “Unsupervised Deep Change Vector Analysis for Multiple-Change Detection in VHR Images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 57, no. 6, pp. 3677–3693, Jun. 2019

    2019

  5. [13]

    An Efficient Protocol to Process Landsat Images for Change Detection With Tasselled Cap Transformation,

    T. Han, M. A. Wulder, J. C. White, N. C. Coops, M. F. Alvarez, and C. Butson, “An Efficient Protocol to Process Landsat Images for Change Detection With Tasselled Cap Transformation,” IEEE Geoscience and Remote Sensing Letters , vol. 4, no. 1, pp. 147–151, Jan. 2007

    2007

  6. [14]

    Spectral–Spatial-Aware Unsupervised Change Detection With Stochastic Distances and Support Vector Machines,

    R. G. Negri, A. C. Frery, W. Casaca, S. Azevedo, M. A. Dias, E. A. Silva, and E. H. Alc ˆantara, “Spectral–Spatial-Aware Unsupervised Change Detection With Stochastic Distances and Support Vector Machines,” IEEE Transactions on Geoscience and Remote Sensing , vol. 59, no. 4, p...

    2021

  7. [15]

    Iterative Robust Graph for Unsupervised Change Detection of Heterogeneous Remote Sensing Images,

    Y . Sun, L. Lei, D. Guan, and G. Kuang, “Iterative Robust Graph for Unsupervised Change Detection of Heterogeneous Remote Sensing Images,” IEEE Transactions on Image Processing , vol. 30, pp. 6277– 6291, 2021

    2021

  8. [16]

    Fusion of SAR and Multispectral Images Using Random Forest Regression for Change Detection,

    D. K. Seo, Y . H. Kim, Y . D. Eo, M. H. Lee, and W. Y . Park, “Fusion of SAR and Multispectral Images Using Random Forest Regression for Change Detection,” ISPRS International Journal of Geo-Information , vol. 7, no. 10, p. 401, Oct. 2018

    2018

  9. [17]

    Domain contrast for domain adaptive object detection,

    F. Liu, X. Zhang, F. Wan, X. Ji, and Q. Ye, “Domain contrast for domain adaptive object detection,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 12, pp. 8227–8237, 2021

    2021

  10. [18]

    Mscaf-net: A general framework for camouflaged object detection via learning multi-scale context-aware features,

    Y . Liu, H. Li, J. Cheng, and X. Chen, “Mscaf-net: A general framework for camouflaged object detection via learning multi-scale context-aware features,” IEEE Transactions on Circuits and Systems for Video Tech- nology, vol. 33, no. 9, pp. 4934–4947, 2023

    2023

  11. [19]

    Mifnet: Multi-scale interaction fusion network for remote sensing image change detection,

    W. Xie, W. Shao, D. Li, Y . Li, and L. Fang, “Mifnet: Multi-scale interaction fusion network for remote sensing image change detection,” IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  12. [20]

    Cross-level attentive feature aggregation for change detection,

    G. Wang, G. Cheng, P. Zhou, and J. Han, “Cross-level attentive feature aggregation for change detection,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 7, pp. 6051–6062, 2023

    2023

  13. [21]

    Catfpn: Adaptive feature pyramid with scale-wise concatenation and self- attention,

    J. Huang, Z. Chen, Q. J. Wu, C. Liu, H. Yuan, and W. He, “Catfpn: Adaptive feature pyramid with scale-wise concatenation and self- attention,” IEEE Transactions on Circuits and Systems for Video Tech- nology, vol. 32, no. 12, pp. 8142–8152, 2021

    2021

  14. [22]

    Spatial- temporal based multihead self-attention for remote sensing image change detection,

    Y . Zhou, F. Wang, J. Zhao, R. Yao, S. Chen, and H. Ma, “Spatial- temporal based multihead self-attention for remote sensing image change detection,” IEEE Transactions on Circuits and Systems for Video Tech- nology, vol. 32, no. 10, pp. 6615–6626, 2022

    2022

  15. [23]

    Change detection based on deep siamese convolutional network for optical aerial images,

    Y . Zhan, K. Fu, M. Yan, X. Sun, H. Wang, and X. Qiu, “Change detection based on deep siamese convolutional network for optical aerial images,” IEEE Geoscience and Remote Sensing Letters , vol. 14, no. 10, pp. 1845–1849, 2017

    2017

  16. [24]

    Fully Convolutional Siamese Networks for Change Detection,

    R. Caye Daudt, B. Le Saux, and A. Boulch, “Fully Convolutional Siamese Networks for Change Detection,” in 2018 25th IEEE Inter- national Conference on Image Processing (ICIP) , Oct. 2018, pp. 4063– 4067

    2018

  17. [25]

    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 Com- puting and Computer-Assisted Intervention – MICCAI 2015 , N. Navab, J. Hornegger, W. M. Wells, and A. F. Frangi, Eds. Cham: Springer Internationa...

    2015

  18. [26]

    Double u-net (w-net): A change detection network with two heads for remote sensing imagery,

    X. Wang, X. Yan, K. Tan, C. Pan, J. Ding, Z. Liu, and X. Dong, “Double u-net (w-net): A change detection network with two heads for remote sensing imagery,” International Journal of Applied Earth Observation and Geoinformation, vol. 122, p. 103456, 2023

    2023

  19. [27]

    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, 2022

    2022

  20. [28]

    VcT: Visual Change Transformer for Remote Sensing Image Change Detection,

    B. Jiang, Z. Wang, X. Wang, Z. Zhang, L. Chen, X. Wang, and B. Luo, “VcT: Visual Change Transformer for Remote Sensing Image Change Detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–14, 2023

    2023

  21. [29]

    Summator– Subtractor Network: Modeling Spatial and Channel Differences for Change Detection,

    L. Wang, Y . Fang, Z. Li, C. Wu, M. Xu, and M. Shao, “Summator– Subtractor Network: Modeling Spatial and Channel Differences for Change Detection,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–12, 2024

    2024

  22. [30]

    Deep- learning for change detection using multi-modal fusion of remote sensing images: A review,

    S. Saidi, S. Idbraim, Y . Karmoude, A. Masse, and M. Arbelo, “Deep- learning for change detection using multi-modal fusion of remote sensing images: A review,” Remote Sensing , vol. 16, no. 20, p. 3852, 2024

    2024

  23. [31]

    ChangeCLIP: Remote sensing change detection with multimodal vision-language representation learn- ing,

    S. Dong, L. Wang, B. Du, and X. Meng, “ChangeCLIP: Remote sensing change detection with multimodal vision-language representation learn- ing,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 208, pp. 53–69, Feb. 2024

    2024

  24. [32]

    A deep convolutional coupling network for change detection based on heterogeneous optical and radar images,

    J. Liu, M. Gong, K. Qin, and P. Zhang, “A deep convolutional coupling network for change detection based on heterogeneous optical and radar images,” IEEE transactions on neural networks and learning systems , vol. 29, no. 3, pp. 545–559, 2016

    2016

  25. [33]

    Learning Transferable Visual Models From Natural Language Supervi- sion,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning Transferable Visual Models From Natural Language Supervi- sion,” in Proceedings of the 38th International Conference on Machine...

    2021

  26. [34]

    A Feature Difference Convolutional Neural Network-Based Change Detection Method,

    M. Zhang and W. Shi, “A Feature Difference Convolutional Neural Network-Based Change Detection Method,” IEEE Transactions on Geo- science and Remote Sensing, vol. 58, no. 10, pp. 7232–7246, Oct. 2020

    2020

  27. [35]

    Very Deep Convolutional Networks for Large-Scale Image Recognition,

    K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” Apr. 2015

    2015

  28. [36]

    Change Detection Based on Deep Siamese Convolutional Network for Optical Aerial Images,

    Y . Zhan, K. Fu, M. Yan, X. Sun, H. Wang, and X. Qiu, “Change Detection Based on Deep Siamese Convolutional Network for Optical Aerial Images,” IEEE Geoscience and Remote Sensing Letters , vol. 14, no. 10, pp. 1845–1849, Oct. 2017

    2017

  29. [37]

    A Survey on Vision Transformer,

    K. Han, Y . Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y . Tang, A. Xiao, C. Xu, Y . Xu, Z. Yang, Y . Zhang, and D. Tao, “A Survey on Vision Transformer,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 1, pp. 87–110, Jan. 2023

    2023

  30. [38]

    MPViT: Multi-Path Vision Transformer for Dense Prediction,

    Y . Lee, J. Kim, J. Willette, and S. J. Hwang, “MPViT: Multi-Path Vision Transformer for Dense Prediction,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2022, pp. 7277– 7286

    2022

  31. [39]

    BiFormer: Vision Transformer with Bi-Level Routing Attention,

    L. Zhu, X. Wang, Z. Ke, W. Zhang, and R. Lau, “BiFormer: Vision Transformer with Bi-Level Routing Attention,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , Jun. 2023, pp. 10 323–10 333

    2023

  32. [40]

    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

  33. [41]

    Swin Transformer: Hierarchical Vision Transformer using Shifted Win- dows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin Transformer: Hierarchical Vision Transformer using Shifted Win- dows,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV). Montreal, QC, Canada: IEEE, Oct. 2021, pp. 9992–10 002

    2021

  34. [42]

    CSTSUNet: A Cross Swin Transformer-Based Siamese U- Shape Network for Change Detection in Remote Sensing Images,

    Y . Wu, L. Li, N. Wang, W. Li, J. Fan, R. Tao, X. Wen, and Y . Wang, “CSTSUNet: A Cross Swin Transformer-Based Siamese U- Shape Network for Change Detection in Remote Sensing Images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–15, 2023

    2023

  35. [43]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV , USA: IEEE, Jun. 2016, pp. 770–778

    2016

  36. [44]

    A Multiscale Cascaded Cross- Attention Hierarchical Network for Change Detection on Bitemporal Remote Sensing Images,

    X. Zhang, L. Wang, and S. Cheng, “A Multiscale Cascaded Cross- Attention Hierarchical Network for Change Detection on Bitemporal Remote Sensing Images,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–16, 2024

    2024

  37. [45]

    DGMA2-Net: A Difference-Guided Multiscale Aggregation Attention Network for Remote Sensing Change Detection,

    Z. Ying, Z. Tan, Y . Zhai, X. Jia, W. Li, J. Zeng, A. Genovese, V . Piuri, and F. Scotti, “DGMA2-Net: A Difference-Guided Multiscale Aggregation Attention Network for Remote Sensing Change Detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1– 16, 2024

    2024

  38. [46]

    DS-HyFA-Net: A Deeply Su- pervised Hybrid Feature Aggregation Network With Multiencoders for Change Detection in High-Resolution Imagery,

    Z. Ying, T. Xian, Y . Zhai, X. Jia, H. Zhang, J. Pan, P. Coscia, A. Genovese, V . Piuri, and F. Scotti, “DS-HyFA-Net: A Deeply Su- pervised Hybrid Feature Aggregation Network With Multiencoders for Change Detection in High-Resolution Imagery,” IEEE Transactions on Geoscience a...

    2024

  39. [47]

    Y . Zhai, W. Li, T. Xian, X. Jia, H. Zhang, Z. Tan, J. Zhou, J. Zeng, and C. L. Philip Chen, “CAS-Net: Comparison-Based Attention Siamese JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 15 Network for Change Detection With an Open High-Resolution UA V Image Dataset,”...

    2021

  40. [48]

    Efficient Adjacent Feature Harmonizer Network With UA V-CD+ Dataset for Remote Sensing Change Detection,

    Y . Zhai, J. Pan, H. Zhang, T. Xian, Y . Xu, P. Coscia, A. Genovese, V . Piuri, F. Scotti, and C. L. Philip Chen, “Efficient Adjacent Feature Harmonizer Network With UA V-CD+ Dataset for Remote Sensing Change Detection,” IEEE Transactions on Geoscience and Remote Sensing, vol....

    2024

  41. [49]

    Urbansam: Learning invariance-inspired adapters for segment anything models in urban construction,

    C. Li, D. Hong, B. Zhang, Y . Li, G. Camps-Valls, X. X. Zhu, and J. Chanussot, “Urbansam: Learning invariance-inspired adapters for segment anything models in urban construction,” arXiv preprint arXiv:2502.15199, 2025

    2025 arXiv

  42. [50]

    Tinyllava: A framework of small-scale large multimodal models,

    B. Zhou, Y . Hu, X. Weng, J. Jia, J. Luo, X. Liu, J. Wu, and L. Huang, “Tinyllava: A framework of small-scale large multimodal models,”arXiv preprint arXiv:2402.14289, 2024

    2024 arXiv

  43. [51]

    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

  44. [52]

    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

  45. [53]

    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, May 2020

    2020

  46. [54]

    Fully Convolutional Networks for Multisource Building Extraction From an Open Aerial and Satellite Imagery Data Set,

    S. Ji, S. Wei, and M. Lu, “Fully Convolutional Networks for Multisource Building Extraction From an Open Aerial and Satellite Imagery Data Set,” IEEE Transactions on Geoscience and Remote Sensing , vol. 57, no. 1, pp. 574–586, Jan. 2019

    2019

  47. [55]

    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

  48. [56]

    ICIF-Net: Intra-Scale Cross-Interaction and Inter-Scale Feature Fusion Network for Bitempo- ral Remote Sensing Images Change Detection,

    Y . Feng, H. Xu, J. Jiang, H. Liu, and J. Zheng, “ICIF-Net: Intra-Scale Cross-Interaction and Inter-Scale Feature Fusion Network for Bitempo- ral Remote Sensing Images Change Detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–13, 2022

    2022

  49. [57]

    A Transformer-Based Siamese Net- work for Change Detection,

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

    2022

  50. [58]

    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

  51. [59]

    Spatiotemporal Enhancement and Interlevel Fusion Network for Remote Sensing Images Change Detection,

    Y . Huang, X. Li, Z. Du, and H. Shen, “Spatiotemporal Enhancement and Interlevel Fusion Network for Remote Sensing Images Change Detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–14, 2024

    2024

  52. [60]

    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 , vol. 62, pp. 1–17, 2024

    2024

  53. [61]

    An augmented linear mixing model to address spectral variability for hyperspectral unmixing,

    D. Hong, N. Yokoya, J. Chanussot, and X. X. Zhu, “An augmented linear mixing model to address spectral variability for hyperspectral unmixing,” IEEE Transactions on Image Processing, vol. 28, no. 4, pp. 1923–1938, 2018

    1923

  54. [62]

    Image caption generator with clip interrogator,

    A. A. Jagtap, J. Musale, S. Nawale, P. Takate, I. Kale, and S. Wagh- mare, “Image caption generator with clip interrogator,” in 2025 IEEE International Conference on Interdisciplinary Approaches in Technology and Management for Social Innovation (IATMSI) , vol. 3. IEEE, 2025, pp. 1–6

    2025

  55. [63]

    mplug: Effective and efficient vision-language learning by cross-modal skip-connections,

    C. Li, H. Xu, J. Tian, W. Wang, M. Yan, B. Bi, J. Ye, H. Chen, G. Xu, Z. Cao et al., “mplug: Effective and efficient vision-language learning by cross-modal skip-connections,” arXiv preprint arXiv:2205.12005, 2022

    2022 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.