Pith. sign in

REVIEW 4 major objections 6 minor 60 references

Change Captioning in Remote Sensing: Evolution to SAT-Cap -- A Single-Stage Transformer Approach

T0 review · 4 major / 6 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 The paper proposes SAT-Cap, a single-stage transformer that jointly models spatial and channel attention to improve change captioning in remote sensing, reporting state-of-the-art CIDEr scores.

desk verdict A well-ablated single-stage change captioning model whose SOTA claim needs matched baselines and error bars, plus two internal fixes. read the letter →

arxiv 2501.08114 v1 pith:CS5SKNS2 submitted 2025-01-14 cs.CV

Yuduo Wang , Weikang Yu , Pedram Ghamisi This is my paper · ORCID
classification cs.CV
keywords remotesensingchangecaptioningvision-languagemodelstransformerspatial-channelattentionbi-temporalimagesimagedeeplearning
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 sets out to show that change captioning for remote sensing—turning a before-and-after image pair into a natural-language description of what changed—does not require the multi-stage feature fusion used by prior transformer models. It claims that a simpler single-stage architecture, SAT-Cap, built on a Spatial-Channel Attention Encoder that models both spatial relationships and channel-wise object semantics, extracts richer information and can fuse the two images with only a cosine-similarity-and-concatenation step. If correct, this means complex fusion is not a necessary ingredient for high-quality change captioning, and more compact models can match or beat heavier baselines. The paper reports CIDEr-D scores of 140.23% on LEVIR-CC and 97.74% on DUBAI-CCD, surpassing previous state-of-the-art methods on both benchmarks.

What carries the argument

The Spatial-Channel Attention Encoder (SCE) is a stack of transformer blocks that each pass the feature map first through a Spatial Attention Module (SAM), which applies multi-head self-attention over the $H \times W$ spatial positions, and then through a Channel Attention Module (CAM), which applies self-attention over the $C$ channels using a cross-covariance matrix and adds a depth-wise convolution branch on the value to capture local context. A convolutional feed-forward network (ConvFFN) with a split-and-multiply design halves the parameters of the second linear projection. The other load-bearing piece is Difference-Guided Fusion, which computes cosine similarity between the two encoder outputs, adds it into each feature map, concatenates them along the channel dimension, and passes the result through three convolutions plus a residual connection. The argument is that because channel attention captures object semantics and attributes, the fused representation is precise enough that no multi-stage fusion is needed.

What would settle it

Re-train SAT-Cap and Chg2Cap with identical code, backbone, training schedule, and evaluation script for at least five random seeds; if the mean CIDEr difference on either dataset falls below the standard deviation across seeds, or if re-evaluating the published baselines with the same script erases the gap, the claim that SAT-Cap surpasses the state of the art would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that jointly modeling spatial and channel dependencies in the transformer encoder is what allows a simple fusion strategy to outperform multi-stage fusion designs. SAT-Cap's encoder stacks Spatial Attention over image positions followed by Channel Attention over the channel dimension, so the refined feature maps carry both where objects are and what attributes they have. The Difference-Guided Fusion module then computes cosine similarity between the two feature maps, adds it back into each, concatenates them along the channel dimension, and refines the result with convolutional layers and a residual block. Against six prior methods on LEVIR-CC and DUBAI-CCD, the paper reports the best CIDEr-D, BLEU-4, and METEOR scores in most comparisons, including a 3.62-point CIDEr gain over Chg2Cap on LEVIR-CC and a 5.25-point gain on DUBAI-CCD. The paper interprets these results as evidence that channel attention supplies the semantic and attribute information that multi-stage fusion was previously used to recover.

Load-bearing premise

The load-bearing premise is that the reported CIDEr improvements over Chg2Cap (3.62 points on LEVIR-CC and 5.25 points on DUBAI-CCD) are genuine differences rather than run-to-run noise or artifacts of comparing against baselines trained under different schedules, backbones, or evaluation settings.

Editorial extensions

If this is right

  • If the central claim holds, multi-stage fusion modules in existing change-captioning models can be replaced by a cosine-similarity-plus-concatenation step without losing caption quality, reducing parameters and computational cost.
  • Future change-captioning encoders should include channel attention alongside spatial attention, since the paper's ablations show that SAM-only or CAM-only versions underperform the sequential SAM-to-CAM combination.
  • The reported numbers set a new reference point for the two standard benchmarks: 140.23% CIDEr-D on LEVIR-CC and 97.74% CIDEr-D on DUBAI-CCD.
  • The split-and-multiply ConvFFN shows that parameter count can be cut in the feed-forward block while improving accuracy, suggesting other transformer components may also be simplified without hurting performance.

Reading between the lines

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

  • The same single-stage fusion recipe may transfer to other bi-temporal vision-language tasks, such as change-focused visual question answering or change retrieval, because the mechanism is not caption-specific; the paper does not test that transfer.
  • The paper's attribution of gains to 'attributes' suggests a testable probe: if channel attention really encodes object attributes, then attention maps or channel importance scores should correlate with the adjectives and object nouns in the generated captions.
  • Because the comparisons use single runs against published baseline numbers, a fair re-run with identical training schedules and evaluation scripts across multiple seeds would reveal whether the 3.62- and 5.25-point CIDEr gaps are robust or within run-to-run noise.
  • DUBAI-CCD has only 500 image pairs, so the larger gain there (5.25 CIDEr points) may partly reflect small-data instability; validating on a third, larger dataset would strengthen the generalization claim.
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 / 6 minor

Summary. The paper introduces SAT-Cap, a transformer-based architecture for remote sensing change captioning. The model uses a ResNet-101 backbone, a spatial-channel attention encoder with semantic-enhanced mapping, spatial and channel attention modules, a convolutional feed-forward network, cosine-similarity-based difference-guided fusion, and a standard transformer decoder. Experiments on LEVIR-CC and DUBAI-CCD report CIDEr-D of 140.23% and 97.74%, respectively, and the paper claims state-of-the-art performance with simpler single-stage fusion than prior multi-stage approaches. The paper includes ablations of components, fusion methods, feature map sizes, encoder/decoder depth, local enhancement choices, and attention module order.

Significance. The proposed architecture is clearly described and the ablation study is unusually thorough, covering each module, fusion variant, feature size, and depth configuration. The central efficiency idea—single-stage cosine-similarity fusion instead of multi-stage fusion—is worth testing, and if the reported results are reproducible and robust, SAT-Cap would be a competitive RSICC baseline. The paper also promises code and pre-trained models, which will help the community. However, the paper's main claim of surpassing the state of the art is currently supported only by single-run test-set scores against baselines from other papers with different training protocols, so the significance is conditional on resolving the comparison issues below.

major comments (4)
  1. [§3.2, §3.3, Tables 1 and 2] The state-of-the-art comparison is not protocol-matched. Section 3.2 states that ResNet-101 is first fine-tuned on each target dataset and then used as the backbone for SAT-Cap, while the DUDA, MCCFormer, RSICCFormer, and Chg2Cap numbers in Tables 1 and 2 are quoted from their original papers, which do not use this protocol. Without re-running the baselines under the same training schedule, data splits, backbone initialization, and evaluation implementation, the reported CIDEr gains of 3.62 on LEVIR-CC and 5.25 on DUBAI-CCD cannot be attributed to the architecture. The DUBAI-CCD test set has only 150 pairs, so these margins may be within run-to-run noise; please provide matched baselines and multi-seed mean and standard deviation statistics.
  2. [§3.4.2, Table 4] The design choice for the cosine-similarity dimension is internally inconsistent. The text states that computing cosine similarity solely along the channel dimension achieves better results and selects that configuration, but Table 4 shows that 'along channel, height, and width dims' achieves CIDEr-D 140.53, which is higher than the selected 'along channel dim' value of 140.23. Please correct the table, the text, or the selection criterion; as written, the reported choice does not follow from the evidence.
  3. [§3.3, Table 1] The text claims that SAT-Cap outperforms the RSCaMa method by 0.58% in BLEU-4 score, but RSCaMa is not listed in Table 1 or anywhere in the quantitative comparisons. Either add the RSCaMa results to Table 1 with a proper citation, or remove the claim.
  4. [§2.4, Eq. (16)] The residual connection in the difference-guided fusion block has incompatible dimensions as written. F_f is defined as [Z1+Sim; Z2+Sim], so it has 2C channels, while F'_f is stated to be in R^{C×H×W}; adding F_f to the output of Conv3 is therefore not shape-consistent. Please specify the channel dimensions of Conv1–Conv3 and the residual branch, or revise the equation.
minor comments (6)
  1. [§3.3, Table 2, Abstract] The phrase 'surpassing current state-of-the-art methods' should be qualified, since SAT-Cap's BLEU-4 on DUBAI-CCD (40.80) is lower than Chg2Cap's 41.70; the SOTA claim is specifically for CIDEr-D.
  2. [§3.1.1, §2.2, §2.4, Table 8] Several typos should be fixed: 'iamge' should be 'image' in Section 3.1.1; 'SM module' should be 'SEM module' in Section 2.2; 'kernel sizeas' should be 'kernel sizes' in Section 2.4; and Table 8 contains the malformed value '.64.47' in the SE (r=16) BLEU-4 row.
  3. [§2.4, Figure 2] The module is called 'Difference-guided Fusion' although the described operation is cosine similarity followed by concatenation and a residual block; please clarify how the module is difference-guided or rename it for consistency.
  4. [Abstract, §1, Table 3] The computational-efficiency claim is supported only by parameter counts in Table 3; please add FLOPs, inference time, or a direct complexity comparison against multi-stage fusion baselines.
  5. [§3.3, Tables 1 and 2] Metric names are used inconsistently: the text refers to CIDEr while the tables report CIDEr-D; please standardize the terminology.
  6. [§3.4] All ablations are performed only on LEVIR-CC; reporting the key ablations on DUBAI-CCD as well would strengthen the generality of the conclusions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: SAT-Cap's claims are empirical benchmark results, not consequences of fitted parameters or self-cited theorems.

full rationale

The paper's central claim is an experimentally measured test-set score (CIDEr 140.23/97.74) obtained by training SAT-Cap and evaluating on fixed LEVIR-CC/DUBAI-CCD splits. No analytic prediction is derived from an input assumption that already contains the output. The fusion module is described as 'Inspired by Change2Captions [39]', and [39] shares an author (Pedram Ghamisi), but this is a design inspiration, not a load-bearing premise; the paper does not invoke [39] to define success or to force its conclusions. Ablation studies compare configurations empirically rather than by construction. The only concerns—baseline protocol mismatch, absent error bars, Table 4 inconsistency—are correctness/robustness issues, not circularity, and do not involve an equation reducing to its inputs. Under the stated rules, no circular step can be exhibited, so the appropriate score is 0.

Assumptions & free parameters 10 free parameters · 5 assumptions · 0 invented entities

This paper makes no analytic derivation; its central claim is a measured performance result. The ledger therefore records the training choices and background assumptions that the result rests on. The model itself has millions of learned weights, which are not counted as free parameters in the ledger sense; instead we list the architectural hyperparameters chosen via validation, since they directly affect the reported scores.

free parameters (10)
  • Number of encoder layers = 3
    Selected via ablation in Table 7; deeper encoders increased CIDEr up to 3 layers.
  • Number of decoder layers = 1
    Selected via ablation in Table 7; more decoder layers lowered CIDEr.
  • Feature map spatial size = 8x8
    Default ResNet-101 output; Table 6 shows 7x7 and 9x9 perform worse.
  • Cosine similarity computation dimension = channel
    Selected via ablation in Table 4; channel dimension yields the highest CIDEr.
  • Feature dimension = 512
    Set in Section 3.2; used in encoder and decoder.
  • FFN hidden dimension = 2048
    Set in Section 3.2.
  • Number of attention heads = 8
    Set in Section 3.2.
  • Batch size = 32 (LEVIR-CC), 8 (DUBAI-CCD)
    Set in Section 3.2; no ablation.
  • Learning rate = 1e-4
    Adam optimizer, Section 3.2.
  • Training epochs = 50
    Set in Section 3.2.
assumptions (5)
  • standard math Transformer self-attention and cross-attention work as described in Vaswani et al. [51]
    Used in SAM and decoder, standard in the field.
  • domain assumption Pre-trained ResNet-101 on ImageNet transfers to remote sensing features after fine-tuning
    Sections 2.2 and 3.2; if this transfer fails, the whole model degrades.
  • domain assumption CIDEr-D, BLEU, METEOR, and ROUGE-L are valid and comparable measures of caption quality
    Metrics used in Tables 1 and 2; comparability across papers is assumed.
  • domain assumption The LEVIR-CC and DUBAI-CCD pre-defined splits are random and representative
    Uses pre-defined partitions from prior work.
  • domain assumption Depth-wise convolution improves local feature modeling as in prior work
    Used in SEM, CAM, and ConvFFN; not derived here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Change Captioning in Remote Sensing: Evolution to SAT-Cap -- A Single-Stage Transformer Approach." pith.science (2026). https://pith.science/paper/CS5SKNS2

@misc{pith2026250108114,
  author       = {Pith},
  title        = {Pith review of: Change Captioning in Remote Sensing: Evolution to SAT-Cap -- A Single-Stage Transformer Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CS5SKNS2}},
  note         = {Machine review of arXiv:2501.08114}
}
read the original abstract

Change captioning has become essential for accurately describing changes in multi-temporal remote sensing data, providing an intuitive way to monitor Earth's dynamics through natural language. However, existing change captioning methods face two key challenges: high computational demands due to multistage fusion strategy, and insufficient detail in object descriptions due to limited semantic extraction from individual images. To solve these challenges, we propose SAT-Cap based on the transformers model with a single-stage feature fusion for remote sensing change captioning. In particular, SAT-Cap integrates a Spatial-Channel Attention Encoder, a Difference-Guided Fusion module, and a Caption Decoder. Compared to typical models that require multi-stage fusion in transformer encoder and fusion module, SAT-Cap uses only a simple cosine similarity-based fusion module for information integration, reducing the complexity of the model architecture. By jointly modeling spatial and channel information in Spatial-Channel Attention Encoder, our approach significantly enhances the model's ability to extract semantic information from objects in multi-temporal remote sensing images. Extensive experiments validate the effectiveness of SAT-Cap, achieving CIDEr scores of 140.23% on the LEVIR-CC dataset and 97.74% on the DUBAI-CC dataset, surpassing current state-of-the-art methods. The code and pre-trained models will be available online.

Figures

Figures reproduced from arXiv: 2501.08114 by the authors.

Figure 1
Figure 1. Examples of common remote sensing vision-language downstream tasks: (a) image captioning, (b) visual question [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of SAT-Cap for RSICC. 2 Methodology 2.1 Overview Our proposed SAT-Cap method comprises four main components: a feature extraction module, a spatial-channel attention encoder, a difference-guided fusion module, and a standard transformer-based caption decoder. The model takes as input a bi-temporal image pair along with the corresponding change caption. The overall flow of the model architecture is illus… view at source ↗
Figure 3
Figure 3. Illustration of SAM and CAM. where Convi(·)(i = 1, 2) are two standard 1 × 1 convolutional layers with BN and ReLU. DWConv(·) represents the 3 × 3 depth-wise convolution layer. Through SM module, we obtain a pair of locality-enhanced feature maps, which enhances the model’s ability to capture and model small objects.. 2.3 Spatial-Channel Attention Encoder The spatial-channel attention encoder is composed of several … view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Illustration of ConvFFN. where Cos(·, ·) can calculate the cosine similarity between two tensors. Convi(·)(i = 1, 2, 3) are three convolutional layers with BN and ReLU, their kernel sizeas are 1 × 1, 3 × 3, and 1 × 1. After obtaining the fused feature F ′ f ∈ R C×H×W ,…
Figure 5
Figure 5. Figure 5: Captioning results on the DUBAI-CCD dataset. Black sentence is one of the five ground truth sentences. The orange [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Captioning results on the LEVIR-CC dataset. Black sentence is one of the five ground truth sentences. The orange [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Typical captioning results on the DUBAI-CCD dataset. From left to right, the figure includes the pre-change image, [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Typical captioning results on the LEVIR-CC dataset. From left to right, the figure includes the pre-change image, post [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: The visualization of attended images along with the caption generation for processes for our method on the DUBAI-CCD [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: The visualization of attended images along with the caption generation for processes for our method on the LEVIR-CC [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 36 canonical work pages

  1. [1]

    Vision-language models in remote sensing: Current progress and future trends,

    X. Li, C. Wen, Y . Hu, Z. Yuan, and X. X. Zhu, “Vision-language models in remote sensing: Current progress and future trends,” IEEE Geoscience and Remote Sensing Magazine , 2024

    2024

  2. [2]

    Geochat: Grounded large vision- language model for remote sensing,

    K. Kuckreja, M. S. Danish, M. Naseer, A. Das, S. Khan, and F. S. Khan, “Geochat: Grounded large vision- language model for remote sensing,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 27 831–27 840

    work page 2024

  3. [3]

    Rs5m and georsclip: A large scale vision-language dataset and a large vision-language model for remote sensing,

    Z. Zhang, T. Zhao, Y . Guo, and J. Yin, “Rs5m and georsclip: A large scale vision-language dataset and a large vision-language model for remote sensing,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    work page 2024

  4. [4]

    Remote sensing temporal vision-language models: A comprehensive survey,

    C. Liu, J. Zhang, K. Chen, M. Wang, Z. Zou, and Z. Shi, “Remote sensing temporal vision-language models: A comprehensive survey,”arXiv preprint arXiv:2412.02573, 2024

    arXiv 2024

  5. [5]

    Earthgpt: A universal multi-modal large language model for multi-sensor image comprehension in remote sensing domain,

    W. Zhang, M. Cai, T. Zhang, Y . Zhuang, and X. Mao, “Earthgpt: A universal multi-modal large language model for multi-sensor image comprehension in remote sensing domain,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    work page 2024

  6. [6]

    Rsvg: Exploring data and models for visual grounding on remote sensing data,

    Y . Zhan, Z. Xiong, and Y . Yuan, “Rsvg: Exploring data and models for visual grounding on remote sensing data,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–13, 2023

    2023

  7. [7]

    Nwpu-captions dataset and mlca-net for remote sensing image captioning,

    Q. Cheng, H. Huang, Y . Xu, Y . Zhou, H. Li, and Z. Wang, “Nwpu-captions dataset and mlca-net for remote sensing image captioning,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–19, 2022

    work page 2022

  8. [8]

    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

Show all 60 references
  1. [9]

    Changeclip: Remote sensing change detection with multimodal vision- language representation learning,

    S. Dong, L. Wang, B. Du, and X. Meng, “Changeclip: Remote sensing change detection with multimodal vision- language representation learning,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 208, pp. 53–69, 2024. 15 SAT-Cap A PREPRINT

    2024

  2. [10]

    Show and tell: A neural image caption generator,

    O. Vinyals, A. Toshev, S. Bengio, and D. Erhan, “Show and tell: A neural image caption generator,” in Proceed- ings of the IEEE conference on computer vision and pattern recognition , 2015, pp. 3156–3164

    2015

  3. [11]

    High-resolution remote sensing image captioning based on structured attention,

    R. Zhao, Z. Shi, and Z. Zou, “High-resolution remote sensing image captioning based on structured attention,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–14, 2021

    2021

  4. [12]

    Sca-cnn: Spatial and channel-wise attention in convolutional networks for image captioning,

    L. Chen, H. Zhang, J. Xiao, L. Nie, J. Shao, W. Liu, and T.-S. Chua, “Sca-cnn: Spatial and channel-wise attention in convolutional networks for image captioning,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 5659–5667

    2017

  5. [13]

    Generating images from captions with attention,

    E. Mansimov, E. Parisotto, J. L. Ba, and R. Salakhutdinov, “Generating images from captions with attention,” arXiv preprint arXiv:1511.02793, 2015

    2015 arXiv

  6. [14]

    Txt2img-mhn: Remote sensing image generation from text using modern hopfield networks,

    Y . Xu, W. Yu, P. Ghamisi, M. Kopp, and S. Hochreiter, “Txt2img-mhn: Remote sensing image generation from text using modern hopfield networks,” IEEE Transactions on Image Processing, 2023

    2023

  7. [15]

    Text2earth: Unlocking text-driven remote sensing image genera- tion with a global-scale dataset and a foundation model,

    C. Liu, K. Chen, R. Zhao, Z. Zou, and Z. Shi, “Text2earth: Unlocking text-driven remote sensing image genera- tion with a global-scale dataset and a foundation model,” arXiv preprint arXiv:2501.00895, 2025

    2025 arXiv

  8. [16]

    Vqa: Visual question answer- ing,

    S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L. Zitnick, and D. Parikh, “Vqa: Visual question answer- ing,” in Proceedings of the IEEE international conference on computer vision , 2015, pp. 2425–2433

    2015

  9. [17]

    Rsadapter: Adapting multimodal models for remote sensing visual question answer- ing,

    Y . Wang and P. Ghamisi, “Rsadapter: Adapting multimodal models for remote sensing visual question answer- ing,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    2024

  10. [18]

    Rsvqa: Visual question answering for remote sensing data,

    S. Lobry, D. Marcos, J. Murray, and D. Tuia, “Rsvqa: Visual question answering for remote sensing data,” IEEE Transactions on Geoscience and Remote Sensing , vol. 58, no. 12, pp. 8555–8566, 2020

    2020

  11. [19]

    A spatial hierarchical reasoning network for remote sensing visual question answering,

    Z. Zhang, L. Jiao, L. Li, X. Liu, P. Chen, F. Liu, Y . Li, and Z. Guo, “A spatial hierarchical reasoning network for remote sensing visual question answering,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–15, 2023

    2023

  12. [20]

    Deep correlation for matching images and text,

    F. Yan and K. Mikolajczyk, “Deep correlation for matching images and text,” in Proceedings of the IEEE con- ference on computer vision and pattern recognition , 2015, pp. 3441–3450

    2015

  13. [21]

    Remote sensing cross-modal text- image retrieval based on global and local information,

    Z. Yuan, W. Zhang, C. Tian, X. Rong, Z. Zhang, H. Wang, K. Fu, and X. Sun, “Remote sensing cross-modal text- image retrieval based on global and local information,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–16, 2022

    2022

  14. [22]

    A deep semantic alignment network for the cross-modal image- text retrieval in remote sensing,

    Q. Cheng, Y . Zhou, P. Fu, Y . Xu, and L. Zhang, “A deep semantic alignment network for the cross-modal image- text retrieval in remote sensing,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 14, pp. 4284–4297, 2021

    2021

  15. [23]

    Generative adversarial text to image synthe- sis,

    S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and H. Lee, “Generative adversarial text to image synthe- sis,” in International conference on machine learning . PMLR, 2016, pp. 1060–1069

    2016

  16. [24]

    Diffusion models in vision: A survey,

    F.-A. Croitoru, V . Hondru, R. T. Ionescu, and M. Shah, “Diffusion models in vision: A survey,” IEEE Transac- tions on Pattern Analysis and Machine Intelligence , vol. 45, no. 9, pp. 10 850–10 869, 2023

    2023

  17. [25]

    National tree species map- ping using sentinel-1/2 time series and german national forest inventory data,

    L. Blickensd ¨orfer, K. Oehmichen, D. Pflugmacher, B. Kleinschmit, and P. Hostert, “National tree species map- ping using sentinel-1/2 time series and german national forest inventory data,” Remote Sensing of Environment , vol. 304, p. 114069, 2024

    2024

  18. [26]

    Evaluating the potential of sentinel-2 time series imagery and machine learning for tree species classification in a mountainous forest,

    P. Liu, C. Ren, Z. Wang, M. Jia, W. Yu, H. Ren, and C. Xia, “Evaluating the potential of sentinel-2 time series imagery and machine learning for tree species classification in a mountainous forest,” Remote Sensing, vol. 16, no. 2, p. 293, 2024

    2024

  19. [27]

    Gee xtract: High-quality remote sensing data preparation and extraction for multiple spatio-temporal ecological scaling,

    F. Valerio, S. Godinho, A. T. Marques, T. Crispim-Mendes, R. Pita, and J. P. Silva, “Gee xtract: High-quality remote sensing data preparation and extraction for multiple spatio-temporal ecological scaling,” Ecological In- formatics, vol. 80, p. 102502, 2024

    2024

  20. [28]

    Minenetcd: A benchmark for global mining change detection on remote sensing imagery,

    W. Yu, X. Zhang, X. X. Zhu, R. Gloaguen, and P. Ghamisi, “Minenetcd: A benchmark for global mining change detection on remote sensing imagery,” arXiv preprint arXiv:2407.03971, 2024

    2024 arXiv

  21. [29]

    Maskcd: A remote sensing change detection network based on mask classification,

    W. Yu, X. Zhang, S. Das, X. Xiang Zhu, and P. Ghamisi, “Maskcd: A remote sensing change detection network based on mask classification,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–16, 2024

    2024

  22. [30]

    Multilevel deformable attention-aggregated networks for change detection in bitemporal remote sensing imagery,

    X. Zhang, W. Yu, and M.-O. Pun, “Multilevel deformable attention-aggregated networks for change detection in bitemporal remote sensing imagery,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–18, 2022. 16 SAT-Cap A PREPRINT

    2022

  23. [31]

    Remote sensing image change captioning with dual-branch transformers: A new method and a large scale dataset,

    C. Liu, R. Zhao, H. Chen, Z. Zou, and Z. Shi, “Remote sensing image change captioning with dual-branch transformers: A new method and a large scale dataset,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–20, 2022

    2022

  24. [32]

    Change captioning: A new paradigm for multitemporal remote sensing image analysis,

    G. Hoxha, S. Chouaf, F. Melgani, and Y . Smara, “Change captioning: A new paradigm for multitemporal remote sensing image analysis,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–14, 2022

    2022

  25. [33]

    Robust change captioning,

    D. H. Park, T. Darrell, and A. Rohrbach, “Robust change captioning,” in Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 2019, pp. 4624–4633

    2019

  26. [34]

    Image change captioning by learning from an auxiliary task,

    M. Hosseinzadeh and Y . Wang, “Image change captioning by learning from an auxiliary task,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 2725–2734

    2021

  27. [35]

    Agnostic change captioning with cycle consistency,

    H. Kim, J. Kim, H. Lee, H. Park, and G. Kim, “Agnostic change captioning with cycle consistency,” in Proceed- ings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 2095–2104

    2021

  28. [36]

    Clip4idc: Clip for image difference captioning,

    Z. Guo, T.-J. J. Wang, and J. Laaksonen, “Clip4idc: Clip for image difference captioning,” AACL-IJCNLP 2022, p. 33, 2022

    2022

  29. [37]

    Remote sensing of alpine lake water environment changes on the tibetan plateau and surroundings: A review,

    C. Song, B. Huang, L. Ke, and K. S. Richards, “Remote sensing of alpine lake water environment changes on the tibetan plateau and surroundings: A review,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 92, pp. 26–37, 2014

    2014

  30. [38]

    Supporting global environmental change research: A review of trends and knowledge gaps in urban remote sensing,

    E. A. Wentz, S. Anderson, M. Fragkias, M. Netzband, V . Mesev, S. W. Myint, D. Quattrochi, A. Rahman, and K. C. Seto, “Supporting global environmental change research: A review of trends and knowledge gaps in urban remote sensing,” Remote Sensing, vol. 6, no. 5, pp. 3879–3905, 2014

    2014

  31. [39]

    Changes to captions: An attentive network for remote sensing change captioning,

    S. Chang and P. Ghamisi, “Changes to captions: An attentive network for remote sensing change captioning,” IEEE Transactions on Image Processing, 2023

    2023

  32. [40]

    Interactive change-aware transformer network for remote sensing image change captioning,

    C. Cai, Y . Wang, and K.-H. Yap, “Interactive change-aware transformer network for remote sensing image change captioning,” Remote Sensing, vol. 15, no. 23, p. 5611, 2023

    2023

  33. [41]

    Progressive scale-aware network for remote sensing image change captioning,

    C. Liu, J. Yang, Z. Qi, Z. Zou, and Z. Shi, “Progressive scale-aware network for remote sensing image change captioning,” in IGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium . IEEE, 2023, pp. 6668–6671

    2023

  34. [42]

    A decoupling paradigm with prompt learning for remote sensing image change captioning,

    C. Liu, R. Zhao, J. Chen, Z. Qi, Z. Zou, and Z. Shi, “A decoupling paradigm with prompt learning for remote sensing image change captioning,” IEEE Transactions on Geoscience and Remote Sensing , 2023

    2023

  35. [43]

    Pixel-level change detection pseudo-label learning for remote sensing change captioning,

    C. Liu, K. Chen, Z. Qi, H. Zhang, Z. Zou, and Z. Shi, “Pixel-level change detection pseudo-label learning for remote sensing change captioning,” arXiv preprint arXiv:2312.15311, 2023

    2023 arXiv

  36. [44]

    Rscama: Remote sensing image change captioning with state space model,

    C. Liu, K. Chen, B. Chen, H. Zhang, Z. Zou, and Z. Shi, “Rscama: Remote sensing image change captioning with state space model,” IEEE Geoscience and Remote Sensing Letters , 2024

    2024

  37. [45]

    Enhancing perception of key changes in remote sensing image change captioning,

    C. Yang, Z. Li, H. Jiao, Z. Gao, and L. Zhang, “Enhancing perception of key changes in remote sensing image change captioning,” arXiv preprint arXiv:2409.12612, 2024

    2024 arXiv

  38. [46]

    Ccexpert: Advancing mllm capability in remote sensing change captioning with difference-aware integration and a foundational dataset,

    Z. Wang, M. Wang, S. Xu, Y . Li, and B. Zhang, “Ccexpert: Advancing mllm capability in remote sensing change captioning with difference-aware integration and a foundational dataset,”arXiv preprint arXiv:2411.11360, 2024

    2024 arXiv

  39. [47]

    Change-agent: Towards interactive comprehensive remote sensing change interpretation and analysis,

    C. Liu, K. Chen, H. Zhang, Z. Qi, Z. Zou, and Z. Shi, “Change-agent: Towards interactive comprehensive remote sensing change interpretation and analysis,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    2024

  40. [48]

    Changeminds: Multi-task framework for detecting and describing changes in remote sensing,

    Y . Wang, W. Yu, M. Kopp, and P. Ghamisi, “Changeminds: Multi-task framework for detecting and describing changes in remote sensing,” arXiv preprint arXiv:2410.10047, 2024

    2024 arXiv

  41. [49]

    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,” IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  42. [50]

    Deep residual learning for image recognition,

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

    2016

  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]

    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

  45. [53]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018. 17 SAT-Cap A PREPRINT

    2018 arXiv

  46. [54]

    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

  47. [55]

    Adam: A method for stochastic optimization,

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

    2014 arXiv

  48. [56]

    Bleu: a method for automatic evaluation of machine trans- lation,

    K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu, “Bleu: a method for automatic evaluation of machine trans- lation,” in Proceedings of the 40th annual meeting of the Association for Computational Linguistics , 2002, pp. 311–318

    2002

  49. [57]

    Rouge: A package for automatic evaluation of summaries,

    C.-Y . Lin, “Rouge: A package for automatic evaluation of summaries,” in Text summarization branches out , 2004, pp. 74–81

    2004

  50. [58]

    Meteor: An automatic metric for mt evaluation with improved correlation with human judgments,

    S. Banerjee and A. Lavie, “Meteor: An automatic metric for mt evaluation with improved correlation with human judgments,” in Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, 2005, pp. 65–72

    2005

  51. [59]

    Cider: Consensus-based image description evaluation,

    R. Vedantam, C. Lawrence Zitnick, and D. Parikh, “Cider: Consensus-based image description evaluation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2015, pp. 4566–4575

    2015

  52. [60]

    Describing and localizing multiple changes with transformers,

    Y . Qiu, S. Yamamoto, K. Nakashima, R. Suzuki, K. Iwata, H. Kataoka, and Y . Satoh, “Describing and localizing multiple changes with transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1971–1980. 18

    2021

Pith tools

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