Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Self-Supervised Cross-Modal Text-Image Time Series Retrieval in Remote Sensing

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Self-supervised contrastive learning can retrieve bitemporal remote-sensing images from text queries and vice versa, with transformer-based temporal fusion giving the best results.

desk verdict Genuine new task and honest qualitative results, but the headline TFF superiority claim rests on n-gram overlap metrics that the authors themselves show can be gamed. read the letter →

arxiv 2501.19043 v2 pith:INHSQFTQ submitted 2025-01-31 cs.CV

Genc Hoxha , Olivér Angyal , Begüm Demir This is my paper · ORCID
classification cs.CV
keywords text-imagetimeseriesretrievalbitemporalimageremotesensingself-supervisedlearningcontrastivetransformerfeaturefusionchangecaptioningcross-modal
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces a new retrieval task for remote-sensing archives: cross-modal text-image time series retrieval, shortened to text-ITSR. The goal is to let a user search a collection of bitemporal image pairs (two images of the same area at different times) with a natural-language sentence, and conversely to retrieve sentences that describe a query image pair. To solve it, the authors propose a self-supervised method that encodes images and text with separate pretrained transformers and aligns them in a shared embedding space with a bidirectional contrastive loss. They compare three ways of combining the two images of a pair and report that a transformer-based fusion strategy, which lets patches from one date attend to patches from the other, gives the most accurate retrievals on two benchmark change-captioning datasets. The paper argues that explicit modeling of temporal change is what separates this method from an early-fusion baseline.

What carries the argument

The load-bearing component is the bitemporal fusion module: after a Vision Transformer extracts patch embeddings from each date, TFF computes the element-wise patch difference, feeds it as key and value into multi-head cross-attention with the original patch embeddings as query, and stacks several such fusion stages; a parallel global-feature fusion (GFF) option simply subtracts or concatenates class-token features. The fused image representation and a BERT sentence embedding are then projected by two feed-forward heads into a shared space trained with the bidirectional temperature-scaled contrastive loss used by CLIP. This machinery ties temporal change modeling to cross-modal alignment, which is the mechanism the paper credits for the retrieval gains.

What would settle it

Re-run the same experiments replacing n-gram caption-overlap metrics with a retrieval metric that is insensitive to caption phrasing, such as recall at rank k judged by human relevance or by matching the image-pair identity of the captions; if the early-fusion baseline's perfect no-change scores persist while its true relevance ranking is at chance, the reported superiority of the transformer-based fusion over the other strategies on change and no-change subsets would not be established.

Watch

Extended reading notes

Core claim

The central claim is that text-to-bitemporal-image and bitemporal-image-to-text retrieval can be learned without any labeled retrieval data, using only paired image sequences and captions, and that the quality of retrieval depends on how the two dates are fused. The authors show that a parameter-free subtraction of global image features is already a competitive fusion, but the strongest results come from transformer-based feature fusion (TFF), where patch-level embeddings of the two dates are differenced and then passed through multi-head cross-attention whose query is the concatenated bitemporal patch embeddings. With TFF the method outperforms both global-feature fusion variants and an early-fusion baseline across BLEU-1/4, METEOR, and ROUGE-L scores on the LEVIR-CC and Dubai CCD archives, with the largest gains on change-focused queries.

Load-bearing premise

The whole comparison rests on the assumption that BLEU, METEOR, and ROUGE-L between the query and the reference captions of the top-five retrieved items measure retrieval relevance; the paper itself shows that on the LEVIR-CC no-change subset the early-fusion baseline scores a perfect 1.000 on text-to-image retrieval by exploiting the fact that only five distinct no-change captions exist, so n-gram overlap can reward frequency rather than semantic match.

Editorial extensions

If this is right

  • Users can query change-focused image archives with free-text sentences such as "new buildings have been constructed" without needing an example image pair.
  • Retrieval in the reverse direction, returning captions for a query bitemporal image pair, works with the same trained model by ranking text embeddings by cosine similarity.
  • The transformer-based fusion strategy yields the largest advantage on change-related queries, suggesting that attending across dates is what captures semantic change.
  • The architecture and contrastive training are presented as a potential starting point for an image-time-series-language foundation model that could be fine-tuned for change captioning or change question answering.

Reading between the lines

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

  • The same contrastive framework could be extended to time series of more than two images by feeding each date's patch embeddings through the same cross-attention fusion, which the paper does not test but its architecture suggests.
  • Because n-gram metrics reward lexical overlap, a semantically driven evaluation such as human relevance judgments or embedding similarity might change the ranking among TFF, GFF, and EF, especially on no-change queries.
  • A testable extension is to add temporal or date embeddings to the text and image encoders so that queries like "the forest was gradually cleared" could distinguish gradual from abrupt changes; the present method is designed for short-term changes.
  • The method's ceiling on a given archive is set by caption diversity: LEVIR-CC's five repeated no-change captions make no-change retrieval artificially easy, and a better-balanced benchmark would be needed to trust cross-dataset conclusions.
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 / 5 minor

Summary. This paper introduces the task of cross-modal text-image time series retrieval (text-ITSR) for bitemporal remote sensing images, in which text sentences and bitemporal image pairs are embedded into a shared space through contrastive learning on paired caption data. The pipeline uses CLIP/BERT encoders and compares three fusion strategies for the bitemporal image content: global feature concatenation, global feature subtraction, and a transformer-based feature fusion (TFF) based on cross-attention over patch-level difference features; an early-fusion (EF) baseline is also evaluated. Experiments on LEVIR-CC and Dubai CCD report BLEU-1/4, METEOR, and ROUGE-L scores for text-to-image and image-to-text retrieval, and the authors conclude that TFF is the most effective fusion strategy. The paper includes publicly available code, a complexity analysis, and qualitative retrieval examples.

Significance. If the results hold, the paper makes a useful contribution by defining a new task (text-ITSR for bitemporal images) and by showing that a transformer-based temporal fusion improves cross-modal retrieval over simple feature stacking. The work is unusually transparent: it releases code, reports training/inference complexity, and explicitly acknowledges the degenerate no-change caption structure of LEVIR-CC in Section V-B-3. The task formulation and the qualitative examples (Figs. 3-6) are convincing evidence that the method retrieves semantically relevant items. The main weakness is that the quantitative foundation for the headline claim is narrower than the conclusions suggest: the evaluation metrics are lexical-overlap scores that the authors themselves show can be gamed, no uncertainty quantification is provided despite five random caption rounds, and the only comparator is a self-constructed early-fusion baseline.

major comments (4)
  1. [Section IV-C and Section V-B-3, Tables IV-VI] The quantitative claim that TFF is the most effective fusion strategy rests entirely on BLEU-1/4, METEOR, and ROUGE-L computed on the top-5 retrieved items, which are lexical-overlap caption metrics rather than retrieval-relevance scores. The authors themselves demonstrate the failure mode in Section V-B-3: on LEVIR-CC no-change T→I the EF baseline scores 1.000 because only five distinct no-change captions exist in the archive, and the same structural degeneracy inflates TFF's 0.951 score; these inflated values contaminate the full-query Table IV, where the TFF-vs-EF margins (e.g., T→I BLEU-4 0.568 vs 0.565) are not meaningful. The change-only results in Tables II and V mitigate this but exhibit small margins of roughly 1-5% and are still n-gram based, and no variance is reported across the five random caption selection rounds, so the observed differences may not be significant. The paper should report per-round variability and significance tests, add a retrieval-oriented metric such as recall@k based on the paired ground-truth annotations, and report the degenerate no-change T→I subset separately or with reweighted aggregation.
  2. [Section III-B and III-C, Eqs. (10)-(14)] For the TFF strategy, the fused representation f^X_i of Eq. (12) is a sequence of patch features of shape R^{T×2d}, but the projection head in Eq. (13) is stated to output a single vector F^X_i in R^{1×dF}; the aggregation step (e.g., mean pooling over patches, or a class token) that reduces the sequence to one vector is never described, and the initial fusion stage f^X_0 in Eq. (12) is not defined. Since the retrieval step in Section III-C ranks archive items by cosine similarity between single vectors, the method as written is incomplete and not fully reproducible without inspecting the code; please specify the pooling and the initialization.
  3. [Section V-A-3, Table III] The narrative misreads its own results: the text states that with TFF 'in this task... our method yields the highest retrieval performance (e.g., BLEU-4 = 0.408)' for the Dubai I→T no-change query set, but Table III shows the EF baseline achieving 0.460 BLEU-4 (and higher METEOR, ROUGE-L, and BLEU-1) for I→T in that same scenario. The paragraph is also mislabeled as the 'change query set scenario' and later asserts that TFF achieves the highest scores 'across the two tasks,' which holds only for the averaged columns. These contradictions should be corrected, and the no-change I→T behavior should be discussed consistently with the table.
  4. [Sections IV-B and V] The experimental comparison includes only a single, self-constructed baseline (early fusion), so the evidence supports the conclusion 'better than early feature stacking' rather than 'effective for text-ITSR.' Given the novelty claim and the importance of the new task, the paper should also compare with a stronger reference, such as CLIP applied directly to the concatenated or differenced bitemporal input without the proposed fusion, or an existing remote-sensing text-image retrieval method (e.g., [8], [12], [14]) extended to process bitemporal pairs; at least one such comparison is needed to substantiate the central effectiveness claim.
minor comments (5)
  1. [Section V-A-3] The no-change query set scenario is repeatedly described as the 'change query set scenario' in the opening sentence and in the Table III analysis.
  2. [Section III-B, Eqs. (7)-(10)] Equation (10) calls MultiHead with a single argument whereas Eq. (8) defines it with three arguments, and the roles of Q, K, and V are described inconsistently between Eqs. (7) and (9).
  3. [Section V-B-1 and V-B-3] Typos: 'GGF: Subtraction' should be 'GFF: Subtraction'; '4rth' should be '4th'; 'using TTF' should be 'using TFF'; and 'GFF concatenation' should be consistent with the 'GFF: Concatenation' form used elsewhere.
  4. [Section I and Fig. 4] Section I contains a stray punctuation artifact '(see Fig. 1, )' and unbalanced quotation marks around the example queries, and Fig. 4's caption lists two 'b)' panels.
  5. [Reference [22]] Reference [22] should carry the full title 'Attention is all you need' with the leading article.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the retrieval system is trained on paired data and evaluated on held-out archives, with no fitted quantity re-presented as a prediction and no load-bearing self-citation.

full rationale

The paper's central claim is empirical: a self-supervised contrastive model trained on paired bitemporal-image/text data generalizes to held-out retrieval, with the TFF fusion strategy scoring highest. The derivation chain is explicit: modality encoders (Eqs. 1-3), fusion operators (Eqs. 4-12), projection heads and the CLIP-style contrastive loss (Eqs. 13-17), followed by cosine-similarity ranking. Each block uses openly specified external backbones (CLIP, BERT, ViT) and standard losses; no training quantity is reused as the reported evaluation signal in a way that forces the conclusion by construction. Evaluation uses predefined splits and a leave-one-out archive (Section IV-B), and the BLEU/METEOR/ROUGE-L scores are computed between queries and reference captions of top-5 retrievals. The authors' own disclosure in Section V-B-3 that the EF baseline's perfect LEVIR no-change T→I scores follow from only five distinct no-change captions is a transparent metric-validity caveat, not evidence that TFF's scores are fitted inputs or that a self-citation supplies the conclusion. The self-citations ([1], [7], [36]) are contextual prior work and dataset provenance; none is invoked as a uniqueness theorem or to forbid an alternative approach. Therefore no specific circular reduction can be exhibited.

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

The paper introduces no new physical or mathematical entities. Its free parameters are standard network design choices and one data-sampling ratio; the central empirical claim depends on these choices, especially the 15% no-change sampling and the top-5 evaluation. The key domain assumptions are dataset representativeness, CLIP transferability, and the validity of n-gram metrics as retrieval performance measures.

free parameters (6)
  • LEVIR-CC no-change training fraction = 15%
    Only 15% of no-change image-text pairs are used in training to stabilize training; this hand-chosen ratio affects the no-change retrieval results and is a data selection decision.
  • TFF fusion stages = 3
    The number of fusion stages l in TFF is set to 3 without reported ablation, affecting TFF performance.
  • Projection head dimensions = hidden 256, output 128
    FFN hidden and output sizes are chosen by hand; no ablation or sensitivity analysis is reported.
  • Top-k retrieved = 5
    Evaluation computes metrics on top-5 retrieved items; the choice of k affects reported scores.
  • Learnable temperature kappa = trained from init 0.07
    The contrastive temperature is learned during training; it is a fitted parameter in the loss, though standard for CLIP-style objectives.
  • Batch size and learning rate = 32 / 0.01
    Standard hyperparameters, but chosen by hand without reported sensitivity analysis.
assumptions (5)
  • domain assumption The paired image-text datasets provide the supervision needed for contrastive alignment.
    The method assumes captions correctly describe changes in the bitemporal pairs, which is the training signal.
  • domain assumption CLIP encoders pre-trained on natural images transfer to remote sensing imagery.
    The method uses CLIP's ViT and text encoders as backbones without extensive RS-specific adaptation, relying on transferability.
  • domain assumption N-gram overlap metrics are valid proxies for cross-modal retrieval quality.
    BLEU/ROUGE/METEOR measure caption similarity, which the paper uses as retrieval evaluation; the authors note limitations in the no-change case.
  • domain assumption Images are co-registered and represent the same area at two times.
    The fusion strategies assume spatial correspondence between the two images in a pair.
  • standard math The transformer attention mechanism and contrastive loss formulas (Eq. 15-17) are standard and are used without modification.
    The paper relies on the standard scaled dot-product attention and CLIP-style bidirectional InfoNCE loss as background.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Supervised Cross-Modal Text-Image Time Series Retrieval in Remote Sensing." pith.science (2026). https://pith.science/paper/INHSQFTQ

@misc{pith2026250119043,
  author       = {Pith},
  title        = {Pith review of: Self-Supervised Cross-Modal Text-Image Time Series Retrieval in Remote Sensing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/INHSQFTQ}},
  note         = {Machine review of arXiv:2501.19043}
}
read the original abstract

The development of image time series retrieval (ITSR) methods is a growing research interest in remote sensing (RS). Given a user-defined image time series (i.e., the query time series), ITSR methods search and retrieve from large archives the image time series that have similar content to the query time series. Existing ITSR methods in RS are designed for unimodal retrieval problems, relying on an assumption that users always have access to a query image time series in the considered image modality. In operational scenarios, this assumption may not hold. To overcome this issue, as a first time in RS we introduce the task of cross-modal text-image time series retrieval (text-ITSR). In detail, we present a self-supervised cross-modal text-ITSR method that enables the retrieval of image time series using text sentences as queries, and vice versa. We focus our attention on text-ITSR in pairs of images (i.e., bitemporal images). Our text-ITSR method consists of two key components: 1) modality-specific encoders to model the semantic content of bitemporal images and text sentences with discriminative features; and 2) modality-specific projection heads to align textual and image representations in a shared embedding space. To effectively model the temporal information in the bitemporal images, we exploit two fusion strategies: i) global feature fusion (GFF) strategy that combines global image features through simple yet effective operators; and ii) transformer-based feature fusion (TFF) strategy that leverages transformers for fine-grained temporal integration. Extensive experiments conducted on two benchmark RS archives demonstrate the effectiveness of our method in accurately retrieving semantically relevant bitemporal images (or text sentences) to a query text sentence (or bitemporal image). The code of this work is publicly available at https://git.tu-berlin.de/rsim/cross-modal-text-tsir .

Figures

Figures reproduced from arXiv: 2501.19043 by the authors.

Figure 1
Figure 1. Overall architecture of the proposed text-ITSR method. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the feature fusion of bitemporal images. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. T → I retrieval example for Dubai CCD. a) query text. The retrieved bitemporal images by b) EF baseline method and our method when the: b) GFF: Concatenation, c) GFF: Subtraction and d) TFF strategies are used to model the semantic content of the bitemporal images. 3) No-change query set scenario: Table III shows the retrieval performances of the EF baseline method and our method under different fusion strategies on… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: I → T retrieval example for Dubai CCD. a) query bitemporal images. The retrieved text sentences by b) EF baseline method and our method when the: b) GFF: Concatenation, c) GFF: Subtraction and d) TFF strategies are used to model the semantic content of bitemporal image…
Figure 5
Figure 5. Figure 5: T → I retrieval example for LEVIR CC. a) query text and the retrieved bitemporal images by b) the EF baseline method and our method when the: c) GFF: Concatenation, d) GFF: Subtraction and e) TFF strategies are used to model the semantic content of bitemporal images. 3…
Figure 6
Figure 6. Figure 6: I → T retrieval example for LEVIR CC. a) query bitemporal images and retrieved text sentences by b) EF baseline method and our method when: c) GFF: Concatenation, d) GFF: Subtraction and e) TFF strategies are used to model the semantic content of bitemporal images. Wor…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Finding Change in Satellite Archives from Text: How to Combine Before-and-After Images Efficiently

    cs.CV 2026-07 accept novelty 5.5 of 10

    A training-free subtraction-then-attention cascade matches or beats full fusion recall on LEVIR-CC at 10–15× lower query cost; Mamba is no faster than attention at L=196; TBF cuts parameters 2.3× for a 0.007 BLEU-1 cost.

Reference graph

Works this paper leans on

39 extracted references · 32 canonical work pages · cited by 1 Pith paper

  1. [8]

    Textrs: Deep bidirectional triplet network for matching text to remote sensing images,

    T. Abdullah, Y . Bazi, M. M. Al Rahhal, M. L. Mekhalfi, L. Rangarajan, and M. Zuair, “Textrs: Deep bidirectional triplet network for matching text to remote sensing images,” Remote Sensing , vol. 12, no. 3, 2020

    work page 2020

  2. [12]

    Multilanguage transformer for improved text to remote sensing image retrieval,

    M. M. A. Rahhal, Y . Bazi, N. A. Alsharif, L. Bashmal, N. Alajlan, and F. Melgani, “Multilanguage transformer for improved text to remote sensing image retrieval,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 15, pp. 9115–9126, 2022

    work page 2022

  3. [14]

    Knowledge-aware text–image retrieval for remote sensing images,

    L. Mi, X. Dai, J. Castillo-Navarro, and D. Tuia, “Knowledge-aware text–image retrieval for remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–13, 2024

    work page 2024

  4. [1]

    A cluster-based appraoch to content based time series retrieval (cbtsr),

    F. Bovolo, B. Demir, and L. Bruzzone, “A cluster-based appraoch to content based time series retrieval (cbtsr),” in 2015 IEEE International Geoscience and Remote Sensing Symposium (IGARSS) , pp. 2793–2796, 2015

    work page 2015

  5. [2]

    Deep learning driven content-based image time-series retrieval in remote sens- ing archives,

    O. Vuran, O. Akcin, M. Ravanbakhsh, B. Sankur, and B. Demir, “Deep learning driven content-based image time-series retrieval in remote sens- ing archives,” in IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium , pp. 1940–1943, 2022

    work page 2022

  6. [3]

    A content-based remote sensing image change information retrieval model,

    C. Ma, W. Xia, F. Chen, J. Liu, Q. Dai, L. Jiang, J. Duan, and W. Liu, “A content-based remote sensing image change information retrieval model,” ISPRS International Journal of Geo-Information , vol. 6, no. 10, p. 310, 2017

    work page 2017

  7. [4]

    Deep learning in remote sensing: A comprehensive review and list of resources,

    X. X. Zhu, D. Tuia, L. Mou, G.-S. Xia, L. Zhang, F. Xu, and F. Fraundorfer, “Deep learning in remote sensing: A comprehensive review and list of resources,” IEEE Geoscience and Remote Sensing Magazine, vol. 5, no. 4, pp. 8–36, 2017

    2017

  8. [5]

    Artificial intelligence to advance earth observation: A review of models, recent trends, and pathways forward,

    D. Tuia, K. Schindler, B. Demir, X. X. Zhu, M. Kochupillai, S. D ˇzeroski, J. N. van Rijn, H. H. Hoos, F. Del Frate, M. Datcu, V . Markl, B. Le Saux, R. Schneider, and G. Camps-Valls, “Artificial intelligence to advance earth observation: A review of models, recent trends, and pathways forward,” IEEE Geoscience and Remote Sensing Magazine , pp. 2–25, 2024

    2024

Show all 39 references
  1. [6]

    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 Transac- tions on Geoscience and Remote Sensing , vol. 57, no. 6, pp. 3677–3693, 2019

    2019

  2. [7]

    Toward remote sensing image retrieval under a deep image captioning perspective,

    G. Hoxha, F. Melgani, and B. Demir, “Toward remote sensing image retrieval under a deep image captioning perspective,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 13, pp. 4462–4475, 2020

    2020

  3. [9]

    Deep unsupervised embedding for remote sensing image retrieval using textual cues,

    M. M. A. Rahhal, Y . Bazi, T. Abdullah, M. L. Mekhalfi, and M. Zuair, “Deep unsupervised embedding for remote sensing image retrieval using textual cues,” Applied Sciences , vol. 10, no. 24, p. 8931, 2020

    2020

  4. [10]

    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

  5. [11]

    Deep unsupervised contrastive hashing for large-scale cross-modal text-image retrieval in remote sensing,

    G. Mikriukov, M. Ravanbakhsh, and B. Demir, “Deep unsupervised contrastive hashing for large-scale cross-modal text-image retrieval in remote sensing,” arXiv preprint arXiv:2201.08125 , 2022

    2022 arXiv

  6. [13]

    Unsupervised con- trastive hashing for cross-modal retrieval in remote sensing,

    G. Mikriukov, M. Ravanbakhsh, and B. Demir, “Unsupervised con- trastive hashing for cross-modal retrieval in remote sensing,” in ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 4463–4467, 2022

    2022

  7. [15]

    Strong and weak prompt engineering for remote sensing image-text cross-modal retrieval,

    T. Sun, C. Zheng, X. Li, Y . Gao, J. Nie, L. Huang, and Z. Wei, “Strong and weak prompt engineering for remote sensing image-text cross-modal retrieval,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , pp. 1–12, 2025

    2025

  8. [16]

    Long Short-Term Memory,

    S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Computation, vol. 9, pp. 1735–1780, 11 1997

    1997

  9. [17]

    Dimensionality reduction by learning an invariant mapping,

    R. Hadsell, S. Chopra, and Y . LeCun, “Dimensionality reduction by learning an invariant mapping,” in 2006 IEEE Computer Society Con- ference on Computer Vision and Pattern Recognition (CVPR’06) , vol. 2, pp. 1735–1742, 2006

    2006

  10. [18]

    Deep metric learning using triplet network,

    E. Hoffer and N. Ailon, “Deep metric learning using triplet network,” in Similarity-Based Pattern Recognition (A. Feragen, M. Pelillo, and M. Loog, eds.), (Cham), pp. 84–92, Springer International Publishing, 2015

    2015

  11. [19]

    Exploring a fine-grained multiscale method for cross-modal remote sensing image retrieval,

    Z. Yuan, W. Zhang, K. Fu, X. Li, C. Deng, H. Wang, and X. Sun, “Exploring a fine-grained multiscale method for cross-modal remote sensing image retrieval,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–19, 2022

    2022

  12. [20]

    A lightweight multi-scale crossmodal text-image retrieval method in remote sensing,

    Z. Yuan, W. Zhang, X. Rong, X. Li, J. Chen, H. Wang, K. Fu, and X. Sun, “A lightweight multi-scale crossmodal text-image retrieval method in remote sensing,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–19, 2022

    2022

  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]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

    2017

  15. [23]

    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 International Conference on Machine Learning , 2021

    2021

  16. [24]

    Parameter-efficient transfer learning for remote sensing image–text retrieval,

    Y . Yuan, Y . Zhan, and Z. Xiong, “Parameter-efficient transfer learning for remote sensing image–text retrieval,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–14, 2023

    2023

  17. [26]

    Global–local information soft-alignment for cross-modal remote-sensing image–text retrieval,

    G. Hu, Z. Wen, Y . Lv, J. Zhang, and Q. Wu, “Global–local information soft-alignment for cross-modal remote-sensing image–text retrieval,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1– 15, 2024

    2024

  18. [27]

    Masking-based cross-modal remote sensing image–text retrieval via dynamic contrastive learning,

    Z. Zhao, X. Miao, C. He, J. Hu, B. Min, Y . Gao, Y . Liu, and K. Phark- suwan, “Masking-based cross-modal remote sensing image–text retrieval via dynamic contrastive learning,”IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–15, 2024

    2024

  19. [28]

    Integrating multisubspace joint learning with multilevel guidance for cross-modal retrieval of remote sensing images,

    Y . Chen, J. Huang, S. Xiong, and X. Lu, “Integrating multisubspace joint learning with multilevel guidance for cross-modal retrieval of remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–17, 2024

    2024

  20. [29]

    Text-to-event retrieval in aerial videos,

    L. Bashmal, S. M. A. Mehmadi, Y . Bazi, M. M. A. Rahhal, and M. Zuair, “Text-to-event retrieval in aerial videos,” IEEE Geoscience and Remote Sensing Letters , vol. 21, pp. 1–5, 2024

    2024

  21. [30]

    BERT: pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M. Chang, K. Lee, and K. Toutanova, “BERT: pre-training of deep bidirectional transformers for language understanding,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,...

    2019

  22. [31]

    An image is worth 16x16 words: Trans- formers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Trans- formers for image recognition at scale,” ArXiv, vol. abs/2010.11929, 2020

    2010 arXiv

  23. [32]

    Fully convolutional siamese networks for change detection,

    R. Caye Daudt, B. Le Saux, and A. Boulch, “Fully convolutional siamese networks for change detection,” in2018 25th IEEE International Conference on Image Processing (ICIP) , pp. 4063–4067, 2018. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 15

    2018

  24. [33]

    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 , pp. 2–27, 2025

    2025

  25. [34]

    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

  26. [35]

    Improved deep metric learning with multi-class n-pair loss objective,

    K. Sohn, “Improved deep metric learning with multi-class n-pair loss objective,” Advances in neural information processing systems , vol. 29, 2016

    2016

  27. [36]

    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

  28. [37]

    Urban change detection for multispectral earth observation using convolutional neural networks,

    R. C. Daudt, B. Le Saux, A. Boulch, and Y . Gousseau, “Urban change detection for multispectral earth observation using convolutional neural networks,” in IGARSS 2018 - 2018 IEEE International Geoscience and Remote Sensing Symposium , pp. 2115–2118, 2018

    2018

  29. [38]

    Bleu: a method for automatic evaluation of machine translation,

    K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu, “Bleu: a method for automatic evaluation of machine translation,” in Proceedings of the 40th Annual Meeting on Association for Computational Linguistics , ACL ’02, (USA), p. 311–318, Association for Computational Linguistics, 2002

    2002

  30. [39]

    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 Pro- ceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization (J. Goldstein, A. Lavie, C...

    2005

  31. [40]

    ROUGE: A package for automatic evaluation of summaries,

    C.-Y . Lin, “ROUGE: A package for automatic evaluation of summaries,” in Text Summarization Branches Out , (Barcelona, Spain), pp. 74–81, Association for Computational Linguistics, July 2004

    2004

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.