Pith. sign in

REVIEW 3 major objections 4 minor 52 references

DARAD: Dual Adapters and Ranking-Aware Distillation for Continual Remote Sensing Image-Text Retrieval

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Continual satellite-data retrieval can learn new stages without losing old ranking order.

desk verdict A solid continual-retrieval paper with a confounded headline forgetting metric; the direct ranking-consistency evidence (Table 4) is the strongest part. read the letter →

arxiv 2608.06059 v1 pith:E7IMJ3V7 submitted 2026-08-06 cs.CV

Xi Chen , Xu Chen , Xiangyang Jia , Wei Wang , Xu Zhang , Zhenyuan Sun This is my paper · ORCID
classification cs.CV
keywords continuallearningremotesensingimage-textretrievalrankingdistillationadapterscatastrophicforgettingvision-languagemodelscross-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

The paper tries to establish that a continual image-text retrieval system for remote sensing can absorb newly arriving data without erasing the retrieval rankings it learned earlier. It proposes DARAD, which combines a spatial fusion adapter for scale-aware visual adaptation, multi-expert semantic routing for text adaptation, and bidirectional ranking distillation that preserves historical cross-modal ranking structure. On a six-stage benchmark built from remote sensing data, DARAD reports the best current-stage recall from task 2 onward while keeping its historical-change score positive throughout, unlike feature-level distillation baselines that go negative. A reader should care because this is the difference between a retrieval archive that degrades as it grows and one that can be updated indefinitely.

What carries the argument

The argument is carried by three components. The spatial fusion adapter pools fine patch tokens into coarse regional tokens and lets a global token attend to both, producing a scale-aware residual that is injected into the base visual representation through a zero-initialized learnable gate. The multi-expert semantic routing keeps one always-active expert for shared textual semantics and routes each query to the top-2 of seven auxiliary low-rank experts, so new terminology updates only specialized residuals. The load-bearing piece is bidirectional ranking distillation: for normalized feature matrices $X$ and $Y$ it forms the relation matrix $S(X,Y)=XY^\top/\tau$, then minimizes mean-squared error between current-model and frozen-teacher relation matrices for text-to-image, image-to-text, and anchor-internal pairs. This is what preserves historical ranking order rather than absolute features.

What would settle it

Run DARAD with a stage-1 schedule strong enough to match the strongest baseline's Task 1 recall, then re-evaluate F-mR at tasks 2–6; if any value turns negative, the claimed forgetting-free behavior is an artifact of slow-start underfitting.

Watch

Extended reading notes

Core claim

The central claim is that what must be preserved during continual retrieval is not the exact feature values but the relative ranking of cross-modal candidates. DARAD's bidirectional ranking distillation (BRD) aligns three temperature-scaled relation matrices — text-to-image, image-to-text, and anchor-internal — between the current model and a frozen teacher, using a compact bank of historical anchors. Aligning these matrices, rather than Euclidean feature distances, allows features to drift while retrieval order stays intact. Evidence is the stage-wise protocol on RST2I-110K: DARAD obtains positive F-mR values from +0.24 to +1.01 across tasks 2–6 and the best C-mR from task 2 onward, and it outperforms replay combined with feature-level distillation on Kendall's tau, Spearman's rho, and Overlap@10.

Load-bearing premise

The forgetting score is measured against the model's own earlier recall on that stage's queries, so if an early stage underfits, later improvements count as 'no forgetting' even when they are just catch-up.

Editorial extensions

If this is right

  • A retrieval archive can be updated stage by stage without a negative historical-change score: DARAD reports F-mR of +0.24 to +1.01 across tasks 2–6 on the six-stage remote sensing protocol.
  • Ranking-aware distillation preserves historical retrieval order better than feature-level distillation: over replay plus feature KD, BRD raises Kendall's tau by 0.077 in T2I and 0.130 in I2T at the same anchor budgets.
  • The two adapters are responsible for new-data adaptation: removing SFA or MSR keeps positive F-mR only because BRD remains, while lowering current-stage recall.
  • The approach transfers across semantic-stage orders and to a cross-dataset stream, where average F-mR stays near zero (+0.01).

Reading between the lines

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

  • If ranking preservation, not feature preservation, is the operative mechanism, the same BRD objective could apply to other ranking-based retrieval domains, such as video or product search, wherever candidate order matters.
  • A testable extension: match stage-1 recall to the strongest baseline before measuring F-mR, to separate true retention from catch-up after a slow start.
  • The anchor bank is selected by herding; alternative exemplar-selection strategies might improve ranking preservation with smaller budgets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes DARAD, a continual remote sensing image-text retrieval (RS-ITR) framework that combines a spatial fusion adapter (SFA) for scale-aware visual adaptation, a multi-expert semantic routing (MSR) module for text adaptation, and a bidirectional ranking distillation (BRD) loss that aligns relation matrices between a frozen teacher and the current model on historical anchors. The method is evaluated on RST2I-110K with a six-stage semantic continual protocol, plus three standard benchmarks, and compared against retrieval-specific and continual-learning baselines. The central claim is that BRD preserves historical cross-modal ranking structure, allowing DARAD to obtain the best current-stage mR (C-mR) from Task 2 to Task 6 while maintaining positive historical-change mR (F-mR) values across all stages.

Significance. If the retention claim holds, the paper addresses a real and under-studied problem: continual retrieval under evolving RS archives, where feature-level distillation is insufficient because retrieval depends on ranking, not absolute feature values. The work has concrete strengths: a clearly specified multi-stage protocol, five-run standard deviations, ablation studies, two additional stage orders, a cross-dataset stream, and a direct same-gallery ranking-consistency evaluation (Table 4) showing statistically significant gains of BRD over replay plus feature distillation. The mechanism analysis with Kendall's tau, Spearman's rho, and Overlap@10 provides evidence that BRD preserves ranking order better than feature-level distillation. However, the headline retention metric F-mR (Eq. 7) conflates model change with gallery expansion and is further affected by DARAD's low Task 1 C-mR, so the central claim needs additional gallery-controlled evidence before the results can be fully accepted.

major comments (3)
  1. [Section 4.1, Eq. (7)] The F-mR metric compares a_{t,k} = mR(M_t, Q_k, G_{\le t}) with a_{k,k} = mR(M_k, Q_k, G_{\le k}). Since the retrieval gallery expands from G_{\le k} to G_{\le t}, adding candidates typically reduces recall for a fixed model, so a positive F-mR is a joint statement about model improvement and gallery enlargement, not a clean measure of forgetting or retention. This issue is compounded by DARAD's Task 1 C-mR of 12.16 versus SeqFT's 15.33 (Table 2), suggesting that the stage-1 baseline a_{1,1} may reflect underfitting rather than a mature retrieval model; later increases on Q_1 may therefore be catch-up learning rather than preservation of a historical ranking. Please report a gallery-controlled version of the retention metric (e.g., evaluate M_t on Q_k against G_{\le k} and compare with M_k on G_{\le k}) and, if possible, quantify the slow-start effect. Table 4 provides same-gallery ranking consistency, but the headline claim in Section 4.2 relies on Eq. (7) and Table 2, where the metric conflates model change with gallery change.
  2. [Section 4.3, Table 3] The ablation removing BRD reports negative F-mR values, but given the metric issue in Eq. (7), this does not directly isolate ranking preservation: the negative values could reflect feature drift, catch-up effects, or interactions with the expanding gallery. The same-gallery ranking metrics in Table 4 are more direct evidence for BRD, but the ablation table and the accompanying text state that 'BRD is critical for preserving the historical ranking structure' based on F-mR alone. Please provide a gallery-controlled recall comparison (for example, a_{t,k} with G_{\le k}) for the w/o BRD and full DARAD variants, so that the retention claim is not dependent on the confounded F-mR definition.
  3. [Section 4.2, Table 1 and Table 2] The comparison with prior continual learning methods is informative, but the reader should be told whether the baselines use the same base model (RemoteCLIP ViT-B/32) and whether their training budgets (epochs, batches, learning rates) are matched. In Table 2, SeqFT and SeqLoRA are reported with RemoteCLIP-like C-mR values, while other VLM-CL methods (C-CLIP, GIFT, MG-CLIP, ConDU) show much lower Task 1 C-mR, which suggests substantial differences in initialization or training protocols. A short paragraph stating the base encoder, initialization, and any hyper-parameter matching for each baseline would strengthen the comparison and make the reported gains in C-mR and F-mR more interpretable.
minor comments (4)
  1. [Section 4.4, Figure 4] The interpretation of expert routing ("E1 and E8 contain mixed colors, suggesting that they capture shared and relation-oriented semantics") is qualitative; adding quantitative routing statistics, such as the distribution of routed queries across experts per stage, would make the claim more concrete.
  2. [Section 4.2, Appendix C.5] The cross-dataset stream result (47.46 mR and +0.01 Avg. F-mR) is reported in prose but not shown in a table; please include the full per-task table and the definition of the cross-dataset stream, since this is a valuable robustness check.
  3. [Section 3.4, Eq. (5)-(6)] BRD uses mean squared error between temperature-scaled relation matrices. This is a sensible surrogate, but the paper should state explicitly that MSE on scores does not directly optimize ranking order; a brief remark that ranking metrics (Table 4) confirm the surrogate's effect would preempt confusion.
  4. [Section 4.1] The definition of Avg. C-mR and Avg. F-mR across the six stages is not written out; please define these aggregates explicitly in the metrics paragraph.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivational circularity: DARAD's components and BRD are defined as training objectives, and the central claims are empirical measurements against external baselines; the author-defined F-mR metric conflates model change with gallery expansion, but this is a validity caveat, not a circular proof.

full rationale

The paper's central claims are empirical comparisons under a fixed continual retrieval protocol, with DARAD's objectives (Eqs. 1-6) defined independently of the evaluation metrics (Eq. 7). BRD minimizes mean squared error between temperature-scaled relation matrices on current batches and historical anchors; it does not directly optimize C-mR or F-mR, so the reported positive F-mR values are not forced by construction. No parameter is fitted to a subset of data and then renamed as a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work. The F-mR metric in Eq. 7 does compare models under different gallery sizes (G<=t versus G<=k), which weakens the retention claim but is a measurement-validity issue rather than circularity. The main benchmark RST2I-110K is cited to Zhang, Huang, and Zhang 2026, which may overlap with an author, but the paper also evaluates on external benchmarks (RSICD, RSITMD, UCM-Captions), tests additional stage orders, and reports a cross-dataset stream, so any self-citation is not load-bearing. Overall, the derivation chain is self-contained and the central results rest on measured behavior, not on equivalence to inputs.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on assumptions about the pretrained alignment space, the validity of MSE-based relation distillation as a ranking proxy, the meaningfulness of the F-mR forgetting metric, the sufficiency of herded anchor banks, and the representativeness of the semantic-stage protocol. No new physical or external entities are introduced. The free-parameter list captures the hand-chosen hyperparameters that the ablation and comparison results depend on; standard network weights are omitted.

free parameters (7)
  • BRD loss weights lambda_T2I, lambda_I2T, lambda_A = not stated in main text (Appendix B.3)
    Manually chosen balancing coefficients for the three relation-matrix distillation losses; central to the final L_BRD objective in Eq. 6.
  • Temperature tau for relation matrix = not stated in main text
    Temperature in Eq. 4 scales the similarity matrices before MSE distillation; affects the softness of the ranking constraint.
  • Anchor bank size (1,000 per stage, herding) = 1000
    Number of historical samples stored per stage; the paper tests fixed budgets of 1,000 and 2,500, so the choice matters for the memory-retention trade-off.
  • Number of experts and routing choice = 8 experts, 1 main + 7 auxiliary, Top-2 routing
    Architecture hyperparameter chosen by hand; routing sparsity directly controls how text residuals are separated.
  • Low-rank dimension for MSR experts = 16
    Rank of the low-rank residual transformations in the text branch; capacity control for textual adaptation.
  • SFA coarse spatial grid = 3x3 with 8-head cross-attention
    Coarse-to-fine pooling and attention configuration in the visual adapter; affects scale-aware cue integration.
  • Learning rates and batch size = 3e-6 base, 1e-5 adapters, batch 64
    Optimization hyperparameters chosen by hand; not reported with sensitivity analysis in the main text.
assumptions (5)
  • domain assumption The pretrained CLIP/RemoteCLIP dual encoders provide a valid shared retrieval space, and residual adapters zero-initialized to the base representation limit visual and textual drift.
    Invoked in Eq. 1 and Eq. 3: the updated representation is the base representation plus learned residuals. If the base alignment space is not a good retrieval anchor, the whole adapter design loses its justification.
  • domain assumption Mean squared error between temperature-scaled relation matrices is a sufficient proxy for preserving retrieval ranking.
    Section 3.4 instantiates D_rel as MSE over S(X,Y). MSE on similarity scores does not directly optimize pairwise ranking, so the claim that ranking structure is preserved rests on this proxy holding empirically.
  • domain assumption The F-mR signed-change metric correctly quantifies historical forgetting.
    Eq. 7 defines F-mR as the average signed change from each stage's original checkpoint. If the stage-k baseline is under-trained, a positive F-mR can be achieved by later catch-up rather than by true retention.
  • domain assumption A historical anchor bank of 1,000 herded samples per stage represents the old-query distribution well enough for distillation.
    Section 4.1 stores 1,000 anchors per stage by herding. The ranking distillation only constrains relations involving these anchors, so the fidelity of the anchor set to the full historical gallery is load-bearing.
  • domain assumption The RST2I-110K semantic-stage splits and the unified all-seen gallery are representative of real evolving RS archives.
    Section 4.1 defines the six semantic stages and the all-seen gallery. The paper's main conclusions are evaluated on this author-arranged protocol, so the protocol's representativeness is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DARAD: Dual Adapters and Ranking-Aware Distillation for Continual Remote Sensing Image-Text Retrieval." pith.science (2026). https://pith.science/paper/E7IMJ3V7

@misc{pith2026260806059,
  author       = {Pith},
  title        = {Pith review of: DARAD: Dual Adapters and Ranking-Aware Distillation for Continual Remote Sensing Image-Text Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E7IMJ3V7}},
  note         = {Machine review of arXiv:2608.06059}
}
read the original abstract

With the rapid growth of Earth observation technologies, remote sensing archives are rapidly expanding, making remote sensing image-text retrieval (RS-ITR) increasingly important. However, continual RS-ITR remains challenging because scale variation and distribution shifts in RS aggravate cross-modal alignment space distortion, making it difficult for existing continual learning (CL) methods to support reliable continual retrieval. To address this challenge, we propose DARAD, a dual-adapter and ranking-aware distillation framework that preserves the historical cross-modal ranking structure while learning new visual and textual concepts from evolving archives. Specifically, the visual branch introduces a spatial fusion adapter, which integrates coarse regional cues and fine-grained patch cues to accommodate RS scale variation while anchoring visual updates to the pretrained alignment space. The textual branch employs multi-expert semantic routing, which separates shared textual semantics from semantically specialized residuals to absorb newly emerging descriptions while constraining global text embedding drift. Furthermore, bidirectional ranking distillation uses a frozen teacher model and historical anchors to preserve the historical cross-modal ranking structure, thereby mitigating alignment space distortion across continual stages. Experiments under a multi-stage continual retrieval protocol show that DARAD achieves superior performance over existing CL methods, improving adaptation to newly arrived data while maintaining effectiveness on historical data.

Figures

Figures reproduced from arXiv: 2608.06059 by the authors.

Figure 1
Figure 1. Current-stage mR (C-mR) over stages, with green [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of DARAD. SFA and MSR provide scale-aware visual and semantically routed textual adaptation, respec [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Spatial response comparison of the original images, [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 47 canonical work pages

  1. [1]

    International Conference on Machine Learning , pages =

    Learning Transferable Visual Models From Natural Language Supervision , author =. International Conference on Machine Learning , pages =

    work page

  2. [2]

    International Conference on Learning Representations , year =

    LoRA: Low-Rank Adaptation of Large Language Models , author =. International Conference on Learning Representations , year =

  3. [3]

    European Conference on Computer Vision , pages =

    Visual Prompt Tuning , author =. European Conference on Computer Vision , pages =

    work page

  4. [4]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    MaPLe: Multi-Modal Prompt Learning , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    work page

  5. [5]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2024 , doi =

    work page 2024

  6. [6]

    International Conference on Learning Representations , year =

    C-CLIP: Multimodal Continual Learning for Vision-Language Model , author =. International Conference on Learning Representations , year =

    work page

  7. [7]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Synthetic Data is an Elegant GIFT for Continual Vision-Language Models , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    work page

  8. [8]

    IEEE/CVF International Conference on Computer Vision , pages =

    Mind the Gap: Preserving and Compensating for the Modality Gap in CLIP-Based Continual Learning , author =. IEEE/CVF International Conference on Computer Vision , pages =

    work page

Show all 52 references
  1. [9]

    International Conference on Learning Representations , year =

    Enhanced Continual Learning of Vision-Language Models with Model Fusion , author =. International Conference on Learning Representations , year =

  2. [10]

    IEEE/CVF International Conference on Computer Vision , pages =

    Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models , author =. IEEE/CVF International Conference on Computer Vision , pages =

  3. [11]

    International Conference on Machine Learning , pages =

    Continual Vision-Language Representation Learning with Off-Diagonal Information , author =. International Conference on Machine Learning , pages =

  4. [12]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    GeoChat: Grounded Large Vision-Language Model for Remote Sensing , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  5. [13]

    IEEE Transactions on Geoscience and Remote Sensing , year =

    EarthGPT: A Universal Multi-modal Large Language Model for Multi-sensor Image Comprehension in Remote Sensing Domain , author =. IEEE Transactions on Geoscience and Remote Sensing , year =

  6. [14]

    arXiv preprint arXiv:2401.09712 , year =

    SkyEyeGPT: Unifying Remote Sensing Vision-Language Tasks via Instruction Tuning with Large Language Model , author =. arXiv preprint arXiv:2401.09712 , year =

    arXiv

  7. [15]

    Remote Sensing , volume =

    RS-LLaVA: A Large Vision-Language Model for Joint Captioning and Question Answering in Remote Sensing Imagery , author =. Remote Sensing , volume =. 2024 , doi =

    2024

  8. [16]

    Advances in Neural Information Processing Systems , volume =

    CogVLM: Visual Expert for Pretrained Language Models , author =. Advances in Neural Information Processing Systems , volume =

  9. [17]

    IEEE Geoscience and Remote Sensing Magazine , year =

    Foundation Models for Remote Sensing: A Survey , author =. IEEE Geoscience and Remote Sensing Magazine , year =

  10. [18]

    arXiv preprint arXiv:2508.04227 , year =

    Continual Learning for VLMs: A Survey and Taxonomy Beyond Forgetting , author =. arXiv preprint arXiv:2508.04227 , year =

    arXiv

  11. [19]

    International Journal of Computer Vision , volume =

    Learning to Prompt for Vision-Language Models , author =. International Journal of Computer Vision , volume =. 2022 , doi =

    2022

  12. [20]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    RemoteCLIP: A Vision Language Foundation Model for Remote Sensing , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2024 , doi =

    2024

  13. [21]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    RS5M and GeoRSCLIP: A Large-Scale Vision- Language Dataset and a Large Vision-Language Model for Remote Sensing , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2024 , doi =

    2024

  14. [22]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    SkyScript: A Large and Semantically Diverse Vision-Language Dataset for Remote Sensing , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2024 , doi =

    2024

  15. [23]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Any2RSI: Controllable Remote Sensing Text-to-Image Generation via Any Control and Enriched Description , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2026 , doi =

    2026

  16. [24]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Exploring Models and Data for Remote Sensing Image Caption Generation , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2018 , doi =

    2018

  17. [25]

    2016 International Conference on Computer, Information and Telecommunication Systems (CITS) , pages =

    Deep semantic understanding of high resolution remote sensing image , author =. 2016 International Conference on Computer, Information and Telecommunication Systems (CITS) , pages =. 2016 , doi =

    2016

  18. [26]

    Proceedings of the 18th SIGSPATIAL International Conference on Advances in Geographic Information Systems , pages =

    Bag-of-visual-words and spatial extensions for land-use classification , author =. Proceedings of the 18th SIGSPATIAL International Conference on Advances in Geographic Information Systems , pages =. 2010 , doi =

    2010

  19. [27]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Exploring a Fine-Grained Multiscale Method for Cross-Modal Remote Sensing Image Retrieval , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2022 , doi =

    2022

  20. [28]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Remote Sensing Cross-Modal Text-Image Retrieval Based on Global and Local Information , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2022 , doi =

    2022

  21. [29]

    Proceedings of the 31st ACM International Conference on Multimedia , pages =

    A Prior Instruction Representation Framework for Remote Sensing Image-text Retrieval , author =. Proceedings of the 31st ACM International Conference on Multimedia , pages =. 2023 , doi =

    2023

  22. [30]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages =

    Accurate and Lightweight Learning for Specific Domain Image-Text Retrieval , author =. Proceedings of the 32nd ACM International Conference on Multimedia , pages =. 2024 , doi =

    2024

  23. [31]

    IEEE Geoscience and Remote Sensing Letters , volume =

    Toward Efficient and Accurate Remote Sensing Image-Text Retrieval With a Coarse-to-Fine Approach , author =. IEEE Geoscience and Remote Sensing Letters , volume =. 2025 , doi =

    2025

  24. [32]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages =

    Eliminate Before Align: A Remote Sensing Image-Text Retrieval Framework with Keyword Explicit Reasoning , author =. Proceedings of the 32nd ACM International Conference on Multimedia , pages =. 2024 , doi =

    2024

  25. [33]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Remote Sensing Image-Text Retrieval With Implicit-Explicit Relation Reasoning , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2024 , doi =

    2024

  26. [34]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Global-Local Information Soft-Alignment for Cross-Modal Remote-Sensing Image-Text Retrieval , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2024 , doi =

    2024

  27. [35]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Fine-Grained Visual-Language Alignment for Remote Sensing Image-Text Retrieval , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2025 , doi =

    2025

  28. [36]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Context-Aware Local-Global Semantic Alignment for Remote Sensing Image-Text Retrieval , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2025 , doi =

    2025

  29. [37]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Frequency- and Spatial-Domain Saliency Network for Remote Sensing Cross-Modal Retrieval , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2025 , doi =

    2025

  30. [38]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , volume =

    Prompt-Based Granularity-Unified Representation Network for Remote Sensing Image-Text Matching , author =. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , volume =. 2025 , doi =

    2025

  31. [39]

    IEEE Transactions on Multimedia , volume =

    Cross-Modal Progressive Perspective Matching Network for Remote Sensing Image-Text Retrieval , author =. IEEE Transactions on Multimedia , volume =. 2025 , doi =

    2025

  32. [40]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , volume =

    Strong and Weak Prompt Engineering for Remote Sensing Image-Text Cross-Modal Retrieval , author =. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , volume =. 2025 , doi =

    2025

  33. [41]

    IEEE Transactions on Neural Networks and Learning Systems , volume =

    Cross-Modal Remote Sensing Image-Text Retrieval via Context and Uncertainty-Aware Prompt , author =. IEEE Transactions on Neural Networks and Learning Systems , volume =. 2025 , doi =

    2025

  34. [42]

    Neurocomputing , volume =

    Representation discrepancy bridging method for remote sensing image-text retrieval , author =. Neurocomputing , volume =. 2025 , doi =

    2025

  35. [43]

    Expert Systems with Applications , volume =

    iEBAKER: Improved remote sensing image-text retrieval framework via eliminate before align and keyword explicit reasoning , author =. Expert Systems with Applications , volume =. 2026 , doi =

    2026

  36. [44]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    TriSim: Tri-Dimensional Similarity Modeling with Extreme Value Theory for False-Negative Mitigation in Remote Sensing Image-Text Retrieval , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  37. [45]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Toward Fine-Grained Remote Sensing Image-Text Retrieval via Hierarchical Multigranularity Alignment , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2026 , doi =

    2026

  38. [46]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Explicit-Implicit Semantic Collaborative Guidance for Cross-Modal Remote Sensing Image-Text Retrieval , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2026 , doi =

    2026

  39. [47]

    ISPRS Journal of Photogrammetry and Remote Sensing , volume =

    Mind the modality gap: Towards a remote sensing vision-language model via cross-modal alignment , author =. ISPRS Journal of Photogrammetry and Remote Sensing , volume =. 2025 , doi =

    2025

  40. [48]

    IEEE Geoscience and Remote Sensing Letters , volume =

    Deep Multiscale Fine-Grained Hashing for Remote Sensing Cross-Modal Retrieval , author =. IEEE Geoscience and Remote Sensing Letters , volume =. 2024 , doi =

    2024

  41. [49]

    International Journal of Digital Earth , volume =

    Transferring CLIP for visual grounding in remote sensing images , author =. International Journal of Digital Earth , volume =. 2025 , doi =

    2025

  42. [50]

    Remote Sensing , volume =

    DDFAV: Remote Sensing Large Vision Language Models Dataset and Evaluation Benchmark , author =. Remote Sensing , volume =. 2025 , doi =

    2025

  43. [51]

    Remote Sensing , volume =

    A Review of Deep Learning-Based Remote Sensing Image Caption: Methods, Models, Comparisons and Future Directions , author =. Remote Sensing , volume =. 2024 , doi =

    2024

  44. [52]

    IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium , pages =

    Vigen500k: A Sustainable-Expansion Image-Text Aligned Dataset For Remote Sensing , author =. IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium , pages =. 2024 , doi =

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.