Pith. sign in

REVIEW 3 major objections 5 minor 62 references

Vision-Language Model Purified Semi-Supervised Semantic Segmentation for Remote Sensing Images

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A vision-language model that purifies a teacher network's pseudo-labels lifts semi-supervised remote sensing segmentation to state-of-the-art accuracy, with mean IoU gains of up to 11 points.

desk verdict Plausible VLM-purification idea, strong reported gains, but the described equations would filter away the very labels the module claims to fix. read the letter →

arxiv 2602.00202 v1 pith:TVOKNBWN submitted 2026-01-30 cs.CV cs.AI

Shanwen Wang , Xin Sun , Danfeng Hong , Fei Zhou This is my paper
classification cs.CVcs.AI
keywords semi-supervisedsemanticsegmentationremotesensingpseudo-labelpurificationvision-languagemodelteacher-studentinterpretabilityboundaryrefinementmeanIoU
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 claims that the bottleneck in semi-supervised semantic segmentation (S4) of remote sensing images is not the amount of unlabeled data but the quality of the pseudo-labels a teacher network generates, especially near multi-class boundaries. It proposes SemiEarth, a teacher-student model that inserts a vision-language model (VLM) into the loop: the VLM inspects each unlabeled image, lists the visible classes, and, where the teacher is uncertain, verifies or overrides the teacher's pixel labels. The authors report that this VLM pseudo-label purifying (VLM-PP) module improves mean IoU on two standard remote sensing datasets by 4-11 percentage points over the prior state of the art at 1%, 5%, and 10% labeled ratios, and that the approach remains interpretable because each purification step corresponds to an explicit visual-linguistic check.

What carries the argument

The load-bearing component is the VLM-PP module: a confidence-gated purifier. When the teacher's softmax confidence for a pixel is below a threshold, VLM-PP asks a vision-language model to list and locate the classes visible in the whole image; it then fuses the teacher's label with the VLM's class-level answer, using a purifying weight that grows as the teacher's confidence falls, and it replaces the teacher's label outright when the two disagree. Because the VLM is frozen and external to the segmentation network, the module is a drop-in appendage to any teacher-student S4 architecture.

What would settle it

One concrete test: take a set of unlabeled images, run the VLM-PP pipeline, and compare the resulting VLM-derived pixel labels to ground truth on a holdout set. If those labels in low-confidence regions match ground truth at or below the rate of the teacher's original pseudo-labels, the purification step is not adding signal and the reported gains would not replicate.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that an off-the-shelf vision-language model, asked only to list and locate the visible classes in an unlabeled image, can serve as a reliable external referee for a teacher-student segmentation framework. Whenever the teacher's confidence for a pixel falls below a threshold, VLM-PP queries the VLM; if the teacher's label matches the VLM's answer, the pixel's pseudo-label confidence is raised adaptively; if it conflicts, the VLM's label replaces the teacher's. The authors show that this purification not only raises the quality of pseudo-labels used to train the student but also prevents the error-amplification loop that makes teacher-student

Load-bearing premise

The load-bearing premise is that the vision-language model's image-level text output can be parsed into correct pixel-level class labels for low-confidence regions; if that parsing is unreliable, the purifier would inject noise rather than remove it.

Editorial extensions

If this is right

  • Semi-supervised remote sensing systems can be upgraded by attaching a VLM-PP-style purifier to an existing teacher-student model, without retraining the segmentation backbone.
  • Low-confidence pixels near class boundaries, which prior methods discard, become usable training signal instead, so annotation savings do not come at the cost of boundary accuracy.
  • The VLM's open-world vocabulary lets the pipeline correct categories that were absent or rare in the labeled set, as demonstrated by the correct segmentation of the Clutter class in Potsdam.
  • Since the VLM is used only at inference for pseudo-label purification, training cost is dominated by the segmentation model; the VLM adds a fixed overhead per image rather than per training step.
  • Each purification or rectification decision traces to an explicit textual statement by the VLM, giving practitioners a way to audit why a given pseudo-label was changed.

Reading between the lines

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

  • If VLM-PP works as described, the same purification scheme could be applied to other weakly-supervised or domain-adaptive segmentation settings where pseudo-labels are noisy, such as cross-city or cross-sensor domain adaptation for remote sensing.
  • The 5-11 point mIoU gains over specialized RS S4 methods suggest that a frozen, general-purpose VLM can inject semantic common sense that pixel-level features never capture, pointing toward using VLMs as a supervision oracle rather than as a feature extractor.
  • Because the paper does not specify how the VLM's text output is parsed into pixel coordinates, a natural test is to measure sensitivity to VLM prompt wording: if a different phrasing of 'list and locate classes' changes accuracy substantially, the method's robustness is tied to prompt engineering rather than to the fusion rule.
  • The purifying weight depends only on the teacher's confidence and is set per-pixel; an inference worth testing is whether making the threshold adaptive per class (rare classes need more purification) would improve results at very low label ratios.
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 / 5 minor

Summary. The paper proposes SemiEarth, a teacher-student semi-supervised semantic segmentation framework for remote sensing images. The central novelty is a VLM-PP module that uses a frozen vision-language model (Qwen-VL) and SAM to purify and rectify low-confidence teacher pseudo-labels before they supervise the student. Experiments on ISPRS-Potsdam and LoveDA at 1%, 5%, and 10% labeled ratios report state-of-the-art mIoU, with gains over the previous best method MUCA of up to 5.69/4.39/5.13 points on Potsdam and 6.47/10.37/11.0 points on LoveDA. Ablations attribute 4.59 and 5.01 points of mIoU to VLM-PP at the 5% labeled ratio.

Significance. If the mechanism worked as described, this would be a noteworthy contribution: it is the first attempt to inject external VLM knowledge into RS S4, it is architecturally simple, and it tries to recover, rather than discard, low-confidence boundary pixels. The release of code, the use of frozen external models (which avoids circularity from end-to-end training), and the interpretability of the module are genuine strengths. However, the central reported gains depend entirely on the VLM-PP equations, and those equations, as written, cannot produce supervision in the multi-class boundary regions that are the paper's target. Because this is a load-bearing internal inconsistency, the significance of the claimed results cannot be assessed without either corrected equations or a clear statement that the implemented code differs from the text.

major comments (3)
  1. The equations are internally inconsistent with the stated thresholding rule. Eq. (8) assigns confidence γ=0.95 to every class mentioned by the VLM, and Eq. (9) normalizes these values. If m classes are mentioned, each gets c̃ ≈ 1/m (e.g., 1/2=0.5 for m=2), which is below the adopted threshold τ_conf=0.7 whenever m≥2. In the purification branch, Eq. (10) is a convex combination of the teacher confidence c<0.7 and c̃≤0.5 with weight α=c/τ_conf∈(0,1); this sum is strictly below 0.7 (for m=2, it is at most 0.5+0.2α<0.7). In the rectification branch, Eq. (11) sets the final confidence to c̃, also below 0.7. The text after Eq. (11) then says that low-confidence purified pseudo-labels are filtered out. Thus, for every multi-class RS image, every pixel in R_low would be removed by VLM-PP, and the module would have no effect on training. The claimed 4.59–5.01 point gains in Table III therefore ca
  2. The mapping from VLM text output and SAM masks to per-pixel class labels is unspecified. Eq. (7) returns a free-form text sequence w containing class names and coordinates, but the paper does not describe how these are parsed, how each name is associated with a SAM mask, how overlapping or partially incorrect masks are resolved, how classes absent from the prompt but present in the teacher prediction are treated, or how the open-world VLM vocabulary is aligned to the fixed K classes (including 'background' and the excluded Potsdam 'Clutter' class). Without these details, the reader cannot evaluate whether the 'purified' or 'rectified' pseudo-labels are accurate pixel-level labels or merely noisy image-level guesses. This is not a cosmetic omission; it is central to the claim that VLM-PP injects external knowledge rather than noise.
  3. The confidence threshold τ_conf is selected by evaluating values 0.5–0.9 and adopting 0.7 'for all comparative experiments.' The text does not state whether this selection is performed on the validation split or on the same test split used for the SOTA comparisons in Tables I and II. If the latter, the SOTA comparisons are tuned on the evaluation data. The hand-set constant γ=0.95 in Eq. (8) is also not subject to sensitivity analysis. The authors should clarify the selection protocol and, if necessary, re-run comparative experiments with a threshold fixed before test-set evaluation.
minor comments (5)
  1. The section title contains a typo: 'Moudle' should be 'Module'.
  2. The figure label 'After VLM-PF' appears to be a typo for 'After VLM-PP'.
  3. Eq. (1) uses y^{u,t}_i for the pseudo-label, but Algorithm 1 and the surrounding text introduce y^{u,t}_p for the purified pseudo-label. Please use a consistent notation.
  4. No variance or repeated-run information is reported. Given that many baseline numbers are taken from prior papers with possibly different training configurations, the reported margins should be accompanied by error bars or at least a statement of the number of runs.
  5. Eq. (7) writes the VLM output as an argmax over a full sequence, but autoregressive generation is sampling-based or greedily decoded. The paper should state the decoding strategy used in practice.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: VLM-PP uses external pretrained Qwen-VL/SAM; self-citations are contextual; τ_conf tuning is an evaluation-selection issue, not a derivation loop.

full rationale

The claimed derivation chain is not circular. Teacher pseudo-labels and confidences are produced by the teacher network (Eqs. 2-4), while the purification signal comes from Qwen-VL and SAM, which are pretrained externally and explicitly not trained on the target datasets: 'the proposed SemiEarth does not train the VLM model itself but only uses it for the purification of pseudo-labels during inference' (Sec. II-B). The supervised/unsupervised loss (Eq. 1) is the standard S4 objective, and the reported mIoU numbers are empirical measurements on held-out evaluation splits, not algebraic consequences of the method's definitions. Self-citations ([10] MUCA, [50] RSCLIP) appear as prior-work comparisons, protocol references, or background; no load-bearing uniqueness theorem or ansatz is imported from them. Two non-circular concerns are worth flagging but do not amount to circularity. First, the purification threshold τ_conf is selected by sweeping mIoU on the same benchmarks used for the final comparison (Sec. IV-E-4, Fig. 10), which is a test-set selection issue that can inflate reported gains, but it is a configuration choice rather than a fitted quantity renamed as a prediction. Second, the skeptic's internal-consistency objection is real: applying Eqs. (8)-(11) literally, softmax-normalized VLM confidence for multi-class images falls below τ_conf=0.7, so the module would filter out the very low-confidence pixels it claims to purify. That could invalidate the mechanism, but it is an inconsistency, not a reduction of the claimed result to its inputs. Because the central derivation is independent of any self-citation chain, the circularity score is low; the minor self-citations justify a score of 2 rather than 0.

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

The central method leans on the unvalidated assumption that image-level VLM text plus SAM can produce reliable pixel labels in boundary regions. The only real fitted scalar is τ_conf, chosen via a test-set sweep, and γ is hand-set. No new physical or model entities are invented.

free parameters (4)
  • τ_conf (VLM-PP purification confidence threshold) = 0.7
    Swept over 0.5-0.9 on the Potsdam and LoveDA benchmark data (Fig. 10); best in 0.7-0.8 and set to 0.7 for all final comparisons. Tuning on evaluation data inflates reported mIoU.
  • γ (VLM confidence constant) = 0.95
    Assigned to any class mentioned by Qwen-VL in Eq. 8; hand-chosen, not derived from data or calibration.
  • EMA decay α = not specified
    Teacher update θ_t = αθ_t + (1-α)θ_s in Sec. III-A; value not reported, and it affects pseudo-label stability.
  • ε (numerical stability constant) = not specified
    Small constant in Eq. 9 to avoid division by zero; not fitted.
assumptions (5)
  • domain assumption Teacher-student EMA consistency learning is an adequate base for semi-supervised segmentation.
    Used throughout Sec. III-A; the proposed method builds on this framework, and the paper does not independently justify it beyond prior work.
  • domain assumption Qwen-VL can reliably list and locate all visible semantic classes in RS images from a single image-level prompt.
    Eqs. 6-7 assume the VLM output contains the correct categories and coordinates; no quantitative evaluation of VLM detection accuracy on these datasets is given.
  • domain assumption SAM can convert class mentions into accurate pixel-level masks.
    Sec. III-C states 'we refine their pixel-level alignments using the SAM model' without specifying prompt generation, mask selection, or failure handling.
  • domain assumption Low-confidence pseudo-label pixels are predominantly in multi-class boundaries and are worth purifying rather than discarding.
    Motivation in Sec. III-B; plausible but not formally evidenced, and the equations later suggest the proposed filtering may discard these pixels anyway.
  • domain assumption Excluding Clutter/Ignore classes from evaluation is appropriate for fair comparison.
    Sec. IV-B adopts this protocol from prior work; class exclusion changes the metric and can affect comparability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vision-Language Model Purified Semi-Supervised Semantic Segmentation for Remote Sensing Images." pith.science (2026). https://pith.science/paper/TVOKNBWN

@misc{pith2026260200202,
  author       = {Pith},
  title        = {Pith review of: Vision-Language Model Purified Semi-Supervised Semantic Segmentation for Remote Sensing Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TVOKNBWN}},
  note         = {Machine review of arXiv:2602.00202}
}
read the original abstract

The semi-supervised semantic segmentation (S4) can learn rich visual knowledge from low-cost unlabeled images. However, traditional S4 architectures all face the challenge of low-quality pseudo-labels, especially for the teacher-student framework.We propose a novel SemiEarth model that introduces vision-language models (VLMs) to address the S4 issues for the remote sensing (RS) domain. Specifically, we invent a VLM pseudo-label purifying (VLM-PP) structure to purify the teacher network's pseudo-labels, achieving substantial improvements. Especially in multi-class boundary regions of RS images, the VLM-PP module can significantly improve the quality of pseudo-labels generated by the teacher, thereby correctly guiding the student model's learning. Moreover, since VLM-PP equips VLMs with open-world capabilities and is independent of the S4 architecture, it can correct mispredicted categories in low-confidence pseudo-labels whenever a discrepancy arises between its prediction and the pseudo-label. We conducted extensive experiments on multiple RS datasets, which demonstrate that our SemiEarth achieves SOTA performance. More importantly, unlike previous SOTA RS S4 methods, our model not only achieves excellent performance but also offers good interpretability. The code is released at https://github.com/wangshanwen001/SemiEarth.

Figures

Figures reproduced from arXiv: 2602.00202 by the authors.

Figure 1
Figure 1. Structure comparison of S4 frameworks: (a) Self-training semi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall structure of our VLM-purified RS S4, i.e., SemiEarth, consists of unsupervised learning with unlabeled data and supervised learning with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The core logic of VLM purifying low-quality pseudo-labels from [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Samples from the two RS Datasets. code, dataset split files, and detailed experimental setup are all released at https://github.com/wangshanwen001/SemiEarth. A. RS Datasets and Data Augmentation LoveDA: The LoveDA RS dataset [51] consists of 5987 im￾ages containing a t…
Figure 5
Figure 5. Figure 5: Qualitative results with different SOTA S4 methods on the ISPRS-Potsdam dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results with different SOTA S4 methods on the LoveDA dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Before and after the VLM-PP module: comparison of the quality of [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: VLM-PP significantly improves the quality of pseudo-labels generated [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 10
Figure 10. Figure 10: Values of the different purifying confidence thresholds. [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 6 linked inside Pith

  1. [1]

    Pamsnet: A point annotation- driven multi-source network for remote sensing semantic segmentation,

    Y . Zhao, M. Jia, G. Sun, and A. Zhang, “Pamsnet: A point annotation- driven multi-source network for remote sensing semantic segmentation,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 229, pp. 1–16, 2025

    2025

  2. [2]

    Hyperspectral imaging,

    D. Hong, C. Li, N. Yokoya, B. Zhang, X. Jia, A. Plaza, P. Gamba, J. A. Benediktsson, and J. Chanussot, “Hyperspectral imaging,”Nature Reviews Methods Primers, 2026

    2026

  3. [3]

    Domain generalization for semantic segmentation of remote sensing images via vision foundation model fine-tuning,

    M. Luo, Y . Zan, K. Khoshelham, and S. Ji, “Domain generalization for semantic segmentation of remote sensing images via vision foundation model fine-tuning,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 230, pp. 126–146, 2025

    2025

  4. [4]

    Rsprotosemiseg: Semi-supervised semantic segmentation of high spatial resolution re- mote sensing images with probabilistic distribution prototypes,

    W. Sun, Y . Lei, D. Hong, Z. Hu, Q. Li, and J. Zhang, “Rsprotosemiseg: Semi-supervised semantic segmentation of high spatial resolution re- mote sensing images with probabilistic distribution prototypes,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 228, pp. 771–784, 2025

    2025

  5. [5]

    Spectralgpt: Spectral remote sensing foun- dation model,

    D. Hong, B. Zhang, X. Li, Y . Li, C. Li, J. Yao, N. Yokoya, H. Li, P. Ghamisi, X. Jiaet al., “Spectralgpt: Spectral remote sensing foun- dation model,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 08, pp. 5227–5244, 2024

    2024

  6. [6]

    Cross-city matters: A multimodal remote sensing benchmark dataset for cross-city semantic segmentation using high-resolution domain adaptation networks,

    D. Hong, B. Zhang, H. Li, Y . Li, J. Yao, C. Li, M. Werner, J. Chanussot, A. Zipf, and X. X. Zhu, “Cross-city matters: A multimodal remote sensing benchmark dataset for cross-city semantic segmentation using high-resolution domain adaptation networks,”Remote Sensing of Envi- ronment, vol. 299, p. 113856, 2023

    2023

  7. [7]

    Semi-meshseg: A semi-supervised semantic segmentation network for large-scale urban textured meshes using all pseudo-labels,

    W. Zi, J. Li, H. Chen, and Q. Jia, “Semi-meshseg: A semi-supervised semantic segmentation network for large-scale urban textured meshes using all pseudo-labels,”International Journal of Applied Earth Obser- vation and Geoinformation, vol. 142, p. 104674, 2025

    2025

  8. [8]

    Improving semi-supervised remote sensing scene classification via multilevel feature fusion and pseudo-labeling,

    J. Feng, H. Luo, and Z. Gu, “Improving semi-supervised remote sensing scene classification via multilevel feature fusion and pseudo-labeling,” International Journal of Applied Earth Observation and Geoinforma- tion, vol. 136, p. 104335, 2025

    2025

Show all 62 references
  1. [9]

    Pseudo labeling methods for semi-supervised semantic segmentation: A review and future perspec- tives,

    L. Ran, Y . Li, G. Liang, and Y . Zhang, “Pseudo labeling methods for semi-supervised semantic segmentation: A review and future perspec- tives,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 4, pp. 3054–3080, 2025

    2025

  2. [10]

    Semi-supervised semantic segmentation for remote sensing images via multiscale uncer- tainty consistency and cross-teacher–student attention,

    S. Wang, X. Sun, C. Chen, D. Hong, and J. Han, “Semi-supervised semantic segmentation for remote sensing images via multiscale uncer- tainty consistency and cross-teacher–student attention,”IEEE Transac- tions on Geoscience and Remote Sensing, vol. 63, pp. 1–15, 2025

    2025

  3. [11]

    Self-supervised learning in remote sensing: A review,

    Y . Wang, C. M. Albrecht, N. A. A. Braham, L. Mou, and X. X. Zhu, “Self-supervised learning in remote sensing: A review,”IEEE Geoscience and Remote Sensing Magazine, vol. 10, no. 4, pp. 213–247, 2022

    2022

  4. [12]

    Semi- supervised privacy-preserving eeg-based motor imagery classification via self and adversarial training,

    J. Zhu, G. Xu, Z. Lin, J. Long, T. Zhou, B. Sheng, and X. Yang, “Semi- supervised privacy-preserving eeg-based motor imagery classification via self and adversarial training,”IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 20 679–20 690, 2025

    2025

  5. [13]

    Fixmatch: Simplifying semi- supervised learning with consistency and confidence,

    K. Sohn, D. Berthelot, N. Carlini, Z. Zhang, H. Zhang, C. A. Raffel, E. D. Cubuk, A. Kurakin, and C.-L. Li, “Fixmatch: Simplifying semi- supervised learning with consistency and confidence,”Advances in neural information processing systems, vol. 33, pp. 596–608, 2020

    2020

  6. [14]

    Unimatch v2: Pushing the limit of semi-supervised semantic segmentation,

    L. Yang, Z. Zhao, and H. Zhao, “Unimatch v2: Pushing the limit of semi-supervised semantic segmentation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  7. [15]

    Dual-level masked semantic inference for semi-supervised semantic segmentation,

    Q. Ma, Z. Zhang, P. Qiao, Y . Wang, R. Ji, C. Liu, and J. Chen, “Dual-level masked semantic inference for semi-supervised semantic segmentation,”IEEE Transactions on Multimedia, 2025

    2025

  8. [16]

    Semi-supervised bidirectional alignment for remote sensing cross-domain scene classifi- cation,

    W. Huang, Y . Shi, Z. Xiong, Q. Wang, and X. X. Zhu, “Semi-supervised bidirectional alignment for remote sensing cross-domain scene classifi- cation,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 195, pp. 192–203, 2023

    2023

  9. [17]

    Uncertainty- aware semi-supervised learning segmentation for remote sensing im- ages,

    X. Lu, L. Li, L. Jiao, X. Liu, F. Liu, W. Ma, and S. Yang, “Uncertainty- aware semi-supervised learning segmentation for remote sensing im- ages,”IEEE Transactions on Multimedia, 2025

    2025

  10. [18]

    Transrefine: Transformer-augmented feature refinement for zero-shot scene classi- fication in remote sensing images,

    R. Damalla, P. A. Bendre, R. Datla, V . Chalavadiet al., “Transrefine: Transformer-augmented feature refinement for zero-shot scene classi- fication in remote sensing images,”Pattern Recognition, vol. 162, p. 111406, 2025

    2025

  11. [19]

    Water areas segmentation from remote sensing images using a separable residual segnet network,

    L. Weng, Y . Xu, M. Xia, Y . Zhang, J. Liu, and Y . Xu, “Water areas segmentation from remote sensing images using a separable residual segnet network,”ISPRS international journal of geo-information, vol. 9, no. 4, p. 256, 2020

    2020

  12. [20]

    Uiu-net: U-net in u-net for infrared small object detection,

    X. Wu, D. Hong, and J. Chanussot, “Uiu-net: U-net in u-net for infrared small object detection,”IEEE Transactions on Image Processing, vol. 32, pp. 364–376, 2022

    2022

  13. [21]

    Difference-complementary learning and label reassignment for multimodal semi-supervised se- mantic segmentation of remote sensing images,

    W. Han, W. Jiang, J. Geng, and W. Miao, “Difference-complementary learning and label reassignment for multimodal semi-supervised se- mantic segmentation of remote sensing images,”IEEE Transactions on Image Processing, 2025

    2025

  14. [22]

    Semisupervised semantic segmen- tation of remote sensing images with consistency self-training,

    J. Li, B. Sun, S. Li, and X. Kang, “Semisupervised semantic segmen- tation of remote sensing images with consistency self-training,”IEEE JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11 Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–11, 2021

    2021

  15. [23]

    Decouple and weight semi- supervised semantic segmentation of remote sensing images,

    W. Huang, Y . Shi, Z. Xiong, and X. X. Zhu, “Decouple and weight semi- supervised semantic segmentation of remote sensing images,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 212, pp. 13–26, 2024

    2024

  16. [24]

    Tse-net: Semi-supervised monocular height estimation from single remote sensing images,

    S. Chen and X. X. Zhu, “Tse-net: Semi-supervised monocular height estimation from single remote sensing images,”arXiv preprint arXiv:2511.13552, 2025

    2025

  17. [25]

    Clr-dlr: A semi-supervised framework for high-fidelity remote sensing segmentation,

    T. Ni, J. Wang, X. Zi, K. Thiyagarajan, S. Kodagoda, and M. Prasad, “Clr-dlr: A semi-supervised framework for high-fidelity remote sensing segmentation,”IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1–10, 2025

    2025

  18. [26]

    Confidence-weighted dual- teacher networks with biased contrastive learning for semi-supervised semantic segmentation in remote sensing images,

    Y . Xin, Z. Fan, X. Qi, Y . Zhang, and X. Li, “Confidence-weighted dual- teacher networks with biased contrastive learning for semi-supervised semantic segmentation in remote sensing images,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  19. [27]

    Mcmcnet: A semi- supervised road extraction network for high-resolution remote sensing images via multiple consistency and multi-task constraints,

    L. Gao, Y . Zhou, J. Tian, W. Cai, and Z. Lv, “Mcmcnet: A semi- supervised road extraction network for high-resolution remote sensing images via multiple consistency and multi-task constraints,”IEEE Trans- actions on Geoscience and Remote Sensing, 2024

    2024

  20. [28]

    Semiroadexnet: A semi- supervised network for road extraction from remote sensing imagery via adversarial learning,

    H. Chen, Z. Li, J. Wu, W. Xiong, and C. Du, “Semiroadexnet: A semi- supervised network for road extraction from remote sensing imagery via adversarial learning,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 198, pp. 169–183, 2023

    2023

  21. [29]

    Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model,

    Y . Zhan, Z. Xiong, and Y . Yuan, “Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 221, pp. 64–77, 2025

    2025

  22. [30]

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

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

    2024

  23. [31]

    Toward open-world remote sensing imagery interpretation: Past, present, and future,

    C. Lang, G. Cheng, J. Wu, Z. Li, X. Xie, J. Li, and J. Han, “Toward open-world remote sensing imagery interpretation: Past, present, and future,”IEEE Geoscience and Remote Sensing Magazine, 2024

    2024

  24. [32]

    Rsgpt: A remote sensing vision language model and benchmark,

    Y . Hu, J. Yuan, C. Wen, X. Lu, Y . Liu, and X. Li, “Rsgpt: A remote sensing vision language model and benchmark,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 224, pp. 272–286, 2025

    2025

  25. [33]

    Skyscript: A large and semantically diverse vision-language dataset for remote sens- ing,

    Z. Wang, R. Prabha, T. Huang, J. Wu, and R. Rajagopal, “Skyscript: A large and semantically diverse vision-language dataset for remote sens- ing,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 6, 2024, pp. 5805–5813

    2024

  26. [34]

    Skysense-o: Towards open-world remote sensing inter- pretation with vision-centric visual-language modeling,

    Q. Zhu, J. Lao, D. Ji, J. Luo, K. Wu, Y . Zhang, L. Ru, J. Wang, J. Chen, M. Yanget al., “Skysense-o: Towards open-world remote sensing inter- pretation with vision-centric visual-language modeling,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025...

    2025

  27. [35]

    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

  28. [36]

    Rotated multi-scale interaction network for referring remote sensing image seg- mentation,

    S. Liu, Y . Ma, X. Zhang, H. Wang, J. Ji, X. Sun, and R. Ji, “Rotated multi-scale interaction network for referring remote sensing image seg- mentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 26 658–26 668

    2024

  29. [37]

    Rs-moe: A vision-language model with mixture of experts for remote sensing image captioning and visual question answering,

    H. Lin, D. Hong, S. Ge, C. Luo, K. Jiang, H. Jin, and C. Wen, “Rs-moe: A vision-language model with mixture of experts for remote sensing image captioning and visual question answering,”IEEE Transactions on Geoscience and Remote Sensing, 2025

    2025

  30. [38]

    Meta captioning: A meta learning based remote sensing image captioning framework,

    Q. Yang, Z. Ni, and P. Ren, “Meta captioning: A meta learning based remote sensing image captioning framework,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 186, pp. 190–200, 2022

    2022

  31. [39]

    Transforming remote sensing images to textual descriptions,

    U. Zia, M. M. Riaz, and A. Ghafoor, “Transforming remote sensing images to textual descriptions,”International Journal of Applied Earth Observation and Geoinformation, vol. 108, p. 102741, 2022

    2022

  32. [40]

    An interpretable fusion siamese network for multi-modality remote sensing ship image retrieval,

    W. Xiong, Z. Xiong, Y . Cui, L. Huang, and R. Yang, “An interpretable fusion siamese network for multi-modality remote sensing ship image retrieval,”IEEE Transactions on Circuits and Systems for Video Tech- nology, vol. 33, no. 6, pp. 2696–2712, 2022

    2022

  33. [41]

    Textir: A simple framework for text-based editable image restoration,

    Y . Bai, C. Wang, S. Xie, C. Dong, C. Yuan, and Z. Wang, “Textir: A simple framework for text-based editable image restoration,”IEEE Transactions on Visualization and Computer Graphics, 2025

    2025

  34. [42]

    Evaluating language biases in remote sensing visual question answer- ing: The role of spatial attributes, language diversity, and the need for clearer evaluation,

    C. Chappuis, E. Walt, V . Mendez, S. Lobry, B. Le Saux, and D. Tuia, “Evaluating language biases in remote sensing visual question answer- ing: The role of spatial attributes, language diversity, and the need for clearer evaluation,”IEEE Geoscience and Remote Sensing Magazine, 2025

    2025

  35. [43]

    Earthvqa: Towards queryable earth via relational reasoning-based remote sensing visual question answering,

    J. Wang, Z. Zheng, Z. Chen, A. Ma, and Y . Zhong, “Earthvqa: Towards queryable earth via relational reasoning-based remote sensing visual question answering,” inProceedings of the AAAI conference on artificial intelligence, vol. 38, no. 6, 2024, pp. 5481–5489

    2024

  36. [44]

    Prompt- rsvqa: Prompting visual context to a language model for remote sensing visual question answering,

    C. Chappuis, V . Zermatten, S. Lobry, B. Le Saux, and D. Tuia, “Prompt- rsvqa: Prompting visual context to a language model for remote sensing visual question answering,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 1372–1381

    2022

  37. [45]

    Rsrefseg: Referring remote sensing image segmentation with foundation models,

    K. Chen, J. Zhang, C. Liu, Z. Zou, and Z. Shi, “Rsrefseg: Referring remote sensing image segmentation with foundation models,”arXiv preprint arXiv:2501.06809, 2025

    2025 arXiv

  38. [46]

    Open-vocabulary remote sensing image semantic segmentation,

    Q. Cao, Y . Chen, C. Ma, and X. Yang, “Open-vocabulary remote sensing image semantic segmentation,”arXiv preprint arXiv:2409.07683, 2024

    2024 arXiv

  39. [47]

    Towards open-vocabulary remote sens- ing image semantic segmentation,

    C. Ye, Y . Zhuge, and P. Zhang, “Towards open-vocabulary remote sens- ing image semantic segmentation,”arXiv preprint arXiv:2412.19492, 2024

    2024 arXiv

  40. [48]

    Exploring efficient open-vocabulary segmentation in the remote sensing,

    B. Li, H. Dong, D. Zhang, Z. Zhao, J. Gao, and X. Li, “Exploring efficient open-vocabulary segmentation in the remote sensing,”arXiv preprint arXiv:2509.12040, 2025

    2025

  41. [49]

    Segearth-ov: Towards training-free open-vocabulary segmentation for remote sensing images,

    K. Li, R. Liu, X. Cao, X. Bai, F. Zhou, D. Meng, and Z. Wang, “Segearth-ov: Towards training-free open-vocabulary segmentation for remote sensing images,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 10 545–10 556

    2025

  42. [50]

    Rsclip for training- free open-vocabulary remote sensing image semantic segmentation,

    S. Wang, X. Sun, D. Hong, and X. Zhu, “Rsclip for training- free open-vocabulary remote sensing image semantic segmentation,” September 2025. [Online]. Available: https://www.techrxiv.org/doi/10. 36227/techrxiv.175790902.28615776/v1

    2025

  43. [51]

    Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation,

    J. Wang, Z. Zheng, A. Ma, X. Lu, and Y . Zhong, “Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation,” inProceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, J. Vanschoren and S. Yeung, Eds., vol. 1. Curra...

    2021

  44. [52]

    ISPRS Potsdam Dataset,

    ISPRS, “ISPRS Potsdam Dataset,” https://www.isprs.org/education/ benchmarks/UrbanSemLab/2d-sem-label-potsdam.aspx?utm source= chatgpt.com, 2018, accessed: 2024-10-8

    2018

  45. [53]

    Cutmix: Reg- ularization strategy to train strong classifiers with localizable features,

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y . Yoo, “Cutmix: Reg- ularization strategy to train strong classifiers with localizable features,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6023–6032

    2019

  46. [54]

    Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results,

    A. Tarvainen and H. Valpola, “Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results,”Advances in neural information processing systems, vol. 30, 2017

    2017

  47. [55]

    Semi-supervised semantic segmen- tation with cross-consistency training,

    Y . Ouali, C. Hudelot, and M. Tami, “Semi-supervised semantic segmen- tation with cross-consistency training,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 12 674–12 684

    2020

  48. [56]

    Semi-supervised semantic segmentation with cross pseudo supervision,

    X. Chen, Y . Yuan, G. Zeng, and J. Wang, “Semi-supervised semantic segmentation with cross pseudo supervision,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 2613–2622

    2021

  49. [57]

    Simple and efficient: A semisupervised learning framework for remote sensing image semantic segmentation,

    X. Lu, L. Jiao, F. Liu, S. Yang, X. Liu, Z. Feng, L. Li, and P. Chen, “Simple and efficient: A semisupervised learning framework for remote sensing image semantic segmentation,”IEEE Transactions on Geo- science and Remote Sensing, vol. 60, pp. 1–16, 2022

    2022

  50. [58]

    Revisiting weak- to-strong consistency in semi-supervised semantic segmentation,

    L. Yang, L. Qi, L. Feng, W. Zhang, and Y . Shi, “Revisiting weak- to-strong consistency in semi-supervised semantic segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7236–7246

    2023

  51. [59]

    Allspark: Reborn labeled features from unlabeled in transformer for semi-supervised semantic segmenta- tion,

    H. Wang, Q. Zhang, Y . Li, and X. Li, “Allspark: Reborn labeled features from unlabeled in transformer for semi-supervised semantic segmenta- tion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 3627–3636

    2024

  52. [60]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Noubyet al., “Dinov2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  53. [61]

    Qwen2. 5-1m technical report,

    A. Yang, B. Yu, C. Li, D. Liu, F. Huang, H. Huang, J. Jiang, J. Tu, J. Zhang, J. Zhouet al., “Qwen2. 5-1m technical report,”arXiv preprint arXiv:2501.15383, 2025

    2025 arXiv

  54. [62]

    Qwen3 technical report,

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.