Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

GeoMag: A Vision-Language Model for Pixel-level Fine-Grained Remote Sensing Image Parsing

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read GeoMag is an end-to-end vision-language model that claims, for the first time, to handle pixel-level remote sensing parsing together with region- and image-level tasks, using prompt-guided cropping to zoom on small objects and cut compute.

desk verdict GeoMag's MagCrop is a genuinely useful idea with large ablation gains, but the paper never explains how the gradient-weighted attention is computed at inference, which is the load-bearing detail. read the letter →

arxiv 2507.05887 v2 pith:5ZOUT47S submitted 2025-07-08 cs.CV

classification cs.CV
keywords remotesensingvision-languagemodelpixel-levelparsingreferringsegmentationgroundedconversationgenerationattentionmechanismprompt-guidedcroppingmulti-granularityimage
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

GeoMag is a vision-language model for remote sensing that claims to handle all three parsing granularities in a single end-to-end pipeline: image-level scene classification and captioning, region-level visual question answering and object classification, and pixel-level referring segmentation and grounded conversation generation. The paper's central claim is that a preprocessing strategy called MagCrop—task-driven resolution adjustment combined with prompt-guided semantic cropping—lets the model concentrate resolution on the regions a prompt actually refers to, which improves small-object parsing and cuts compute on high-resolution satellite and aerial images. The authors report top results on pixel-level benchmarks and competitive performance on image- and region-level benchmarks across ten datasets. If true, this would move remote sensing multimodal understanding toward a unified all-granularity model rather than separate specialists.

What carries the argument

MagCrop is the load-bearing mechanism, and within it the central object is the gradient-weighted attention heatmap of Equation (4), $H = \frac{1}{N}\sum_{i=1}^{N} \mathrm{ReLU}\left(\frac{\partial L}{\partial A_i}\right)\odot A_i$, which converts the VLM's own loss gradient over last-layer attention into a spatial prior for where the prompt is pointing. Task-driven Multi-granularity Resolution Adjustment routes each query through a three-way classifier that picks image-, region-, or pixel-level processing and applies the corresponding resize/crop/stitch operations; Prompt-guided Semantic-aware Cropping searches that heatmap with a sliding-window candidate strategy to select a bounding box, crops it, and stitches it back at full resolution. For mask generation, GeoMag feeds the adjusted image through a frozen vision encoder and combines its multi-scale features with weighted [SEG] tokens in a trainable pixel decoder, with the fusion expressed as a weighted sum over scales and tokens. The argument of the paper is that this machinery is what lets a single model preserve global context, magnify small objects, and save computation.

What would settle it

Run GeoMag on prompts that name one among several identical small objects in a high-resolution image, for example 'the second ship from the left' on a 4K scene with many ships. If the selected bounding box from the Equation (4) heatmap fails to contain the named instance in more than a small fraction of cases, the cropped model's P@0.5 should fall below the no-crop baseline, directly contradicting the reported ablation. This is testable on the RRSIS-D or RefSegRS test sets with the reported metrics.

Watch

Extended reading notes

Core claim

The paper claims that one vision-language model can generate text and segmentation masks together, and that the decisive ingredient is not a heavier decoder but a prompt-aware preprocessing step. GeoMag adds a trainable pixel decoder over a frozen general-purpose vision encoder and introduces [SEG] tokens into the language model's output, so a single forward pass can answer a question, describe a scene, and emit masks. MagCrop then decides how the input image should be treated: a lightweight query classifier routes each prompt to image-, region-, or pixel-level processing; image-level tasks get a 100×100 downsampled image, while finer tasks run one or two rounds of gradient-weighted attention cropping that finds the prompt-relevant window and stitches it back into a downsampled base. Evaluated on ten benchmarks, GeoMag reports state-of-the-art referring segmentation (P@0.5 of 81.30 on RRSIS-D and 77.50 on RefSegRS) and grounded conversation generation (CIDEr of 235.17 on GeoPixelD), with competitive VQA, captioning, and scene classification, and the ablation attributes most of this to MagCrop.

Load-bearing premise

The whole method rests on the model's internal attention signal pointing at the object the prompt names, even for tiny or ambiguous targets, at the exact moment of inference; if that signal misfires, the cropping step will cut away the very region the model is supposed to magnify.

Editorial extensions

If this is right

  • One GeoMag instance could replace separate specialist models for scene classification, captioning, visual question answering, referring object classification, referring segmentation, and grounded conversation generation on remote sensing imagery.
  • High-resolution inputs become cheaper to process because image-level tasks run at 100×100 and only prompt-relevant windows keep full resolution for region- and pixel-level tasks.
  • Small-object parsing improves substantially: the ablation attributes P@0.5 gains of 15.85 points on RRSIS-D and 32.57 points on RefSegRS to MagCrop, and the 4K case study shows ships and aircraft being segmented.
  • The pixel-decoder-plus-[SEG]-token design can be added on top of existing remote sensing VLMs without retraining the entire language model, since the vision encoder is frozen and only the pixel decoder, text projection, and low-rank adapter are trained.

Reading between the lines

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

  • The same gradient-weighted attention heatmap could serve as a zero-shot localization diagnostic, letting users see which image region drove the model's answer; that would be a natural byproduct of MagCrop rather than an extra module.
  • A testable extension is to make the crop schedule adaptive to the estimated target size, since the fixed two-crop protocol for pixel-level tasks may over-crop very large objects or under-zoom very small ones.
  • The separate granularity classifier introduces a routing step the paper does not stress-test: prompts that mix granularities, such as 'count the cars and describe the overall scene', could be misrouted, and the error would propagate into resolution and cropping choices.
  • If the compute savings generalize, the same prompt-guided cropping idea could transfer to other high-resolution multimodal domains where relevant content is sparse, such as medical imaging or satellite video.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. GeoMag is a remote sensing vision-language model built on LLaVA-v1.5-7B with a frozen SAM2 vision encoder and a trainable pixel decoder, designed to handle image-level, region-level, and pixel-level tasks in a single pipeline. Its main novelty is MagCrop, a preprocessing scheme that classifies each query into image/region/pixel granularity, adjusts the input resolution accordingly, and crops prompt-relevant regions using a gradient-weighted attention map defined in Eq. (4). The paper evaluates GeoMag on 10 benchmarks, reporting strong or competitive results on referring segmentation (RRSIS-D, RefSegRS), grounded conversation generation (GeoPixelD), VQA, referring object classification, captioning, and scene classification, and presents ablations showing large performance drops without MagCrop.

Significance. The paper addresses a genuine gap in RS-VLMs: pixel-level parsing, and it does so with a relatively simple and modular architecture that keeps the LLM and vision encoder frozen and trains only lightweight components. The benchmark coverage is broad, and the ablation in Table 7 shows large, directionally consistent gains from MagCrop on referring segmentation and referring object classification. If the method works as described, it would be a useful general-purpose RS-VLM. The main risks are that the inference-time loss for the gradient-weighted cropping is not specified, the claimed compute savings are not measured, and the comparisons lack variance estimates; these issues must be resolved before the central claims can be accepted.

major comments (4)
  1. [§3.2.2, Eq. (4)] The heatmap H in Eq. (4) is defined from ∂L/∂A_i for a cross-entropy loss L, but the manuscript never specifies what L is at inference time for a referring-segmentation, VQA, or captioning query, where no ground-truth token or mask exists. If L is computed against hidden ground-truth labels during evaluation, then Table 7's large ablation gaps (P@0.5 -15.85 and -32.57 on RRSIS-D/RefSegRS; SIoU -32.18 on DIOR-RSVG) measure an oracle-guided preprocessing step rather than the deployable method. If L is instead computed against the model's own sampled output, the gradient signal can vanish or be uninformative for small objects, and the crop may discard the prompt-relevant region. Please state how L is formed at inference and report an ablation with a label-free localization signal (e.g., attention-only or GradCAM-style without loss gradients) on the same benchmarks.
  2. [§4 and §3.2] The abstract and contribution 2 claim that MagCrop reduces the computational cost of interpreting high-resolution RS imagery, and Section 3.2 states that it reduces memory, GPU usage, and time. However, no experiment in Section 4 reports latency, FLOPs, GPU memory, or number of visual tokens processed with and without MagCrop. Figure 3 reports file-size and resolution ratios, which are not model-level compute measurements. Please add quantitative efficiency comparisons on at least one high-resolution benchmark (e.g., RRSIS-D at 4K) so the compute-saving claim is testable.
  3. [§4.2, Tables 3 and 5] All reported results appear to come from a single evaluation run, with no error bars, multiple seeds, or significance tests. This matters for the 'competitive/outperforms' claims because several margins are small: in Table 3, GeoMag's RSVQA-LR average accuracy is 91.59 vs RSGPT's 92.29, and in Table 5, GeoMag's UCM CIDEr is 346.42 vs RS-CapRet's 354.80. Without variance estimates, the rank ordering on these benchmarks is not established. Please report mean±std over at least three runs for the main comparisons, or otherwise justify that the numbers are stable.
  4. [§3.2.1 and §4.1] The Task Granularity Classifier is a load-bearing component: it decides whether the input is downsampled to 100×100 or passed through one or two semantic crops, which changes both accuracy and compute. The manuscript says BERT-base is frozen while the FCs are 'trained', but it does not specify the training set, the labels, the loss, the number of epochs, or the classifier's accuracy on the ten benchmarks. Please provide these details and an analysis of how classification errors affect downstream task performance.
minor comments (6)
  1. [Figure 1] The label 'Image Sence' should be 'Image Scene'.
  2. [§3.2.2] The candidate bounding-box generation step ('multiple grid cells', 'comparing with the average total attention in neighboring regions') lacks concrete parameter values; include pseudo-code or a parameter table for reproducibility.
  3. [§4.3 and Figure 4] The ablation section says Grounded Conversation Generation was excluded because it 'did not use MagCrop', but the Figure 4 caption states that during GCG 'MagCrop provides a detailed description'. Please clarify whether MagCrop is used in the GCG evaluation and how the Figure 4 example was produced.
  4. [Eq. (1)] The weighting factors ω_l and β_n are not tied to any training objective; specify the loss used for the pixel decoder and Text Projection and the training schedule.
  5. [§4.2] In the image-level results paragraph, 'MLMMs' appears to be a typo for 'MLLMs'.
  6. [Tables 1 and 2] The metrics OIoU, MIoU, and P@0.5 are used without definitions; define them or cite the metric sources.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark results are external and the attention-gradient cropping is a design choice, not a circular derivation.

full rationale

The paper's central claims are empirical and are evaluated against external benchmarks (RRSIS-D, RefSegRS, GeoPixelD, DIOR-RSVG, RSVQA-LR/HR, UCM-Captions, Sydney-Captions, AID, RESISC45). No equation is fitted to those benchmark numbers: GeoMag is a supervised model trained with LoRA, and its results are compared with published methods rather than derived from its own assumptions. The MagCrop mechanism is self-referential in the sense that Eq. (4) uses the VLM's own attention gradients to choose crops, but this is an inference-time design choice, not a circularity in the argument. The paper does leave unspecified what cross-entropy loss L is available at inference time for Eq. (4); that is a specification gap and a correctness risk, not a circular reduction, because the paper does not claim to derive the crop from the final answer or to fit the benchmarks through this mechanism. The only self-citations ([23], [32], [33]) support background statements about SDGs and road extraction and are not load-bearing for the paper's novelty or its comparative results. No step in the derivation chain reduces, by construction or by self-citation, to its own inputs.

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

The paper adds no new physical entities, but the empirical method rests on several hand-chosen constants (100x100, 4x compression, crop depth) and unverified assumptions about the model's self-attention as a localizer and the availability of a loss at inference. The Task Granularity Classifier and PSC's selection rules are effectively learned or chosen components whose training data and hyperparameters are not disclosed.

free parameters (6)
  • Image-level target resolution = 100x100
    Eq. (2) sends all image-level tasks to a 100x100 image; chosen by hand, no ablation for this value.
  • Downsampling factor for background regions = 4 (H/4 x W/4)
    Eq. (3) compresses by 4x in each dimension; this constant is not ablated or justified.
  • Cropping depth per granularity = 1 crop for Region, 2 crops for Pixel
    Eq. (2) hardcodes the number of MagCrop iterations; no study of optimal depth.
  • Task Granularity Classifier (BERT+FC) weights = not disclosed
    Trained classifier whose training data, labels, and accuracy are not reported in Section 4.1.
  • PSC bounding-box selection parameters = not disclosed
    Grid cell sizes, candidate box generation, and the greatest-difference comparison in Section 3.2.2 are left unspecified; these effectively determine the crop.
  • LoRA rank = 8
    Section 4.1 states rank 8 for fine-tuning; standard but unablated.
assumptions (5)
  • domain assumption The VLM's internal attention, weighted by gradients of the cross-entropy loss, identifies the image regions most relevant to the prompt.
    PSC (Section 3.2.2) builds the crop around this heatmap; the paper does not validate localization quality independently of the final benchmark scores.
  • domain assumption A cross-entropy loss is available at inference time, or equivalently that the model's own outputs can serve as targets for gradient computation.
    Eq. (4) requires dL/dA_i, but the paper does not explain what L refers to when no ground truth exists at test time.
  • domain assumption BERT-based prompt classification into Image, Region, and Pixel is accurate enough to route resolution adjustment correctly.
    No accuracy or confusion matrix is reported for the Task Granularity Classifier (Section 3.2.1).
  • domain assumption SAM2's frozen features and LLaVA-v1.5 with LoRA transfer to remote sensing imagery.
    GeoMag relies on these pretrained models (Section 4.1) as the backbone for segmentation and language; this is plausible but not established beyond benchmark results.
  • domain assumption Benchmark splits and protocols are comparable across the models compared in Tables 1 through 6.
    The paper does not state the exact training sets for GeoMag versus each baseline, which matters for fair comparison, notably in Table 2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GeoMag: A Vision-Language Model for Pixel-level Fine-Grained Remote Sensing Image Parsing." pith.science (2026). https://pith.science/paper/5ZOUT47S

@misc{pith2026250705887,
  author       = {Pith},
  title        = {Pith review of: GeoMag: A Vision-Language Model for Pixel-level Fine-Grained Remote Sensing Image Parsing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5ZOUT47S}},
  note         = {Machine review of arXiv:2507.05887}
}
read the original abstract

The application of Vision-Language Models (VLMs) in remote sensing (RS) image understanding has achieved notable progress, demonstrating the basic ability to recognize and describe geographical entities. However, existing RS-VLMs are mostly limited to image-level and region-level tasks, lacking the capability to handle pixel-level tasks and performing poorly in small-object recognition scenarios. Moreover, RS-VLMs consume significant computational resources when processing high-resolution RS images, further restricting their practical applicability. In this context, we propose GeoMag (Geographical Magnifier), an end-to-end general-purpose large model framework for RS. GeoMag dynamically focuses the attention scope based on prompt semantics to effectively perform remote sensing image parsing across multiple levels of granularity. This method introduces Task-driven Multi-granularity Resolution Adjustment (TMRA) and Prompt-guided Semantic-aware Cropping (PSC), which adaptively reduce the spatial resolution of task-irrelevant regions while enhancing the visual representation of task-relevant areas. This approach improves the model's perception of critical target regions, suppresses background redundancy, and reduces the computational cost of interpreting high-resolution RS imagery. Extensive comparative experiments on 10 benchmarks demonstrate that GeoMag not only excels in handling pixel-level tasks but also maintains competitive performance across tasks of other granularities compared to existing RS-VLMs.

Figures

Figures reproduced from arXiv: 2507.05887 by the authors.

Figure 1
Figure 1. Overview of GeoMag’s capabilities. In addition to [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. GeoMag overview. The framework classifies the input query by task granularity, adjusts the resolution of the input [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The impact of different image preprocessing methods on tasks of varying granularity. In the figure, (1) represents [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: A case study of GeoMag on pixel-level tasks. On the left, GeoMag performs the Referring Segmentation task, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Vision-and-Language Navigation for UAVs: Progress, Challenges, and a Research Roadmap

    cs.RO 2026-04 unverdicted novelty 4.0 of 10

    A survey of UAV vision-and-language navigation that establishes a methodological taxonomy, reviews resources and challenges, and proposes a forward-looking research roadmap.

Reference graph

Works this paper leans on

74 extracted references · 57 canonical work pages · cited by 1 Pith paper

  1. [1]

    Yakoub Bazi, Mohamad Mahmoud Al Rahhal, Mohamed Lamine Mekhalfi, Man- sour Abdulaziz Al Zuair, and Farid Melgani. 2022. Bi-Modal Transformer-Based Approach for Visual Question Answering in Remote Sensing Imagery. IEEE Transactions on Geoscience and Remote Sensing 60 (2022), 1–11

  2. [2]

    Yuxiang Cai, Yongheng Shang, and Jianwei Yin. 2024. MultiDAN: Unsupervised, Multistage, Multisource and Multitarget Domain Adaptation for Semantic Seg- mentation of Remote Sensing Images. InProceedings of the 32nd ACM International Conference on Multimedia. 1168–1177

  3. [3]

    Huan Chen, Tingfa Xu, Zhenxiang Chen, Peifu Liu, Huiyan Bai, and Jianan Li

  4. [4]

    Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. 2023. MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning. (2023). arXiv preprint arXiv:2310.09478

  5. [5]

    Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao

  6. [6]

    Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G. Schwing. 2021. Mask2Former for Video Instance Segmentation. (2021). arXiv preprint arXiv:2112.10764

  7. [7]

    Gong Cheng, Junwei Han, and Xiaoqiang Lu. 2017. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proc. IEEE 105, 10 (2017), 1865–1883

  8. [8]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. (2018). arXiv preprint arXiv:1810.04805

Show all 74 references
  1. [9]

    Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, Huimei He, Jian Wang, Jingdong Chen, Ming Yang, Yongjun Zhang, and Yansheng Li. 2024. SkySense: A Multi- Modal Remote Sensing Foundation Model Towards Univers...

  2. [10]

    Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xiuping Jia, Antonio Plaza, Paolo Gamba, Jon Atli Benediktsson, and Jocelyn Chanussot. 2024. SpectralGPT: Spectral Remote Sensing Foundation Model. IEEE Transactions on P...

  3. [11]

    Genc Hoxha, Giacomo Scuccato, and Farid Melgani. 2023. Improving Image Cap- tioning Systems With Postprocessing Strategies. IEEE Transactions on Geoscience and Remote Sensing 61 (2023), 1–13

  4. [12]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In The Tenth International Conference on Learning Representa- tions, ICLR

  5. [13]

    Jiayun Hu, Yueyi He, Tianyi Liang, Changbo Wang, and Chenhui Li. 2025. Mu- sic2Palette: Emotion-aligned Color Palette Generation via Cross-Modal Repre- sentation Learning. arXiv preprint arXiv:2507.04758

  6. [14]

    Yuan Hu, Jianlong Yuan, Congcong Wen, Xiaonan Lu, and Xiang Li. 2023. RSGPT: A Remote Sensing Vision Language Model and Benchmark. (2023). arXiv preprint arXiv:2307.15266

  7. [15]

    Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. 2020. Bi- Directional Relationship Inferring Network for Referring Image Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  8. [16]

    Ling Huang, Wenqian Dong, Song Xiao, Jiahui Qu, Yuanbo Yang, and Yunsong Li

  9. [17]

    Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. 2020. Referring Image Segmentation via Cross-Modal Progressive Comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  10. [18]

    Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. 2020. Linguistic Structure Guided Context Modeling for Referring Image Segmentation. In Computer Vision – ECCV 2020 , Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm (Eds.). 59–75

  11. [19]

    In Proceedings of the 32nd ACM International Conference on Multimedia

    Language-Guided Visual Prompt Compensation for Multi-Modal Remote Sensing Image Classification with Modality Absence. In Proceedings of the 32nd ACM International Conference on Multimedia . 5161–5170

  12. [20]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. 2023. Segment Anything. In2023 IEEE/CVF International Conference on Computer Vision (ICCV) . ...

  13. [21]

    Kartik Kuckreja, Muhammad Sohail Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad Shahbaz Khan. 2024. GeoChat:Grounded Large Vision- Language Model for Remote Sensing. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 27831–27840

  14. [22]

    Zhong Ji, Changxu Meng, Yan Zhang, Haoran Wang, Yanwei Pang, and Jungong Han. 2024. Eliminate Before Align: A Remote Sensing Image-Text Retrieval Framework with Keyword Explicit Reasoning. In Proceedings of the 32nd ACM International Conference on Multimedia . 1662–1671

  15. [23]

    Jianhui Li, Chao Wu, Yingchao Piao, Yuchu Qin, Xiaoping Du, Lili Zhang, and Huadong Guo. 2023. How can we support the UN Sustainable Development Goals when open data is stagnant? Science Bulletin 68, 12 (2023), 1216–1218

  16. [24]

    Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. 2018. Referring Image Segmentation via Recurrent Refinement Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

  17. [25]

    Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. 2024. LISA: Reasoning Segmentation via Large Language Model. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 9579– 9589

  18. [26]

    Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. 2024. RemoteCLIP: A Vision Language Foundation Model for Remote Sensing. IEEE Transactions on Geoscience and Remote Sensing 62 (2024), 1–16

  19. [27]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024. Improved Baselines with Visual Instruction Tuning. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 26286–26296

  20. [28]

    Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, Jiaming Han, Siyuan Huang, Yichi Zhang, Xuming He, Hongsheng Li, and Yu Qiao. 2023. SPHINX: The Joint Mixing of Weights, Tasks, and Visual Embeddings for Multi-modal...

  21. [29]

    Sihan Liu, Yiwei Ma, Xiaoqing Zhang, Haowei Wang, Jiayi Ji, Xiaoshuai Sun, and Rongrong Ji. 2024. Rotated Multi-Scale Interaction Network for Referring Remote Sensing Image Segmentation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 26648–26658

  22. [30]

    Sylvain Lobry, Diego Marcos, Jesse Murray, and Devis Tuia. 2020. RSVQA: Visual Question Answering for Remote Sensing Data. IEEE Transactions on Geoscience and Remote Sensing 58, 12 (2020), 8555–8566

  23. [31]

    Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. 2022. Cross-Modal Progressive Comprehension for Referring Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 9 (2022), 4761–4775

  24. [32]

    Xianzhi Ma, Jianhui Li, Mingyang Lv, Xiaofeng Yang, and Hao Liu. 2025. SSCNet: Synchronous Stratification and Cross-Level Feature Fusion for Road Extraction. IEEE Transactions on Intelligent Transportation Systems (2025), 1–17

  25. [33]

    Xianzhi Ma, Xiaokai Zhang, Daoxiang Zhou, and Zehua Chen. 2024. StripUnet: A Method for Dense Road Extraction From Remote Sensing Images. IEEE Trans- actions on Intelligent Vehicles 9, 11 (2024), 7097–7109

  26. [34]

    Xiaoqiang Lu, Binqiang Wang, and Xiangtao Zheng. 2020. Sound Active Atten- tion Framework for Remote Sensing Image Captioning. IEEE Transactions on Geoscience and Remote Sensing 58, 3 (2020), 1985–2000

  27. [35]

    Dilxat Muhtar, Zhenshi Li, Feng Gu, Xueliang Zhang, and Pengfeng Xiao. 2025. LHRS-Bot: Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model. In Computer Vision – ECCV 2024 . 440–457

  28. [36]

    Ruizhe Ou, Yuan Hu, Fan Zhang, Jiaxin Chen, and Yu Liu. 2025. GeoPix: Multi- Modal Large Language Model for Pixel-level Image Understanding in Remote Sensing. (2025). arXiv preprint arXiv:2501.06828

  29. [37]

    Utkarsh Mall, Cheng Perng Phoo, Meilin Kelsey Liu, Carl Vondrick, Bharath Hariharan, and Kavita Bala. 2024. Remote Sensing Vision-Language Foundation Models without Annotations via Ground Remote Alignment. In The Twelfth International Conference on Learning Representations, ICLR

  30. [38]

    Chao Pang, Xingxing Weng, Jiang Wu, Jiayu Li, Yi Liu, Jiaxing Sun, Weijia Li, Shuai Wang, Litong Feng, Gui-Song Xia, and Conghui He. 2024. VHM: Versatile and Honest Vision Language Model for Remote Sensing Image Analysis. (2024). arXiv preprint arXiv:2403.20213

  31. [39]

    Bo Qu, Xuelong Li, Dacheng Tao, and Xiaoqiang Lu. 2016. Deep semantic understanding of high resolution remote sensing image. In 2016 International Conference on Computer, Information and Telecommunication Systems (CITS). 1–5

  32. [40]

    Yuwen Pan, Rui Sun, Yuan Wang, Tianzhu Zhang, and Yongdong Zhang. 2024. Rethinking the Implicit Optimization Paradigm with Dual Alignments for Re- ferring Remote Sensing Image Segmentation. In Proceedings of the 32nd ACM MM ’25, October 27–31, 2025, Dublin, Ireland. Xianzhi Ma...

  33. [41]

    Girshick, Piotr Dollár, and Christoph Feichtenhofer

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloé Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dollár, and Christoph Feicht...

  34. [42]

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. 2016. You Only Look Once: Unified, Real-Time Object Detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . 779–788

  35. [43]

    Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S

    Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. 2024. GLaMM: Pixel Grounding Large Multimodal Model. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognit...

  36. [44]

    Akashah Shabbir, Mohammed Zumri, Mohammed Bennamoun, Fahad Khan, and Salman H. Khan. 2025. GeoPixel: Pixel Grounding Large Multimodal Model in Remote Sensing. (2025). arXiv preprint arXiv:2501.13925

  37. [45]

    João Daniel Silva, João Magalhães, Devis Tuia, and Bruno Martins. 2024. Large Language Models for Captioning and Retrieving Remote Sensing Images. (2024). arXiv preprint arXiv:2402.06475

  38. [46]

    Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. 2024. PixelLM: Pixel Reasoning with Large Multimodal Model. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 26364–26373

  39. [47]

    Di Wang, Jing Zhang, Bo Du, Gui-Song Xia, and Dacheng Tao. 2023. An Empirical Study of Remote Sensing Pretraining. IEEE Transactions on Geoscience and Remote Sensing 61 (2023), 1–20

  40. [48]

    Zhecheng Wang, Rajanie Prabha, Tianyuan Huang, Jiajun Wu, and Ram Rajagopal

  41. [49]

    Gencer Sumbul, Sonali Nayak, and Begüm Demir. 2021. SD-RSIC: Summarization- Driven Deep Remote Sensing Image Captioning. IEEE Transactions on Geoscience and Remote Sensing 59, 8 (2021), 6922–6934

  42. [50]

    Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang

  43. [51]

    Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip H.S. Torr. 2022. LAVT: Language-Aware Vision Transformer for Referring Image Segmentation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 18134–18144

  44. [52]

    In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI

    SkyScript: A Large and Semantically Diverse Vision-Language Dataset for Remote Sensing. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI. 5805–5813

  45. [53]

    Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. 2017. AID: A Benchmark Data Set for Perfor- mance Evaluation of Aerial Scene Classification. IEEE Transactions on Geoscience and Remote Sensing 55, 7 (2017), 3965–3981

  46. [54]

    Bo Yuan, Danpei Zhao, Zhuoran Liu, Wentao Li, and Tian Li. 2024. Continual Panoptic Perception: Towards Multi-modal Incremental Interpretation of Remote Sensing Images. In Proceedings of the 32nd ACM International Conference on Multimedia. 2117–2126

  47. [55]

    In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    GSVA: Generalized Segmentation via Multimodal Large Language Models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 3858–3869

  48. [56]

    Zhenghang Yuan, Lichao Mou, Qi Wang, and Xiao Xiang Zhu. 2022. From Easy to Hard: Learning Language-Guided Curriculum for Visual Question Answering on Remote Sensing Data. IEEE Transactions on Geoscience and Remote Sensing 60 (2022), 1–11

  49. [57]

    Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. 2019. Cross-Modal Self- Attention Network for Referring Image Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  50. [58]

    Xinlei Yu, Changmiao Wang, Hui Jin, Ahmed Elazab, Gangyong Jia, Xiang Wan, Changqing Zou, and Ruiquan Ge. 2025. CRISP-SAM2: SAM2 with Cross-Modal Interaction and Semantic Prompting for Multi-Organ Segmentation. (2025). arXiv preprint arXiv:2506.23121

  51. [59]

    Yang Zhan, Zhitong Xiong, and Yuan Yuan. 2023. RSVG: Exploring Data and Models for Visual Grounding on Remote Sensing Data. IEEE Transactions on Geoscience and Remote Sensing 61 (2023), 1–13

  52. [60]

    Zhenghang Yuan, Lichao Mou, Yuansheng Hua, and Xiao Xiang Zhu. 2024. RRSIS: Referring Remote Sensing Image Segmentation. IEEE Transactions on Geoscience and Remote Sensing 62 (2024), 1–12

  53. [61]

    Jinxiao Zhang, Runmin Dong, Juepeng Zheng, Mengxuan Chen, Lixian Zhang, Yi Zhao, and Haohuan Fu. 2024. Spatial-Temporal Context Model for Remote Sensing Imagery Compression. In Proceedings of the 32nd ACM International Conference on Multimedia. 6063–6072

  54. [62]

    Zhiqiang Yuan, Wenkai Zhang, Kun Fu, Xuan Li, Chubo Deng, Hongqi Wang, and Xian Sun. 2022. Exploring a Fine-Grained Multiscale Method for Cross-Modal Remote Sensing Image Retrieval. IEEE Transactions on Geoscience and Remote Sensing 60 (2022), 1–19

  55. [63]

    Zhi Zeng, Minnan Luo, Xiangzheng Kong, Huan Liu, Hao Guo, Hao Yang, Zihan Ma, and Xiang Zhao. 2024. Mitigating World Biases: A Multimodal Multi-View Debiasing Framework for Fake News Video Detection. In Proceedings of the 32nd ACM International Conference on Multimedia . 6492–6500

  56. [64]

    Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. 2024. Earth- GPT: A Universal Multimodal Large Language Model for Multisensor Image Comprehension in Remote Sensing Domain. IEEE Transactions on Geoscience and Remote Sensing 62 (2024), 1–20

  57. [65]

    Yang Zhan, Zhitong Xiong, and Yuan Yuan. 2025. SkyEyeGPT: Unifying remote sensing vision-language tasks via instruction tuning with large language model. ISPRS Journal of Photogrammetry and Remote Sensing 221 (2025), 64–77

  58. [66]

    Zilun Zhang, Tiancheng Zhao, Yulong Guo, and Jianwei Yin. 2024. RS5M and GeoRSCLIP: A Large-Scale Vision- Language Dataset and a Large Vision- Language Model for Remote Sensing. IEEE Transactions on Geoscience and Remote Sensing 62 (2024), 1–23

  59. [67]

    Sheng Zhang and Xi Yang. 2024. Information Fusion with Knowledge Distillation for Fine-grained Remote Sensing Object Detection. In Proceedings of the 32nd ACM International Conference on Multimedia . 8081–8089

  60. [68]

    Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, Jun Li, and Xuerui Mao. 2025. EarthMarker: A Visual Prompting Multimodal Large Language Model for Remote Sensing. IEEE Transactions on Geoscience and Remote Sensing 63 (2025), 1–19

  61. [69]

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2024. MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models. In The Twelfth International Conference on Learning Represen- tations, ICLR

  62. [70]

    Zixiao Zhang, Licheng Jiao, Lingling Li, Xu Liu, Puhua Chen, Fang Liu, Yuxuan Li, and Zhicheng Guo. 2023. A Spatial Hierarchical Reasoning Network for Remote Sensing Visual Question Answering. IEEE Transactions on Geoscience and Remote Sensing 61 (2023), 1–15

  63. [72]

    Siru Zhong, Xixuan Hao, Yibo Yan, Ying Zhang, Yangqiu Song, and Yuxuan Liang. 2024. UrbanCross: Enhancing Satellite Image-Text Retrieval with Cross- Domain Adaptation. In Proceedings of the 32nd ACM International Conference on Multimedia. 6307–6315

  64. [73]

    Yue Zhou, Litong Feng, Yiping Ke, Xue Jiang, Junchi Yan, Xue Yang, and Wayne Zhang. 2024. Towards Vision-Language Geo-Foundation Model: A Survey. (2024). arXiv preprint arXiv:2406.09385

  65. [2023]

    Shikra: Unleashing Multimodal LLM’s Referential Dialogue Magic. (2023). arXiv preprint arXiv:2306.15195

  66. [2024]

    In Proceedings of the 32nd ACM International Conference on Multimedia

    Multi-scale Change-Aware Transformer for Remote Sensing Image Change Detection. In Proceedings of the 32nd ACM International Conference on Multimedia. 2992–3000

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.