pith. sign in

arxiv: 2509.21976 · v3 · submitted 2025-09-26 · 💻 cs.CV · cs.AI

Geo-R1: Improving Few-Shot Geospatial Referring Expression Understanding with Reinforcement Fine-Tuning

Zilun Zhang , Zian Guan , Tiancheng Zhao , Haozhan Shen , Tianyu Li , Yuxiang Cai , Zhonggen Su , Zhaojun Liu
show 2 more authors
Jianwei Yin Xiang Li
This is my paper

Pith reviewed 2026-05-18 14:04 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords few-shot learninggeospatial referring expressionsreinforcement fine-tuningremote sensingmultimodal language modelsobject localizationreasoning chains
0
0 comments X

The pith

Geo-R1 uses reinforcement fine-tuning to make models first generate explicit reasoning chains that decompose referring expressions before localizing geospatial targets.

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

The paper proposes a new training approach called Geo-R1 for understanding referring expressions in remote sensing images when only a few labeled examples are available. Instead of directly teaching the model to point to objects, it uses reinforcement fine-tuning to require the model to first produce clear step-by-step reasoning that breaks down the language description. This reasoning then guides the localization step, allowing the model to extract more value from limited data and to perform better when tested on new datasets. The method is shown to outperform standard supervised fine-tuning on three specially built few-shot benchmarks while also providing human-readable explanations of its decisions.

Core claim

Geo-R1 is a reasoning-centric reinforcement fine-tuning paradigm that requires the model to first generate explicit, interpretable reasoning chains decomposing the referring expression and then use those rationales to localize the target object, enabling more effective use of limited annotations and stronger generalization in geospatial referring tasks.

What carries the argument

The 'reason first, then act' process in which reinforcement fine-tuning enforces generation of explicit reasoning chains as intermediates for object localization.

If this is right

  • Models achieve higher accuracy on few-shot geospatial referring benchmarks than supervised fine-tuning baselines.
  • Performance remains strong when the trained model is evaluated on data from different sources or distributions.
  • Model decisions become more interpretable because each output includes an explicit reasoning chain.
  • Scarce labeled remote-sensing data can be leveraged more efficiently without requiring massive additional annotations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same reasoning-first structure could be tested on other vision-language tasks that suffer from limited labels, such as medical image description or urban scene understanding.
  • Combining the approach with larger base models or richer reward functions might further widen the gap over direct fine-tuning.
  • Evaluating the method on real operational remote-sensing workflows rather than controlled benchmarks would reveal whether the gains hold under practical constraints like varying image resolutions.

Load-bearing premise

That enforcing explicit reasoning chains through reinforcement fine-tuning will produce better few-shot performance and generalization than supervised fine-tuning on the same limited geospatial data.

What would settle it

Running the same few-shot training data through both Geo-R1 and a supervised fine-tuning baseline on the three designed geospatial referring benchmarks and finding no measurable gain in accuracy or cross-dataset transfer for the reinforcement approach.

Figures

Figures reproduced from arXiv: 2509.21976 by Haozhan Shen, Jianwei Yin, Tiancheng Zhao, Tianyu Li, Xiang Li, Yuxiang Cai, Zhaojun Liu, Zhonggen Su, Zian Guan, Zilun Zhang.

Figure 1
Figure 1. Figure 1: Geo-R1 method overview. Geo-R1 is trained on a few labeled samples with reinforcement [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Learning curves of GRPO vs. SFT on FS-REC. We fine-tune Qwen2.5-VL-3B with both SFT and GRPO on the FS-REC task using same batch size and evaluate checkpoints every 100 steps to sketch the learning curve. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Few-shot Learning Upper-Bound. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Few-shot Learning Meets Model Size. We then examine how model size influences the per￾formance under different post-training paradigms. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Geo-R1 inference samples (success case for GRES). [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Geo-R1 inference samples (success case for GREC). [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Geo-R1 inference samples (failure case). [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
read the original abstract

Referring expression understanding in remote sensing poses unique challenges, as it requires reasoning over complex object-context relationships. While supervised fine-tuning (SFT) on multimodal large language models achieves strong performance with massive labeled datasets, they struggle in data-scarce scenarios, leading to poor generalization. To address this limitation, we propose Geo-R1, a reasoning-centric reinforcement fine-tuning (RFT) paradigm for few-shot geospatial referring. Geo-R1 enforces the model to first generate explicit, interpretable reasoning chains that decompose referring expressions, and then leverage these rationales to localize target objects. This "reason first, then act" process enables the model to make more effective use of limited annotations, enhances generalization, and provides interpretability. We validate Geo-R1 on three carefully designed few-shot geospatial referring benchmarks, where our model consistently and substantially outperforms SFT baselines. It also demonstrates strong cross-dataset generalization, highlighting its robustness. Code and data will be released at: https://github.com/Geo-R1/geo-r1.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces Geo-R1, a reasoning-centric reinforcement fine-tuning (RFT) paradigm for few-shot geospatial referring expression understanding in remote sensing. The approach requires models to first generate explicit, interpretable reasoning chains that decompose referring expressions before localizing target objects, with the claim that this 'reason first, then act' process enables more effective use of limited annotations, improves generalization, and provides interpretability over standard supervised fine-tuning (SFT). Validation is reported on three few-shot geospatial referring benchmarks with consistent outperformance versus SFT baselines, plus strong cross-dataset generalization; code and data release is promised.

Significance. If the empirical gains prove robust and attributable to the explicit reasoning component, the work would offer a practical advance for data-scarce remote-sensing vision-language tasks by combining reinforcement learning with interpretable intermediate reasoning. The promised code and data release would further strengthen reproducibility and enable follow-on studies in geospatial few-shot settings.

major comments (1)
  1. [§5 (Experiments) and §4 (Method)] §5 (Experiments) and the method description in §4: The central claim that explicit reasoning chains enable more effective use of limited annotations and enhanced generalization (abstract and §1) is load-bearing, yet no ablation is presented that holds the RFT reward function, training data, and output format fixed while removing the requirement to produce explicit reasoning chains. Without this control, it remains unclear whether observed gains over SFT arise from the reasoning decomposition itself or from the RL objective and format constraints alone.
minor comments (2)
  1. [Abstract] Abstract: The claim of 'consistent and substantial' outperformance would be strengthened by briefly naming the three benchmarks, the exact few-shot settings (e.g., 1-shot, 5-shot), and the primary SFT baselines used.
  2. [§3 and §4] §3 and §4: Notation for the reward function and the reasoning-chain generation step could be clarified with a short pseudocode or explicit equation to make the 'reason first, then act' pipeline easier to replicate.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback. The suggested ablation directly addresses a key aspect of our central claim, and we outline our planned response below.

read point-by-point responses

  1. Referee: The central claim that explicit reasoning chains enable more effective use of limited annotations and enhanced generalization (abstract and §1) is load-bearing, yet no ablation is presented that holds the RFT reward function, training data, and output format fixed while removing the requirement to produce explicit reasoning chains. Without this control, it remains unclear whether observed gains over SFT arise from the reasoning decomposition itself or from the RL objective and format constraints alone.

    Authors: We agree that the current experiments do not include an ablation that isolates the explicit reasoning requirement while keeping the RFT reward function, training data, and output format fixed. Such a control would strengthen the attribution of gains specifically to the reasoning decomposition rather than the RL objective or formatting alone. In the revised manuscript we will add this ablation: we will train a variant under identical RFT conditions (same reward model, data, and output constraints) but modify the prompt and reward signal to remove any requirement or incentive for generating explicit reasoning chains, allowing direct comparison to the full Geo-R1 pipeline. This will clarify whether the observed improvements over SFT baselines stem from the reasoning step itself. revision: yes

Circularity Check

0 steps flagged

Empirical RFT method with external benchmark validation exhibits no circularity

full rationale

The paper describes Geo-R1 as an empirical reinforcement fine-tuning procedure that trains models to produce explicit reasoning chains before localization on few-shot geospatial data. Claims of improved generalization rest on performance comparisons against SFT baselines on three external benchmarks, with no mathematical derivations, fitted parameters renamed as predictions, or load-bearing self-citations that reduce the central result to its own inputs by construction. The method is self-contained against held-out test sets and does not invoke uniqueness theorems or ansatzes from prior author work to justify its design.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are specified in the abstract; the work relies on standard components of reinforcement learning and multimodal fine-tuning.

pith-pipeline@v0.9.0 · 5740 in / 1053 out tokens · 53000 ms · 2026-05-18T14:04:56.026174+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 3 Pith papers

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

  1. RemoteAgent: Bridging Vague Human Intents and Earth Observation with RL-based Agentic MLLMs

    cs.CV 2026-04 unverdicted novelty 7.0

    RemoteAgent uses RL fine-tuning on VagueEO to align MLLMs for vague EO intent recognition, handling simple tasks internally and routing dense predictions to tools via Model Context Protocol.

  2. RemoteZero: Geospatial Reasoning with Zero Human Annotations

    cs.CV 2026-05 unverdicted novelty 6.0

    RemoteZero replaces coordinate supervision with intrinsic semantic verification to enable box-free GRPO training and self-evolution for geospatial reasoning.

  3. RemoteShield: Enable Robust Multimodal Large Language Models for Earth Observation

    cs.CV 2026-04 unverdicted novelty 6.0

    RemoteShield improves robustness of Earth observation MLLMs by training on semantic equivalence clusters of clean and perturbed inputs via preference learning to maintain consistent reasoning under noise.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages · cited by 3 Pith papers · 13 internal anchors

  1. [1]

    Qwen2.5-VL Technical Report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report.ar...

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning

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

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Gong Cheng, Junwei Han, Peicheng Zhou, and Lei Guo

    Accessed: 2025-02-02. Gong Cheng, Junwei Han, Peicheng Zhou, and Lei Guo. Multi-class geospatial object detection and geographic image classification based on collection of part detectors.ISPRS Journal of Pho- togrammetry and Remote Sensing, 98:119–132,

    work page 2025

  4. [4]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al

    doi: 10.1109/TGRS.2021.3078507. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1 incentivizes reasoning in llms through reinforce- ment learning.Nature, 645(8081):633–638,

    work page doi:10.1109/tgrs.2021.3078507 2021

  5. [5]

    GREC: Generalized Referring Expression Comprehension, 2023

    Shuting He, Henghui Ding, Chang Liu, and Xudong Jiang. Grec: Generalized referring expression comprehension.arXiv preprint arXiv:2308.16182,

    work page arXiv

  6. [6]

    Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    OpenAI o1 System Card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Xufeng Jiang, Nan Zhou, and Xiang Li

    doi: 10.1109/ TGRS.2025.3531874. Xufeng Jiang, Nan Zhou, and Xiang Li. Few-shot segmentation of remote sensing images using deep metric learning.IEEE Geoscience and Remote Sensing Letters, 19:1–5,

    work page arXiv 2025

  9. [9]

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

    11 Preprint Kaiyu Li, Ruixun Liu, Xiangyong Cao, Xueru Bai, Feng Zhou, Deyu Meng, and Zhi Wang. Segearth-ov: Towards training-free open-vocabulary segmentation for remote sensing images. InProceedings of the Computer Vision and Pattern Recognition Conference, pp. 10545–10556, 2025a. Kaiyu Li, Zepeng Xin, Li Pang, Chao Pang, Yupeng Deng, Jing Yao, Guisong ...

    work page doi:10.1109/tgrs.2024.3417965 2024

  10. [10]

    Mini-Gemini: Mining the Potential of Multi-modality Vision Language Models

    Xiang Li, Jian Ding, and Mohamed Elhoseiny. Vrsbench: A versatile vision-language benchmark dataset for remote sensing image understanding.Advances in Neural Information Processing Systems, 37:3229–3242, 2024b. Xiang Li, Congcong Wen, Yuan Hu, Zhenghang Yuan, and Xiao Xiang Zhu. Vision-language mod- els in remote sensing: Current progress and future trend...

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Seg-Zero: Reasoning-Chain Guided Segmentation via Cognitive Reinforcement

    Yuqi Liu, Bohao Peng, Zhisheng Zhong, Zihao Yue, Fanbin Lu, Bei Yu, and Jiaya Jia. Seg- zero: Reasoning-chain guided segmentation via cognitive reinforcement.arXiv preprint arXiv:2503.06520, 2025a. Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning.arXiv preprint ...

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Jiancheng Pan, Yanxing Liu, Yuqian Fu, Muyuan Ma, Jiahao Li, Danda Pani Paudel, Luc Van Gool, and Xiaomeng Huang

    Accessed: 2025-08-30. Jiancheng Pan, Yanxing Liu, Yuqian Fu, Muyuan Ma, Jiahao Li, Danda Pani Paudel, Luc Van Gool, and Xiaomeng Huang. Locate anything on earth: Advancing open-vocabulary object detection for remote sensing community. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp. 6281–6289,

    work page 2025

  13. [13]

    LMM-R1: Empowering 3B LMMs with Strong Reasoning Abilities Through Two-Stage Rule-Based RL

    Yingzhe Peng, Gongrui Zhang, Miaosen Zhang, Zhiyuan You, Jie Liu, Qipeng Zhu, Kai Yang, Xingzhong Xu, Xin Geng, and Xu Yang. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl.arXiv preprint arXiv:2503.07536,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    PoseNet: A convolutional network for real-time 6-dof camera relocalization,

    doi: 10.1109/ICCV .2015.303. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pp. 8748–8763. PmLR,

    work page doi:10.1109/iccv 2015

  15. [15]

    Proximal Policy Optimization Algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathemati- cal reasoning in open language models.arXiv preprint arXiv:2402.03300,

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    VLM-R1: A Stable and Generalizable R1-style Large Vision-Language Model

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, et al. Vlm-r1: A stable and generalizable r1-style large vision-language model.arXiv preprint arXiv:2504.07615,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Xian Sun, Bing Wang, Zhirui Wang, Hao Li, Hengchao Li, and Kun Fu

    doi: 10.1109/TGRS.2024.3435086. Xian Sun, Bing Wang, Zhirui Wang, Hao Li, Hengchao Li, and Kun Fu. Research progress on few-shot learning for remote sensing image interpretation.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:2387–2402,

    work page doi:10.1109/tgrs.2024.3435086 2024

  19. [19]

    A billion- scale foundation model for remote sensing images

    doi: 10.1109/JSTARS. 2021.3052869. Yuxi Sun, Shanshan Feng, Xutao Li, Yunming Ye, Jian Kang, and Xu Huang. Visual grounding in remote sensing images. InProceedings of the 30th ACM International Conference on Multimedia, MM ’22, pp. 404–412, New York, NY , USA,

    work page doi:10.1109/jstars 2021

  20. [20]

    Visual Grounding in Remote Sensing Images

    Association for Computing Machinery. ISBN 9781450392037. doi: 10.1145/3503161.3548316. 13 Preprint Xingxing Weng, Chao Pang, and Gui-Song Xia. Vision-language modeling meets remote sensing: Models, datasets, and perspectives.IEEE Geoscience and Remote Sensing Magazine,

    work page doi:10.1145/3503161.3548316

  21. [21]

    doi: 10.1109/TGRS.2024. 3385655. Ling Yang, Liangliang Li, Zilun Zhang, Xinyu Zhou, Erjin Zhou, and Yu Liu. Dpgn: Distribution propagation graph network for few-shot learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June

    work page doi:10.1109/tgrs.2024 2024

  22. [22]

    R1-Onevision: Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization

    Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, et al. R1-onevision: Advancing generalized multimodal rea- soning through cross-modal formalization.arXiv preprint arXiv:2503.10615,

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476,

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma

    doi: 10.1109/LGRS.2021.3116858. Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models.arXiv preprint arXiv:2403.13372,

    work page doi:10.1109/lgrs.2021.3116858 2021

  25. [25]

    R1-Zero's "Aha Moment" in Visual Reasoning on a 2B Non-SFT Model

    Hengguang Zhou, Xirui Li, Ruochen Wang, Minhao Cheng, Tianyi Zhou, and Cho-Jui Hsieh. R1- zero’s “aha moment” in visual reasoning on a 2b non-sft model.arXiv preprint arXiv:2503.05132,

    work page internal anchor Pith review arXiv

  26. [26]

    Towards vision-language geo-foundation model: A survey.arXiv preprint arXiv:2406.09385, 2024a

    Yue Zhou, Litong Feng, Yiping Ke, Xue Jiang, Junchi Yan, Xue Yang, and Wayne Zhang. Towards vision-language geo-foundation model: A survey.arXiv preprint arXiv:2406.09385, 2024a. Yue Zhou, Mengcheng Lan, Xiang Li, Litong Feng, Yiping Ke, Xue Jiang, Qingyun Li, Xue Yang, and Wayne Zhang. Geoground: A unified large vision-language model for remote sensing v...

    work page arXiv