Pith. sign in

REVIEW 4 major objections 6 minor 25 references

FUSAR-R1 proposes that SAR image interpretation be recast as explicit chain-of-thought reasoning followed by reinforcement-learning optimization, and reports consistent gains over general-purpose multimodal models on detection, counting, an

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Training a SAR vision-language model with chain-of-thought cold start plus GRPO reinforcement learning yields large reported gains over zero-shot general-purpose baselines.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A plausible domain-specific adaptation of the R1 recipe to SAR, with a genuinely useful CoT construction pipeline — but the central evaluation claim is unsupported until the authors show the test data are independent of the training corpus. the 4 major comments →

arxiv 2607.16819 v1 pith:PDWNTFKL submitted 2026-07-18 cs.AI

FUSAR-R1: A Large-Scale Reasoning Model for Intelligent Interpretation of SAR Images

Yi Yang , Xiaokun Zhang , Yuxuan Li , Ruyi Zhang , Xinpeng Zhou , Haipeng Wang This is my paper
classification cs.AI
keywords SAR image interpretationvision-language modelchain-of-thoughtreinforcement learningGRPOremote sensingobject detectionland-cover classification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

FUSAR-R1 sets out to show that SAR image interpretation can be turned from an end-to-end visual matching problem into an explicit reasoning problem. The paper builds expert-style chain-of-thought data from an existing SAR image-text corpus, fine-tunes a multimodal backbone on that data as a cold start, then optimizes the model with reinforcement learning under a multi-task reward that checks format, detection boxes, counting, classification, and land-cover proportions. The model is claimed to beat general-purpose multimodal models on all six evaluated tasks, e.g., 67.33% object-counting accuracy versus 45.45% for the best baseline and land-cover proportion error of 7.67 versus 39.58. If this holds, the implication is that explicit reasoning plus physical priors, not raw model scale, is what makes a vision-language model reliable on radar imagery.

Core claim

The central claim, stated in the paper's own terms, is that FUSAR-R1 consistently outperforms existing multimodal large-scale models across SAR interpretation tasks — target detection, object classification and counting, and land-cover category recognition — by replacing direct image-to-answer prediction with a chain of thoughts that explicitly tracks observation cues, scattering-physics evidence, spatial scale, and geographic context, and by using GRPO reinforcement learning to reward correct, well-structured inference. The strongest numbers are an object-counting accuracy of 67.33% against 45.45% for the best general-purpose baseline, all-land-cover-category accuracy of 93.44% against 75.8

What carries the argument

The mechanism that carries the argument is the chain-of-thought cold-start corpus plus GRPO (Group Relative Policy Optimization, which estimates reward advantages within a sampled group of outputs). The corpus is generated by reconstructing the FUSAR-GEOVL-1M dataset with a local large language model that turns fragmented labels into structured reasoning chains constrained by physical hard rules (geographic anchoring, scattering-feature library, scale-prior pixel-to-meter conversion). The model is trained in two stages: supervised fine-tuning on this CoT corpus, then reinforcement learning with a weighted multi-task reward covering output format, detection IoU, counting/category match, and l

Load-bearing premise

The load-bearing premise is that the six evaluation tasks are independent of the FUSAR-R1 training corpus: the paper does not state a train/test split, and both the cold-start CoT data and the Table III evaluation tasks derive from the same FUSAR-GEOVL-1M data family, so overlapping images or labels could inflate the reported gains over zero-shot baselines (Section IV-A).

What would settle it

Run the six tasks on a geographically and temporally disjoint set of SAR images that were never part of FUSAR-GEOVL-1M or its CoT reconstruction, and check whether FUSAR-R1 keeps its counting, detection, and land-cover margins; a drop to near-baseline levels would falsify the claim that the model reasons rather than recalls.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • SAR interpretation can be made explainable: the model outputs a readable reasoning chain before each answer, so users can audit why a target was counted or classified.
  • A two-stage recipe of CoT cold start followed by RL transfers to radar-specific tasks: the paper's ablations show RL alone collapses, CoT alone helps, and the combination is best (F1 0.676 vs 0.528 for the base model).
  • Physical metadata embedded as text — especially ground resolution — can substitute for large amounts of task-specific labeled data.
  • The unified multi-task reward lets one model handle detection, counting, classification, and land-cover estimation without task-specific heads.
  • Comparisons across five general models indicate that model size is not the decisive factor; reasoning structure and domain priors contribute more.

Where Pith is reading between the lines

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

  • Editorial inference: the recipe should generalize to other sensor modalities with sparse semantic labels, such as hyperspectral or LiDAR imagery, provided a reconstruction pipeline can turn existing labels into expert-style reasoning chains.
  • Editorial inference: the strong effect of the scale-prior ablation suggests a cheap and testable extension — feeding FUSAR-R1 deliberately wrong resolution metadata and checking whether its conclusions shift accordingly; if they do not, the model may be relying on texture shortcuts despite the CoT style.
  • Editorial inference: the counting ceiling of 67.33% on the target-counting task hints that the current reward scheme does not fully close dense-target errors; a counting-specific loss or a verification step could push this higher.
  • Editorial inference: if overlapping train/test data is ruled out, the approach may reduce the need for large SAR image-text pairs, since RL explores beyond the annotated distribution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes FUSAR-R1, a reasoning-oriented vision-language model for SAR image interpretation. The method reconstructs the authors' FUSAR-GEOVL-1M dataset into chain-of-thought cold-start data using Qwen3-8B, performs SFT on this corpus, and then applies GRPO reinforcement learning with a multi-task reward function that combines format, detection, counting/classification, and land-cover rewards. The authors report experiments on six evaluation tasks and claim consistent superiority over mainstream general-purpose multimodal models, e.g., object counting accuracy 67.33% vs 45.45% for the best baseline and land-cover proportion MAE 7.67 vs 39.58.

Significance. If the reported gains are valid and reproducible, FUSAR-R1 would be a meaningful advance in applying reasoning models to SAR interpretation, combining physical priors, chain-of-thought supervision, and reinforcement learning. The paper is well motivated and provides detailed descriptions of the data reconstruction pipeline, the reward design, and the GRPO training setup. The ablation study (Table IV) is a useful addition. However, the central empirical claim is conditional on evaluation independence and baseline fairness, which are currently not established; the absence of a reported train/test split, the use of the same data family for training and evaluation, and the comparison only against zero-shot general-purpose baselines all leave the quantitative conclusions undersupported.

major comments (4)
  1. [§III-B, §IV-A, Table III, §IV-B-4a] No train/test split is reported anywhere in the paper. The cold-start CoT corpus is reconstructed from FUSAR-GEOVL-1M (III-B), and the six evaluation tasks in Table III are defined on the same data family. Section IV-B-4a refers to a 'test set' without explaining how it was separated from the training data. This is load-bearing: if the evaluation images or labels overlap the CoT training distribution, even shared geographic scenes or label distributions could inflate the Table V gains (e.g., 67.33% vs 45.45% counting accuracy; regional land-cover 64.75% vs <15% for baselines) via memorization rather than reasoning. The authors must report an explicit, disjoint split and/or evaluate on an external benchmark.
  2. [Table V] The comparison is not apples-to-apples. FUSAR-R1 is fine-tuned and RL-optimized on SAR data, while the five baselines (InternVL3.5, LLaVA 1.5, Qwen2.5-VL) are evaluated zero-shot with no task-specific fine-tuning. The conclusion that FUSAR-R1 'consistently outperforms existing multimodal large-scale models' is therefore overstated. The authors should fine-tune the baselines on the same task data (or at least provide matched prompt/format adaptation and report zero-shot limitations explicitly) to support the claimed superiority.
  3. [Table V, Table IV] Tables IV and V report point estimates only, with no error bars, repeated runs, or significance tests. Given the sample sizes (600–9000 per task), the differences between FUSAR-R1 and baselines in Table V may be significant, but the reader cannot assess variance. In Table IV, the improvement from Experiment 5 to 6 is only 0.676 vs 0.665 F1; without variance or significance testing, the statement that GRPO 'significantly outperforms' CISPO is unsupported. Reporting confidence intervals and a small number of seeds would address this.
  4. [§III-B] The quality of the reconstructed CoT corpus is asserted but not validated. The text states that Qwen3-8B-generated chains are 'expert-level' and 'physically valid' after automatic filtering, but no human evaluation, filtering statistics, or inter-annotator agreement are reported. Since this corpus is the basis of the cold-start SFT and the reward functions in §III-D-3 use ground-truth labels from the same FUSAR-GEOVL-1M family, corpus quality and label correctness are crucial. The authors should quantify filtering rates, provide a human-validated sample, or otherwise demonstrate that the CoT data and reward ground truths are reliable.
minor comments (6)
  1. [Abstract] Typo: 'chain-of-thought(coT)' should be 'chain-of-thought (CoT)'.
  2. [§III-D-3, Eq. (13)] The reward weights α, β, γ, δ are said to be in [0.8, 1.0] in Table II, but the actual values per sub-reward are not specified. This makes the unified reward in Eq. (13) ambiguous and hinders reproducibility.
  3. [Table II] The setting 'Training batch size 1' with 'Gradient accumulation 2' is unusual and should be clarified; also clarify the relationship between the 8-A100 cluster and the 'Rollout GPUs 2 / RLHF GPUs 6 / Inference GPUs 8' configuration.
  4. [References] Reference [21] contains a duplicated author name 'Baiyun Baiyun' and should be corrected.
  5. [Figures 1–4] The figures are referenced but not described in enough detail in the text; ensure they are legible and that captions are self-contained.
  6. [§II-B] FUSAR-GEOVL-1M is central to this work, but its composition is only described qualitatively. Adding basic statistics (number of images, resolution distribution, scene types) would help the reader assess the dataset and the difficulty of the tasks.

Circularity Check

2 steps flagged

Training CoT and evaluation tasks both trace to the authors' FUSAR-GEOVL-1M with no reported train/test split, so the headline gains are not shown to be out-of-distribution; the core 'reasoning' claim partly reduces to reproducing the dataset's own labels.

specific steps
  1. self citation load bearing [Section III-A/III-B (CoT construction) and Section IV-A/Table III (evaluation); cf. Section IV-B-4a]
    "This paper reconstructs the thought process based on the previously constructed FUSAR-GEOVL-1M dataset. ... this study designed six evaluation tasks covering target perception, semantic analysis, and quantitative reasoning (as shown in Table III). ... As the training steps increase, the loss values of both tasks on the test set exhibit a steady and significant decreasing trend."

    The CoT cold-start corpus is a 'deep logical reconstruction' of FUSAR-GEOVL-1M (III-A), and Table III's six tasks—object detection, object classification/counting, main land-cover category, comprehensive land-cover recognition, regional land-cover analysis, and land-cover proportion—are exactly the annotation dimensions FUSAR-GEOVL-1M provides. The paper never defines a train/test split, and the only 'test set' reference (IV-B-4a) gives no separation procedure. If the same FUSAR-GEOVL-1M family supplies both the CoT training labels and the evaluation labels, the large Table V margins over zero-shot baselines are in-distribution reproduction rather than independent reasoning. The authors' own dataset is therefore load-bearing for the central comparison, without external benchmark validation

  2. renaming known result [Section III-B (CoT data construction) and III-C (Eq. 1 cold-start objective)]
    "the original FUSAR-GEOVL-1M mainly provides descriptive labels corresponding to images. ... To compensate for this deficiency, this paper utilizes the locally deployed Qwen3-8B large model as the logic reconstruction engine ... to transform the originally fragmented semantic information into a structured reasoning path. ... the final answer corresponds to the specific interpretation result, such as target category, target number, spatial location, or surface object type."

    The 'expert-level' reasoning chains are not collected from human experts; they are synthesized by Qwen3-8B from the existing FUSAR-GEOVL-1M labels and metadata, and each chain's conclusion is the original label. Cold-start SFT (Eq. 1) then trains the model to autoregressively predict these LLM-written chains and conclusions, and the RL rewards (Eqs. 7-13) compare against the same ground-truth labels. Thus the claimed 'expert reasoning' is a post-hoc textual re-packaging of the dataset's known annotations; evaluation on the same task types does not independently validate physical scattering reasoning. The contribution is partly a renaming of existing labels as chain-of-thought.

full rationale

The paper's own derivation chain does not contain an equation-level equivalence: the model is genuinely optimized with SFT and GRPO, and external baselines are reported. However, the central claim that FUSAR-R1 'consistently outperforms' is supported by an evaluation whose independence from the training corpus is never established. The CoT corpus is reconstructed from the authors' FUSAR-GEOVL-1M, and the six evaluation tasks mirror that dataset's annotation types; the only mention of a 'test set' (IV-B-4a) lacks any split definition. This makes the headline result vulnerable to in-distribution memorization. I do not score this as full circularity because the paper does not explicitly state that the evaluation images are the same instances used in training, and the model does have an independently trainable architecture; but the missing independence guarantee, combined with the load-bearing self-citation of FUSAR-GEOVL-1M, justifies a moderate score.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The 'common scattering feature library' is a prompt-engineering construct inside the CoT generation pipeline, not an independently testable entity. The main unevidenced input is the synthetic CoT corpus and the unverified independence of the evaluation split.

free parameters (4)
  • Reward weights α, β, γ, δ = [0.8, 1.0] (range)
    Weighting of sub-rewards in Eq. (13); chosen by hand, reported only as a range in Table II.
  • Error tolerance threshold τ = not reported
    Appears in Eq. (12) land-cover proportion reward; value is unspecified, affecting the reward scale.
  • KL coefficient β = 0.001
    KL regularization coefficient in Eq. (4); hand-set.
  • GRPO group size G = 4
    Number of sampled trajectories per instruction in Eq. (3); hand-set.
axioms (5)
  • domain assumption A valid SAR interpretation can be modeled as a chain of thought with steps 'observation cues → physical analysis → spatial verification → conclusion generation'.
    Entire method rests on this decomposition (Section III-A, III-C); if SAR interpretation is not faithfully decomposable this way, the CoT corpus teaches a fictional reasoning process.
  • ad hoc to paper The Qwen3-8B-generated CoT corpus is expert-level and physically valid after automatic filtering.
    Section III-B uses a locally deployed Qwen3-8B as the logic reconstruction engine; no human expert validation of the generated chains is reported, and the filtering rules are described qualitatively.
  • domain assumption Ground-truth annotations in FUSAR-GEOVL-1M are correct and complete for detection, counting, and land-cover tasks.
    Both the CoT reconstruction and the RL rewards inherit these labels; label errors propagate into model training and evaluation.
  • domain assumption Metadata (resolution, band, lat/lon, satellite) is available at inference time and is reliable.
    The model's reasoning and ablation gains depend heavily on metadata priors (Table VI, e.g., removing image size drops F1 from 0.676 to 0.325); if metadata is missing at deployment, performance will drop.
  • standard math GRPO provides an unbiased policy optimization signal under the designed reward.
    Equations (3)-(4) adopt the standard GRPO estimator from DeepSeekMath [25]; the paper adds no formal justification.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of FUSAR-R1: A Large-Scale Reasoning Model for Intelligent Interpretation of SAR Images." pith.science (2026). https://pith.science/paper/PDWNTFKL

@misc{pith2026260716819,
  author       = {Pith},
  title        = {Pith review of: FUSAR-R1: A Large-Scale Reasoning Model for Intelligent Interpretation of SAR Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PDWNTFKL}},
  note         = {Machine review of arXiv:2607.16819}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In recent years, large-scale vision-language models have been driving a paradigm shift in intelligent remote sensing image interpretation. By incorporating textual semantic information, the cognitive expression, semantic understanding, and human-computer interaction capabilities of interpretation models have been significantly improved, achieving initial progress in the field of Synthetic Aperture Radar (SAR) image interpretation. However, SAR images are affected by factors such as coherent imaging mechanisms, complex scattering characteristics, speckle noise interference, and target-background coupling, resulting in complex and variable image features with significant uncertainties and specializations. Existing SAR vision-language models do not yet possess the step-by-step analysis, logical judgment, and self-correction capabilities of human experts, making it difficult to support reliable intelligent interpretation in complex scenarios. To address this issue, this paper proposes a large-scale reasoning model, FUSAR-R1, for intelligent interpretation of SAR images. The model first constructs explicit chain-of-thought reasoning data by simulating the interpretation process of human experts and uses this data to guide instruction learning, thereby endowing the model with basic reasoning capabilities. Subsequently, a reinforcement learning strategy is introduced to optimize the model's outputs based on inference results, enabling self-correction and more reliable reasoning. Experimental results demonstrate that FUSAR-R1 consistently outperforms existing multimodal large-scale models across various SAR interpretation tasks, including target detection, target counting and classification, and land-cover category recognition.

Figures

Figures reproduced from arXiv: 2607.16819 by Haipeng Wang, Ruyi Zhang, Xiaokun Zhang, Xinpeng Zhou, Yi Yang, Yuxuan Li.

Figure 1
Figure 1. Figure 1: overall framework of FUSAR-R1 III. METHODS To address the aforementioned issues, this paper proposes a large-scale inference model, FUSAR-R1, for intelligent interpretation of SAR images. This model explicitly models the SAR interpretation process as a chain of thoughts and introduces Group Relative Policy Optimization (GRPO) [25] as a reinforcement learning training paradigm. A. Overall Framework This pap… view at source ↗
Figure 2
Figure 2. Figure 2: SAR interpretation mind chain data construction [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: GRPO reinforcement learning process same task instruction, multiple candidate outputs generated by the model through stochastic sampling may differ in logical rigor and physical consistency. Instead of relying on a predefined absolute scoring criterion, GRPO determines the reward gradient by computing the relative position of the target sample oi within the current sam￾pling group (G), thereby enabling ite… view at source ↗
Figure 6
Figure 6. Figure 6: Airport aircraft reasoning examples real-world interpretation tasks, this section selects two typical scenarios: a civil aviation hub ( [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Reasoning examples for surface ships and artificial infrastructures, thereby identifying the scene as a “clean sea surface background”. During the subsequent search process, the model detects three isolated high-intensity scattering targets. According to the inference process, the model estimates their physical lengths in real time based on the 1 m spatial resolution, obtaining values of 62 m, 67 m, and 37… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

25 extracted references · 5 linked inside Pith

  1. [1]

    Radar imaging technology, 2005

    Zheng Bao, Mengdao Xin, and Tong Wang.Radar imaging technology. Radar imaging technology, 2005

    2005

  2. [2]

    Microwave vision and intelligent interpretation of sar images.Journal of Radar, 13(2):285–306, 2024

    Feng Xu and Yaqiu Jin. Microwave vision and intelligent interpretation of sar images.Journal of Radar, 13(2):285–306, 2024

    2024

  3. [3]

    Intelligent SAR Image Processing and Interpretation, 2008

    Licheng Jiao.Intelligent SAR Image Processing and Interpretation. Intelligent SAR Image Processing and Interpretation, 2008

    2008

  4. [4]

    Transfer learning with deep convolutional neural network for sar target classification with limited labeled data.Remote sensing, 9(9):907, 2017

    Zhongling Huang, Zongxu Pan, and Bin Lei. Transfer learning with deep convolutional neural network for sar target classification with limited labeled data.Remote sensing, 9(9):907, 2017

    2017

  5. [5]

    Analysis and effects of speckle noise in sar images

    Prabhishek Singh and Raj Shree. Analysis and effects of speckle noise in sar images. In2016 2nd International Conference on Advances in Computing, Communication, & Automation (ICACCA)(Fall), pages 1–5. IEEE, 2016

    2016

  6. [6]

    Layover solution in sar imaging: A statistical approach.IEEE Geoscience and Remote Sensing Letters, 6(3):577–581, 2009

    Fabio Baselice, Alessandra Budillon, Giampaolo Ferraioli, and Vito Pascazio. Layover solution in sar imaging: A statistical approach.IEEE Geoscience and Remote Sensing Letters, 6(3):577–581, 2009

    2009

  7. [7]

    Sarlang-1m: A benchmark for vision-language modeling in sar image understanding.arXiv preprint arXiv:2504.03254, 2025

    Yuxiang Wei et al. Sarlang-1m: A benchmark for vision-language modeling in sar image understanding.arXiv preprint arXiv:2504.03254, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Sarclip: A vision language foundation model for semantic understanding and target recognition in sar imagery

    Qiwei Ma, Zhiyu Wang, Wang Liu, Xukun Lu, Bin Deng, Puhong Duan, Xudong Kang, and Shutao Li. Sarclip: A vision language foundation model for semantic understanding and target recognition in sar imagery. arXiv preprint arXiv:2510.22665, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Ssl-lip: A two-stage pre-training foundation model for sar images

    Yi Yang, Qingchen Fang, Xiaokun Zhang, and Haipeng Wang. Ssl-lip: A two-stage pre-training foundation model for sar images. InIGARSS 2025 - 2025 IEEE International Geoscience and Remote Sensing Symposium, 2025

    2025

  10. [10]

    OpenAI o1 System Card

    OpenAI. OpenAI o1 System Card. Technical report, OpenAI, December 2024

    2024

  11. [11]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforce- ment learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforce- ment learning.arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Visual-rft: Visual reinforcement fine- tuning

    Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine- tuning. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2034–2044, 2025. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14

    2034

  13. [13]

    Remoteclip: A vision language foundation model for remote sensing.IEEE Transactions on Geoscience and Remote Sensing, 62:1–16, 2024

    Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing.IEEE Transactions on Geoscience and Remote Sensing, 62:1–16, 2024

    2024

  14. [14]

    Skyscript: A large and semantically diverse vision-language dataset for remote sensing

    Zhecheng Wang, Rajanie Prabha, Tianyuan Huang, Jiajun Wu, and Ram Rajagopal. Skyscript: A large and semantically diverse vision-language dataset for remote sensing. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5805–5813, 2024

    2024

  15. [15]

    Vhm: Versatile and honest vision language model for remote sensing image analysis

    Chao Pang, Xingxing Weng, Jiang Wu, Jiayu Li, Yi Liu, Jiaxing Sun, Weijia Li, Shuai Wang, Litong Feng, Gui-Song Xia, et al. Vhm: Versatile and honest vision language model for remote sensing image analysis. InProceedings of the AAAI Conference on Artificial Intelligence, vol- ume 39, pages 6381–6388, 2025

    2025

  16. [16]

    Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model.ISPRS Journal of Photogrammetry and Remote Sensing, 221:64– 77, 2025

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

    2025

  17. [17]

    Geochat: Grounded large vision-language model for remote sensing

    Kartik Kuckreja, Muhammad Sohail Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad Shahbaz Khan. Geochat: Grounded large vision-language model for remote sensing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27831–27840, June 2024

    2024

  18. [18]

    Bootstrapping interactive image–text alignment for remote sensing image captioning.IEEE Transactions on Geoscience and Remote Sensing, 62:1–12, 2024

    Cong Yang, Zuchao Li, and Lefei Zhang. Bootstrapping interactive image–text alignment for remote sensing image captioning.IEEE Transactions on Geoscience and Remote Sensing, 62:1–12, 2024

    2024

  19. [19]

    A new learning paradigm for foundation model-based remote-sensing change detection.IEEE Transactions on Geoscience and Remote Sensing, 62:1–12, 2024

    Kaiyu Li, Xiangyong Cao, and Deyu Meng. A new learning paradigm for foundation model-based remote-sensing change detection.IEEE Transactions on Geoscience and Remote Sensing, 62:1–12, 2024

    2024

  20. [20]

    FUSAR-KLIP: Towards multimodal foundation models for remote sensing.arXiv preprint arXiv:2509.23927, 2025

    Yi Yang, Xiaokun Zhang, Qingchen Fang, Jing Liu, Ziqi Ye, Rui Li, Li Liu, and Haipeng Wang. FUSAR-KLIP: Towards multimodal foundation models for remote sensing.arXiv preprint arXiv:2509.23927, 2025

    arXiv 2025

  21. [21]

    FUSAR-GPT: A spatiotemporal feature-embedded and two-stage de- coupled visual language model for sar imagery

    Xiaokun Zhang, Yi Yang, Ziqi Ye, Baiyun Baiyun, Xiaorong Guo, Qingchen Fang, Ruyi Zhang, Xinpeng Zhou, and Haipeng Wang. FUSAR-GPT: A spatiotemporal feature-embedded and two-stage de- coupled visual language model for sar imagery. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 42158–42168, June 2026

    2026

  22. [22]

    Towards faithful rea- soning in remote sensing: A perceptually-grounded geospatial chain-of- thought for vision-language models.arXiv preprint arXiv:2509.22221, 2025

    Jiaqi Liu, Lang Sun, Ronghao Fu, and Bo Yang. Towards faithful rea- soning in remote sensing: A perceptually-grounded geospatial chain-of- thought for vision-language models.arXiv preprint arXiv:2509.22221, 2025

    arXiv 2025

  23. [23]

    Geo-r1: Improving few-shot geospatial referring expression understand- ing with reinforcement fine-tuning.arXiv preprint arXiv:2509.21976, 2025

    Zilun Zhang, Zian Guan, Tiancheng Zhao, Haozhan Shen, Tianyu Li, Yuxiang Cai, Zhonggen Su, Zhaojun Liu, Jianwei Yin, and Xiang Li. Geo-r1: Improving few-shot geospatial referring expression understand- ing with reinforcement fine-tuning.arXiv preprint arXiv:2509.21976, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Rsground-r1: Rethinking remote sensing visual grounding through spatial reasoning.arXiv preprint arXiv:2601.21634, 2026

    Shiqi Huang, Shuting He, and Bihan Wen. Rsground-r1: Rethinking remote sensing visual grounding through spatial reasoning.arXiv preprint arXiv:2601.21634, 2026

    arXiv 2026

  25. [25]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    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 mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.