Pith. sign in

REVIEW 4 major objections 7 minor 53 references

FUSAR-GPT : A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery

T0 review · 4 major / 7 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read FUSAR-GPT embeds geospatial and temporal priors into SAR image tokens and, via two-stage decoupled fine-tuning, beats general visual-language baselines by over 10 percentage points on four SAR interpretation tasks.

desk verdict Promising SAR-VLM architecture with a plausible fusion idea, but the empirical claims are not yet testable because the evaluation set appears to overlap the fine-tuning set and there is no external benchmark. read the letter →

arxiv 2602.19190 v4 pith:SBFWYRBT submitted 2026-02-22 cs.CV cs.AI

Xiaokun Zhang , Yi Yang , Ziqi Ye , Baiyun , Xiaorong Guo , Qingchen Fang , Ruyi Zhang , Xinpeng Zhou
show 1 more author
Haipeng Wang
This is my paper
classification cs.CVcs.AI
keywords SARimageryvision-languagemodelspatiotemporalfeatureembeddinggeospatialpriortoken-wiselinearmodulationtwo-stagefine-tuningremotesensinginterpretationtargetdetection
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

FUSAR-GPT is built around one claim: a visual-language model that reads synthetic aperture radar (SAR) images can be made to understand them by supplying what SAR pixels lack — geographic and temporal context. The paper's recipe is to sample a global geospatial embedding field at each image's location and year, convert those embeddings into per-channel modulation parameters, and apply them directly to the visual tokens, followed by a two-stage fine-tuning schedule that first aligns SAR, geospatial, and text semantics and only then tunes task behaviour. On counting, spatial localization, classification, and detection benchmarks, the model beats mainstream vision-language baselines by roughly 10-28 percentage points. If true, this means SAR understanding does not require a new architecture from scratch; it can be achieved by injecting external geospatial priors into an existing VLM and separating knowledge injection from task execution.

What carries the argument

Two mechanisms carry the argument. The first is 'spatiotemporal anchors': for a SAR image, the method computes a geographic bounding box and imaging year, samples a 64-dimensional embedding vector from a global geospatial embedding model on a regular latitude/longitude grid, and linearly maps each geographic coordinate to a pixel coordinate in the image. The second is Token-wise Linear Modulation (TLM): each sampled embedding vector is projected through a two-layer MLP into channel-wise scaling and shifting coefficients, which are then interpolated onto the visual feature grid using Gaussian-distance weights and applied as affine transformations to the visual tokens. The training protocol is

What would settle it

Take a SAR scene with known strong terrain relief, compute the paper's linear geolocation-to-pixel mapping and a rigorous range-Doppler geocoding, and compare where the geospatial priors land; if the two placements differ by more than a few pixels and the model's answers shift accordingly, the alignment assumption is doing real work and currently unchecked. A simpler test: degrade the geospatial embeddings to random vectors at the same coordinates and see whether counting accuracy falls from 52% toward the 42% TLM-only level.

Watch

Extended reading notes

Core claim

The paper's central claim is that the reason general vision-language models fail on SAR imagery is not only the optical-SAR modality gap but a lack of geospatial context and a training objective that conflates knowledge injection with task execution. FUSAR-GPT addresses both by injecting a global geospatial embedding field — sampled at the image's location and year — into the visual token stream via a Token-wise Linear Modulation (TLM) module, and by decoupling alignment from instruction tuning. On four SAR benchmarks (counting, spatial localization, classification, detection) the model reports state-of-the-art numbers, with the clearest margins on detection (F1 74.8% vs 47.1% at IoU 0.25) a

Load-bearing premise

The mechanism assumes that the linear mapping from geographic coordinates to SAR pixel coordinates places the geospatial priors at the right image locations; SAR's geometric distortions (foreshortening, layover, shadow) can break that alignment, and the paper provides no direct validation of the mapping on real SAR data.

Editorial extensions

If this is right

  • On target counting, FUSAR-GPT reaches 52.53% accuracy against a best baseline of 45.45%, and the model's advantage holds across model sizes — larger general baselines do not close the gap.
  • On grid-based spatial localization, the model achieves 52.02% exact-match and 91.41% top-1 accuracy, beating the best baseline by 8-12 points.
  • On target detection, overall F1 at IoU 0.25 rises from 47.1% to 74.8%, with plane F1 up 28 points, indicating large gains for small, low-contrast targets.
  • Ablations attribute the gain specifically to the TLM fusion (34.85% → 41.92% on counting) and to the first-stage knowledge-injection alignment (41.92% → 52.23%), showing both components are needed.
  • The two-stage schedule indicates that SAR domain adaptation can be separated cleanly from downstream task learning, avoiding conflicting optimization in a single fine-tuning pass.

Reading between the lines

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

  • The same spatiotemporal-anchor plus modulation recipe transfers naturally to other sparse or ambiguous sensor modalities (hyperspectral, night-time optical, lidar-derived intensity), where a global geographic prior can disambiguate otherwise unreadable pixels.
  • The reported localization and detection numbers depend on the coordinate conventions and bounding-box preprocessing used in the benchmark; whether these gains survive on other SAR datasets with different geolocation metadata is an open question the paper does not address.
  • The two-stage decoupling suggests a cheap domain-adaptation recipe: if a descriptive corpus exists but task labels are scarce, align the representation first with frozen backbones, then adapt a tiny parameter set to the tasks — this could make SAR VLMs practical with far fewer task annotations.
  • A direct stress test of the geolocation mapping on high-relief terrain would show whether the linear coordinate assumption is a ceiling on real-world SAR scenes with strong foreshortening and shadow.
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 / 7 minor

Summary. The paper presents FUSAR-GPT, a SAR-oriented vision-language model built on Qwen2.5-VL-7B. The method has three pillars: (i) a spatiotemporal feature-extraction step that queries AlphaEarth (AEF) 64-dimensional embeddings on a lon/lat grid and linearly maps them to SAR pixel coordinates; (ii) a Token-wise Linear Modulation (TLM) module that converts these heterogeneous AEF priors into channel-wise affine modulation parameters for visual tokens via Gaussian-weighted spatial interpolation; and (iii) a two-stage SFT scheme that first trains only the AEF embedding MLP on descriptive data (FUSAR-GEOVL-1M) and then trains only LoRA adapters on task-specific data. The authors claim state-of-the-art performance on target counting, spatial localization, target classification, and target detection, with margins of over 10% versus mainstream VLMs, supported by Tables 2-4 and supplementary experiments. The evaluation is conducted entirely on a newly constructed, non-released FUSAR-GPT dataset.

Significance. If the empirical claims are sound, the paper makes a useful contribution: it introduces a triplet data paradigm (SAR image, text, geospatial feature), a lightweight TLM fusion mechanism that avoids early concatenation, and a two-stage decoupling of knowledge injection from task adaptation. The architectural choices are clearly described and the ablations are internally consistent in showing incremental gains. However, the central 'state-of-the-art' claim is currently not testable because the evaluation set appears to overlap with the Stage-2 fine-tuning set and the baselines are not given the same fine-tuning protocol. The absence of a released dataset or an external benchmark further limits verification. The geolocation-to-pixel alignment underlying the whole TLM mechanism is also unvalidated. If these issues are fixed, the paper could be a valuable system paper; in its present form, the evidence does not support the headline quantitative claims.

major comments (4)
  1. [Section 4.1 / Table 1] The paper states that a subset of 2k images was selected for 'downstream task training and evaluation' and that 'Stage 2 fine-tuning and evaluation are conducted on FUSAR-GPT', but no train/evaluation split is described. There are no split sizes, image IDs, or a statement that the evaluation images are disjoint from the LoRA fine-tuning images. If Tables 2-4 are computed on images used in Stage 2, the reported results reflect memorization rather than generalization. This is the load-bearing issue for the paper's main claim and must be resolved by explicitly defining a disjoint split, reporting its construction, and documenting the exact evaluation subset.
  2. [Section 4.2] All baselines were evaluated 'using their default parameters' in MS-SWIFT, i.e., in a zero-shot or near-zero-shot regime, while FUSAR-GPT is fine-tuned on the task data. Even if the evaluation split is disjoint, this is not an apples-to-apples comparison: the baselines have not seen any SAR task instruction examples, whereas FUSAR-GPT has been trained on the same task distribution. The authors should either fine-tune all baselines on the same Stage-2 training data or clearly frame the comparison as 'fine-tuned FUSAR-GPT versus zero-shot baselines' and temper the SOTA claim accordingly.
  3. [Section 3.1, Eqs. (2)-(3)] The core fusion mechanism rests on a linear mapping from geographic coordinates (lon, lat) to SAR image pixel coordinates. The paper provides no validation of this mapping on real SAR data, no discussion of map projection, and no treatment of SAR geometric distortions (foreshortening, layover, shadow). An inaccurate mapping means AEF priors are injected at incorrect spatial locations, so the TLM mechanism could be averaging semantically unrelated embeddings. The authors should report the source of image georeferencing, the assumed projection, and a quantitative alignment check (e.g., tie-point error or qualitative overlay examples).
  4. [Section 4.3 / Table 9] No error bars, number of test samples, or multiple-run statistics are reported. Moreover, the full model's counting accuracy is 52.53% in Table 2 but 52.23% in Table 9 for the same configuration. This inconsistency needs clarification. Given that several reported margins are on the order of a few percentage points, the absence of variance estimates and the internal discrepancy make it difficult to judge whether the gains are statistically significant.
minor comments (7)
  1. [Throughout] Several typos: 'Dateset' (Section 4.1), 'Relate Work' (Section 2), 'Target lassification' (Table 3), and inconsistent 'AFE' versus 'AEF' in Section 3.1.
  2. [Eq. (3)] The tuple notation says '(lon_i, lon_j)' in the coordinate text; this should presumably be '(lon_i, lat_j)' to match Eq. (2).
  3. [Eq. (7)] The Gaussian kernel width sigma is a free parameter, but its value and sensitivity are not reported. A short sensitivity analysis or a chosen value in Section 4.2 would improve reproducibility.
  4. [Table 6] The row labeled 'Base TLM Sum Concat ChatGPT-5.2 Gemini-3' is unclear: it mixes fusion variants with proprietary chat models and does not define what 'Base' means. The caption should be expanded to describe each column.
  5. [Figure 5] The claim of 'faster convergence and higher data efficiency' is based on a figure without axis definitions, metric specifications, or number of evaluation points. Please clarify what is plotted and on which split.
  6. [Abstract / Section 1] The abstract claims 'over 10%' improvement, while Table 2 shows about 7 points over the best baseline on counting. The global claim may be defensible through the detection and classification tables, but the per-task margins should be stated individually to avoid overgeneralization.
  7. [Nomenclature] The name 'FUSAR-GPT' is used for both the model and the evaluation dataset (Table 1). This is confusing; a distinct dataset name would improve clarity.

Circularity Check

2 steps flagged · score 6.0 of 10

Central >10% SOTA claim is an in-sample comparison: Stage-2 LoRA is optimized on the same FUSAR-GPT set used for Tables 2-4, baselines are zero-shot, and the benchmark comes from the authors' own unreleased FUSAR-KLIP line.

  1. fitted input called prediction [Section 4.1 / Table 1 / Section 4.2 / Eq. (14)]
    "Subsequently, a subset of 2k images containing precise ground-truth target annotations was further selected for downstream task training and evaluation ... Stage 2 fine-tuning and evaluation are conducted on FUSAR-GPT ... All baseline models were evaluated using their default parameters ... The loss function is optimized exclusively over the LoRA parameters θlora"

    The LoRA parameters behind Tables 2-4 are fit on D2 = {(I_sar, F_ae, T_inst, T_ans)} drawn from the same 2k FUSAR-GPT images on which the tables report accuracy; no train/eval split is given. Baseline VLMs are not fine-tuned ('default parameters'), so the reported >10% margin is a tuned-vs-untuned comparison on the tuning set. The 'prediction' is therefore in-sample by construction, not a generalization result.

  2. self citation load bearing [Section 3.3 (Stage I) and Section 4.1]
    "This corpus is derived from the FUSAR-GEOVL-1M dataset [42] ... The data used in this study are sourced from the FUSAR-GEOVL dataset within FUSAR-KLIP"

    Reference [42] (FUSAR-KLIP) shares authors with this paper (Yi Yang, Xiaokun Zhang, Qingchen Fang, Ziqi Ye, Haipeng Wang). The training corpus and the evaluation set are both taken from this unreleased self-cited dataset, so the claimed 'inaugural' triplet benchmark and the SOTA results are validated only against the authors' own data line, with no external SAR-VQA benchmark or released split to break the loop.

full rationale

The method itself is not definitionally circular: AEF features are an external prior, the TLM Eqs. (4)-(11) define a standard affine modulation, and the two SFT losses (Eqs. 12 and 14) are ordinary likelihood objectives. The unvalidated linear geolocation-to-pixel mapping (Eqs. 2-3) is a correctness risk, not a circularity. The circularity is in the empirical claim: Stage-2 LoRA is optimized on the same FUSAR-GPT set later reported in Tables 2-5, and the baselines are run zero-shot, so the central '>10% improvement' is forced by the protocol rather than demonstrated on held-out data. This is compounded by the fact that the benchmark itself is drawn from the authors' prior FUSAR-KLIP work and not released. Score 6 reflects one load-bearing in-sample 'prediction' plus a supporting self-citation chain; the architecture retains independent content.

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

The paper introduces no new physical entities. It relies on design choices (grid size, sigma, LoRA rank) and assumptions about the quality and alignment of AEF priors, plus an unvalidated self-constructed benchmark.

free parameters (4)
  • Gaussian kernel width sigma (Eq. 7) = not reported
    Controls the spatial influence radius of each AEF prior on visual tokens; chosen by hand, not specified or tuned with validation.
  • Sampling grid size N_lon x N_lat (Eq. 2) = not reported
    Determines the density of AEF anchor points; chosen by hand, not specified.
  • LoRA rank and alpha = rank=8, alpha=32
    Hyperparameters for the LoRA adapters used in Stage 2 fine-tuning.
  • Learning rates and epochs for Stage 1/Stage 2 = 1e-4 for 30 epochs; 1e-5 for 5 epochs
    Training schedule chosen to stabilize the two-stage SFT; no sensitivity analysis.
assumptions (4)
  • domain assumption AEF embeddings are a stable geospatial prior relevant to SAR interpretation
    Section 3.1 states AEF provides 'cross-modal prior knowledge' and 'macroscopic geographic semantics remain relatively consistent at the annual scale.' This is not validated for SAR-specific tasks.
  • domain assumption Linear mapping from lon/lat to SAR pixel coordinates is a valid approximation
    Equations 2-3 define a linear mapping without considering SAR geometric distortions (foreshortening, layover, shadow) or map projections. No validation is provided.
  • ad hoc to paper Two-stage SFT decoupling improves performance
    The claimed benefit of separating knowledge injection (Stage 1) from task execution (Stage 2) is a design hypothesis, not derived from theory or prior evidence.
  • domain assumption The FUSAR-GPT test subset is a representative benchmark
    The 2k-image subset used for evaluation is drawn from the authors' own FUSAR-GEOVL dataset; no external validation or description of the split (train/test) is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FUSAR-GPT : A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery." pith.science (2026). https://pith.science/paper/SBFWYRBT

@misc{pith2026260219190,
  author       = {Pith},
  title        = {Pith review of: FUSAR-GPT : A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SBFWYRBT}},
  note         = {Machine review of arXiv:2602.19190}
}
read the original abstract

Research on the intelligent interpretation of all-weather, all-time Synthetic Aperture Radar (SAR) is crucial for advancing remote sensing applications. In recent years, although Visual Language Models (VLMs) have demonstrated strong open-world understanding capabilities on RGB images, their performance is severely limited when directly applied to the SAR field due to the complexity of the imaging mechanism, sensitivity to scattering features, and the scarcity of high-quality text corpora. To systematically address this issue, we constructed the inaugural SAR Image-Text-AlphaEarth feature triplet dataset and developed FUSAR-GPT, a VLM specifically for SAR. FUSAR-GPT innovatively introduces a geospatial baseline model as a 'world knowledge' prior and embeds multi-source remote-sensing temporal features into the model's visual backbone via 'spatiotemporal anchors', enabling dynamic compensation for the sparse representation of targets in SAR images. Furthermore, we designed a two-stage SFT strategy to decouple the knowledge injection and task execution of large models. The spatiotemporal feature embedding and the two-stage decoupling paradigm enable FUSAR-GPT to achieve state-of-the-art performance across several typical remote sensing visual-language benchmark tests, significantly outperforming mainstream baseline models by over 10%.

Figures

Figures reproduced from arXiv: 2602.19190 by the authors.

Figure 1
Figure 1. FUSAR-GPT: Embeds spatiotemporal features for SAR [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of FUSAR-GPT. The framework adopts a two-stage training strategy: Stage-1 jointly updates LoRA and the TLM [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. AEF–SAR Visual Comparison. AEF embeddings extracted at different times are visualized by mapping channels 1, 16 and 9 of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Overview of the four main downstream tasks: Target Counting, Spatial Localization, Target Classification, and Target Detection. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Performance of different training stages across down [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Challenges in developing SAR visual language models. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Different Fusion Strategies object detection task. Although our approach is not specif￾ically designed for detection, it demonstrates competitive performance. These results provide an initial indication of the potential of our framework for SAR detection tasks. We leav…
Figure 8
Figure 8. Figure 8: (a) shows the data characteristics of FUSAR-GEOVL, (b) shows the data characteristics of FUSAR-GPT. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 3 linked inside Pith

  1. [1]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025. 7, 2

    arXiv 2025

  2. [2]

    Qwen2.5-vl technical report, 2025

    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, Zhao- hai 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,...

    2025

  3. [3]

    Multi-spectral remote sensing image retrieval using geospatial foundation models

    Benedikt Blumenstiel, Viktoria Moor, Romeo Kienzler, and Thomas Brunschwiler. Multi-spectral remote sensing image retrieval using geospatial foundation models. In IGARSS 2024-2024 IEEE International Geoscience and Re- mote Sensing Symposium, pages 7286–7291. IEEE, 2024. 1, 2

    2024

  4. [4]

    Brown, Michal R

    Christopher F. Brown, Michal R. Kazmierski, Valerie J. Pasquarella, William J. Rucklidge, Masha Samsikova, Chen- hui Zhang, Evan Shelhamer, Estefania Lahera, Olivia Wiles, Simon Ilyushchenko, Noel Gorelick, Lihui Lydia Zhang, Sophia Alj, Emily Schechter, Sean Askay, Oliver Guinan, Rebecca Moore, Alexis Boukouvalas, and Pushmeet Kohli. Alphaearth foundatio...

  5. [5]

    Changeclip: Remote sensing change detection with multi- modal vision-language representation learning.ISPRS Jour- nal of Photogrammetry and Remote Sensing, 208:53–69,

    Sijun Dong, Libo Wang, Bo Du, and Xiaoliang Meng. Changeclip: Remote sensing change detection with multi- modal vision-language representation learning.ISPRS Jour- nal of Photogrammetry and Remote Sensing, 208:53–69,

  6. [6]

    Combining sam with limited data for change detection in remote sensing.IEEE Transactions on Geoscience and Remote Sensing, 2025

    Junyu Gao, Da Zhang, Feiyu Wang, Lichen Ning, Zhiyuan Zhao, and Xuelong Li. Combining sam with limited data for change detection in remote sensing.IEEE Transactions on Geoscience and Remote Sensing, 2025. 1

    2025

  7. [7]

    Deep- learning for radar: A survey.IEEE Access, 9:141800– 141818, 2021

    Zhe Geng, He Yan, Jindong Zhang, and Daiyin Zhu. Deep- learning for radar: A survey.IEEE Access, 9:141800– 141818, 2021. 2

    2021

  8. [8]

    Unpaired image-text match- ing via multimodal aligned conceptual knowledge.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(7):5160–5176, 2024

    Yan Huang, Yuming Wang, Yunan Zeng, Junshi Huang, Zhenhua Chai, and Liang Wang. Unpaired image-text match- ing via multimodal aligned conceptual knowledge.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(7):5160–5176, 2024. 2

    2024

Show all 53 references
  1. [9]

    A fast progressive ship detection method for very large full-scene sar images.IEEE Transactions on Geo- science and Remote Sensing, 62:1–15, 2024

    Hecheng Jia, Xinyang Pu, Qiaoyu Liu, Haipeng Wang, and Feng Xu. A fast progressive ship detection method for very large full-scene sar images.IEEE Transactions on Geo- science and Remote Sensing, 62:1–15, 2024. 2

    2024

  2. [10]

    Sarclip: a multimodal foundation framework for sar imagery via con- trastive language-image pre-training.ISPRS Journal of Pho- togrammetry and Remote Sensing, 231:17–34, 2026

    Chaowei Jiang, Chao Wang, Fan Wu, Peifeng Ma, Lichuan Zou, Tianyang Li, Jing Ning, and Yixian Tang. Sarclip: a multimodal foundation framework for sar imagery via con- trastive language-image pre-training.ISPRS Journal of Pho- togrammetry and Remote Sensing, 231:17–34, 2026. 3, 1, 2

    2026

  3. [11]

    Em-yolo: A fine-grained recognition model for aircraft targets in sar im- ages

    Luewen Jiang, Jing Zang, and Yan Zhang. Em-yolo: A fine-grained recognition model for aircraft targets in sar im- ages. In2025 International Conference on Communication Networks and Smart Systems Engineering (ICCNSE), pages 255–259, 2025. 2

    2025

  4. [12]

    Sar ship detection based on an improved faster r-cnn using deformable convolution

    Xiao Ke, Xiaoling Zhang, Tianwen Zhang, Jun Shi, and Shunjun Wei. Sar ship detection based on an improved faster r-cnn using deformable convolution. In2021 IEEE Interna- tional Geoscience and Remote Sensing Symposium IGARSS, pages 3565–3568, 2021. 2

    2021

  5. [13]

    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, pages 27831– 27840...

    2024

  6. [14]

    Llava-onevision: Easy visual task transfer, 2024

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Zi- wei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 5

    2024

  7. [15]

    BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InProceed- ings of the 39th International Conference on Machine Learn- ing, pages 12888–12900. PMLR, 2022. 1

    2022

  8. [16]

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

    Kaiyu Li, Xiangyong Cao, and Deyu Meng. A new learn- ing paradigm for foundation model-based remote-sensing change detection.IEEE Transactions on Geoscience and Re- mote Sensing, 62:1–12, 2024. 1, 2

    2024

  9. [17]

    Co-training vision-language models for remote sensing multi-task learning.Remote Sensing, 18(2): 222, 2026

    Qingyun Li, Shuran Ma, Junwei Luo, Yi Yu, Yue Zhou, Fengxiang Wang, Xudong Lu, Xiaoxing Wang, Xin He, Yushi Chen, et al. Co-training vision-language models for remote sensing multi-task learning.Remote Sensing, 18(2): 222, 2026. 7

    2026

  10. [18]

    Weijie Li, Wei Yang, Tianpeng Liu, Yuenan Hou, Yuxuan Li, Zhen Liu, Yongxiang Liu, and Li Liu. Predicting gradient is better: Exploring self-supervised learning for sar atr with a joint-embedding predictive architecture.ISPRS Journal of Photogrammetry and Remote Sensing, 218:3...

    2024

  11. [19]

    Star: A first-ever dataset and a large-scale benchmark for scene graph generation in large-size satellite imagery.IEEE Trans

    Yansheng Li, Linlin Wang, Tingzhu Wang, Xue Yang, Jun- wei Luo, Qi Wang, Youming Deng, Wenbin Wang, Xian Sun, Haifeng Li, et al. Star: A first-ever dataset and a large-scale benchmark for scene graph generation in large-size satellite imagery.IEEE Trans. Pattern Anal. Mach. In...

    2025

  12. [20]

    Re- moteclip: 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. Re- moteclip: A vision language foundation model for remote sensing.IEEE Transactions on Geoscience and Remote Sensing, 62:1–16, 2024. 2, 1

    2024

  13. [21]

    Improved baselines with visual instruction tuning, 2023

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 7, 2

    2023

  14. [22]

    Llava-next: Im- proved reasoning, ocr, and world knowledge, 2024

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-next: Im- proved reasoning, ocr, and world knowledge, 2024. 7, 2

    2024

  15. [23]

    Texture classification from random features.IEEE transactions on pattern analysis and machine intelligence, 34(3):574–586, 2012

    Li Liu and Paul Fieguth. Texture classification from random features.IEEE transactions on pattern analysis and machine intelligence, 34(3):574–586, 2012. 2

    2012

  16. [24]

    A causal adjustment module for debiasing scene graph generation.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 47(5): 4024–4043, 2025

    Li Liu, Shuzhou Sun, Shuaifeng Zhi, Fan Shi, Zhen Liu, Janne Heikkil ¨a, and Yongxiang Liu. A causal adjustment module for debiasing scene graph generation.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 47(5): 4024–4043, 2025. 2

    2025

  17. [25]

    Atrnet-star: A large dataset and benchmark to- wards remote sensing object recognition in the wild.IEEE Transactions on Pattern Analysis and Machine Intelligence,

    Yongxiang Liu, Weijie Li, Li Liu, Jie Zhou, Bowen Peng, Yafei Song, Xuying Xiong, Wei Yang, Tianpeng Liu, Zhen Liu, et al. Atrnet-star: A large dataset and benchmark to- wards remote sensing object recognition in the wild.IEEE Transactions on Pattern Analysis and Machine Intelligence,

  18. [26]

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

    Chao Pang, Xingxing Weng, Jiang Wu, Jiayu Li, Yi Liu, Ji- axing 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, pag...

    2025

  19. [27]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceedings of the...

    2021

  20. [28]

    Grounding everything in to- kens for multimodal large language models.arXiv preprint arXiv:2512.10554, 2025

    Xiangxuan Ren, Zhongdao Wang, Liping Hou, Pin Tang, Guoqing Wang, and Chao Ma. Grounding everything in to- kens for multimodal large language models.arXiv preprint arXiv:2512.10554, 2025. 7

    2025 arXiv

  21. [29]

    Earthdial: Turning multi-sensory earth observations to interactive dialogues

    Sagar Soni, Akshay Dudhane, Hiyam Debary, Mustansar Fiaz, Muhammad Akhtar Munir, Muhammad Sohail Dan- ish, Paolo Fraccaro, Campbell D Watson, Levente J Klein, Fahad Shahbaz Khan, and Salman Khan. Earthdial: Turning multi-sensory earth observations to interactive dialogues. In ...

    2025

  22. [30]

    Fully polsar image reconstruction for enhanced land cover mapping.Pattern Recognition, 169:111895, 2026

    Xu Sun, Junyu Gao, and Yuan Yuan. Fully polsar image reconstruction for enhanced land cover mapping.Pattern Recognition, 169:111895, 2026. 2

    2026

  23. [31]

    Vi- sual position prompt for mllm based visual grounding.IEEE Transactions on Multimedia, 2026

    Wei Tang, Yanpeng Sun, Qinying Gu, and Zechao Li. Vi- sual position prompt for mllm based visual grounding.IEEE Transactions on Multimedia, 2026. 7

    2026

  24. [32]

    Re- ferring expressions as a lens into spatial language grounding in vision-language models

    Akshar Tumu, Varad Shinde, and Parisa Kordjamshidi. Re- ferring expressions as a lens into spatial language grounding in vision-language models. InProceedings of the 14th Inter- national Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific...

  25. [33]

    Group equivariant u-net for the semantic segmentation of sar im- ages

    ˙Ilter T¨urkmenli, Erchan Aptoula, and Koray Kayabol. Group equivariant u-net for the semantic segmentation of sar im- ages. In2022 30th Signal Processing and Communications Applications Conference (SIU), pages 1–4, 2022. 2

    2022

  26. [34]

    Annotation-free, high-fidelity sar oil-spill image synthesis via classification-guided diffusion model.IEEE Transactions on Geoscience and Remote Sens- ing, 63:1–11, 2025

    Bin Wang, Song Dai, Dongmei Song, Lei Chen, Weimin Chen, and Jintao Yu. Annotation-free, high-fidelity sar oil-spill image synthesis via classification-guided diffusion model.IEEE Transactions on Geoscience and Remote Sens- ing, 63:1–11, 2025. 1

    2025

  27. [35]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Jun- yang Lin. Qwen2-vl: Enhancing vision-language model’s ...

    2024

  28. [36]

    Internvl3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Sheng- long Ye, Jie Shao, et al. Internvl3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265, 2025. 7, 2

    2025 arXiv

  29. [37]

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

    Zhecheng Wang, Rajanie Prabha, Tianyuan Huang, Jiajun Wu, and Ram Rajagopal. Skyscript: A large and seman- tically diverse vision-language dataset for remote sensing. Proceedings of the AAAI Conference on Artificial Intelli- gence, 38(6):5805–5813, 2024. 2, 1

    2024

  30. [38]

    Sarlang-1m: A benchmark for vision-language modeling in sar image un- derstanding.IEEE Transactions on Geoscience and Remote Sensing, 2026

    Yimin Wei, Aoran Xiao, Yexian Ren, Yuting Zhu, Hongruix- uan Chen, Junshi Xia, and Naoto Yokoya. Sarlang-1m: A benchmark for vision-language modeling in sar image un- derstanding.IEEE Transactions on Geoscience and Remote Sensing, 2026. 3, 1, 2

    2026

  31. [39]

    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. 1, 2

    2024

  32. [40]

    R3det: Refined single-stage detector with feature refinement for ro- tating object

    Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3det: Refined single-stage detector with feature refinement for ro- tating object. InProceedings of the AAAI conference on ar- tificial intelligence, pages 3163–3171, 2021. 2

    2021

  33. [41]

    Yi Yang, Zhengxin Lei, Xiuci Mo, Da Lu, Hecheng Jia, and Haipeng Wang. Sardet-cl: Self-supervised contrastive learn- ing with feature enhancement and imaging mechanism con- straints for sar target detection.IEEE Transactions on Geo- science and Remote Sensing, 63:1–15, 2025. 2

    2025

  34. [42]

    Fusar-klip: Towards multimodal foundation models for remote sensing, 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, 2025. 5

    2025

  35. [43]

    Object fidelity diffu- sion for remote sensing image generation.arXiv preprint arXiv:2508.10801, 2025

    Ziqi Ye, Shuran Ma, Jie Yang, Xiaoyi Yang, Ziyang Gong, Xue Yang, and Haipeng Wang. Object fidelity diffu- sion for remote sensing image generation.arXiv preprint arXiv:2508.10801, 2025. 2

    2025

  36. [44]

    Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multimodal large lan- guage model for multisensor image comprehension in re- mote sensing domain.IEEE Transactions on Geoscience and Remote Sensing, 62:1–20, 2024. 2

    2024

  37. [45]

    Earthgpt-x: A spatial mllm for multi-level multi-source re- mote sensing imagery understanding with visual prompting,

    Wei Zhang, Miaoxin Cai, Yaqian Ning, Tong Zhang, Yin Zhuang, Shijian Lu, He Chen, Jun Li, and Xuerui Mao. Earthgpt-x: A spatial mllm for multi-level multi-source re- mote sensing imagery understanding with visual prompting,

  38. [46]

    A fast training method for sar large scale samples based on cnn for targets recognition

    Yuan Zhang, Yang Song, Yanping Wang, and Hongquan Qu. A fast training method for sar large scale samples based on cnn for targets recognition. In2018 11th International Congress on Image and Signal Processing, BioMedical En- gineering and Informatics (CISP-BMEI), pages 1–5, 2018. 2

    2018

  39. [47]

    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, 62:1–23,

    Zilun Zhang, Tiancheng Zhao, Yulong Guo, and Jianwei Yin. 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, 62:1–23,

  40. [48]

    Geo-r1: Improving few-shot geospatial referring expression understanding with reinforcement fine- tuning, 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 understanding with reinforcement fine- tuning, 2025. 3

    2025

  41. [49]

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

    Yue Zhou, Zhihang Zhong, and Xue Yang. Towards vision- language geo-foundation model: A survey.arXiv preprint arXiv:2406.09385, 2024. 2 FUSAR-GPT : A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery Supplementary Material

    2024

  42. [50]

    6, SAR imagery exhibits inherent lim- itations that constrain visual–semantic understanding

    The specificity of SAR images As illustrated in Fig. 6, SAR imagery exhibits inherent lim- itations that constrain visual–semantic understanding. First, the large modality gap between optical and SAR images leads to systematic misinterpretation: optical models rely on color, t...

  43. [51]

    FUSAR-GPT consistently outperforms all competing methods by a signif- icant margin across counting, grid-based localization, and classification tasks

    Additional Experiments Table 5 presents an extended comparison on three SAR un- derstanding tasks with additional baselines. FUSAR-GPT consistently outperforms all competing methods by a signif- icant margin across counting, grid-based localization, and classification tasks. I...

  44. [52]

    Ablation results on the target counting task

    Ablation Experiment Table 9. Ablation results on the target counting task. Each com- ponent:SFT1, SFT2, and TLM independently contributes to per- formance improvement, while combining all modules yields the highest accuracy. Model SFT1 SFT2 TLM ACC BaseModel✗ ✗ ✗- BaseModel(+S...

  45. [53]

    Data Description As shown in Fig.8, the SFT1 stage is built upon the FUSAR-GEOVL dataset, which provides richly structured and multi-dimensional semantic information that substan- tially enhances the model’s understanding of SAR imagery. Unlike conventional SAR datasets that o...

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.