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

Xiaokun Zhang; Yi Yang; Ziqi Ye; Baiyun; Xiaorong Guo; Qingchen Fang; Ruyi Zhang; Xinpeng Zhou

cs.CVFeb 22, 2026

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

Authors: Xiaokun Zhang, Yi Yang, Ziqi Ye, Baiyun, Xiaorong Guo, Qingchen Fang, Ruyi Zhang, Xinpeng Zhou,

Organizations: Discipline and Technology Center of Microwave Vision Intelligent Sensing, Fudan University · Shanghai Innovation Institute

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 & tables

Figure 1: FUSAR-GPT: Embeds spatiotemporal features for SAR feature compensation and uses two-stage decoupled SFT for target interpretation.

Figure 2: Overview of FUSAR-GPT. The framework adopts a two-stage training strategy: Stage-1 jointly updates LoRA and the TLM-MLP for multimodal prior injection, while Stage-2 fine-tunes only LoRA for task adaptation. The TLM module fuses SAR visual tokens with AEF priors by generating spatially informed modulation parameters, which are applied to the visual tokens to enhance downstream reasoning.

Figure 3: AEF–SAR Visual Comparison. AEF embeddings extracted at different times are visualized by mapping channels 1, 16 and 9 of the 64-dimensional feature vector to RGB.

Panel of Table 1 — Table 1: Training datasets used in this work: Stage 1 fine-tuning is performed on FUSAR-GEOVL-1M, and Stage 2 fine-tuning and evaluation are conducted on FUSAR-GPT.

Figure 4: Overview of the four main downstream tasks: Target Counting, Spatial Localization, Target Classification, and Target Detection.

Model	Year	Size	Counting	Spatial Localization
Model	Year	Size	@Acc	@Acc100	@Acc50	Top1
Qwen2-VL [ 35 ]	2024	2B	35.86	23.23	46.97	75.25
Qwen2-VL [ 35 ]	2024	7B	41.41	35.35	67.68	83.33
Qwen2.5-VL [ 2 ]	2025	3B	35.35	23.23	46.97	68.18
Qwen2.5-VL [ 2 ]	2025	7B	34.85	30.81	57.07	78.28
Qwen3-VL [ 1 ]	2025	4B	45.45	43.94	71.21	82.32

Table 2: Performance comparison on the Target Counting and Spatial Localization

Model	Coarse-grained		Fine-grained-plane													Fine-grained-ship
Model	Plane	Ship	P-1	P-2	P-3	P-4	P-5	P-6	P-7	P-8	P-9	P-10	P-11	P-12	P-13	S-1	S-2	S-3
Qwen2.5-VL-7B [ 2 ]	64.66	55.19	82.42	83.95	66.67	25.00	31.43	57.14	69.05	57.89	53.85	40.00	0.00	85.71	35.00	57.65	51.06	100.00
FUSAR-GPT	76.85	67.42	85.71	85.19	100	75.00	76.47	66.67	71.43	36.84	94.44	39.39	71.43	100	90.48	77.65	57.78	66.67

Table 3: Performance comparison on the Target Classification. Results are reported for both coarse-grained categories and fine-grained classes. FUSAR-GPT consistently outperforms Qwen2.5-VL-7B across all category levels.

Model	IoU = 0.25						IoU = 0.50						IoU = 0.75
	All		Plane		Ship		All		Plane		Ship		All		Plane		Ship
	R	F1	R	F1	R	F1	R	F1	R	F1	R	F1	R	F1	R	F1	R	F1
Qwen2.5-VL-7B [ 2 ]	39.9	47.1	40.3	47.5	32.3	38.5	23.5	27.7	24.3	28.7	6.5	7.7	6.2	7.3	6.3	7.5	3.2	3.8
FUSAR-GPT	66.4	74.8	67.1	75.7	51.6	57.1	52.1	58.7	53.1	59.8	32.3	35.7	21.4	24.1	22.0	24.8	9.7	10.7

Table 4: Performance comparison of Target Detection across different IoU thresholds.

Figure 5: Performance of different training stages across downstream tasks. a: Target Counting; b: Spatial Localization; c: Target Classification; d: Target Detection.

Figure 6: Challenges in developing SAR visual language models.

Model	Count	Grid			Classification
Model	@Acc	Acc@100	Acc@50	Top1	@Acc
BAN-L [ 16 ]	21.21	12.63	23.74	37.37	57.83
ChangeCLIP-B [ 5 ]	27.78	14.14	30.81	43.43	72.59
GeoRSCLIP-B [ 48 ]	30.30	10.61	20.71	37.88	68.98
Prithvi-B [ 3 ]	21.21	8.08	17.17	28.79	40.66
RemoteCLIP-L [ 20 ]	31.31	11.62	21.71	37.88	67.92

Table 5: Performance comparison on SAR understanding tasks. B and L denote ViT-Base and ViT-Large backbones, respectively.

Model	Base	TLM	Sum	Concat	ChatGPT-5.2	Gemini-3
Accuracy	34.85	52.53	36.36	37.88	18.18	30.3

Table 6: Models evaluated on Target Counting task.

Model	Bleu-1	Bleu-2	Bleu-3	Bleu-4	CIDEr	METEOR	SPICE
LLaVA-1.5-7B [ 21 ]	73.54	67.20	61.34	55.59	81.97	42.13	84.14
LLaVA-1.6-7B [ 22 ]	75.03	69.08	63.35	57.83	93.63	43.13	84.94
InternVL-3.5-4B [ 36 ]	76.76	70.76	64.99	59.36	116.21	42.34	84.36
Qwen2-VL-3B [ 35 ]	72.11	65.80	59.80	54.11	91.34	41.20	83.07
Qwen2-VL-7B [ 35 ]	76.11	70.20	64.57	59.13	109.11	42.52	84.62
Qwen2.5-VL-3B [ 2 ]	61.56	56.27	51.17	46.24	81.68	35.88	81.22

Table 7: Corpus-level captioning performance.

Model	P	R	F1
R3Det [ 40 ]	62.77	44.73	52.20
FUSAR-GPT	66.79	52.10	58.70

Table 8: Models evaluated on Target Detection task.

Model	SFT1	SFT2	TLM	ACC
BaseModel	✗	✗	✗	-
BaseModel (+SFT2)	✗	✓	✗	34.85
BaseModel (+SFT1+SFT2)	✓	✓	✗	36.36
BaseModel (+SFT2+TLM)	✗	✓	✓	41.92
FUSAR-GPT	✓	✓	✓	52.53

Table 9: Ablation results on the target counting task. Each component:SFT1, SFT2, and TLM independently contributes to performance improvement, while combining all modules yields the highest accuracy.

Figure 8: (a) shows the data characteristics of FUSAR-GEOVL, (b) shows the data characteristics of FUSAR-GPT.

Vision-Language Models

Vision-language models connect visual inputs with text for retrieval, question answering, generation, and grounded multimodal reasoning.

See all 736 papers

FUSAR-R1: A Large-Scale Reasoning Model for Intelligent Interpretation of SAR Images
Jul 18, 2026Yi Yang, Xiaokun Zhang, Yuxuan Li +3Remote Sensing Image Understanding Vision-Language Models
Key Laboratory for Information Science of Electromagnetic Waves (MoE), Fudan University, Shanghai 200433, China
SAREO-FM: Decoupled Semantic Supervision for SAR-EO Foundation Models
Oct 7, 2026Jeonghyeok Do, Munchurl KimRemote Sensing Image Understanding Multimodal Pretraining
Korea Advanced Institute of Science and Technology (KAIST)
SARLO-80: Worldwide Slant SAR Language Optic Dataset 80cm
Jun 18, 2026Solène Debuysère, Nicolas Trouvé, Nathan Letheule +2Image-Text Retrieval Cross-Modal Retrieval
1DEMR-ONERA – The French Aerospace Lab, Université Paris-Saclay, Palaiseau, France · 2DTIS-ONERA – The French Aerospace Lab, Université Paris-Saclay, Palaiseau, France · 3Hugging Face, London

Jul 18, 2026J/K move · Enter open · S save

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

Yi Yang, Xiaokun Zhang, Yuxuan Li, Ruyi Zhang+2

Key Laboratory for Information Science of Electromagnetic Waves (MoE), Fudan University, Shanghai 200433, China

Remote Sensing Image Understanding Vision-Language Models VLM Reasoning

Open paper AskPDF arXiv

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.

Abstract

Vision-Language Models

Explore similar work

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

SAREO-FM: Decoupled Semantic Supervision for SAR-EO Foundation Models

SARLO-80: Worldwide Slant SAR Language Optic Dataset 80cm