HuiYanEarth-SAR: A Foundation Model for High-Fidelity and Low-Cost Global Remote Sensing Imagery Generation

Yongxiang Liu; Jie Zhou; Yafei Song; Tianpeng Liu; Li Liu

cs.CVApr 13, 2026

HuiYanEarth-SAR: A Foundation Model for High-Fidelity and Low-Cost Global Remote Sensing Imagery Generation

Authors: Yongxiang Liu, Jie Zhou, Yafei Song, Tianpeng Liu, Li Liu

Organizations: College of Electronic Science and Technology, National University of Defense Technology (NUDT), Changsha 410073, China.

Abstract

Synthetic Aperture Radar (SAR) imagery generation is essential for deepening the study of scattering mechanisms, establishing trustworthy electromagnetic scene models, and fundamentally alleviating the data scarcity bottleneck that constrains development in this field. However, existing methods find it difficult to simultaneously ensure high fidelity in both global geospatial semantics and microscopic scattering mechanisms, resulting in severe challenges for global generation. To address this, we propose HuiYanEarth-SAR, the first foundational SAR imagery generation model based on AlphaEarth and integrated scattering mechanisms. By injecting geospatial priors to control macroscopic structures and utilizing implicit scattering characteristic modeling to ensure the authenticity of microscopic textures, we achieve the capability of generating high-fidelity SAR images for global locations solely based on geographic coordinates. This study not only constructs an efficient SAR scene simulator but also establishes a bridge connecting geography, scatter mechanism, and artificial intelligence from a methodological standpoint. It advances SAR research by expanding the paradigm from perception and understanding to simulation and creation, providing key technical support for constructing a high-confidence digital twin of the Earth.

Figures & tables

Fig. 1: (a) We propose HuiYanEarth-SAR, the first generative foundation model conditioned on both geophysical priors and SAR scattering mechanism. The model only takes the geographic coordinates as input with low cost, ultimately generating high-fidelity SAR images that are strictly aligned with the input conditions and physically consistent. (b) HuiYanEarth-SAR shows powerful capabilities in generating global, high-fidelity SAR images. Code is available at https://github.com/JoyeZLearning/HuiYanEarth-SAR .

Fig. 2: Simplified schematic diagram of generation process.

Fig. 3: Global diverse scene images generated by HuiYanEarth-SAR . (a) the optical reference image, (b) the real Sentinel-1 SAR image, and (c) the SAR image generated by our model under the same geographic coordinates. (d)-(f) show the visualization of the 2nd, 7th, and 12th dimensions of the AlphaEarth 64-dimensional embedding vector used in the input conditions, reflecting geographical semantic information at different levels.

Fig. 4: Cross-continental multi-regional images generated by HuiYanEarth-SAR and their style consistency. Six representative locations are selected: (a) Sichuan China, (b) Tokyo Japan (Asia), (c) Rotterdam Netherlands, (d) Venice Italy (Europe), (e) New York USA, and (f) Amazon Brazil (Americas). For each column, from top to bottom, the images show the optical reference, the real SAR image, and the SAR image generated by our model.

Fig. 5: Comparison between (a) HuiYanEarth-SAR and other text-to-image models, including (b) GPT-4V, (c) Stable Diffusion, (d) LibLib F1, (e) Seedream, and (f) Jimeng. The prompt words given to these models are all the same. Taking mountains as an example, the text prompts are givens as:”A 10-meter resolution Synthetic Aperture Radar (SAR) image of mountainous terrain. The image should show bright slopes facing the radar and dark slopes facing away due to layover and shadowing effects. The image has speckle noise and is in grayscale.”

Fig. 6: Visual comparison of the core module ablation study. The blue bounding boxes highlight strong scatterers (e.g., bridges, large buildings). The complete model (c) accurately reconstructs these bright regions, whereas (d) appears significantly dimmer and blurred. The yellow circles indicate point targets, which are clearly visible in (c) but smoothed out or missing in (d). Column (e) demonstrates that the geospatial embedding is the fundamental prerequisite for generating valid imagery.

Methods		Metrics
AlphaEarth	Scattering	Mean/Std	FSIM $\uparrow$	SSIM $\uparrow$
Embedding	Mechanism	Similarity $\downarrow$	FSIM $\uparrow$	SSIM $\uparrow$
$\times$	$\times$	-	-	-
$\times$	$\surd$	-	-	-
$\surd$	$\times$	0.2028/0.0829	0.5693	0.3976
$\surd$	$\surd$	0.1752/0.0070	0.6781	0.5625

TABLE I: Ablation study results of core modules. Mean/Std Similarity measures radiometric distribution consistency between generated and real images. FSIM and SSIM assess structural similarity.

Level	Personnel	Professional	Experience	Head-
Level	Type	Background	(Years)	count
L1	Novice	with SAR basics	1-3	18
L2	Familiar	with SAR processing	3–5	4
L2	Familiar	experience	3–5	4
L3	Expert	SAR image interpreters	$\geq$ 5	2
Total	-	-	-	24

TABLE II: Participant information for the human visual evaluation.

Fig. 7: Hierarchical human visual evaluation results. (a) Misjudgment rate. (b) Quality score. (c) Confidence. L1/L2/L3 denote novice, familiar, and expert groups, respectively. The results demonstrate that HuiYanEarth-SAR successfully achieves cross-level visual deception. The retention of approximately 35% misjudgment even among experts proves that its generation quality approaches the limits of human visual discrimination.

Methods	Training Data	Accuracy
VGG16	Real Data	91.26
VGG16	Augment Data	95.37
ResNet18	Real Data	92.35
ResNet18	Augment Data	96.47
ResNet34	Real Data	92.55
ResNet34	Augment Data	96.79

TABLE III: Accuracy of the SAR scene classification task w/ and w/o augmentation.

Image Generation

Image generation studies models that create or edit images from text, examples, layouts, and other structured controls.

See all 463 papers

PriSAR: 3D Geometric-Prior-Guided Diffusion for Parameter-Controlled SAR Image Generation
Jul 25, 2026Fan Zhang, Xuanting Wu, Fei Ma +2Diffusion Model Guidance Diffusion-Based Image Generation
College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029, China · 2Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China
CrossEarth-SAR: A SAR-Centric and Billion-Scale Geospatial Foundation Model for Domain Generalizable Semantic Segmentation
Mar 12, 2026Ziqi Ye, Ziyang Gong, Ning Liao +10Domain Generalization for Semantic Segmentation Remote Sensing Image Segmentation
Fudan University · Shanghai Innovation Institute · Shanghai Jiao Tong University +7
FUSAR-GPT : A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery
Feb 22, 2026Xiaokun Zhang, Yi Yang, Ziqi Ye +6Vision-Language Models Remote Sensing Image Understanding
Discipline and Technology Center of Microwave Vision Intelligent Sensing, Fudan University · Shanghai Innovation Institute

Jul 25, 2026J/K move · Enter open · S save

PriSAR: 3D Geometric-Prior-Guided Diffusion for Parameter-Controlled SAR Image Generation

Fan Zhang, Xuanting Wu, Fei Ma, Qiang Yin+1

College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029, China · 2Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China

Diffusion Model Guidance Diffusion-Based Image Generation Synthetic Aperture Radar Controllable Image Generation

Open paper AskPDF arXiv

Synthetic aperture radar (SAR) image generation can mitigate data scarcity, but controllablegeneration under sparse observation angles remains difficult. Recent SAR generative studies im-prove texture realism, yet explicit geometry-aware control is still limited. This paper studiesthe focused and verifiable setting of intermediate-azimuth completion: 3D-model-derived geo-metric priors guide a diffusion model to synthesize the views missing from sparse-angle trainingdata. GeoDiff-SAR constructs a lightweight multi-bounce ray-tracing prior, encodes the result-ing point cloud, and fuses it with text conditioning while adapting Stable Diffusion 3.5 Mediumthrough low-rank adaptation. On a real four-category aircraft dataset, GeoDiff-SAR reaches anSSIM of 0.812 and azimuth consistency of 0.940, compared with 0.738 and 0.782 for the text-conditioned SD3.5 Medium baseline. The same sparse-angle protocol on five MSTAR vehicleclasses yields an SSIM of 0.878 and azimuth consistency of 0.917. These results support theconclusion that a lightweight 3D geometric prior improves viewpoint adherence for controllableSAR generation; it is intended as generation guidance rather than high-fidelity electromagneticreconstruction.

Abstract

Image Generation

Explore similar work

PriSAR: 3D Geometric-Prior-Guided Diffusion for Parameter-Controlled SAR Image Generation

CrossEarth-SAR: A SAR-Centric and Billion-Scale Geospatial Foundation Model for Domain Generalizable Semantic Segmentation

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