遥感
计算机科学
合成孔径雷达
计算机视觉
扩散
遥感应用
人工智能
辐射测量
图像处理
雷达成像
图像分辨率
地质学
图像融合
曲面拟合
图像分割
图像(数学)
图像配准
像素
大气模式
作者
Yu-Kyung Kang,Honghu Pan,Hao Liu,Qiang Liu,Weiying Xie,Leyuan Fang,Lorenzo Bruzzone
标识
DOI:10.1109/tgrs.2026.3739963
摘要
Multimodal remote sensing imagery provides complementary information for Earth observation, yet acquiring large-scale, spatially aligned multimodal data is expensive and labor-intensive, motivating data-efficient multimodal image generation. Existing methods are primarily developed for single-modal synthesis or pairwise cross-modal generation and thus cannot explicitly model the scene structure shared by multiple modalities for multimodal set generation. This limitation makes it difficult to simultaneously preserve cross-modal spatial alignment and modality-specific fidelity. To address these issues, we propose anchor diffusion (AnchorDiff), a unified framework that is able to generate spatially aligned multimodal remote sensing images from textual descriptions. To enable text-to-set generation and multimodal spatial alignment, the proposed framework produces an optical image from the given text, which is then used to construct a modality-shared scene anchor as an explicit structural prior. Specifically, we devise a UNet-based AnchorMap extractor to learn a modality-shared anchor map from modality-specific patterns to constrain structural consistency across modalities. For the multimodal generative network, we implement a conditional diffusion with a structure–attention hybrid sampling (SA-Hybrid) strategy that integrates anchor-based structural control and reference-attention guidance, thus enhancing structural alignment while maintaining modality-specific fidelity. Experiments on benchmarks covering optical, multispectral, synthetic aperture radar (SAR), and digital elevation model (DEM) data show that the proposed method consistently outperforms state-of-the-art baselines in terms of both spatial alignment and modality-specific fidelity. Code will be available athttps://github.com/kyh433/Anchor-Diff
科研通智能强力驱动
Strongly Powered by AbleSci AI