多光谱遥感图像可控生成的扩散模型参数高效适配与光谱一致性学习

Key Points

• The goal is to develop a diffusion model that efficiently adapts parameters for generating multispectral remote sensing images while addressing data scarcity and annotation costs.
• Implemented parameter-efficient fine-tuning strategies in a frozen pre-trained diffusion model.
• Introduced physical constraints based on spectral characteristics during the adaptation training.
• Designed a text-perceptual encoding mechanism for accurate control over spatial distribution of land cover.
• Conducted experiments on datasets like FLAIR and Five-Billion-Pixels for validation.
• Demonstrated enhanced spectral fidelity and semantic alignment compared to mainstream models like ControlNet and T2I-Adapter.
• Generated near-infrared bands with clear physical significance.
• Achieved accuracy improvement in downstream tasks using generated data for training.

Abstract