关键词: 时空智能, 遥感影像, 生成模型, 扩散模型, 深度学习, 时空信息

Abstract:

The visual appearance of land cover objects in high-resolution optical remote sensing images is significantly influenced by seasonal evolution and regional differences. Enhancing the spatio-temporal controllability of generation models to reproduce object features under specific spatio-temporal contexts accurately remains a critical challenge. Existing research has limitations in the encoding method of multi-source spatio-temporal information, as well as the interaction guidance method between encoded spatio-temporal features and visual features, making it difficult to accurately model the precise mapping between spatio-temporal conditions and the visual appearance of land cover objects. To address this problem, this paper proposes a framework for spatio-temporal controllable high-resolution optical remote sensing image generation. First, a multi-source spatio-temporal information encoding strategy considering attribute differences is designed, which utilizes heterogeneous frequency encoding and independent projections to transform diverse spatio-temporal information into accurate and decoupled feature representations, thereby modeling the unique properties of diverse spatio-temporal information. Second, an interaction guidance mechanism between spatio-temporal features and visual features based on decoupled attention is designed. This mechanism employs an independent parallel attention branch to facilitate deep interaction between spatio-temporal features and visual features, effectively leveraging the constraining role of spatio-temporal information without interfering with text-guided generation. We adopt low-rank adaptation to efficiently transfer domain knowledge by optimizing only low-rank decomposition matrices, thereby preserving the pre-trained generative priors of the base model. Experiments on a large-scale dataset covering seven typical regions in China demonstrate that the proposed method outperforms state-of-the-art methods by 46.69% and 14.67% in terms of spatio-temporal distribution consistency and structural-textural consistency, respectively. These results confirm the controllability and generalization potential of the proposed framework across diverse spatio-temporal scenarios.

Key words: spatio-temporal intelligence, remote sensing image, generation model, diffusion model, deep learning, spatio-temporal information