Semantic-Controllable UAV Land-Cover Texture Generation for 3D Surface Texturing

UAV imagery provides high-resolution observations of land-cover surfaces and rich texture information for realistic 3D scene construction. However, directly mapping UAV image textures onto large-scale 3D surfaces is affected by viewing angle, illumination, spatial scale, and image mosaicking, while conventional texture tiling methods introduce periodic repetition, visible seams, and unnatural transitions between land-cover categories. Meanwhile, general-purpose diffusion models lack adaptation to the top-down perspective and fine-grained surface characteristics of UAV imagery, making it difficult to maintain semantic consistency and texture realism. Therefore, this study proposes a UAV land-cover texture generation and mapping framework integrating Stable Diffusion and semantic–texture dual-LoRA decoupled fine-tuning. Semantic-level and texture-level textual descriptions are constructed for different land-cover categories, and two independent LoRA branches separately learn category semantics and fine-grained texture characteristics. Low-frequency appearance and high-frequency texture constraints are incorporated to improve the preservation of global color distributions and local surface details. On this basis, a random patch-based texture synthesis strategy is developed to reduce periodic repetition and stitching artifacts in large-scale single-category texture expansion, while semantic color control maps are introduced to constrain the spatial distribution and boundary transitions of multiple land-cover textures on 3D surfaces. Experiments on representative UAV land-cover categories demonstrate that the proposed dual-LoRA strategy effectively improves perceptual quality, distributional consistency, and category-specific texture representation. The texture mapping results further show that the proposed framework effectively improves the spatial continuity, visual naturalness, and semantic consistency of land-cover texture representation, providing a feasible means of integrating UAV-derived texture information into realistic 3D scene construction.