Generative Design of Cold-Region Waterfront Ice and Snow Cultural Scenes Based on LiDAR Point Clouds and Generative Adversarial Network (GAN)

Accurate reconstruction and generation of complex cold-region waterfront environments remain challenging due to the coupled influence of geometric structures, environmental conditions, and physical constraints. This study proposes a multimodal conditional latent-space disentanglement and physically constrained adversarial generation framework (MC-PAG) for LiDAR-based 3D scene modeling and synthesis. The framework integrates LiDAR point-cloud measurements, semantic descriptors, and environmental parameters into a structured latent representation through multimodal feature encoding and latent-space disentanglement. A cascaded generative adversarial architecture is subsequently employed to progressively reconstruct dense spatial point-cloud distributions, while differentiable physical-consistency modules are embedded into the generation process to enforce environmental plausibility and geometric continuity. Furthermore, a multi-scale discrimination strategy is introduced to jointly evaluate spatial fidelity, semantic consistency, and physical-rule compliance. Experimental results on a dedicated cold-region waterfront dataset demonstrate that the proposed framework achieves a semantic segmentation IoU of 0.845 for ice–snow structures, maintains an average physical-rule violation rate of only 3.37%, and generates spatial layouts with an average path curvature of 0.12. The generated scenes further exhibit high geometric fidelity, environmental consistency, and controllable semantic representation. By establishing a unified pipeline for LiDAR sensing information processing, spatial environment reconstruction, and physically constrained scene generation, the proposed framework provides an effective methodology for digital environment modeling and propagation-aware spatial representation in complex outdoor environments.