Semantic-driven parametric 3D geographic scene modeling: Integrating knowledge graphs and large language models

Parametric geographic scene modeling serves as the primary method for achieving large-scale rapid spatial visualization. However, balancing modeling efficiency and specificity of geographic entities poses significant challenges due to the complexity and diversity of real-world geographic environments. This study proposes a novel 3D geographic scene modeling approach that integrates knowledge graphs and large language models (LLMs). The method leverages the extensive pre-trained knowledge and inference capabilities of LLMs to autonomously infer and enhance semantic information of unknown geographic entities. Through progressive knowledge graphs, it transforms the semantic information of geographic entities into modeling parameters, ultimately achieving more intelligent 3D geographic scene modeling. Our approach addresses current limitations in parametric modeling by offering a flexible and adaptive solution capable of efficiently handling diverse geographic entities. Through case studies and comparative analyses, we examine the inference results and modeling effects under various prompt ratios, validating the effectiveness and advantages of this method. • The study uses large language model and knowledge graphs for geographic semantic inference. • Semantic data converts into model parameters for efficient 3D scene generation. • Experiments show improved accuracy and scalability compared to traditional methods.