Research on Multi modal Geographic Mapping Data Governance and Recombination Technology Driven by Large Models

Abstract

In response to the core problems of feature heterogeneity, difficult alignment, low quality control efficiency, and insufficient standardization in the governance and compilation of multimodal geographic surveying and mapping data such as vectors, grids, text, and 3D point clouds, this paper proposes a large model driven full process compilation technology solution. Constructing a four level integration framework of "data access feature processing quality control standardized output", this design is based on Transformer's cross modal feature alignment model (Geo MFTransformer) to achieve unified representation of multi-source features of various surveying and mapping geographic data. The intelligent quality cleaning module is constructed by integrating generative adversarial networks and semantic verification mechanisms, and a dynamic standardization system adapted to large models is established to achieve unified integration and effective governance of multimodal heterogeneous data. This experiment uses 12TB multimodal geographic mapping data as the experimental object. The results show that this scheme improves the accuracy of feature alignment to 90.3%, increases data cleaning efficiency by 62 times compared to manual methods, and shortens the standardization compilation time by 74%. It provides high-quality standardized data support for various geographic data governance scenarios.