地理时空大模型：概念、范式与实践

题目：地理时空大模型：概念、范式与实践

Title: Geospatial-temporal foundation models: concepts, paradigms and practices

geo-spatiotemporal-llm-concepts-paradigms-1
【大模型画不对地图？5分钟看懂地理时空大模型（UrbanComp Lab）】：视频原链接（BiliBili）

摘要 (Abstract)

大语言模型的快速发展为地理信息科学带来了新的机遇，但通用大模型并不天然具备处理坐标参考、拓扑关系和尺度效应等空间约束的能力。地理时空大模型作为在多源、多粒度和多模态地理时空大数据上预训练的大规模模型，从预训练中涌现出地理表征、地理理解、地理推理等核心能力，并可通过智能体机制将这些能力转化为自主执行空间分析的行动能力。本文梳理地理时空大模型的概念内涵与研究框架，沿"感知—认知—推理—行动"主线归纳四类研究范式：地理实体表征、多模态地理场景理解、时空语义驱动的地理推理以及大模型驱动的地理智能体。地理时空大模型不是通用大模型在地理领域的简单应用，而是将地理信息科学领域知识与人工智能方法深度融合形成的新型研究方向。未来应完善多源地理时空数据的组织与评测基准，将空间关系和时空规律融入模型训练与结果校验，并提高GIS任务执行过程及空间结果的可验证性，以增强模型在土地利用分析、城市时空预测和空间决策等任务中的可靠性。

The rapid development of large language models has created new opportunities for geographic information science. However, general-purpose large models do not inherently possess the ability to handle spatial constraints such as coordinate reference systems, topological relationships, and scale effects. Geospatial-temporal foundation models are large-scale models pretrained on multi-source, multi-granularity, and multimodal geospatial-temporal data. Through pretraining, these models acquire core capabilities in geospatial representation, geospatial understanding, and geospatial reasoning, which can be further transformed into the ability to autonomously perform spatial analyses through agent-based mechanisms. This paper reviews the conceptual foundations and research framework of geospatial-temporal foundation models and summarizes four research paradigms along the progression from perception and cognition to reasoning and action: geospatial entity representation, multimodal geospatial scene understanding, spatiotemporal-semantic-driven geospatial reasoning, and large-model-driven geospatial agents. Geospatial-temporal foundation models are not merely applications of general-purpose large models to geographic tasks; rather, they represent an emerging research direction that deeply integrates domain knowledge from geographic information science with artificial intelligence methods. Future research should improve the organization of multi-source geospatial-temporal data and the development of standardized evaluation benchmarks, incorporate spatial relationships and spatiotemporal regularities into model training and output validation, and enhance the verifiability of GIS task execution processes and spatial results. These advances are essential for improving model reliability in applications such as land-use analysis, urban spatiotemporal prediction, and spatial decision-making.

关键词 (Key Words)

地理时空大模型;
GeoAI;
地理实体表征;
多模态地理场景理解;
时空推理;
地理智能体;

geospatial-temporal foundation models;
GeoAI;
geospatial entity representation;
multimodal geospatial scene understanding;
spatiotemporal reasoning;
geospatial agents