【摘要】 【背景】大语言模型（Large Language Model, LLM）的出现推动地理信息系统（Geographic Information System, GIS）空间分析的交互范式发生深刻变革：由早期“用户直接编写代码驱动分析”、中期“通过GUI工具链配置间接调用代码”的模式，逐步转向“自然语言-代码-空间分析”的新型交互范式。在此范式下，自然语言与程序代码并列登台，成为驱动GIS空间分析的媒介。然而，二者在执行结构的稳定性、空间语义的明确性及分析过程的可传播性上存在差异，辨析何者为当下GIS空间分析的本体，有助于明确大语言模型驱动GIS空间分析的优化方向。【分析】在此背景下，本文对GIS空间分析代码的本体定位展开论证：首先界定GIS空间分析代码的概念内涵，将其归纳为5种功能类型，并从本地通用编程环境、空间资源托管型云平台、数据库内嵌环境及知识图谱环境4类执行平台刻画其能力边界。其次，从结构稳定性、语义锚定性与知识传播性3个维度阐释代码的驱动机制，论证代码作为空间分析科学有效性得以成立的最小完备单元所具备的本体地位。【进展与展望】在此基础上，综述LLM时代以GIS空间分析代码为核心的研究进展，并辨证讨论空间语义结构感知、领域自适应学习、自主智能体生态、知识迁移与积累、代码定位修复等具备研究条件但尚未系统展开的方向，以及数据与知识治理、空间计算能力、因果解释能力、地理空间表征模型等仍缺乏基础支撑却亟待探索的方向。【目的】本文围绕驱动机制展开的系统化分析，明确LLM时代下GIS空间分析代码的理论角色与方法论意义，阐明该领域的理论框架与能力边界，为GIS空间智能化研究提供理论参考。更多还原

【Abstract】 [Background] The emergence of Large Language Models(LLMs) has driven a profound transformation in the interaction paradigm of Geographic Information System(GIS) spatial analysis. This evolution has progressed from the early mode in which users directly wrote code to drive the analysis, through an intermediate mode in which code was indirectly invoked via GUI toolchain configuration, and is now shifting toward a new paradigm of "natural language-code-spatial analysis." Under this paradigm, natural language and program code jointly take the stage as the media driving GIS spatial analysis. However, the two differ in the structural stability of execution, the clarity of spatial semantics, and the transmissibility of the analytical process; discerning which of them constitutes the ontology of contemporary GIS spatial analysis is therefore essential for clarifying the optimization direction of LLM-driven GIS spatial analysis. [Analysis] Against this background, this paper advances a systematic argument for the ontological positioning of GIS spatial analysis code. It first defines the conceptual connotation of GIS spatial analysis code, identifies five functional types, and characterizes the capability boundaries across four categories of execution platforms, namely local general-purpose programming environments, spatial resource-hosting cloud platforms, database-embedded environments, and knowledge graph environments. It then explicates the driving mechanism of code from three dimensions—structural stability, semantic anchoring, and knowledge transmissibility—and argues that code constitutes the minimum complete unit upon which the scientific validity of spatial analysis is established, thereby affirming its ontological status within the GIS spatial analysis system. [Progress and Prospects] Building on this foundation, the paper reviews recent research progress centered on GIS spatial analysis code in the LLM era, and dialectically discusses two categories of research directions. The first category comprises directions that are currently feasible yet have not been systematically investigated, including spatial-semantic structural perception, domain-adaptive learning, autonomous agent ecosystems, knowledge transfer and accumulation, and code localization and repair. The second category comprises directions that still lack foundational support but are urgently worth exploring, including data and knowledge governance, spatial computing capability, causal explanation capability, and geospatial representation models. [Purpose] Through a systematic analysis organized around the driving mechanism, this paper delineates the theoretical role and methodological significance of GIS spatial analysis code in the LLM era, elucidates the theoretical framework and capability boundaries of this field, and provides a theoretical reference for research on intelligent GIS.更多还原