摘要:
已有研究探索了生成式人工智能在地图制图任务中的应用,例如应用图像翻译模型进行地图风格转换、使用图像扩散模型尝试将文本描述转换成地图图像以及将大语言模型作为制图助手完成用户需求解析和专业制图工具调用。上述研究验证了生成式人工智能在地图制图任务中的可行性,但地图生成结果存在不准确性和制图知识缺失等问题。提出一种集成制图规则的制图框架,以制图知识图谱为决策支持、大语言模型为决策智能体、制图插件为决策执行器。实验以制作武汉市交通地图为例,实现了制图策略生成、交互式修改、地图结果生成的智能制图流程。验证了智能制图框架和制图知识在人工智能生成地图任务中的有效性
Abstract:
Objectives: Previous research has explored the applications of Generative AI in cartographic tasks, such as using image translation models for map style transformation, employing image diffusion models to convert textual descriptions into map images, and utilizing large language models (LLMs) as cartographic assistants to understand user requirements and operate professional cartographic tools. These studies have demonstrated the feasibility of applying Generative AI to cartographic tasks, although issues such as inaccuracies in map generation and the absence of cartographic knowledge persist. Methods: In response to the problem of the lack of cartographic knowledge in existing generative artificial intelligence for map-making tasks, a cartographic framework that integrates cartographic rules is proposed. The cartographic knowledge graph is taken as an external knowledge base, and based on the ReAct framework, the automatic reasoning of cartographic strategies, map generation, and user-interactive modifications are to be achieved. After analyzing and evaluating the cartographic capabilities of existing large language models, it is found that there are problems such as a scarcity of cartographic concepts, unstable content output, and a lack of computing and operational capabilities when directly using large language models for map-making tasks. Inspired by the existing cartographic process, the overall cartographic process of cartographic strategy generation, interactive modification, map result generation is designed, and a cartographic framework is determined, with the knowledge graph as the decision-making basis, the large language model as the decision-making agent, and Langchain and cartographic plugins as the decisionmaking executors. Results: The framework achieves an intelligent cartographic process encompassing strategy generation, interactive modification, and map result generation. Conclusions: Experimental results validate the effectiveness of incorporating cartographic knowledge in AI-generated map tasks. Future work will focus on enhancing map datasets and exploring spatial representation methods for geospatial data.