Objectives Generative artificial intelligence has introduced new possibilities for map cartography. Existing studies have explored map style transformation, text-to-map generation, and large language model (LLM)-assisted cartography, indicating the potential of generative artificial intelligence for cartographic tasks. However, generated results still suffer from inaccurate map representation, insufficient cartographic knowledge, and limited support for stable cartographic reasoning. Traditional cartographic expert systems also make limited progress because of the difficulty of cartographic knowledge modeling and reasoning construction. Direct use of general LLM for cartographic tasks faces additional problems, including weak understanding of cartographic concepts, unstable output, and limited capability in computation and operation. The intelligent map cartography framework is therefore constructed to integrate cartographic knowledge graph, LLM reasoning, and cartographic execution tools, so that cartographic knowledge could be explicitly embedded into map-making tasks.

Methods The framework is designed according to the logic of practical cartographic workflow and organized into three major stages that are cartographic strategy generation, interactive modification, and map generation. The cartographic knowledge graph is first constructed to formalize cartographic concepts, entities, rules, and relations. Map ontology is used to organize cartographic concepts at different levels. Similarity relations between cartographic regions and themes are added to strengthen associations among map entities and support knowledge retrieval when the exact match is unavailable. The knowledge graph serves as the external knowledge base for cartographic reasoning, and the LLM acts as the decision agent. The fine-tuning strategy based on low-rank adaptation is adopted to improve the model's understanding of cartographic concepts and map-making tasks. During cartographic strategy generation, natural-language requirements are parsed into structured task information, including cartographic region, map theme, and related design conditions. The structured information is used to retrieve relevant data and cartographic configuration knowledge from the knowledge graph. Cartographic reasoning is performed by decomposing the map-making problem into multiple sub-tasks and generating corresponding query sequences, so that symbol, annotation, and other map design configurations could be inferred step by step. During interactive modification, user feedback is transformed into parameterized revision information, and the corresponding configurations are updated. During map generation, the selected data, inferred design configuration, and existing symbol resources are combined to complete data processing and automated map output.

Results The traffic map of Wuhan is used to verify the proposed framework. The framework completes a full intelligent cartographic process covering requirement parsing, cartographic strategy inference, interactive revision, data processing, and automated map output. Natural-language cartographic requirements could be transformed into structured task descriptions and used as the basis for subsequent processing. The knowledge graph supports the retrieval and recommendation of cartographic knowledge related to symbols, annotations, and other map design elements. When exact knowledge for a target map is unavailable, related knowledge could still be inferred through similarity relations between cartographic regions and themes. Interactive modification enables users to adjust map configurations according to revised requirements, and the updated configuration could be directly incorporated into the subsequent cartographic process. The experiment of Wuhan traffic map shows that the framework could connect cartographic knowledge, LLM-based reasoning, and execution tools within one workflow, while reducing operational steps and alleviating problems caused by insufficient professional knowledge. The proposed framework also demonstrats the value of cartographic knowledge in constraining and guiding artificial intelligence (AI)-based map generation.

Conclusions Integration of cartographic knowledge graph, LLM, and cartographic tools provides a feasible framework for intelligent map cartography. Explicit representation of cartographic knowledge improves support for cartographic reasoning, strategy generation, and interactive revision. The framework extends AI-assisted cartography from isolated task execution to coordinated workflow linking requirement understanding, knowledge-guided reasoning, user interaction, and automated map production. Results from the Wuhan traffic map validate the effectiveness and practical applicability of incorporating cartographic knowledge into AI-based map generation.