Objectives With the advancement of artificial intelligence, automatic extraction and interpretation of map information from large-scale statistical thematic maps have become significant research topics in cartography. The aim to develop a method that combines the YOLOv8 model with a syntactic model to extract structured information from statistical thematic maps. The primary objectives are to enhance extraction accuracy, decode semantic meanings based on cartographic syntax, and generate tuple-structured data for applications like knowledge graph construction and intelligent cartographic recommendations.

Methods We employed an improved YOLOv8 model to extract and classify various map elements. First, a dataset of 133 statistical thematic maps was compiled from diverse sources, including online atlases and scanned documents, with 3 512 map elements annotated as the training set. Elements were categorized into six types: thematic symbols, map title, legend, scale, other text, and other charts. To enhance the performance of YOLOv8 model for this specific application, the model was modified by incorporating residual hybrid attention groups (RHAG) in the Backbone and replacing the upsampling part with DySample upsampler. Subsequently, the model detected and classified map surface elements; subsequently, thematic symbols were processed to extract geometric primitives through a second YOLOv8 application. Pattern recognition techniques, such as Hough transform, were then used to extract visual variables, including construction attributes and color attributes. Finally, based on a syntactic model, structured descriptions of the map content were generated, which involved analyzing the relationships between graphical elements and reconstructing the hierarchical and syntactic structure of the symbols.

Results The improved YOLOv8 model demonstrates remarkable effectiveness. In map element extraction, it achieves a mean average precision (mAP)50 of 92.6% and mAP50:95 of 73.8%, with precision rates up to 99.5% for categories like scale and other text. For geometric primitive extraction, the model achieved an mAP50 of 99.4% and mAP50:95 of 90.3%, accurately identifying primitives even in complex symbols. Ablation studies confirmes the contributions of RHAG and DySample, showing improvements over the baseline model. The method successfully captures the syntactic structure of statistical thematic maps, enabling the extraction of semantic information such as thematic categories, spatial relationships, and quantitative values.

Conclusions The combination of object detection and syntactic modeling can efficiently extract structured information from statistical thematic maps. This approach provides an in-depth interpretation of map content and highlights its practical utility in cartographic analysis. Future research will continue to explore and optimize this method for broader applications.