Reasoning cartographic knowledge in deep learning-based map generalization with explainable AI.

Area of Science:

• Geographic Information Science
• Artificial Intelligence
• Computer Vision

Background:

• Automating cartographic map generalization remains challenging despite decades of research.
• Deep neural networks (DNNs) show promise for partial automation but are often treated as black boxes.
• Lack of transparency in DNNs hinders the understanding and refinement of learned cartographic principles.

Purpose of the Study:

• To integrate explainable AI (XAI) into deep learning (DL)-based map generalization.
• To gain insights into the cartographic knowledge learned by DNNs.
• To develop and refine DL models for map generalization through transparency.

Main Methods:

• An empirical case study employing an XAI framework.
• Application of visual analytics and quantitative experiments to a pre-trained ResU-Net model.
• Analysis of input feature importance for map generalization predictions.

Main Results:

• XAI-based visualizations are interpretable by human experts.
• The DNN prioritizes building boundaries over interior building features.
• Boundary intersection over union is proposed as a superior evaluation metric for raster-based generalization.

Conclusions:

• XAI is necessary and feasible for enhancing DL-based map generalization.
• Understanding learned features improves the development of automated generalization systems.
• The study advocates for the integration of XAI into future map generalization frameworks.