A GIS-Integrated Land Use Map Generation Agent: Hybrid Pipeline with GAN-Augmented Diffusion
Authors: Mi, X., Adjeisah, M., Cao, P., Tan, M., Ran, R., Chang, J., Zhang, J.J.
Journal: Proceedings of the International Conference on Computer Aided Architectural Design Research in Asia
Publication Date: 01/01/2026
Volume: 1
Pages: 395-404
eISSN: 2710-4265
ISSN: 2710-4257
Abstract:Urban land use maps represent a fundamental data resource in geographic information science and urban planning. Due to the complexity of urban morphology and pronounced spatiotemporal heterogeneity, the generation of high precision land use maps remains heavily reliant on manual annotation by professional surveying agencies, leading to inefficiency and high costs. We propose a POIconditioned hybrid generative framework (PH-LDM-GAN) for automated land use map generation. The framework integrates Latent Diffusion Models (LDM) with Generative Adversarial Networks (GAN) and incorporates R1 gradient regularization and Relativistic Average (Ra) adversarial learning to stabilize training and enhance spatial and visual coherence. It captures global spatial consistency while refining local details, achieving a balance between computational efficiency and semantic precision. Trained on large-scale datasets from Qingdao, Xi’an, and Shijiazhuang, the model achieves 87.4% accuracy across 25 land use categories and maintains over 85% accuracy in crosscity validation, demonstrating strong generalizability. By being integrated into a GIS platform as a toolbox, the proposed method significantly improves computational efficiency and the accuracy of land use map generation, while simplifying the automated mapping workflow and opening new possibilities for the engineering implementation and practical application of land use mapping.
Source: Scopus