A hybrid framework for regional land valuation using generative intelligence and AutoML techniques

Abstract

Land value is a crucial indicator of economic dynamics and regional development, providing essential information for urban planning and policy development. However, most existing studies estimate a singular land value over large areas, lacking the fine-grained details for urban management. This study therefore develops a RAHGV (relative-to-absolute hybrid generative valuation) framework for regional land valuation, which combines a hybrid learning strategy with deep generative modeling to produce high-resolution, spatially continuous land value distribution across extensive urban areas. In a case study of New York City (NYC), the RAHGV model outperforms typical one-step models by differentiating between local land variations and broader regional tendencies. Its bi-attention bottleneck significantly improves model performance, reducing MAE (Mean Absolute Error) by 45.75% and MSE (Mean Squared Error) by 69.86% compared to conventional deep generative methods. Local physical infrastructure and mixed land-use patterns primarily influence micro-scale land values, while community amenities and economic vibrancy drive macro-scale values. The findings highlight the potential of the RAHGV framework as a powerful tool for promoting sustainable urban development by delivering high-resolution, data-driven insights that support informed decision-making in rapidly evolving urban environments.