Mapping high-resolution real estate value distribution: a multi-attention deep generative model inspired by image inpainting

Accurate real estate valuation is essential for urban planning, investment, and policy development. Traditional point-based methods treat properties as isolated units, oversimplifying complex urban environments and failing to capture spatial interactions and continuous value variation for informed decision-making. This study introduces REIN (Real Estate Inpainting Network), a novel deep generative model that leverages both local and surrounding context to predict high-resolution, spatially continuous property value distributions. By reframing valuation as a spatial inpainting task, REIN transforms multi-source urban data into image-like inputs and employs a hybrid multi-attention architecture—integrating channel–spatial interactions and dense–sparse contextual dependencies—to learn urban spatial structure and infer center values from their surroundings. A relative value estimation strategy further enhances adaptability across diverse regions. Applied to New York City, REIN outperforms existing models in both accuracy and visual coherence, demonstrating the effectiveness of its attention mechanisms and context-to-center inference strategy in property valuation. The model also exhibits strong generalizability under missing spatial context, incomplete features, and cross-region transfer, making it suitable for data-scarce planning scenarios. Feature importance analysis through the Squeeze-and-Excitation block reveals globally consistent and regionally adaptive value drivers across heterogeneous settings. By combining predictive precision, adaptivity, and interpretability, REIN provides an engineering informatics framework that supports planning simulations and data-driven urban policy decisions.