A Hybrid cGAN Framework: For Predictive Spatial Openness Analyses of Urban Open Spaces

This paper introduces a novel approach to extracting spatial information from aerial imagery using a hybrid conditional Generative Adversarial Network (cGAN) model. Aerial images have long been essential tools for mapping large areas, capturing settlement patterns and vegetation. However, in complex, small-scale urban environments, these images often lack detailed volumetric or spatial depth information. To address this limitation, this study proposes an automated methodology that directly transforms aerial imagery into predictive spatial openness analyses. The hybrid framework is structured in three phases: (1) data acquisition and preprocessing, (2) model development and training, and (3) predictive generation using real-world examples. The model consists of two interdependent cGANs: the first translates aerial imagery into segmented plan masks, and the second converts these masks into spatial openness analyses. The architecture integrates U-Net and PatchGAN structures for high resolution outputs. A unique dataset focused on parks and green open spaces in urban contexts was manually curated and used for training. The results demonstrate that the system can accurately predict spatial openness and depth characteristics based on visual similarities, offering a new way to evaluate spatial constraints. This integrated model not only produces reliable spatial analysis but also enhances the urban design decision making process through automated, image-based insights.