Learning from growth regions: Exploring optimal paths for shrinking built-up areas using machine learning and case-based reasoning

In the context of global urbanization, shrinking built-up areas (SBA) characterized by declining nighttime light intensity and reduced utilization efficiency, has emerged as a critical challenge to sustainable urban development. While existing research primarily focuses on identifying shrinkage phenomena, it methodologically lacks systematic reference to growth regions’ experiences. This study proposes an analytical framework combining machine learning and case-based reasoning to reverse-diagnose shrinking regions. Rather than imposing a growth-centric logic on shrinking cities, high-performing regions are utilized to establish a benchmark of the functional balance between infrastructure, economic density, and service provision. By learning these universal conditions of vitality, specific structural deficits in shrinking areas can be identified. Analyzing 572483 newly added built-up patches in China (2012–2019), we find 16.78% exhibited SBA patterns, displaying distinct spatial gradients from Northeast China (33.16%) to East China (11.86%). Three machine learning models achieved prediction accuracy exceeding 68%, with LightGBM reaching 72.64%, demonstrating robust predictive capability. Comparative analysis revealed substantial gaps between SBA and growth regions in infrastructure provision, economic density, and industrial structure balance. Through cluster analysis, five urban optimization types were identified requiring differentiated strategies. This framework provides quantitative foundations for evidence-based urban development optimization.