Urban shrinkage morphology: A quantitative classifying framework using deep learning

Urban shrinkage (US) has emerged as a critical global challenge for the high-quality development of cities, prompting extensive academic discussion. However, the typology of urban shrinkage morphology (USM) has been relatively overlooked, requiring breakthroughs in scientific boundary identification, morphological quantification, fine-grained content, and advanced classification techniques. The study took the urban physical areas (UPA) within the top ten urban agglomerations (UAs) in China as the research object, proposed and validated a comprehensive framework. The results reveal that the enhanced backpropagation neural network (BPNN) excelled in dasymetric mapping, while the optimized convolutional neural network (CNN) achieved 98.7 % accuracy in USM classification. USM indicated the evolutionary trajectory, with Local shrinkage being more likely to transition into shrinkage cities (SCs). The greater the urban scale and expansion intensity, the higher the complexity and diversity of USM. The proposed composite morphology (CM) proved significant, accounting for 22.3 %. The CM proportions in the eastern and southeastern UAs were relatively low, while those in the northeastern UAs exhibited the opposite trend. This study established a new paradigm, highlighting the immense research potential in USM research, providing technical perspectives for urban planning and sustainable development.