Using deep learning to unravel the structural evolution of block-scale green spaces in urban renewal

Abstract

Most urban green space (UGS) studies focus on city-level analysis, which is not conducive to urban renewal. We developed a new deep learning approach for characterising the structural evolution of block-scale UGS in urban renewal scenarios. Taking Beijing and Shanghai as models, we first segmented the study areas into block-scale grids and trained a random forest model based on a local climate zone framework to identify the spatial morphological classes of the grids. Second, we established a VGG16-based deep learning model to identify UGS structural types. Finally, we examined the associated patterns of grid spatial morphology and UGS structure under different renewal scenarios. We found that six distinct types of UGS structure are associated with different urban morphological classes. Spatial and temporal variations of UGS structures were observed in terms of direction of conversion, activity level, and extent of improvement to service level. Holistic and micro-renewal approaches drove positive evolution of block-scale UGS structures mainly by reducing building density and via precise optimization, respectively. Our results provide a reference for urban renewal policies and planning guidelines for block-scale UGSs in high-density cities.