Does Deep Learning Enhance the Estimation for Spatially Explicit Built Environment Stocks through Nighttime Light Data Set? Evidence from Japanese Metropolitans.

Built environment stocks have attracted much attention in recent decades because of their role in material and energy flows and environmental impacts. Spatially refined estimation of built environment stocks benefits city management, for example, in urban mining and resource circularity strategy making. Nighttime light (NTL) data sets are widely used and are regarded as high-resolution products in large-scale building stock research. However, some of their limitations, especially blooming/saturation effects, have hampered performance in estimating building stocks. In this study, we experimentally proposed and trained a convolution neural network (CNN)-based building stock estimation (CBuiSE) model and applied it to major Japanese metropolitan areas to estimate building stocks using NTL data. The results show that the CBuiSE model is capable of estimating building stocks at a relatively high resolution (approximately 830 m) and reflecting spatial distribution patterns, although the accuracy needs to be further improved to enhance the model performance. In addition, the CBuiSE model can effectively mitigate the overestimation of building stocks arising from the blooming effect of NTL. This study highlights the potential of NTL to provide a new research direction and serve as a cornerstone for future anthropogenic stock studies in the fields of sustainability and industrial ecology.