High-resolution quantification of building stock using multi-source remote sensing imagery and deep learning

In recent decades, urbanization has led to an increase in building material stock. The high-resolution quantification of building stock is essential to understand the spatial concentration of materials, urban mining potential, and sustainable urban development. Current approaches rely excessively on statistics or survey data, both of which are unavailable for most cities, particularly in underdeveloped areas. This study proposes an end-to-end deep-learning model based on multi-source remote sensing data, enabling the reliable estimation of building stock. Ground-detail features extracted from optical remote sensing (ORS) and spatiotemporal features extracted from nighttime light (NTL) data are fused and incorporated into the model to improve accuracy. We also compare the performance of our feature-fusion model with that of an ORS-only regression model and traditional NTL regression for Beijing. The proposed model yields the best building-stock estimation, with a Spearman's rank correlation coefficient of 0.69, weighted root mean square error of 0.58, and total error in the test set below 14%. Using gradient-weighted class activation mapping, we further investigate the relationship between ORS features and building-stock estimation. Our model exhibits reliable predictive capability and illustrates the tremendous value of the physical environment for estimating building stock. This research illustrates the significant potential of ORS and deep learning for stock estimation. Large-scale, long-term building-stock investigations could also benefit from the end-to-end predictability and the data availability of the model.