Urban settlements account for a substantial proportion of global energy consumption, embodied carbon emissions, and material use, while simultaneously harboring significant potential resources for circular reuse. However, quantifying their material composition at the urban scale remains challenging due to the diversity and complexity of buildings. With its scalable and automated approach to extracting information from multiple data sources, machine learning presents new opportunities for urban sustainability analysis.
This paper aims to investigate the potential of machine learning adoption in material flow analysis (MFA) for building stock modeling, particularly in the estimation of the existing materials. A systematic review of Scopus and Web of Science databases (2017–2025) was carried out, with the identification of 1,054 papers. Through a multi-stage screening process, 66 representative studies were selected. Four key building stock attributes were considered: construction year, building function, material typology, and structural solution. Unlike building height and footprint area, which are relatively simple to estimate through mature methodologies, these four remain complex, heterogeneous, and poorly standardized. Integrated interdisciplinary visual methods are needed for acquiring spatial and semantic data, which present significant methodological and data challenges.
For each attribute, data sources, algorithm categories, and performance trends were systematically summarized. Findings indicate that Convolutional Neural Network -based visual models dominate in structural and material recognition, while traditional machine learning approaches remain predominant for age and function inference. Government databases and open imagery platforms remain primary data sources, with increasing shifts from single-source classification towards integrated multi-source reasoning fusing remote sensing, street view imagery, points of interest, and cadastral data. However, persistent challenges still persist, including inconsistent data standards, low coupling between recognition outcomes and environmental models, as well as limited cross-city generalization capabilities. Future directions encompass open datasets, multimodal fusion, and integration with life-cycle assessments, thereby combining data-driven intelligence with sustainable urban material.