From satellite imagery to material outflows: master plan and deep learning-based quantification of demolition...

Developing countries, undergoing rapid urbanization, have a keen interest in circular economy and its application in the construction sector in hopes of alleviating stresses on urban resources and waste management systems. This study presents a novel approach to identifying the demolition candidates from a city master plan, estimating natural building turnover from past demolitions, and quantifying the resultant materials with a potential to close the loop. Astana, the capital of Kazakhstan, was selected as a representative case due to its booming construction development along with scarce data on demolition activities. The methodology included (1) using ArcGIS and its fine-tuned Mask RCNN model to extract building footprints, (2) manual identification of demolition activities, (3) characterization of the building stock based on the city's master plan to estimate the demolition waste composition, and using results obtained, (4) the prediction of annual demolition material volumes until 2035. The model for footprint extraction from past satellite imagery achieved 70.1% object-based recall (≥ 50% overlap with ground truth) while the average error for the footprint area estimation was 2.0%. Between 2016 and 2022, there were nearly 4,500 demolition works resulting in over 108 Mt of demolition materials including concrete (57.6%), steel (3.6%), brick (14.0%), wood (7.4%), and others (17.4%). Due to planned land-use changes occurring over a third of the city area by 2035, around 9,000 buildings with cumulative 500 Mt of materials are potential demolition candidates and, by extension, sources for urban mining. The proposed methodology allows to capitalize on the knowledge from city master plans and paves way to preparing a city-scale waste management strategy on how to incorporate demolition materials into new construction and close the material loop.

The online version contains supplementary material available at 10.1007/s40201-026-00978-0.