Autonomous building material stock estimation using 3D modeling and multilayer perceptron

Abstract

Building material stock (BMS) modeling is crucial to promote circular economy in urban environment. This study presents a fully autonomous bottom-up approach for BMS estimation using 3D modeling and Multilayer Perceptron (MLP). Firstly, an archetype-based MLP model for concrete and steel stock estimation is developed and trained using the feature information collected from BIM (Building Information Modeling) models of local buildings, reaching 91.0 % estimation accuracy for concrete and 87.6 % for steel. A novel 3D modeling workflow is then developed using an innovative point cloud reconstruction method, in order to create editable 3D building models. These models facilitate the extraction of key building features, which are then inputted into pre-trained MLP models to estimate the BMS. The full pipeline was validated through three experiments at a neighborhood, with concrete and steel stock estimation errors of 13.7 % and 10.9 %. Such bottom-up method allows for a more precise assessment of building material composition, intensity, and geographic distribution, offering crucial insights for resource management. In summary, the proposed method enables automated, scalable quantification of building material stocks directly from UAV-acquired data, supporting material reuse planning, demolition decision-making, and circular economy initiatives in urban areas.