Modeling interior component stocks of UK housing using exterior features and machine learning techniques

Abstract

Building stock modeling is a vital tool for assessing material inventories in buildings, playing a critical role in promoting a circular economy, facilitating waste management, and supporting socio-economic analyses. However, a major challenge in building stock modeling lies in achieving accurate component-level assessments, as current approaches primarily rely on archetype-based statistical data, which often lack precision. Addressing this challenge requires scalable methods for estimating the dimensions of interior components across large building stocks. In this study, we introduce the UKResi dataset, a novel dataset containing 2000 residential houses in the United Kingdom, designed to predict interior wall systems and room-level spatial configurations using exterior building features. Benchmark experiments demonstrate that the proposed approach achieves high predictive performance, with an $R^2$ score of 0.829 for interior wall length and up to 0.880 for bedroom counts, 0.792 for lounge counts, and 0.943 for the kitchen counts. Contributions of this work also include the introduction of a multi-modal approach into the field of building stock modeling, integrating exterior features and facade imagery. Furthermore, we analyze the driving factors influencing wall length and room predictions using permutation importance and SHapley Additive exPlanations values, providing insights into feature contributions, especially facade opening information being a critical driving factor of modeling interior features. The UKResi dataset serves as a foundation for future component-level building stock modeling, offering a scalable and data-driven solution to assess building interiors. This advancement holds significant potential for improving material inventory assessments, enabling more accurate resource recovery, and supporting sustainable urban planning.