dc.description.abstract
Building analysis applications across architectural and engineering domains, including architectural design, building systems engineering, life safety engineering, and construction management, rely on accurate and specific space classification data. Currently, building analysis software end users mostly enter such classification data manually. Manual entry of classification data poses various challenges. It is prone to errors, leading to inaccurate analysis. It is time-consuming and significantly slows the analysis process, particularly for large buildings. Therefore, automating space classification is desirable. This dissertation explores existing and emerging machine learning (ML) and deep learning (DL) methods for automating space classification tasks. The focus is on classifying space functions and space access elements, such as doors and accessible openings, across entire floors of apartment buildings. To achieve the research goal, a methodology based on a five-step ML and DL workflow was applied across three studies that examine different aspects of space classification using ML and DL. The workflow consists of problem formulation, data collection, pre-processing, model construction, and model evaluation. In the first study, a multi-view image dataset was created to develop image deep learning (IDL) semantic segmentation models for multi-view space function classification. Results show that the predictive performance was higher for the multi-view than the single-view approach, indicating that the former may better capture the shapes of spaces and space elements. Performance was sensitive to view direction, implying view bias. However, view bias decreased as the number of views increased. Contextual space elements, such as windows and doors, were found to be important for enhancing classification accuracy. A post-processing algorithm was developed to convert pixel labels into space-level classes, facilitating the reuse of space function classification data in BIM applications.In the second study, a dataset with tabular and graph representation formats was created to compare ML and graph deep learning models (GDL), including a novel extended GDL model (eEdgeGAT), for space function and space access element classification. Results show that, among GDL models, those with attention mechanisms and featured relationships had the highest predictive performance. Among ML models, tree-based ML models significantly outperformed other models, indicating a hierarchical feature structure and feature interactions. Further analysis indicates that certain features and relationships of spaces and space access elements were particularly important for predictive performance. Prediction time scaled linearly for ML and non-linearly for GDL models.In the third study, a multimodal dataset with tabular, graph, and multi-view image data formats was created to benchmark five classification methods for space function and space access element classification. Results reveal that GDL models achieved the highest predictive performance. Specifically, HGAT, a novel heterophilic and heterogeneous graph attention network, had the best performance. Feature and relationship analysis for the top-performing ML and GDL models confirmed findings from the second study. Results further show that GDL models required fewer training parameters, less pre-processing, and shorter training times than IDL, natural language processing, and generative models. ML models were less accurate than IDL models but had fewer trainable parameters, shorter training times, and fewer pre- and no post-processing requirements.The overall contributions of this dissertation to space classification research include: i) the systematic exploration and benchmarking of existing and emerging ML and DL methods; ii) the expansion of the spatial scope from single functional units to entire floors with multiple units comprised of spaces and space access elements; and iii) the creation of a multimodal and diverse dataset that supports the development and benchmarking of space classification models.
en