Explainable machine learning framework for automated classification of structural objects in complex building façades

Accurate point-wise classification of structural objects on building surfaces is essential for high-fidelity 3D modeling and digital twin applications. However, maintaining spatial coherence in complex façade environments remains a significant challenge due to geometric heterogeneity. This study proposes an explainable machine learning framework for classifying façade-mounted structural elements from terrestrial laser scanning (TLS) point clouds using multiscale geometric features. The dataset, comprising over 10 million labeled points across nine classes, represents dense infrastructural components attached to building surfaces. To capture both fine- and coarse-scale structural details, nine geometric descriptors were evaluated across multiple neighborhood radii ranging from 0.02 m to 1.0 m. Feature relevance was assessed using three filter-based selection methods—Fisher score, Gini importance, and Chi-square—under an 80% cumulative contribution threshold. The resulting subsets were used to train Random Forest, XGBoost, and LightGBM classifiers, with macro-F1-based hyperparameter optimization to mitigate class imbalance. XGBoost combined with Fisher score-selected features yielded the best performance, achieving a 0.920 overall accuracy and 0.811 macro-F1. Class-wise SHAP analysis revealed that vertical positioning (Z) and scale-dependent surface descriptors are the primary drivers of class discrimination. These findings enable diagnosing systematic misclassifications, proving explainable frameworks offer transparent alternatives for enriching complex building façades. Practical Application This explainable framework provides AEC and digital twin professionals with a transparent, efficient tool to process massive TLS data. By identifying scale-dependent geometric features, engineering practitioners can accurately automate the semantic enrichment of complex façade elements (e.g., pipelines and structural components). Ultimately, these diagnostic insights minimize manual annotation costs, mitigate systematic misclassifications, and directly enhance Scan-to-BIM modeling and structural asset monitoring.