Optimized Approach to Feature Selection for Semantic Segmentation Using Random Forest

The study proposes an optimized approach for feature selection in the semantic segmentation of point clouds within the architectural domain of cultural heritage, with a specific focus on historical monastic architecture. The goal is to enhance the automatic recognition and classification of architectural elements using the Random Forest algorithm, by reducing classifier dependency and increasing the model’s generalization capability. The developed method is based on a multiscale statistical selection of features, through p-value analysis and the optimization of influence radii, fully automating the process within a Python environment. The method was tested on a TLS point cloud dataset specifically built for Franciscan cloisters in the Campania region, segmented into ten architectural classes. The new approach builds upon the existing but implemented RF4PCC model, against which it was compared, showing significant improvements in the classification of minority classes, thanks to the adoption of the class_weight=”balanced” parameter and the expansion of the dataset. The analysis of feature_importances_ revealed biases related to class imbalance, which were addressed through regularization strategies and complexity control of the decision trees. Experimental results show an increase in the macro F1-score and greater fairness in class classification. The proposed approach proves effective for applications in the cultural heritage field, offering an interpretable, efficient, and adaptable method for complex architectural contexts.