Toward transparent intelligence: Explainable stacked ensembles learning for LiDAR point cloud segmentation

Abstract

Segmentation of LiDAR point cloud data has various applications, ranging from urban planning to environmental monitoring. Although machine learning approaches have achieved impressive segmentation performance, their black-box nature often limits their interpretability. While stacked ensemble learning improves segmentation accuracy by combining multiple classifiers, it further increases model complexity and obscures decision transparency. To address this gap, this study proposes an explainable stacked ensemble framework for LiDAR point cloud segmentation that integrates multiple base learners with Logistic Regression as a meta-model and incorporates model-agnostic Explainable Artificial Intelligence (XAI) techniques. Experimental results on two benchmark datasets demonstrate segmentation accuracies of 91.13% and 95.71%. The study also identifies the most effective base models within the ensemble to facilitate optimal model selection. Furthermore, XAI-driven feature analysis enables effective feature reduction, achieving a minimum 7% reduction in training time while maintaining consistent accuracy. In addition, variants of the SHAP algorithm are employed to investigate the relevance of features and the impact of neighborhood selection strategies on the segmentation performance of each base model. The experimental results demonstrate that the proposed approach achieves competitive segmentation performance while improving interpretability and computational efficiency.