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

Area of Science:

• Geospatial data analysis
• Machine learning applications
• Computer vision

Background:

• LiDAR point cloud segmentation is crucial for urban planning and environmental monitoring.
• Existing machine learning models offer high accuracy but lack transparency.
• Stacked ensemble methods improve performance but increase complexity and reduce interpretability.

Purpose of the Study:

• To develop an explainable stacked ensemble framework for LiDAR point cloud segmentation.
• To integrate model-agnostic Explainable Artificial Intelligence (XAI) techniques for enhanced transparency.
• To improve segmentation accuracy, interpretability, and computational efficiency.

Main Methods:

• Proposed a stacked ensemble framework using multiple base learners and Logistic Regression as a meta-model.
• Incorporated model-agnostic XAI techniques, including SHAP algorithm variants.
• Conducted experiments on two benchmark datasets for LiDAR point cloud segmentation.

Main Results:

• Achieved high segmentation accuracies of 91.13% and 95.71% on benchmark datasets.
• Identified effective base models within the ensemble for optimal model selection.
• Demonstrated XAI-driven feature reduction, decreasing training time by at least 7% with consistent accuracy.
• Analyzed feature relevance and neighborhood selection impacts on base model performance using SHAP.

Conclusions:

• The proposed explainable stacked ensemble framework offers competitive segmentation performance for LiDAR data.
• The integration of XAI significantly improves model interpretability and transparency.
• The approach enhances computational efficiency through effective feature reduction.