Explainable Artificial Intelligence for Machine Learning-Based Photogrammetric Point Cloud Classification

Point clouds are one of the most widely used data sources for spatial modeling. Artificial intelligence approaches have become an important tool for understanding and extracting semantic information of point clouds. In particular, the explainability of machine learning approaches for 3-D data has not been sufficiently investigated. Moreover, existing studies are generally limited to object classification issues. This is a pioneer study that addresses the classification of photogrammetric point clouds in terms of explainable artificial intelligence. In this study, the explainability of black-box machine learning models in the context of the classification of photogrammetric point clouds was investigated. Each point in the point cloud is defined using geometric and spectral features. In addition, the effect of selecting the most important of these features on the classification performance of ML models such as Random Forest, XGBoost, and LightGBM was examined. The explainability of ML models was analyzed with Shapley additive explanation (SHAP), an explainable artificial intelligence approach. SHAP analysis was compared with filter-based information gain (IG) and ReliefF methods for feature selection. Using the features selected with SHAP analysis, overall accuracy (OA) of 85.50% in the Ankeny dataset, 91.70% in the Building dataset, and 83.28% in the Cadastre dataset was achieved with LightGBM. The evaluation with XGBoost shows an OA of 85.22% for Ankeny, 91.21% for Building, and 82.47% for Cadastre. The evaluation with RF shows an OA of 83.70% for Ankeny, 89.08% for Building, and 79.36% for Cadastre.