Back to Home

Interpretation of Geometric Features Effectiveness for Point Cloud Semantic Segmentation Using Explainable Artificial Intelligence

A Bachelor's thesis exploring the efficiency of geometric features in point cloud semantic segmentation using global and local SHAP analyses on XGBoost, LightGBM, and CatBoost models.

Type Bachelor's Thesis
Domain Point Cloud Processing · Explainable AI
The RMIT dataset from the FOR-instance benchmark dataset

Abstract

Processing 3D point cloud data presents challenges, such as dealing with unstructured and noisy datasets. Semantic segmentation assigns labels to points based on geometric features. Incorporating geometric features into point cloud segmentation techniques improves performance, but selecting effective characteristics requires high processing power.

This study focuses on using geometric features for point cloud semantic segmentation and applying SHapley Additive exPlanations (SHAP) — an Explainable Artificial Intelligence (XAI) technique. A 0.5 meter local neighborhood radius was defined to extract 8 geometric features (Linearity, Planarity, Sphericity, Omnivariance, Anisotropy, Eigenentropy, Sum of λs, Change of curvature) alongside Height (Z), roughness, mean curvature, gaussian curvature, verticality, intensity, normal change rate, number of returns, and volume density.

Machine learning models — XGBoost, LightGBM, and CatBoost — were employed. Global and Local SHAP interpretations assessed the model behavior and feature effectiveness.

I. Introduction

Semantic segmentation of 3D point clouds is required to interpret 3D data in object detection. Recently, machine learning algorithms are commonly employed for these tasks. However, these classifiers have a black-box structure, meaning their decision processes are unknown to the operator.

Explainable Artificial Intelligence (XAI) techniques provide insights into the decision-making process of AI models. XAI helps determine what features or variables the model thinks are most significant when making predictions. In semantic segmentation, a proper neighborhood definition is a necessary precondition for the extraction of geometric characteristics. This study interests in the efficiency of geometric features in point cloud semantic segmentation with XAI by finding spatial relationships with neighborhood points.

Point cloud segmentation methods perform better when geometric elements are included. Thus, there is a growing need in the interpretation of geometric features. XAI techniques better describe transparency, the behavior of the model, and the black-box part inside it. Pointhop was one of the first XAIs used for point cloud classification. While there are studies researching the global analysis of geometric features in point cloud semantic segmentation using XAI, there are very limited studies researching local analysis by implementing XAI methods.

III. Dataset and Materials

Dataset

The FOR-instance dataset, a benchmark dataset created for semantic and instance segmentation of individual trees from dense aerial laser scanning data, was used. It consists of five UAV-based laser scanning data collections representing different forest kinds (Norway, Czech Republic, Austria, New Zealand, Australia). The RMIT collection from Australia was used in this study.

Geometric Features

It is necessary to compute geometric features like covariance and local plane features from a specific local neighborhood surrounding each point. The local neighborhood distance was selected as 0.5 m. The calculated geometric features include:

Machine Learning Classifiers and SHAP

The machine learning techniques were implemented using the Scikit-Learn library:

SHAP (SHapley Additive exPlanations): An XAI method based on cooperative game theory that provides each feature an importance score representing how much of an impact it has on the model's output.

IV. Results

Classified Points for the Dataset

The optimized models achieved highly robust baseline metrics across 6 highly-imbalanced semantic classes. XGBoost reached ~0.82 Accuracy and ~0.80 Macro F1. LightGBM and CatBoost performed similarly with ~0.81 Accuracy and ~0.80 Macro F1 scores, demonstrating consistent performance across different boosting implementations.

XGBoost point cloud classification LightGBM point cloud classification CatBoost point cloud classification Legend for classification classes
Figure 1: Side-by-side Classification Results (XGBoost, LightGBM, CatBoost)

SHAP Analysis Model Interpretation

1. Global Interpretation

Global interpretation focuses on the overall significance and influence of features throughout the entire dataset. It is shown in the plots that the height (Z) has the most feature importance value in each class. Verticality and volume density are also important features. Normal change rate is crucial, indicating that changes in normal vectors play a significant role. The height feature (Z) is analytical for distinguishing between classes, particularly for Terrain, Low-vegetation, and Stem.

Global SHAP Analysis XGBoost Global SHAP Analysis LightGBM Global SHAP Analysis CatBoost
Figure 2: Calculated Global SHAP Summary Plots (XGBoost, LightGBM, CatBoost)

2. Local SHAP Analysis

Local SHAP analysis breaks down each feature's contribution to the prediction for a particular instance. A point from a specific class is selected and Local SHAP decision values for each geometric feature are calculated to see exactly how the model reached its conclusion.

a. XGBoost Local Interpretation

Terrain Stem Live Branches Woody Branches Out Points

b. LightGBM Local Prediction

Terrain Stem Live Branches Woody Branches Out Points

c. CatBoost Local Prediction

Terrain Stem Live Branches Woody Branches Out Points

VI. Conclusion

This study explored the effectiveness of geometric features in point cloud semantic segmentation using Explainable Artificial Intelligence (XAI), specifically SHAP. Three machine learning algorithms (XGBoost, LightGBM, and CatBoost) were used to assess performance and interpret predictions.

The SHAP analysis demonstrates the critical role of the z feature (height) in influencing the model's predictions across all classes for both global and local analysis. Verticality, Volume density, and Normal change rate also had significant impacts. The study highlights that high-accuracy point cloud classification and robust interpretability can be achieved by integrating XAI.