Creation of Training SETS from Airborne LiDAR Data for ML/AI Applications: a Study of Ground Points Classification Using SVM

LIDAR (Light Detection and Ranging) is a leading remote sensing technology that provides highly accurate, three-dimensional data about the Earth’s surface quickly. It excels in high point density, precision, automation, multi-return capability, and day-night operation. However, extracting meaningful features such as ground, buildings, and poles from LIDAR data can be challenging. This study extracts thirteen geometrical and shape features from airborne LIDAR data and uses Principal Components Analysis (PCA) to reduce dimensionality, improving training and testing efficiency. Support Vector Machines (SVM), a robust classification technique, are employed to classify these features. To optimize SVM performance for binary classification, Bayesian Optimization is used to fine-tune the $\gamma$ and C hyperparameters of the Gaussian RBF kernel. The method achieves a classification accuracy of 95.54% and an F1 score of 0.9135, outperforming linear and Gaussian SVM methods. These results can enhance AI and ML applications, such as point cloud classification and object detection.