Comparison of machine learning algorithm performances in digital terrain model generation

Authors : Abdullah Can Özen, Özgül Vupa Çilengiroğlu

Pages : 179-193

Doi:10.9733/JGG.2025R0013.E

View : 60 | Download : 181

Publication Date : 2025-11-04

Article Type : Research Paper

Abstract :LiDAR technology enables precise distance measurements by emitting laser pulses that reflect off surface objects, allowing for the calculation of spatial coordinates. Alongside spatial data associated color values of LiDAR points can be extracted from images captured by onboard cameras. As the laser beams reflect upon their initial contact with surfaces, the resulting point cloud must be appropriately classified to support specific analytical or operational objectives. This study uses different machine learning methods to sort and label LiDAR point cloud data into ground and non-ground points, then compares how well each method works. For this purpose, a dataset acquired by an unmanned aerial vehicle over the Democratic Republic of Congo was utilized. The dataset comprises 114,557 points, each described by three geometric features (DeltaH, Verticality, 3rd Eigenvalue) and two normalized color attributes (Red and Green Ratios), derived from RGB values. A total of ten machine learning algorithms were implemented and assessed. Among them, the XGBoost algorithm demonstrated the highest classification accuracy at 84.1%, while the Naive Bayes algorithm yielded the lowest accuracy, at 72.4%.
Keywords : Uzaktan algılama, LiDAR, Fotogrametri, Makine öğrenmesi, Sınıflandırma

ORIGINAL ARTICLE URL