Effectiveness of Training Sample and Features for Random Forest on Road Extraction from Unmanned Aerial Vehicle-Based Point Cloud

Abstract

The accuracy of random forest (RF) classification depends on several inputs. In this study, two primary inputs—training sample and features—are evaluated for road classification from an unmanned aerial vehicle-based point cloud. Training sample selection is a challenging step since the machine learning stage of the RF classification depends greatly on it. That is, an imbalanced training sample might dramatically decrease classification accuracy. Various criteria are defined to generate different types of training samples to evaluate the effectiveness of the training sample. There are several point features that can be used in RF classification under different circumstances. More features might increase the classification accuracy, however, in that case, the processing time is also increased. Point features such as RGB (red/green/blue), surface normals, curvature, omnivariance, planarity, linearity, surface variance, anisotropy, verticality, and ground/non-ground class are investigated in this study. Different training samples and sets of features are used in the RF to extract the road surface. The experiment is conducted on a local road without a raised curb located on a relatively steep hill. The accuracy assessment is conducted by comparing the model classification results with the manually extracted road surface point cloud. It is found that the accuracy increases up to around 4%–13%, and 95% overall accuracy was obtained when using convenient training samples and features.