LiDAR Data Classification Using Machine Learning Methods

Abstract:

Light Detection and Ranging (LiDAR) is a state of the art technique (a part of active remote sensing) that provides three dimensional (3D) point cloud of an object. Each point of this cloud has definite coordinates (X, Y and Z) in some coordinate systems (local or global depending on the availability of navigational sensors). The post-processing of these point clouds is at times challenging because of the large number of points making the dataset bulky. Often these large datasets are not well-read and supported. To overcome these challenges, we demonstrated the use of Machine learning (ML) methods for LiDAR data classification. ML methods are known for their capability to handle large volumes of data and identifying the patterns within. We acquired the aerial laser-scanned point cloud using Leica Airborne Laser Scanner (ALS 70-CM) from a flying height of 1 Km during the field survey. There are various algorithms in ML that can be used for classification and having their pros and cons. In this study, we have applied six different ML algorithms (Random Forest, Logistic Regression, Support Vector Machine (SVM), Artificial Neural Network (ANN), Naïve Bayes and AdaBoost) to perform the classification. A comparative analysis among these classifiers has been shown to identify their potential applicability for LiDAR data classification. We took three classes (vegetation, building and ground) that are dominant in this region, and performed manual segmentation to generate the training sample. The parameters collected for the sample are RGB values, number of returns and elevation. We estimated the Confusion matrix, kappa coefficient, overall accuracy, F1 score and recall to perform the accuracy assessment. We found that Random Forest and AdaBoost provide the highest classification accuracy, followed by ANN, logistic regression and Naïve Bayes. The accuracy of SVM is least among all these classifiers, probably because of the linear kernel. However, we did not perform the SVM classification with various other kernels that may improve the accuracy, and hence the applicability for Lidar data classification.