Estimation of Building Height from ICESat-2/ATLAS and Airborne LiDAR Data Using Machine Learning Algorithms

The height of buildings in cities are of critical importance to urban planning and sustainable growth, as these factors are closely linked to environmental, economic, social, and aesthetic elements that shape the structure and character of cities. The main purpose of this study is to calculate building heights using machine learning algorithms based on airborne LiDAR and ICESat-2 data in urban areas. Sub-objectives of the study to achieve the main objective are: (a) investigate the relationships between spaceborne and airborne LiDAR data and integrate the data, (b) investigate the performance of various machine learning algorithms that can be used in point cloud analysis when ICESat-2/ATLAS and airborne LiDAR data are analyzed, and (c) comparing the accuracy of building height data obtained from both spaceborne and airborne LiDAR systems with data obtained from ground-based measurement techniques. Airborne LiDAR data provided by GDM were used for the study. LiDAR points corresponding to the projection of each ICESat-2/ATLAS photon in the study area were selected. Python libraries were used to analyze the LiDAR data. Machine learning algorithms (K-Nearest Neighbors (K-NN), Random Forest (RF), Support Vector Machines (SVM), Artificial Neural Networks (ANNs) and RANSAC) were used to obtain highly accurate data by removing ground points. The results of the iv classification processes indicate that the classification accuracy of the ICESat-2 data using the RF algorithm (strong beams RF = 96.83%, weak beams RF = 95.72%) is higher than that of the other algorithms used (strong beams K-NN = 94.08%, SVM = 75.43%, ANN = 77.60%, RANSAC = 71.68%, weak beams K-NN = 94.26%, SVM = 72.11%, ANN = 75.46%, RANSAC = 66.17%). In the classification of airborne LiDAR data, the classification accuracy of the K-NN and RF algorithms (K-NN = 99.99%, RF = 99.98%) was found to be higher than that of the other algorithms (SVM = 91.11%, ANN = 98.48%, RANSAC = 53.66%). After an initial analysis of the point clouds, digital elevation and terrain models were created for the study area. The resulting point clouds and elevation models were used to calculate building heights in the area. A regression model was created to examine the relationship between the systems used within the scope of the study, and statistical metrics were computed. After removing outlier values, the relationship between field measurement-airborne LiDAR (R2 = 0.9985, RMSE = 0.1354, MSE = 0.0183, MAE = 0.0942, ME = 0.0388), field measurement-ICESat-2 (R2 = 0.9905, RMSE = 0.3527, MSE = 0.1244, MAE = 0.2548, ME = -0.0610) and airborne LiDAR- ICESat-2 (R2 = 0.9921, RMSE = 0.3209, MSE = 0.1029, MAE = 0.2314, ME = -0.0978) proved to be superior.