Integrating Machine Learning with Open-Source GIS for Reproducible High-Resolution Terrain Classification and Hazard Mapping from UAV LiDAR Data

Abstract. This study provides an approachable, open-source methodology for terrain classification and hazard-susceptibility mapping based on UAV LiDAR, geomorphometry, and machine learning. First, the terrain model is calculated from UAV LiDAR data after ground-point classification, outliers’ exclusion, and data interpolation. Geomorphometry variables are then computed based on the terrain model, including slope, aspect, plan and profile curvatures, TPI, terrain roughness, flow accumulation, and LS-factor. Reference data required for supervised classification come from the geomorphological analysis of LiDAR-based terrain model and engineering-geological information. Stable and unstable terrain units are used to build and test Random Forest, SVM, and GB models. Metrics such as the confusion matrix, overall accuracy, precision, recall, F1-score, ROC-AUC allow for evaluating the results of models, and feature importance determines the key predictors. The main novelty lies in an easily reproducible methodology using only open-source software, including QGIS, PDAL, and Python. The introduced methodology makes it possible to perform validation and code reuse, thereby transferring the UAV LiDAR terrain hazard analysis to any area of research in the open geospatial environment.