A COMPARATIVE ANALYSIS ON GIS BASED METHODS AND MACHINE LEARNING ALGORITHM IN LANDSLIDE SUSCEPTIBILITY MODELING, A case study of Bududa, Uganda

Abstract

Landslide susceptibility modeling is of critical importance to landslide risk management, urban planning, understanding landscape evolution, and identifying landslide spatial and temporal signatures. In this study, the GIS-based weight of evidence model and the Support Vector Machine learning algorithms were compared in landslide susceptibility assessments using a case study of Bududa located in the eastern part of Uganda. The inventory of landslides applied in the study was created using satellite imagery and historical maps of the region. The causative factors were derived from the STRM DEM of 30m resolution and the Geological characteristics of the region were obtained from the Ministry of Geological Survey and Mines. The Weight of Evidence model revealed 5 factors with a positive spatial association to landslide occurrences and these were the Slope Angle, Profile Curvature, Plan Curvature, Geology and Distance to Rivers. The SVM model identified Slope Angle, Profile Curvature, Stream Power Index and Elevation as triggering factors. The comparative analysis between the two methods was conducted using the Confusion Matrix, the Receiver Operating Curves (ROC) and the developed susceptibility maps. The ROC curves gave an accuracy of 87% of the Area under Curve (AUC) for the Optimizable Support Vector Machines and a 79% (AUC) for Weight of Evidence in performance. The field validation data placed ten landslide points in the mapped susceptibility zones for both models which indicated a good overall performance for both methodologies, however the study evinced potential of more efficient and accurate results with the integration of both approaches in landslide susceptibility modeling.