Automatic Landslides Mapping in the Principal Component Domain

The availability of digital elevation model (DEM) delivered by airborne laser scanning (ALS) opens new horizons in the geomorphological research, especially in the landslide studies. This detailed geomorphological information allows for mapping of landslide affected areas using DEM data only. In order to map landslide areas in the automatic manner using machine learning classification algorithms and only DEM, generation of several DEM derivatives is needed. These first and second order derivatives provide information about specific properties of the terrain. However, involving a set of topographic features in the machine learning process increases significantly time of computations. Moreover, the topographic features are correlated since they are generated using the same DEM. The objective of this study is an in-depth exploration of the topographic information provided by the DEM data as well as the reduction of the computational time while the automatic landslide mapping. For this reason, a set of DEM derivatives have been generated and transformed into the principal component domain. The Principal Component Analysis (PCA) is a procedure that converts the set of correlated features into a set of linearly uncorrelated components using the orthogonal transformation. For the automatic landslide detection, the support vector machine (SVM) algorithm was used. The achieved results were compared with the existing landslide inventory map and overall accuracy and kappa coefficient were calculated. For the non-reduced original topographic model, we received 73% of overall accuracy. For the PCA-reduced models, accuracy parameters are not significantly worse. For instance, using only 7 principal components, which provide 90% of the total variability of the original topographic features, we received the overall accuracy of 72% while the computation time was reduced.