Investigating the effects of different data classification methods on landslide susceptibility mapping

Abstract

In this study, landslide susceptibility maps (LSMs) were produced for three regions where landslides are common in the Eastern Black Sea Region of Türkiye. The regions studied include the districts of Trabzon, Rize and Artvin. The eXtreme Gradient Boosting (XGBoost) machine learning algorithm was used to generate the LSMs. Ten different factors that can affect landslides including lithology, land cover, topographic wetness index (TWI), plan and profile curvature, slope, elevation, aspect, distance to roads and drainages were used for the research. The study tested various spatial data classification methods for these factors. Specifically, the data was categorized using five distinct classification methods: “geometric interval,” “equal interval,” “manual interval,” “natural breaks,” and “quantile.” The main objective of the study was to see how these classification methods affect the accuracy of LSMs. For this purpose, six different models using the XGBoost algorithm were created. In the first model, continuous data was used for most of the factors, while some factors (aspect, land cover and lithology) were used as discrete data. The other five models categorized the data using the different classification methods mentioned above. The receiver operating characteristic (ROC) curve and area under the curve (AUC) approach were used to measure how well each model performed. The results showed that the Model_1 using mostly continuous data performed the best among all three study areas with the highest AUC value. The model with the lowest AUC value was the model using the equal interval classification method (Model_3). The most important finding gained from this study was that when producing LSMs, it is preferable to maintain continuous data as is rather than reclassifying it, as this improves the accuracy of the susceptibility model.