Landslide susceptibility assessment in Karanganyar regency - Indonesia - Comparison of knowledge-based and Data-driven Models

Abstract

Disaster management requires spatial information as a backbone of preparedness and mitigation process. In that context, an assessment of landslide susceptibility becomes essential in an area that is prone to landslide due to its geographical condition. The Tawangmangu, Jenawi and Ngargoyoso Subdistric in Karanganyar Regency is the one of such areas, and is the area most frequently hit by landslides in the Central Java Province of Indonesia. In this study, three different methods were applied to examine landslide susceptibility in that area: heuristic, statistical logistic regression and Artificial Neural Network (ANN). Heuristic method is a knowledge-based approach whereas the latter two are categorized as data-driven methods due to the involvement of landslide inventory in their analysis. Eight site-specific available and commonly used landslide influencing factors (slope, aspect, topographical shape, curvature, lithology, land use, distance to road and distance to river) were preprocessed in a GIS environment and then analyzed using statistical and GIS tools to understand the relationship and significance of each to landslide occurrence, and to generate landslide susceptibility maps. ILWIS, Idrisi and ArcGIS software were used to prepare the dataset and visualize the model while PASW was employed to run prediction models (logistic regression for statistical method and multi-layer perceptron for ANN). The study employed degree of fit and Receiving Operating Characteristic (ROC) to assess the models performance. The region was mapped into five landslide susceptibility classes: very low, low, moderate, high and very high class. The results also showed that lithology, land use and topographical are the three most influential factors (i.e., significant in controlling the landslide to take place). According to degree of fit analysis applied to all models, ANN performed better than the other models when predicting landslide susceptibility of the study area. Meanwhile, according to ROC analysis applied to data-driven methods, ANN shows better performance (AUC 0,988) than statistical logistic regression (AUC 0,959).