LANDSLIDE SUSCEPTIBILITY MODELING USING WOE-RF-SBS (WEIGHT OF EVIDENCE-RANDOM FOREST-WITH SPATIALLY BALANCED SAMPLING) METHOD BASED ON MACHINE LEARNING AND BIVARIATE STATISTICAL ANALYSIS (CASE STUDY: CISANGKUY SUB-WATERSHED)

• Main Author: Sukristiyanti
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/70973
• Institution: Institut Teknologi Bandung

Landslide susceptibility modeling (LSM) is necessary as an initial effort in disaster mitigation. State of the art in study of LSM is an integrated method of bivariate statistics and machine learning. In machine learning-based modeling, the response variable has two classes, i.e., landslide presence data and landslide absence data (non-landslide). A representative approach to determining non-landslide data does not yet exist and has not been widely discussed, and then this research was conducted to answer these problems. The novelty of this research is to develop a new approach to obtain more representative non-landslide data. The new approach is to take non-landslide samples randomly in the lowest zone of the LSM bivariate statistic. The non-landslide samples of the new approach were compared with the non-landslide samples of the previous approaches by testing them on machine learning-based LSM modeling integrated with bivariate statistics. Two machine learning methods, namely random forest (RF) and support vector machine (SVM), were integrated with the weight of evidence (WoE). The WoE was chosen from the other three bivariate statistical methods used here (frequency ratio, information value model, and Shannon entropy) then it was chosen as an input in this integrated modeling. Therefore, WoE values were used not only to convert predictor variables which are nominal ones, into numeric ones but also to obtain the non-landslide samples. The results showed that using non-landslide samples from the new approach could improve the LSM model compared to using non-landslide samples from the previous approaches. The model's accuracy by using the non-landslide samples from the new approach increased significantly (±10% – 40%) on balanced data both in number and location (spatial).