High-Resolution Landslide Susceptibility Map Generation using Machine Learning (Case Study in Pacitan, Indonesia)

Landslide, one of the most disastrous natural hazards, causes damage to infrastructure worldwide and local communities. Pacitan, Indonesia is one city with high susceptibility to landslides occurrence. The conditions of landslide occurrence are assumed to be the same in the future. This study’s objective is to produce a landslide susceptibility map by using machine learning methods based on topographical factors including elevation, slope, aspect, profile curvature, plan curvature, Topographic Wetness Index (TWI), distance to the river, and geological map as independent variables, whereas the landslide inventory map derived from Sentinel-2A and Landsat 7 were used as the dependent variables in the model construction. This study's datasets were constructed in three different compositions where each composition was treated as input in Random Forest, Decision Tree, and Logistic regression model. The first dataset was composed of a 70:30 ratio for training and testing sample points, the second dataset with a 60:40 ratio, and the third with a 50:50 ratio. The performance of each model using each dataset composition was analyzed using various accuracy assessments. This study also considered each topographical factor's effect on model performance by excluding several factors in model construction. From the results, random forest with the first dataset appeared to give the best performance for mapping landslide susceptibility area, shown by the highest Area Under Curve (AUC) value, Coefficient Correlation (CC), and Cohen’s Kappa of 0.96, 0.92 (92%) and 0.84, respectively. Elevation and geological maps were considered as essential variables shown by significant drops in model accuracy assessment when these two factors were separately excluded, while profile curvature was the least essential variable based on the insignificant drop in the model accuracy assessment result.