Relationships among shallow landslides, geomorphology, and geology inferred from decision-tree models for mountains in Taiwan and Japan

This study analyzes landslide susceptibility in the Shihmen Reservoir watershed, Taiwan, using a statistical decision-tree model, and compares the result with the model for the Akaishi Mountains, Japan. A decision-tree model is advantageous in that the estimation process of susceptibility and the importance of each explanatory variable are explicitly represented by the tree structure. For the landslide susceptibility modeling, topographic characteristics (elevation, slope angle, profile curvature, plan curvature, and estimated dissection and undissection heights) and geology were used as explanatory variables. The topographic characteristics were calculated from 40 m grid cell DEMs (Digital Elevation Models). The objective variable was the distribution of shallow landslides interpreted and mapped from aerial photographs. The overall correctness of the obtained decision-tree model by a 10-fold cross-validation is 78.3%, which is comparable to values reported in previous studies. Therefore, the model can predict landslide susceptibility in the Shihmen Reservoir watershed with high accuracy. The tree-structure indicates that shallow landslides occurred in specific locations where mean slope angle exceeds 34 degrees, and dominant geology is not the Shihti Formation (alternations of sandstone and shale, with intercalated coal seams), especially in the Baichi sub-watershed. This result agrees well with that for the Akaishi Mountains, indicating that hillslopes steeper than ca. 35 degrees are unstable and easily eroded by shallow landslides in tectonically active mountainous regions with heavy storms such as Taiwan and Japan.