Slope failure prediction using a decision tree: A case of engineered slopes in South Korea

摘要

The South Korean-engineered slope database consists of 6828 slope observations. In this study, general slope factors were analyzed and classified by using a decision tree algorithm to evaluate the validity of the Korean slope database. The decision tree technique was used to automatically extract significant rules from a massive amount of data. These rules can be used to predict the failure possibility of new slopes. In order to rank the importance of the slope factors in the database, attribute evaluators such as ReliefF, Information Gain, Gain Ratio, Symmetrical Uncertainty and Chi square were used. These are independent computer algorithms that can automatically assess the importance of individual attributes. The most relevant attributes to slope failure, in order of importance, are fracture orientation, lithology, presence of seepage, rainfall, engineered slope angle, notches, tectonic domain, topography and weathering. Natural slope angle, height and length were found to be less important. The output of the decision tree consisted of 174 slope rules. The prediction rate of the decision tree was 72 percent. These rules could be grouped into two types. The first type suggested that failure is likely to occur when seepage is present, when slope angle is higher than 61 deg, when precipitation amount is high. These are commonly accepted characteristics of a slope failure, and such rules obtained from the constructed decision tree from Korean highway data suggested that the analyzed data and the methodology were statistically valid. The other type suggested that slope failure is highly dependent on the interaction of various slope factors. As an example, slopes with granitic gneiss failed when rainfall amount was high and did not fail, when low. However, this safe slope could fail at high weathering. In order to strengthen the validity of these rules, it should be applied to a new set of engineered slopes.