Application of Self-Organizing Maps (SOM) to the Soundness Evaluation of Rock Slopes Along National Roads

In the conventional way of soundness evaluation performed by Ministry of Land, Infrastructure and Transport (MLIT), Japan, the soundness of road slopes (for both rock and soil slopes) have been evaluated based on the weighted-sum of risk factors including, rock type, geological structures, conditions of slopes (such as surface condition and groundwater etc.) using regular inspection table to decide the priority for preventive measures. However, it has been reported that slope failures often occurred in many road slopes which had been judged as "sound" in the regular inspections carried out in the year 1996, 2001, and 2006. Another problem is that a large number of the paper based inspection results have not always been used effectively in decision-making process for maintenance and/or repair of road slopes. Furthermore, considering recent increase of local torrential rainfall, it has become important to take into account the triggers of slope failure, which have not considered in the conventional soundness evaluation of road slopes. In this study, therefore, Kohonen's Self-Organizing Maps (SOM) was applied to the soundness evaluation of road slopes. The SOM is one of the artificial neural network using unsupervised learning and an effective tool for clustering and visualizing complex high-dimensional data on a single low-dimensional (typically two and/or three dimensional) map, saving information among original data without any external criteria. The main objective of this study is to extract unstable road slopes whose features are similar to the ones collapsed in the past using SOM. To draw maps, the amount of rainfall, which is one of the triggers of slope failure, was added as one of the components in the input vectors. By adding the data for newly inspected slopes and drawing maps using SOM, the soundness for newly inspected slopes can be evaluated.