In recent years, many reports of damage caused by debris flows have been published. However, sediment disaster hazard areas are designated only in areas where there is a risk of human damage, and the risk of damage throughout the country cannot be exhaustively evaluated. In this study, we developed a method to estimate the location of slope failure and distance of debris flow occurrence along a mountain stream using deep learning, which is computationally inexpensive, and to assess the risk of damage from debris flow. As a result, a hierarchical neural network model was constructed by learning the topographical information at the collapse site, and it enabled classification of the collapse with an F value of over 0.7. The LSTM model, which learned the traces of debris flow and topographic information along the stream, enabled easy estimation of the probability of debris flow arrival along the mountain streams, including the uncertainty of the streambed condition and the physical state of the generated debris flow. Furthermore, our results indicate that the combination of these methods can be used to exhaustively assess the risk of debris flow damage over a wide area.
View full abstract