2024 Volume 5 Issue 2 Pages 33-40
In recent years, heavy rains have caused frequent sediment disasters in Japan, and it is essential to improve the reliability of sediment disaster risk estimation to reduce the damage caused by these disasters. In this study, we selected the input data in which only cells with high reliability for the slope failure occurrence were used to learn a sediment disaster risk estimation method. 60-minute total rainfall and soil rainfall index were inputted as triggers to a fully coupled deep neural network, and maximum slope angle, forest area, surface geology, and topographic classification were inputted as inherent factors. For training, validation, and threshold determination, we used the cases of sediment disasters caused by typhoons on September 6, 2007, September 21, 2011, and September 3, 2011. We conducted an estimation experiment and quantitative evaluation for the case of Typhoon Hagibis (No. 19) on October 12, 2019; we created a confusion matrix by counting each administrative district, calculated accuracy indices, and evaluated the results. The results showed that the estimation was performed without false negatives, and the accuracy was improved to 0.614 compared to the previous study (0.314).
