10:45 AM - 12:15 PM
[HCG20-P02] Inverse Analysis of Debris Flow Deposits Using Convolutional Neural Network
Keywords:Convolutional Neural Network, Debris Flow, Machine Learning
This study proposes a new method for estimating the hydraulic conditions of debris flows from the horizontal distribution of their deposit thickness using the convolutional neural network (CNN).
Although the forward models for debris flows have been established to reproduce the flow behavior from the initial conditions, it has not been easy to reconstruct the flow conditions from the actual distribution of deposits.
In this method, the distributions of deposits are firstly calculated using the forward model with randomly generated initial flow conditions of debris flows.
Then, the CNN model explores the relationship between the initial flow conditions and the resulting deposits via training based on this artificial dataset.
The CNN model adjusts its internal parameters to estimate the model parameters, including initial flow concentration, yield stress, and viscosity of material.
The Herschel-Bulkley fluid model was adopted as the forward model to generate the artificial data sets.
The region of southern Hokkaido, Japan, at 42.736°−42.748° north latitude, 141.872°-141.884° east longitude, was selected to test the inverse analysis method.
The Eastern Iburi earthquake occurred in this region on 6 September 2018, and destructive debris flows were triggered by this event.
The CNN trained with 1000 training data sets, and the performance of the trained CNN model was evaluated with 20 test data sets. As a result,
inverse analysis on test data sets exhibited that parameters were precisely reconstructed from depositional properties, and the forward model reproduced the thickness distribution of debris-flow deposits with sufficient accuracy.
Thus, the application of the CNN model to field data sets is anticipated to provide information for assessing the risk of debris flows.
Although the forward models for debris flows have been established to reproduce the flow behavior from the initial conditions, it has not been easy to reconstruct the flow conditions from the actual distribution of deposits.
In this method, the distributions of deposits are firstly calculated using the forward model with randomly generated initial flow conditions of debris flows.
Then, the CNN model explores the relationship between the initial flow conditions and the resulting deposits via training based on this artificial dataset.
The CNN model adjusts its internal parameters to estimate the model parameters, including initial flow concentration, yield stress, and viscosity of material.
The Herschel-Bulkley fluid model was adopted as the forward model to generate the artificial data sets.
The region of southern Hokkaido, Japan, at 42.736°−42.748° north latitude, 141.872°-141.884° east longitude, was selected to test the inverse analysis method.
The Eastern Iburi earthquake occurred in this region on 6 September 2018, and destructive debris flows were triggered by this event.
The CNN trained with 1000 training data sets, and the performance of the trained CNN model was evaluated with 20 test data sets. As a result,
inverse analysis on test data sets exhibited that parameters were precisely reconstructed from depositional properties, and the forward model reproduced the thickness distribution of debris-flow deposits with sufficient accuracy.
Thus, the application of the CNN model to field data sets is anticipated to provide information for assessing the risk of debris flows.