Long Short-Term Memory Based Subsurface Drainage Control for Rainfall-Induced Landslide Prevention

Aynaz Biniyaz,Ye Sun,Zhen Liu,Behnam Azmoon

doi:10.3390/geosciences12020064

Aynaz Biniyaz, Ye Sun + Show 2 more

Open Access

https://doi.org/10.3390/geosciences12020064

Copy DOI

Abstract

Subsurface drainage has been widely accepted to mitigate the hazard of landslides in areas prone to flooding. Specifically, the use of drainage wells with pumping systems has been recognized as an effective short-term solution to lower the groundwater table. However, this method has not been well considered for long-term purposes due to potentially high labor costs. This study aims to investigate the idea of an autonomous pumping system for subsurface drainage by leveraging conventional geotechnical engineering solutions and a deep learning technique—Long-Short Term Memory (LSTM)—to establish a geotechnical cyber-physical system for rainfall-induced landslide prevention. For this purpose, a typical soil slope equipped with three pumps was considered in a computer simulation. Forty-eight cases of rainfall events with a wide range of varieties in duration, total rainfall depths, and different rainfall patterns were generated. For each rainfall event, transient seepage analysis was performed using newly proposed Python code to obtain the corresponding pump’s flow rate data. A policy of water pumping for maintaining groundwater at a desired level was assigned to the pumps to generate the data. The LSTM takes rainfall event data as the input and predicts the required pump’s flow rate. The results from the trained model were validated using evaluation metrics of root mean square error (RMSE), mean absolute error (MAE), and R2. The R2-scores of 0.958, 0.962, and 0.954 for the predicted flow rates of the three pumps exhibited high accuracy of the predictions using the trained LSTM model. This study is intended to make a pioneering step toward reaching an autonomous pumping system and lowering the operational costs in controlling geosystems.

Highlights

Natural disasters cause severe human and economic losses all over the world [1].Landslides account for approximately 5% of natural disasters and result in economic losses of $1.6 to $3.2 billion in the United States and approximately 25–50 deaths annually [2].Researchers have conducted landslide susceptibility assessments in which landslide susceptibility is defined as the probability of landslide occurrence in an area considering various geoenvironmental conditions, and rainfall has been identified as one of the main factors influencing landslide occurrence [3,4,5,6,7]
The goal of this study is to investigate the idea of an autonomous pumping system for subsurface drainage by leveraging conventional geotechnical engineering solutions and deep learning techniques to establish a geotechnical cyberphysical system (CPS) for rainfall-induced landslide prevention
It is noted that each Long-Short Term Memory (LSTM) model was trained using an adaptive optimization learning algorithm, (i.e., Adam), with a learning rate of 0.001

Summary

Introduction

Natural disasters cause severe human and economic losses all over the world [1].Landslides account for approximately 5% of natural disasters and result in economic losses of $1.6 to $3.2 billion in the United States and approximately 25–50 deaths annually [2].Researchers have conducted landslide susceptibility assessments in which landslide susceptibility is defined as the probability of landslide occurrence in an area considering various geoenvironmental conditions, and rainfall has been identified as one of the main factors influencing landslide occurrence [3,4,5,6,7]. Slope failures induced by rainfall events and the above mechanisms can cause massive economic losses and fatalities [2]; for example, [22] reported that monsoon flooding and landslides in the 2019 summer impacted more than 7 million people in Nepal, India, and Bangladesh. Memory input and the output from the previous time step This dependency is transferred by the hidden layer units connected through time. Where Xt is the input at time-step t, U is the weight matrix between the input and the hidden layer, W is the weight matrix between hidden units, ht−1 is the hidden state in the previous time-step, bh is the bias vector of the hidden layer, and tanh is a hyperbolic tangent function that transforms the input to a value between −1 and 1.

Objectives

Methods

Results

Conclusion