Three-phase data augmentation for the prediction of sediment flux in mountain basins during typhoon events

ANN

Artificial neural network

ARIMA

Auto regressive integrated moving average

BPNN

Back-propagation neural networks

CC

Correlation coefficient

CN

Curve number

EnKF

Ensemble Kalman filter

HEC-HMS

Hydrologic Engineering Center – Hydrologic Modeling System

LVMDR

Landslide volume equation modified for the Dahan River basin

MAE

Mean absolute error

ML

Machine learning

RBFNN

Radial basis function neural network

SEEDR

Soil erosion empirical formula modified for the Dahan River basin

SRH-2D

Sediment and river hydraulic-two dimension model

SVM

Support Vector machines

RMSE

Root mean square error

USBR

United States Bureau of Reclamation

USLE

Universal soil loss equation

Watershed sedimentation has been identified as a severe problem worldwide (Fredriksen 1970; Colosimo & Wilcock 2007; Kim et al. 2013; Huang et al. 2018; Siswanto & Francés 2019), resulting in river erosion, excess sediment, reservoir deposition, and a compromised water supply for downstream populations (Nagle et al. 1999; Dang et al. 2010; Huang et al. 2019b). A reasonable early-warning system needs to be developed to face the significant serious sedimentation disaster due to climate change. Sedimentation is generally investigated using empirical analysis, numerical simulations, and/or field measurements (Aksoy & Kavvas 2005; Garcia 2008; Walling 2017), and the collected references are described below.

The empirical methods used to estimate debris volume, sediment discharge, and river deposition are generally tied to historical events (Boomer et al. 2008; Gartner et al. 2014). Numerical simulations are widely used to estimate sediment yield and erosion volume (Morgan et al. 1998; Chou 2010; Alatorre et al. 2012). Field measurements are generally implemented using aerial photos, satellite images, UAVs, and Lidar (Jain & Kothyari 2000; Green & Westbrook 2009; Mackey & Roering 2011; Peter et al. 2014). These methods are well suited to assessing the scale of disasters and formulating mitigation strategies after the fact (Koi et al. 2008; Hsu et al. 2012); however, it is difficult to predict upcoming sedimentation hazards. In this study, we sought to establish a practical early-warning system for watershed sedimentation based on empirical, numerical, and machine learning (ML) methods.

ML methods have been widely applied, for example, Suwarno et al. (2021) used a decision tree to develop a lava flood early-warning system; Chutiman et al. (2022) presented a forecasting model for daily maximum rainfall during tropical cyclones based on grey theory; Ekwueme (2022) built an ARIMA (auto regressive integrated moving average) to forecast the flood intensity under the climate change scenario. Furthermore, fusions of numerical modeling with ML (e.g., artificial neural network (ANN)) have also been mentioned in the past few years. ANN and fuzzy models have also been used to predict long-term flow discharge and water levels. Yu et al. (2004) combined chaos theory and support vector machines (SVM) with big data to improve daily flow predictions. Sannasiraj et al. (2004) combined deterministic and stochastic models to facilitate predictions of daily tidal variation. Karri et al. (2013) used an Ensemble Kalman filter (EnKF)-based method to improve the forecast ability of deterministic models. Loukas & Vasiliades (2014) implemented a hydrological model in conjunction with ANNs to forecast rainfall runoff. Young & Liu (2015) developed a hybrid model based on numerical models and ANNs to train and test flow discharge during typhoon events. Humphrey et al. (2016) integrated a hydrological model output with a Bayesian ANN to predict the monthly flow.

ANNs have also been used to predict daily sediment concentration (Zounemat-Kermani et al. 2016; Malik et al. 2017; Huang et al. 2019a) and sediment yields along rivers (Babovic 2009). Predicting watershed sedimentation should account for precipitation, river discharge, landslide volume, soil erosion, and river sediment load. However, few studies have successfully predicted sediment concentrations using multiple factors.

This study collected hydrological data on water elevation, flowrate, and sediment discharge within the Dahan River basin from 2008 to 2016, including six typhoon events: Fungwong, Sinlaku, Jangmi, Morakot, Fitow, and Soudelor. Four other typhoon events were selected as test cases. This information is listed in Table 1.

The watershed sediment yield was defined as the sum of the landslide volume and soil erosion in this study, and this information was treated as an inflow boundary condition for SRH-2D. The landslide volume equation modified for the Dahan River basin (LVMDR) and soil erosion empirical formula modified for the Dahan River basin (SEEDR) were used to estimate the watershed sediment yield as shown in Equations (1)–(3). The LVMDR was based on Equation (1) and implemented with Equation (2) to estimate the landslide volume under different precipitation scenarios.

This study employed three machine learning methods (SVM, BPNN, and RBFNN) for the third phase. SVM is a supervised learning approach developed by Vapnik (1999) for statistical analysis. The structure of SVM is transformed into a quadratic planning issue to minimize computing time, and structured risk minimization is used to maintain accuracy. SVM has proven highly effective in dealing with classification and prediction problems. BPNNs are multilayer feed-forward ANNs using one input and output layer with one or more hidden layers. In the current study, we employed the gradient descent method to solve the energy function and thereby minimize error. RBFNN is another ML method with a simple topological structure used to solve nonlinear system identification.

The prediction performance of ANNs depends on a careful training and testing process. ANNs have been used to predict variations in rainfall-runoff, runoff-discharge, and spatial and temporal sediment concentrations (Afan et al. 2015; Knighton et al. 2019; Pakdaman et al. 2020; Jhong et al. 2021). In the proposed early-warning system, these algorithms are applied to historical data related to precipitation and simulation data related to sediment flow.

The relationship between rainfall and discharge is a critical factor in estimating river sedimentation due to the reliance of sediment threshold velocity on the accuracy of the discharge value. In other words, the reliability of sediment simulations depends on the accurate river discharge estimation. Hence, the discharge and sediment concentration simulation outcomes need to be confirmed by comparing it with the measured data.

Sediment transport is affected by inflow discharge and sediment volume. The estimation approach of the inflow boundary condition was confirmed in Sections 3.1 and 3.2. Figure 7(e) and 7(f) presents a comparison of measured and simulated sediment concentration values associated with typhoons Fitow (sediment calibration) and Trami (sediment validation) at Xiayun station based on SRH-2D modeling. We can see that SRH-2D effectively described sedimentation as long as accurate inflow discharge data were available. The trends in peak concentration were very close during the two typhoons with an absolute difference of less than 20%. These results demonstrate the efficacy of these simulations in capturing temporal variations in sediment flow.

In Scenario I, combination 4 resulted in the best performance, indicating that sediment prediction accuracy depended on the availability of input discharge data. In Scenario II, combination 1 resulted in the best performance. Note that the overall performance in Scenario II was much better than in Scenario I. The prediction results obtained using BPNN, RBFNN, and SVM were much better in Case II-1 than in Case I-4. For example, the comparison of SVM between two cases shows RMSE reduces from 17.81 to 8.47, MAE reduces from 11.78 to 2.88, and CC increases from 0.94 to 0.99. The sediment term was indicated as a representative factor because the evolution indexes, including RMSE, MAE, and CC of BPNN, RBFNN, and SVM, were showed comprehensiveness better results in scenario II.

Figure 8(a) compares the prediction results with the measured data. Overall, BPNN, RBFNN, and SVM provided satisfactory results under both scenarios regardless of the input data combination, as evidenced by the strong correlation between predictions and observations. These results demonstrate that the discharge term must be considered to capture sediment transport under flood conditions, and the sediment term must also be used to improve predictions pertaining to the sedimentation process. All three prediction models (BPNN, RBFNN, and SVM) presented a similar trend, making them suitable for predicting sediment movement patterns and generating sediment discharge hydrographs (see Figure 8(b)). In conjunction with sediment yield estimation, our numerical models (HEC-HMS and SRH-2D) provided high-quality input data for the prediction models reflecting the mechanism underlying flow and sedimentation along the river.

In this section, we discuss the performance of the three prediction models (SVM, BPNN, and RBFNN) (see Table 4). During the training stage, the RMSE and MAE values decreased with an increase in forecast lead time, whereas CC presented the opposite trend. The three models achieved the best performance when given a 1-h lead time. The RMSE, MAE, and CC values gradually degraded as the lead time was extended from 1 to 6 h, as follows: 1 h (0.99, 0.99, 0.99), 3 h (0.94, 0.92, 0.91), 4 h (0.89, 0.87, 0.86), and 6 h (0.82, 0.78, 0.75).

In the testing phase, the RMSE, MAE, and CC of the models were similar to that observed in the training stage. The RMSE, MAE, and CC values gradually degraded as the lead time was extended from 1 to 6 h, as follows: 1 h (0.99, 0.99, 0.99), 3 h (0.93, 0.91, and 0.90), 4 h (+0.7 for all three), and 6 h (+0.7 for all three). These results demonstrate that the proposed prediction model would be able to sound an alarm for disaster prevention/evacuation from roughly 3 h in advance, which is quite reasonable for mountain basins under the effects of extreme weather conditions.

Insufficient data are the major factor in the difficulty in improving the prediction ability. This interdisciplinary study integrates hydrology, hydraulic, and sediment transportation methods to produce sufficient data to develop the three-stage early-warning system. The critical meaning is that the physical mechanism should be embedded in the prediction model to solve the uncertainty problem of the unstable measured data. This study combines HEC-HMS, LVMDR, SEEDR, and SRH-2D to establish the precipitation–discharge–sedimentation relationship and confirms that reasonable discharge can improve the accuracy of sediment concentration. Furthermore, the input optimization shows that multiple inputs (discharge and sediment concentration) obtain better prediction results than the single input (sediment concentration). Note that the historical references major used a single input to establish the prediction model, and this study proves that effective, sufficient, and believable multiple inputs can improve the prediction accuracy. The primary strength of this study is to obtain better prediction accuracy in mountain rivers during large-scale sedimentation events. The effectiveness of the prediction models in terms of temporal and spatial variation indicates their reliability in forecasting adverse events. The error associated with long lead times can be attributed to the inclusion of irrelevant information.

This paper presents a multi-phase early-warning system based on ML models that account for fluid mechanics and sedimentation dynamics. We obtained a reasonable conversion of precipitation into flow discharge using a numerical model of the hydrology in the study area for use in estimating soil erosion based on empirical formulas. This made it possible to establish a well-calibrated hydraulic model tasked with generating flow data (for training and testing) to compensate for the lack of data in typical field measurements. Our robust results demonstrated that the accuracy of sediment flow predictions could be improved by adopting various data sources, including rainfall, inflow discharge, landslide volume, soil erosion, and sediment. This approach greatly improves the results of conventional models that rely on a single variable, such as inflow discharge. This research demonstrates that sedimentation can be predicted with a high degree of accuracy, even in mountain basins and during intense short-term events, such as typhoons. This proof-of-concept is a feasible approach to dealing with sedimentation issues in other fields. The most important is that the proposed early-warning system not only strives for the government to be the reference of emergency evacuation strategy but also provides reservoir sediment information.

The financial support for this study was partially provided by a grant from the Ministry of Science and Technology (project number: 110-2625-M-002-024-MY3; 111-2625-M-035-007-MY3) and partially provided by the Research Centre of Climate Change and Sustainable Development of National Taiwan University.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.