Download as pdf or txt
Download as pdf or txt
You are on page 1of 37

Natural Hazards

https://doi.org/10.1007/s11069-021-04582-3

ORIGINAL PAPER

Comparative study of very short‑term flood forecasting using


physics‑based numerical model and data‑driven prediction
model

Fiaz Hussain1,2 · Ray‑Shyan Wu1 · Jing‑Xue Wang1

Received: 4 October 2019 / Accepted: 21 January 2021


© The Author(s), under exclusive licence to Springer Nature B.V. part of Springer Nature 2021

Abstract
Reliable hourly flood forecasting using weather radar rainfall data for early warning system
is essential for reducing natural disaster risk during extreme typhoon events. This study
proposed a novel approach integrated with physics-based WASH123D and HEC-HMS
models to forecast 1 h ahead flood level in the Fengshan Creek basin, northern Taiwan. The
comparison was done with data-driven support vector machine (SVM) model, and perfor-
mances were assessed by using statistical indicators (root mean square error, correlation
coefficient, the error of time to peak flood level, the error of peak flood). Four typhoons
and two plum rain events (with 620 data sets) were selected for the process of model cali-
bration and validation. The model performs better when it used quantitative precipitation
estimate radar data rather than rain gauge data. Results of using 1 h ahead quantitative
precipitation forecast (QPF) as input for flood forecasting were encouraging but not fea-
sible to use directly for early flood warning system due to errors in peak flood levels and
timing. Therefore, the improvement in accuracy of 1 h ahead flood forecasting was done
using physics-based approach and SVM model. The systematic comparison revealed that
the SVM model is an attractive way out to improve the accuracy of QPF forecasted flood
levels but unable to fully describe the flood level patterns in terms of timings and flood
peaks, while the results obtained by the physics-based approach were accurate and much
better than the SVM model. The approach fully described the physics of hydrograph pat-
terns and outputs have exactly the same 1 h ahead predictions, in excellent agreement with
observations. The reliable and accurate reflections of timing and amount of flood peaks in
all selected typhoons by a newly developed physics-based approach with its operational
nature are recommended to use by the government in the future for early warning to reduce
the flood impacts during typhoon events.

Keywords Flood forecasting · Early warning · Radar rainfall · SVM model · WASH123D ·
HEC-HMS

* Ray‑Shyan Wu
raywu@ncu.edu.tw
Extended author information available on the last page of the article

13
Vol.:(0123456789)
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

1 Introduction

According to Central Weather Bureau (CWB), the official meteorological service in Tai-
wan, typically more than three typhoons coupled with heavy rainstorm during May to
October season hit the island each year and causing downstream flooding within few hours
due to high mountains and steep slope of rivers (Chang et al. 2018). The hourly rainfall
estimation and forecasting are essential for real-time flood forecasting. Therefore, the accu-
rate and reliable hourly flood forecasting is the key part of the early warning system during
the extreme typhoon events. The real-time flood forecasting, warning and response system
provides a timely warning to people living in the floodplain. This alert or warning strategy
helps in the alleviation of massive damages to human life and the socioeconomic system
during heavy storm events. In Taiwan, the flood forecasting system generally comprises
of a rainfall forecasting model, a rainfall–runoff forecasting model and a flood dynamic
routing model. The radar-derived very short-term rainfall, quantitative precipitation esti-
mation/forecasting (QPE/QPF) reliability tends to decrease with greater forecast lead time
due to the mountainous terrain and the island nature (Wu and Lin 2017). However, the
accurate prediction of occurrence location and flood lead time is fundamentally complex
due to the dynamic nature of rainfall and watershed characteristics. To date, the mimicry of
physical processes and complex basin behavior of watershed are still under the challenges
for modeling framework of flood prediction. Therefore, many flood prediction models are
highly data specific and involved various simplified assumptions such as empirical black
box, deterministic and stochastic, lumped and distributed, event base or continuous mod-
els (Lohani et al. 2014). The associated uncertainties such as (input, model structure and
parameter uncertainties) must be focused in flood forecasting and warning operations. In
flood forecast, the QPF input uncertainty is more significant than others (Wu et al. 2012);
therefore, improving the very short-term rainfall forecasting is essential and continuously
focused in Taiwan. To counter these uncertainties during hydrodynamic modeling, we pro-
posed flood forecasting update in real time using error correction approaches. More impor-
tantly, due to these approaches, the simulation results not being updated to the hydrological
model and hydraulic routings to save computational time by recalibrating the parameters of
the proposed methods with real-time observation.
Generally, flood forecasting models are categorized into two major groups: the physical-
based models (Zhao and Hendon 2009) and/or data-driven approaches. With the advance-
ment of computational capabilities and numerical modeling techniques, the hydrodynamic
simulations using distributed physical-based models become an attractive way-out for flood
forecasting. A typical flood forecasting model comprises of a hydrological component (i.e.,
flood discharges are estimated from rainfall measurements and forecasts) and a hydrody-
namic component (the water levels along the entire river network are simulated and fore-
cast). The selection of flood variable is very important according to the application of flood
warning, and, in this study, river stage (flood level) is considered as the most effective vari-
able to be forecasted. The flood stage is chosen based on its practical usefulness in flood
forecasting and its direct measurement create easiness for flood warning issuance authori-
ties. Many studies around the world used river stage variable in flood forecasting mod-
eling techniques, such as Japan (Kim et al. 2016), India (Panda et al. 2010), Brazil (Pereira
Filho and dos Santos 2006), Taiwan (Yu et al. 2006), Canada (Adamowski 2008), Ireland
(Leahy et al. 2008), Bangladesh (Liong and Sivapragasam 2002), to name a few. There
are various studies for real-time flood forecasting that used hydrological models for rain-
fall–runoff modeling and hydrodynamic models for river stage/level routing. For example,

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

European flood forecasting system used TOPKAPI (TOPographic Kinematic APproxima-


tion and Integration) distributed and HBV (Hydrologiska Byråns Vattenbalansavdelning)
conceptual hydrological models for rainfall–runoff simulation, while river routing is car-
ried out using LISFLOOD-2D hydrodynamic model (De Roo et al. 2003). The operational
real-time flood forecasting in Netherland Rivers (Rhine and Meuse) is achieved via the
HBV hydrological model and SOBEK hydrodynamic model (Werner and Dijk 2005). The
real-time flood forecasting system of Yangtze and Han Rivers and the Dongting Lake in
China incorporates three hydrological models (Xinanjiang, URBS and API), and MIKE
11 hydraulic model (Markar et al. 2005). Flood early warning system in Pakistan applied
Sacramento model for rainfall–runoff and SOBEK model for river routing (WMO 2010).
The web-based flood forecasting system of Philippines uses HEC-HMS for hydrological
modeling and Flo2D and ISIS 2D/ISIS FAST for hydrodynamic modeling (Lagmay 2017).
MIKE NAM hydrological model and MIKE 11 hydrodynamic model applied in Bagmati
basin, Nepal, for flood early warning system (Government of Nepal 2018). MIKE RR was
applied for rainfall–runoff modeling, while MIKE 11, MIKE FLOOD WATCH was used
for hydrodynamic routing modeling in Bangladesh (BWDB 2015). Flood early warning
system—Taiwan platform, is developed by Water Resources Agency (WRA) integrated
with numerous hydrological and hydrodynamic models such as 1D/2D SOBEK model
(Chang et al. 2018). In real-time flood forecasting, the hydrological and hydrodynamic
models with constant parameters may not be able to completely represent the complex pro-
cesses in a watershed or basin due to uncertainties in input data, differences between basin
physics and model structure, model calibration and changes in catchment characteristic
over time (Jian et al. 2018). Generally, the input updating, parameter updating and out/error
updating are recommended to counter these uncertainties, but input updating and param-
eter updating are costly and unsuitable for very short-term flood forecasting (Wu et al.
2012). Therefore, error updating procedure using data-driven machine learning models is
an attractive alternative that can be can be trained and make predictions on data with low
computational cost and great flexibility in modeling nonlinear processes (Yu et al. 2017).
A comprehensive review by Mosavi et al. (2018) elaborated the state of the art of machine
learning (ML) models in both long-term and short-term floods prediction. This review arti-
cle helps in understanding the flood prediction modeling frameworks. Similarly, a brief
review by Jain et al. (2018) demonstrated flood forecasting techniques and different aspects
in terms of applications and warnings.
The comparative statement for accurate prediction is more common between physical-
based and data-driven models, i.e., “physically based models require high-quality data,
in-depth knowledge and modeling expertise regarding hydrological parameters, while data-
driven prediction models need minimal inputs without a prior knowledge about catchment
characteristics and physical processes” (Mosavi et al. 2018; Rezaeianzadeh et al. 2014). Physi-
cal models capabilities are encouraging for flooding scenarios but confront highly challenges
to reach required accuracy as reported in numerous studies (Costabile and Macchione 2015;
Nayak et al. 2005; Van den Honert and McAneney 2011; Shrestha et al. 2013). In comparison,
data-driven prediction models with an extensive application using ML for flood forecasting
over the last two decades showed higher accuracy and suitability (Abbot and Marohasy 2014;
Xu and Li 2002; Ortiz-García et al. 2014). Nevertheless, the performance of ML as well as
physical models could be improved through the hybridization of models (Bellos and Tsakiris
2016; Young et al. 2017). Rainfall is a key input data for the short-term, long-term and/or real-
time flood stage forecasting, but inherent errors in radar rainfall data can lead to the poor per-
formance of predictive hydrological models. The accuracy of flood lead time can be improved
with more accurate rainfall forecasts. The wide use of short-term rainfall forecasting, such as

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

quantitative precipitation estimate (QPE) and quantitative precipitation forecast (QPF), contin-
uously becomes more reliable for practical applications of flood stage mitigation and warning
system in Taiwan. Chiang and Chang (2009) used QPF as input to the recurrent neural net-
work (RNN)-based flood forecasting model and supported the capability of forecasted rainfall.
Similarly, the studies reported by Yu et al. (2017) and Lin et al. (2013) encouraged the use of
radar-derived rainfall (QPE and QPF) for flood forecasting.
In this study, a comparative analysis of a very short-term (1-h) flood stage forecasting was
done with support vector machine (regression) and a physics-based correction approach in
Fengshan Creek basin. For that purpose, we used the HEC-HMS hydrological model for rain-
fall–runoff modeling in mountainous regions of Fengshan Creek basin using hourly rain-gauge
observed data, QPE and QPF radar data, separately. Then, WASH123D hydrodynamic rout-
ing was used to predict the flood stage during typhoon events. HEC-HMS is a rainfall–runoff
model developed by US Army Corp of Engineers—Hydrologic Engineering Center. It is a
software-based computer model comprised of variety of modeling tools for simulating each
segment of the hydrological cycle in the context of engineering problem solving (Fedelman
2000). More importantly, this model can handle mixed complex development situation and
gave an equal value on both urbanized and natural watersheds. WASH123D is from “WAter-
SHed systems of 1-D river network, 2-D overland regime and 3-D subsurface media.” It is
integrated multimedia, multi-processes, the physics-based watershed model applied to sim-
ulate flow in different components of the watershed system at various spatiotemporal scale
for watershed planning and management, and support decision-making activities (Yeh et al.
1998). It is a useful protocol for research-related issues of watershed hydrology such as predic-
tion and forecasting. It is helpful in simulating the water flow using Saint–Venant equations in
the form of fully dynamic wave (DYW), diffusive wave (DIW), kinematic wave (KIW) mod-
els. One of the key points for reasonable and stable conditions for transient simulation, either
water depth or stage based on observed measurements, must be given as an initial condition
for flood routing during 1-D and 2-D case. WASH123D has been used to simulate coastal
flood inundation, flood estimation and predication during typhoons, groundwater flow simula-
tion, and effect of land use and climate change on flow regimes by different researchers in Tai-
wan river watersheds (Shih and Yeh 2011; Shih et al. 2012a,b, 2014, 2019; Hsiao et al. 2013;
Hsu et al. 2017; Wu and Shih 2018). The model performance prediction is examined by four
statistical indices, i.e., coefficient of correlation (CC), root mean square error (RMSE), the
error of time to peak flood (ETp) and error of peak flood (EFp ). These models were selected
because of the many successful applications in hydrological modeling and forecasting. For
example, SVM (Suykens and Vandewalle 1999) approach is successfully used for river
stage, streamflow and typhoon flood forecasting by (Lin et al. 2013; Yu et al. 2006; Liong
and Sivapragasam 2002); a physically based HEC-HMS and hybrid model (HEC-HMS-SVR)
for rainfall–runoff modeling during extreme typhoon events by (Young et al. 2017); a case
report on WASH123D simulation for river stages in Gaoping river Taiwan (Shih et al. 2019).
Wu and Shih (2018) did watershed modeling of Fengshan Creek basin using HEC-HMS and
WASH123D, where the flow from mountainous regions was estimated using HEC-HMS and
WASH123D model to simulate the river stage.

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

2 Study area description and data collection

The steep slope of rivers in Taiwan along with short and abrupt typhoon rainfall events
creates the challenges to predict and forecast correct peak time and stage of rivers which
are very important for early flood warning. For this purpose, this study was conducted in
Fengshan Creek basin located in Hsinchu County northern Taiwan as shown in Fig. 1. The
main Fengshan Creek is 45.45 km long, has 1/225 average gradient which originates from
1320 m high Najie Mountain and comprised of 250.1 km2 basin area. The average annual
rainfall range and runoff are approximately 1754–2715 mm and 376 Mm3, respectively.
The selected study basin is important in terms of water resources availability and manage-
ment because of many semiconductor industries located in science park Hsinchu. In addi-
tion, the construction of golf course and land reclamation activities in the hillside region of
Fengshan basin enhanced the importance of subjected study (Wu and Shih 2018).
The input data required for HEC-HMS and WASH123D modeling include topograph-
ical data (digital elevation model (DEM), land use and soil type), hydro-meteorological
data (river stage, rain gauge and radar rainfall) and river bathymetry data such as river
cross sections and Manning’s roughness are presented in Table 1.
The surface topography of the study area is mountainous with slope range from 5 to
35% and elevation ranges 91 to 1261 m as shown in Fig. 2a. Landuse data from field inves-
tigations conducted by NLSC comprised of nine coverage groups and it was investigated
that 76% of the overall area is covered with agricultural and forest land as shown in Fig. 2b.
Hydro-stratigraphic investigations were conducted by the Central Geological Survey
(CGS) Ministry of Economic Affairs (MOEA) Taiwan. The investigations indicated that
Tamaopu conglomerate and Yangmei formation (Chaochin) covered the upstream part
of Fengshan Creek basin, while the inter-spreading of terrace deposits with alluvial lay-
ers and Tientzhu formation were explored among the other part of watershed (Wu and
Shih 2018). The rainfall data of four typhoon events (Dujuan, Soudelor, Megi, Mala-
kas) and two heavy rainstorms called plum rain during 2015–2017 were collected at the
Hsin-Pu and Guan-Xi stations. According to the past 100 years records from the Central
Weather Bureau (CWB), an average of three typhoons strike Taiwan each year during
May to October (Chang et al. 2018). The waiting time for an extreme event to occur has

Fig. 1  Location map of Fengshan Creek basin

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
13
Table 1  Data source and type used in this study
Data type Source Resolution

DEM from DTM data Center for space and remote sensing research (CSRSR), National Central University. http://www.csrsr​ 40 × 40 m (grid data)
.ncu.edu.tw/
Landuse map National Land Surveying and Mapping Center, Ministry of Interior, Taiwan (NLSC) https​://www. Polygon shapefile (vector data)
nlsc.gov.tw/En
Soil type Central Geological Survey (CGS), MOEA https​://www.moeac​gs.gov.tw/main.jsp Polygon shapefile (vector data)
Rain gauge data Water Resource Agency (WRA) https​://gweb.wra.gov.tw/Hydro​Info/?id=Index​ Hourly rainfall data of selected typhoon events
Water level data Water Resource Agency (WRA) https​://gweb.wra.gov.tw/Hydro​Info/?id=Index​ Hourly water level data

Radar rainfall data Central Weather Bureau (CWB) of Taiwan and the National Severe Storms Laboratory (NSSL) of 0.0125 (1.25 km) spatial resolution and
National Oceanic and Atmospheric Agency (NOAA) of the USA 10 min temporal resolution
River cross sections Water Resource Agency (WRA) https​://gweb.wra.gov.tw/Hydro​Info/?id=Index​ Along the river length at different interval
Natural Hazards

Content courtesy of Springer Nature, terms of use apply. Rights reserved.


Natural Hazards

Fig. 2  a Surface elevation of Fengshan Creek basin, and b Landuse classification map

shortened considerably in recent years (Chu et al. 2018). The occurrence date, duration
and total rainfall amount for each typhoon are listed in Table 2 along with the return
period according to data analysis from Water Resources Agency (WRA 1990).
The true distribution of rainfall on the spatial scale is a key for mountainous catch-
ment hydrology, while this situation is limited in Fengshan Creek basin due to only
two rain gauge stations (Hsin-Pu and Guan-Xi). Therefore, hourly radar-derived rain-
falls were used for model simulations and analysis. The radar rainfall was collected
from the Central Weather Bureau (CBW) of Taiwan. The CWB and NSSL of NOAA
of USA developed the system called Quantitative Precipitation Estimation and Segre-
gation Using Multiples Sensors (QPESUMS) for radar rainfall data. This system com-
prises four weather Doppler radars which covered the whole of Taiwan and up to some
range of adjacent ocean. The QPESUMS system provides corrected (more precise) past
72 h up to real-time quantitative precipitation estimate (QPE) grid-based radar-derived
rainfalls and 1 h quantitative precipitation forecast (QPF). The correction of grid-based
QPE radar rainfall is assured with the integration of satellite, the radar network and rain
gauge data. The system covered the whole area of Taiwan (by 441 × 561 grid points)
in the domain of (118 E − 123.5 E , 20 N − 27 N ) with a spatial resolution of 0.0125
◦ ◦ ◦ ◦ ◦

(around 1.25 km) in both latitude and longitude directions with 10 min of temporal res-
olution since 2012. The study area domain has 15 × 30 grid points.

Table 2  Typhoon events used in the study


Event Occurrence date (Taipei time) Duration (h) Total Return
rainfall period
(mm) (Year)

Dujuan September 27 + 00:00 to September 29 + 23:00, 2015 73 336 2


Soudelor August 6 + 00:00 to August 8 + 23:00, 2015 73 504 5
Malakas September 14 + 00:00 to September 18 + 23:00, 2016 97 193 2
Megi September 26 + 00:00 to September 28 + 23:00, 2016 73 513 5
Plum Rain June 1 + 00:00 to June 5 + 23:00, 2017 96 350 2
Plum Rain July 1 + 00:00 to July 9 + 23:00, 2017 208 574 2

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

3 Methodology

This study applied machine learning and physically based watershed models for flood stage
forecasting using radar rainfall and observed rain gauge rainfall data. The methodology
adopted in this research consists of the following sections. Section 3.1 talks about the brief
description of the study area, while the next step deals with the preparation of radar-derived
rainfall data for real-time (QPE) and 1 h ahead forecasting (QPF) for watershed modeling. The
next Sect. 3.3 described the assembling of watershed models (HEC-HMS and WASH123D)
for flood stage simulation using rain gauge and radar rainfall data. Sections 3.4 and 3.5 explain
the support vector machine, preparation of training data and description of physical correction
approach using the numerical model (WASH123D). In the last, the description of performance
evaluation measures is presented in Sect. 3.6 for comparative analysis of results obtained by
both modeling approaches. The step-wise methodology proposed in this study is presented in
Fig. 3.

3.1 Study area hydrological configuration

To efficiently conduct the simulations, the watershed modeling structure was divided into
three zones (A, B and C) as shown in Fig. 4. Zone A and zone B are the mountainous
region, and discharge simulation was done using HEC-HMS hydrological model. Then,
WASH123D was applied to simulate the river stage of zone C. The discharge from zone
A and B is specified as boundary conditions. The discharge from zone A was considered
as upstream incoming boundary flow, while zone B discharge was taken as source flow in
WASH123D for river stage routing of zone C.

3.2 Preparation of Radar‑derived rainfall data

The QPESUMS system provides grid data of radar all around Taiwan. The radar ( reflectivity
)
factor (Z) is converted to rainfall rate (R) using the power-law relationship Z = 32.5R1.65 .
The radar-rain gauge correction method is used to correct the bias present in radar accumu-
lations using real-time rain gauge accumulations as the true rainfall value. The entire radar
field is used as background and is simply adjusted by adding a correction factor. After get-
ting the radar rainfall data from CWB, the accuracy and ground-truthing of radar data were
assessed using statistical techniques. There are two kinds of adjustment for radar rainfall
data, i.e., temporal adjustment and spatial (grid size) adjustment. The QPESUMS system
uses UTC (Coordinated Universal Time) and time corresponding is adjusted according to
Taipei time UTC + 8. The grid size of QPESUMS radar data does not coincide with the
model grid size as identified after overlapping the two. Therefore, the area weight is used
to divide the radar grid. The Fengshan Creek basin covers 95 grid points which then con-
verted into mesh (15 × 30) using ArcGIS 10.1 for areal rainfall of whole study region as
shown in Fig. 4a. The new mesh coverage is done using a weighted average of four rain
centre points as shown in Fig. 5. The weighting average equations for X and Y location are
as follows:
( )
XN1 = x1 + x2 + x3 + x4 ∕4 (1)

( )
YN1 = y1 + y2 + y3 + y4 ∕4 (2)

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Fig. 3  Step by step methodology scheme for developing flood forecasting and early warning

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Fig. 4  Study area modeling configuration, zones (A and B) in HEC-HMS and zone C in WASH123D a
Mesh construction from radar grid points. b grid constructions of WAH123D

Fig. 5  Procedure for mesh coverage using four rainfall center points of radar data in ArcGIS

3.3 Watershed models

3.3.1 HEC‑HMS rainfall–runoff model

The Hydrologic Engineering Center-Hydrologic Modeling System (HEC-HMS) is widely


applied for the simulations of rainfall–runoff processes in natural and controlled systems
(Feldman 2000; USACE 2017). The major applications coverage ranges from large basin
water supply to small urban watershed runoff as well as flood hydrology, flow forecasting,
urban drainage design, flood damages reductions and many more (Gumindoga et al. 2016;
USACE 2018). The main objective for using HEC-HMS in this study is to determine run-
off contributions from mountainous region, zones (A and B) of Fengshan Creek basin.
The HEC-HMS simulation setup was comprised of the basin, meteorological and con-
trol specification modules, and parameters were selected based on local topography and
watershed characteristics. The runoff estimation was done using the Soil Conservation
Service-Curve Number (SCS-CN) method (SCS 1964). The SCS-CN is based on a water
balance equation as given below.
Ia = 0.2S (3)

1000
S= − 10 (4)
CN

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

(P − 0.2S)2
Q= (5)
(P + 0.8S)
where Q = runoff discharge; P = rainfall; S = potential maximum retention after runoff
begins; and Ia = initial abstraction before runoff begins. The SCS-CN method was selected
because of its simplicity and minimal data requirements which suggest good estimation for
initial abstraction and potential maximum retention based on CN. The parameters of the
initial loss, CN value and impervious percentage are evaluated based on the land use type.
The sub-basin routings were performed using SCS unit hydrograph method. The lag time
was calculated using historically recorded data. For the meteorological module, Thies-
sen polygons weighted average method was used to obtain the regional rainfall distribu-
tion. A 1-h time interval in the control specification module was selected for hydrological
simulations.

3.3.2 WASH123D physical‑based hydrological model

The water stage at different sections in river reaches is simulated using flood routing, and
physics behind the flow of water wave is de Saint Venant equations. The governing equa-
tions are based on the law of conservation of water mass and linear momentum, presented
in Eqs. 6 and 7.
The continuity equation:
𝜕A 𝜕Q
+ = ql (6)
𝜕t 𝜕x
The momentum equation:
( )
𝜕Q 𝜕QV 𝜕 Z0 + h ( )
+ = −gA + gA S0 − Sf + Vx ql (7)
𝜕t 𝜕x 𝜕x
where: t is time [t]; x is axis along river direction [L]; A is cross-sectional area of river ­[L2];
Q is flow rate of river ­[L3/t]; ql is source/sink ­[L3/t]; V is river velocity [L/t]; g is gravity
[L/t2]; Z0 is bottom elevation [L]; h is water depth [L]; S0 is river bed slope; and Sf is fric-
tion slope.
The governing Eq. (7) is comprised of various forces such as local acceleration, convective
acceleration, pressure force, gravity force and friction force. The solution of an equation is
known as dynamic routing either fully dynamic wave, diffusive wave and/or kinematic wave.
The variables are approximated using numerical techniques such as WASH123D is a numeri-
cal model, developed by Yeh et al. (1998) to conceptualize the watershed system as a 1-D
stream-river network, 2-D overland regime and 3-D subsurface media. WASH123D model is
the most appropriate protocol for this study because it can simulate all flows that comprise a
watershed. Moreover, it is capable of simulating problems related to various spatial and tem-
poral scales as long as the continuum assumptions are valid. WASH123D has been applied
to many regionally important projects and was chosen by the US Army Corps as the core
computational code for coastal and watershed studies (Yeh et al. 2011). A revamped version of
the WASH123D model has been applied to many flood mitigation and groundwater resource
problems in Taiwan (Wu and Shih 2018; Hsu et al. 2015; Shih et al. 2012a; Shih and Yeh
2011). 1-D model setup and simulation for river stage need five input files (1) super file (.sup),
(2) global file (.glb), (3) geometry file (1 dm), (4) B.C. file (1bc) and (5) initial file (1dep). The
preparation of these files is done in GMS 6.5 using conceptual model properties with type

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

WASH123D and 1D flow. The Fengshan Creek 1D river model setup was done using 82 river
cross sections taken from field geometry measured by the Second River Management Office,
WRA. The Manning’s roughness coefficient (n) values also incorporated in material proper-
ties of each cross section in GMS 6.5. The value of Manning’s roughness along each section is
presented in Table 3.
After completing 1D flow coverage in GMS 6.5, the model was converted into WASH123D
files. To conduct efficient and simplified simulations, 1-D diffusive wave approach was
selected. A direct band matrix solver was used for solving linear matrix equations of 1-D flow.
The discharge obtained from zone B using HEC-HMS simulation was considered as a source
flow rate for 1-D river routing, while discharge from zone A was considered as upstream flow
rate at node 1. Specified time-dependent water depth/stage boundary condition was applied at
last node 82. Baseflow is also considered as a source term in open boundary parameters. The
geometry file (1 dm) consists of reach section, Manning’s n values and location of nodes with
XYZ coordinates. The initial file (1dep) having initial river stage was applied as initial bound-
ary conditions for subsequent transient simulations.

3.4 Support vector machine (SVM)

The SVM, first proposed by Vapnik (1999), is known as a powerful and robust supervised
learning machine. The guidelines are based on structural risk minimization (SRM) principle
and the framework of statistical learning theory. The SRM principle acts as a tradeoff between
the quality and multifaceted nature of approximating function to minimize the expected error
and overfitting problems of a learning machine. SVM is a comparatively new approach and
widely adopted for regression, classification and other learning tasks (Vapnik 1999). The SVM
training model doles out new non-probabilistic binary linear classifier, which minimizes the
empirical classification error and maximizes the geometric margin via inverse problem-solv-
ing. Recently, SVM applications are popular in the field of hydrological modeling and flood
forecasting (Yu et al. 2006; Dehghani et al. 2014). The extended version of SVM as a regres-
sion tool alludes as support vector regression (SVR) (Smola et al. 2004). It is a developed and
efficient alternative protocol over the past two decades to take care of regression problems by
presenting alternative loss function (Li et al. 2016; Kang and Li 2016). SVR depends on map-
ping and solving the original data using linear and/or nonlinear regression classification into a
high-dimensional feature space. SVR formulation is based on SRM instead of empirical risk
minimization (ERM) which minimizes an upper bound of the generalization error rather than
minimizing the prediction error on the training set. The unique architect materialized SVR
in various flood forecast problems with promising results, excellent generalization ability and
better performance.
The brief description of SVR is as follows.
Let {(x1, y1), ( x2, y2), …, ( xl , yl )} be a given training set, where xl is an input vector, yl
is its corresponding output, and l is the number of data points. The nonlinear relationship
between the input and the output can be described by a regression function as

Table 3  Manning’s roughness coefficient (n) values used in the 1-D river model of Fengshan Creek
Cross sections 1–10 11–15 16–22 23–41 42–44 45–53 54–62 63–67 68–72 73–77 78–82

n 0.036 0.035 0.034 0.036 0.035 0.034 0.035 0.036 0.035 0.029 0.03

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

f (x) = wT ⋅ Φ(x) + b (8)


where f (x) = forecasting value; Φ(x) = nonlinear function; and wandb = parameters to be
adjusted.
The tolerated errors within the extent of the ε-tube, as well as the penalized losses Lε
when data concern the outside of the tube, are defined by Vapnik’s ε− insensitive loss
function
{ | [ )
( ) 0 for|yi − wT ⋅ Φ(xi + b ]| < 𝜀
L𝜀 yi = [ | ) [ ) (9)
| |
|yi − wT ⋅ Φ(xi + b ]| − 𝜀 for|yi − wT ⋅ Φ(xi + b ]| ≥ 𝜀
| |
The SVR problem is then formulated as the following optimization problem.
l � �
1 T ∑
minw,b,𝜉,𝜉 ∗ 2
w ⋅w+C 𝜉i + 𝜉i∗
� T i=1 �

�yi − w ⋅ Φ(xi )� + b ≤ ε + 𝜉i (10)


subject to wT ⋅ Φ(xi ) + b − yi ≤ ε + 𝜉i∗
𝜉i , 𝜉i∗ ≥ 0, i = 1, 2, ..l

where 𝜉i and𝜉i∗ are slack variables to represent the distance from the actual values to the
corresponding boundary values of the ε-tube, and C is a positive constant that determines
the degree of penalized loss when a training error occurs.
The optimization problem is solved by a dual set of Lagrange multipliers, αi and α∗i .
Therefore, the approximate function can be represented as
l
∑ ( )
f (x) = 𝛼i − 𝛼i∗ Φ(xi )T ⋅ Φ(x) + b (11)
i=1

The data with nonzero Lagrange multipliers (αi − α∗i ) are actively in the regression func-
tion and are the support vectors that “support” the definition of the approximate function.
The final regression function in Eq. (11) can be rewritten as
m
∑ ( ) ( )
f (x) = 𝛼k − 𝛼k∗ K xk , x + b (12)
k=1
( )
where xk is the support vector and m is the number of support vectors; K xk , x is the ker-
nel function.
In this study, nonlinear SVR is used to establish an hourly stage forecasting model.
Libsvm package (Chang and Lin 2011) is used for performing SVR. Libsvm is a simple,
easy-to-use and efficient software for SVM classification and regression. The nonlinear
SVR with error correction of river stage as an input vector is applied to perform 1 h ahead
flood stage forecasting. For this purpose, first, the data are transformed into the format of
Libsvm package and then normalized to the interval from 0 to 1 to represent same quan-
tity scale and avoid numerical problems and difficulties for both training and testing data.
The normalization approach also prevents the model from being dominated by a variable
with large values. The radial basis function (RBF) is selected from several available ker-
nel
( functions.
) The
( RBF kernel ) is most applicable with one parameter (𝛾 ) is adopted as
K xk , x = exp −𝛾 ||xk − x|2 . SVR with RBF kernel has received significant attention
because of its excellent nonlinear forecasting performance in the field of hydrological

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

modeling (Su et al. 2014; Ghosh 2010; Raje and Mujumdar 2011; Yu et al. 2004). Herein,
three independent parameters ( C, ε, 𝛾 ) are to be determined. The straightforward two-
step (coarse and fine) grid search method was adopted to find the best parameters. The
root mean squared error (RMSE) was used to optimize the parameters as a result 𝛾 = 5.6;
C = 1.41; ε = 0.098. The overfitting problem of training data was investigated during grid
search method so K-fold cross-validation scheme was adopted to avoid the problem of
overfitting and increase the accuracy of model. Herein, (K = 10) is selected to randomly
split the data into 10 parts and forecasting is done by using 9 subsets, and the mode is
fixed 10 times with minimum average. The mean squared error (MSE) is used to determine
cross-validation accuracy.
The construction of input vector of SVR model is done based on the error correction
method as given below.
(x, y) = ((Error(t), Error(t − 1), Error(t − 2)), (Error(t + 1)) (13)
where Error(t), Error(t − 1), Error(t − 2)) are support vector and Error(t + 1) is output
vector.
The river flood stage was first simulated using WASH123D model with QPE and
QPF radar rainfall data as input; then, the difference between simulated water levels and
observed water level data was considered as an input variable in SVR model. The simu-
lated error at different time steps is obtained by using Eq. 14.
Pre(t + 1) − Obs(t + 1) = Error(t + 1)
Sim(t) − Obs(t) = Error(t)
Sim(t − 1) − Obs(t − 1) = Error(t − 1) (14)

Sim(t − n) − Obs(t − n) = Error(t − n))

where Pre is QPF forecasted water level, Sim is the QPE simulated water level, and Obs
is the observed water level. The step by step procedure to perform SVM error correction
approach is given as: first developed the training model with training data (Observed) using
following steps.

1. Run WASH123D for QPE water level using QPE (t − n) radar rainfall where n = 0, 1, 2.
2. Use Obs (t − n) water level (Hsin-Pu Bridge) and calculate error (t − n) using step 1 water
level with this relation QPE (t − n)–Obs (t − n) = error (t − n).
3. Run WASH123D again for QPF water level using QPF (t + 1) radar rainfall.
4. Use Obs (t + 1) water level and calculate error (t + 1) using step 3 water level and with
this relation QPF (t + 1)–Obs (t + 1) = error (t + 1).
5. Calculate x = error (t − n) and y = error (t + 1) using SVM method to get function y = f(x)
of Eq. 13.

Model verification or testing was done using following steps.

1. Run WASH123D for QPE water level using QPE (t − n) radar rainfall where n = 0, 1, 2.
2. Use Obs (t − n) water level (Hsin-Pu Bridge) and calculate error (t − n) using step 1 water
level with this relation QPE (t − n)–Obs (t − n) = error (t − n).
3. Calculate error (t + 1) using developed SVM function (y = f(x)) where x = error(t − n)
from step2.
4. Run WASH123D for QPF water level using QPF (t + 1) radar rainfall.

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

5. Step 3 + Step 4 = [error(t + 1) + QPF(t + 1)] = error and QPF water level to get SVM
(t + 1) water level.

The error correction model is applied based on the correlation analysis between the
error of forecast data and other time errors. The error correlation was in decreasing order
up to first four-time steps and then become constant. Therefore, correlation coefficient
close to 0.8 is used as a criterion to select the time points before each forecast. So, the error
generated by three points before each prediction is used in SVR input vector is shown in
Eq. 13.
To avoid the overlapping of training data with testing data used for forecasting model,
we use additionally 300 sets of data from three different rainfall events (two single peak
extreme rainfall events and one multiple peak event) according to the nature of selected
rainfall storms in the study. These additional training data rather than input events serve as
a capability criterion of SVR forecasting model, i.e., how well the SVR model can forecast
beyond the range of training data set. After forecasting model establishment, the model is
simulated using the training data set and six input events to get the error correction. The
flood stage forecasts were obtained by taking the difference of QPF water level and error
obtained from SVR model output that could be compared to the observed water level. The
data-driven ML models such as SVM depend upon the quality and quantity of training data
and model optimization parameters. If the data are scarce and unable to cover the varia-
tions, their learning falls short, and hence, could not obtain reasonable accuracy. The draw-
backs of SVM type ML models to deal with “generalization problem” could be reduced by
the firm and comprehensive knowledge about ML algorithms and user defined practical
cookbooks.

3.5 Flood level forecasting using physics‑based approach

Keeping in view the difficulties observed in SVM forecasting model, we tried to intro-
duce a novel methodology for flood forecasting (1 h ahead) using physics-based correc-
tion approach (PC). In this approach, the WASH123D model outputs were directly read in
MATLAB code and processed as predictor corrector approach using following steps.

1. Read observed water levels at time t (Hsin-Pu Bridge) as input.


2. Using rating curve to find flow discharge at Hsin-Pu Bridge from step1.
3. Run WASH123D in steady condition to find all cross sections water levels at time t from
step 2 inputs.
4. Run HEC-HMS with QPF radar rainfall data and simulate flows.
5. Run WASH123D with step 3 and 4 inputs and forecast 1 h ahead water level.
6. Repeat step 1 to 5 for next hour forecast.

The procedure for performing physics-based correction approach is shown in Fig. 6.


This approach was tested on cross-Sect. 44 (Hsin-Pu Bridge) because of the availability
of observed water level data. The base flow calculation was done using rating curve against
each event separately using water level data as shown in Fig. 7.

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Fig. 6  Procedure for perform-


ing physics-based correction Input water level at time (t)
approach

Use rating curve and find flow rate at Hsin-Pu


Bridge with step 1 water levels

Simulate water level of each section using


WASH123D (Initial conditions)

Input QPF flow data from HEC-HMS

Simulate water level at time (t+1) as output


1-h ahead forecast

800
700
Flow Rate (m3/s)

600
500
400
300
200
100
0
40.5 41 41.5 42 42.5 43 43.5 44 44.5
Water Stage (m)

Fig. 7  Rating curve of Fengshan Creek basin at Hsin-Pu Bridge measuring station for base flow calculation

3.6 Performance evaluation measures

To evaluate the forecast performance of models, the following statistical indices were
selected:

1. Correlation coefficient (CC)


∑ � n
�� �
i=1
Ssi − Ss Soi − So
CC = � (15)
∑n � �2 ∑
n
� �2
i=1
Ssi − Ss i=1
Soi − So

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

  CC range from 0 to 1. If CC = 1 indicates strong positive correlation and forecast is


perfect while CC = 0 indicates no correlation.
2. Root Mean Square Error (RMSE)

∑n � �2
Ssi − Soi
RMSE = i=1 (16)
n
  RMSE is the error representation between two data sets. The smaller RMSE value
indicates better forecast.
3. Error of peak flood ( EFp)
EFp = Ss,p − So,p (17)
4. Error of time to peak flood ( ETp)
ETp = Ts,p − To,p (18)
where Ss and So are the forecasted and observed flood stages, respectively; n is the
total number of data points; Ss,p and So,p are forecasted and observed peak flood
stages, respectively; Ts,p and To,p are the peak time the forecasted and observed peak
flood stages, respectively; Ss and So are mean forecasted and observed flood stages,
respectively.

4 Results and discussion

4.1 Accuracy and ground‑truthing of radar rainfall data

Rainfall data are a key input variable for flood forecasting and its spatial distribution has
a significant impact on model simulation and performance. So the radar rainfall obtained
from QPESUMS system may be encountered with a certain error such as random, sys-
tematic and range-dependent errors; therefore, the determination of accuracy and ground-
truthing is more important on the spatial scale before using in flood forecasting and mod-
eling. The comparison is done based on CC and RMSE against each typhoon event using
rain gauge stations (Guan-Xi and Hsin-Pu) data and QPE radar rainfall grid data as shown
in Table 4. The CC and RMSE range 0.68–0.99 and 1.09–2.32 mm, respectively, for all
events at both stations indicated the credibility of radar data is high and can be used directly
in modeling without any correction factor. Also, the scatter plot relationship between these
two data sets explored correlation is high and data distribution is consistent with 45-degree
line as shown in Fig. 8. The radar rainfall data have high resolution and provide rainfall
distribution throughout the region. Based on geographical location of Guan-Xi station in
the mountainous region, the radar data are underestimated while Hsin-Pu station is rela-
tively in the flat region give consistently with radar rainfall data. The typhoon distribution
is characterized by a strong west–east gradient (Fig. 9). The spatial distribution of radar
rainfall pattern followed the same terrain distribution having gradient and flow of the river
from west to east. Moreover, the accumulated rainfall data obtained at both stations for all
selected events are comparable with radar grid rainfall data. For example, the accumulative
rainfall at Hsin-Pu station for Dujuan typhoon is 104 mm which is comparable with Hsin-
Pu radar grid rainfall (107 mm); similarly, for the same typhoon event at Guan-Xi station,

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Table 4  Correlation of rain Event Station CC RMSE(mm)


gauge and radar rainfall data
based on statistical measures
DUJUAN Guan-Xi 0.77 2.65
Hsin-Pu 0.78 1.63
SOUDELOR Guan-Xi 0.95 2.85
Hsin-Pu 0.95 1.12
MEGI Guan-Xi 0.99 1.71
Hsin-Pu 0.68 2.32
MALAKAS Guan-Xi 0.99 1.19
Hsin-Pu 0.87 1.37
0601–0605 Guan-Xi 0.99 1.35
Hsin-Pu 0.98 1.09
0701–0709 Guan-Xi 0.90 2.29
Hsin-Pu 0.99 1.68

Fig. 8  Comparison of QPE radar rainfall to observed rain-gauge rainfall of all selected events

the accumulative rainfall was 171 mm, while radar grid rainfall was 155 mm. These results
give strength to direct use of radar rainfall data for simulation and forecasting of the flood.
In all four typhoons and two plum rain events, the comparison of hourly QPE radar rainfall
to observed rain-gauge rainfall was satisfactory as shown in Fig. 8.

4.2 Calibration and validation of watershed models

According to the hydrological distribution of Fengshan Creek basin in this study, the
mountainous part is simulated with HEC-HMS, while the remaining region was simulated
with WASH123D for river flood stage. Dujuan and Malakas typhoon events were selected
for calibration of WASH123D for river flood stage compared to observations at Hsin-Pu
bridge river stage station, while the other typhoon events were used for the validation of

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Fig. 9  Areal distribution of accumulated radar rainfall over Fengshan Creek basin during four selected
typhoons and two plum rain events

the model. The parameters used for HEC-HMS modeling with SCS-CN method are given
in Table 5. These parameters were determined based on the geographic properties of area
and land cover survey data. The evaluation of modeling parameters was performed with the
reference study by Wu and Shih (2018).
The model calibration results for river flood stage are shown in Fig. 10. The calibra-
tion results provide reasonable river stage hydrograph during September 27 + 00:00 to Sep-
tember 29 + 23:00, 2015 and September 14 + 00:00 to September 18 + 23:00, 2016 where

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Table 5  Parameters used in Zone Initial loss Non-infiltration Lag time (min.) CN(I)
HEC-HMS modeling with (mm) covers (%)
SCS-CN method for Fengshan
Creek basin (Wu and Shih 2018)
Zone A 4.0 17.6 131.2 57.9
Zone B 4.0 22.1 147.2 61.1

Fig. 10  Comparison of simulated and observed flood stage for Hsin-Pu Bridge gauging station for model
calibration using WASH123D during typhoon Dujuan and Malakas

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Taiwan is suffered by Typhoon Dujuan and Malakas, respectively. Calibration results also
reveal that there is a good resemblance between observed and simulated river stage using
WASH123D model based on statistical measures. The CC is over 0.80 for both typhoon
events indicating a strong positive correlation between simulated and observed water stage;
moreover, the small error of EFp less than -0.4 m with a perfect match with zero ETp indi-
cates the high quality of calibration and applicability of WASH123D for validation pro-
cess. From Fig. 10, it can be seen that there is also a close corresponding between rainfall
intensity and water stage with a time lag of almost 1 to 1.5 h between maximum rainfall
intensity and peak river stage. Meanwhile, it was also analyzed that WASH123D model did
not fully capture the rising and receding limbs of water stage patterns. These inconsisten-
cies with observation are indicated by RMSE which is less than 0.4 m for both typhoons.
The small error indicators signify the validation and forecasting capability of WASH123D
model.
After model calibration, the validation is conducted using all fixed parameters during
the calibration period. The study validated two typhoon events Soudelor and Megi from
August 6th to August 8th of 2015 and September 26th to September 28th of 2016, respec-
tively, and two additional plum rain multi-peak events from June 1st to June 5th and July
1st to July 9th of 2017. Total four events were used to validate the parameters used in the
model. The simulation results compared to observation water stage are shown in Fig. 11.
The simulated hydrograph indicated discrepancies from the observed data, but the gen-
eral trends are comparable. For example, the simulated hydrographs and peaks of typhoon
Megi and plum rain June 2017 agree well with observations deliberated by an acceptable
range of statistical indices, i.e., CC = 0.90, RMSE = 0.36 to 0.49 m for both events. While
for Soudelor typhoon and plum rain event in July 2017, WASH123D model did not fully
capture water stage hydrograph. The simulated water level values were lower than obser-
vations; however, the water stage follows the rainfall intensity pattern, i.e., there is a clear
rise and fall to accompany with rainfall in water stage pattern indicated that the selected

Fig. 11  Comparison of observed and simulated water stage for model validation during typhoon Soude-
lor-2015 and Megi-2016 and Plum rain during June-2017 and July-2017

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Table 6  Performance evaluation Typhoon event CC RMSE (m) ETp(h) EFp(m)


indices during model validation
Soudelor 0.87 0.55 −5 − 0.26
Megi 0.89 0.49 0 − 0.84
0601–0605 0.90 0.36 1 0.04
0701–0709 0.82 0.37 − 141 − 0.41

hydrological parameters are suitable for flood modeling and forecasting. Further, Table 6
evaluates the simulated performance for model validation events. The plum rain event in
June 2017 give best results with CC = 0.90, RMSE = 0.36, ETp = 1 h and EFp = 0.04 m.
Comparisons between numerically simulated results and observation during calibration
and validation showed apparent discrepancies in terms of peak discharge (underestimation),
which is critical to flood forecasting. For example, the typhoon Megi event showed peak
underestimation of 0.84 m; meanwhile, it was also analyzed that the simulations fail to
fully capture the hydrograph during rising and falling stages. These might be due to initial
and boundary conditions in WASH123D. The simulations of WASH123D were performed
using transient simulation. Initial and boundary conditions must be identified for transient
simulation. The initial condition can be obtained from steady-state simulation with con-
stant boundary conditions (Yeh et al. 2011). The steady-state simulation must be based on
field measurements. In this study, the same instructions were used for transient simulations.
The initial conditions were obtained by steady-state simulation before truly beginning to
run in transient mode. For steady-state simulation, the starting conditions must match field
measurement. To force that to happen, we assign a constant known water stage to be input
at each node from measured data using 0.1dep file. This is called the initial enforced water
stage because it is only enforced in that steady-state run. The 0.0 m/s constant velocity and
cold start type were used to the initial conditions. The discharges from zone A and zone
B which HEC-HMS provides and a constant base flow from the rating curve are specified
as boundary conditions. The validity of boundary flows may be one constrain because the
underestimation of inflows from zone A and zone B into the main river may explain the
under-estimation of the peak discharge at Hsin-Pu Bridge. The discrepancies in rising and
recession limbs of stage hydrographs may be also due to non-consideration of subsurface
flow because of the domain of research study. This study is mainly related to peak flow
simulation for flood forecasting using event base typhoon precipitation. In this event base
simulations, we consider only a constant base flow that did not reproduce rising/recession
substantial changes in stage hydrograph. Reconsiderations of subsurface modeling may be
another way to reproduce the rising and recession limbs in a more obvious way, but this is
not in our scope of the study.

4.3 Flood stage simulation using QPE radar rainfall: model application

In this section, the model was applied to simulate the flood stage using radar rainfall (QPE)
data rather than rain-gauge rainfall data. The comparable results of two different rainfall
inputs are shown in Fig. 12, where Obs. is the observed flood stage at Hsin-Pu bridge
gauge station, Sim. is flood stage simulated using rain-gauge rainfall data, and QPE rep-
resents flood stage simulated using radar rainfall (QPE) data. For almost all events, the
model performs well and gives better results using radar rainfall data rather than with

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Fig. 12  a Comparison of water stage simulation using radar rainfall (QPE) and rain-gauge rainfall data dur-
ing four selected typhoon events. b Comparison of water stage using radar rainfall (QPE) and rain-gauge
rainfall data during plum rain events

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Table 7  Relative comparison of Typhoon event CC RMSE (m) ETp(h) EFp(m)


statistical parameters between
QPE and rainfall simulation
DUJUAN 0.79/0.88 0.43/0.37 0/ − 1 − 0.42/ − 0.29
SOUDELOR 0.87/0.94 0.55/0.38 − 5/0 − 0.26/0.22
MEGI 0.89/0.92 0.49/0.39 0/0 − 0.84/ − 0.53
MALAKAS 0.82/0.88 0.23/0.19 0/0 − 0.12/ − 0.16
0601–0605 0.90/0.89 0.36/0.39 1/1 0.04/0.03
0701–0709 0.82/0.88 0.37/0.31 − 141/0 − 0.41/0.42

The left values are rainfall simulation, while the right values are QPE
simulation

rain-gauge rainfall data. Due to the aerial distribution of radar QPE data, the model sim-
ulation trend, accuracy and peak water stage performance are improved as mentioned in
Table 7. The comparative analysis of statistical parameters performance gave better results
and acknowledged that radar rainfall estimation is more representative of watershed rainfall
situation and impacts on hydrological changes. This study demonstrates the good perfor-
mance of QPE radar rainfall grid data against rain gauge data, but calibration and vali-
dation of hydrological models were directly performed with the help of rain gauge data
because the rain gauge data are first-hand primary data and it is a common practice to
use rain gauge data to calibrate the model. More importantly, radar QPE also depends on
rain gauge observations in terms of adjustment and quality of estimated rainfall (He et al.
2013). Although the improvement in hydrograph simulation is more obvious using QPE,
the priority is given to rain gauge data for the calibration and validation of the model in this
study.

4.4 Flood level forecasting using QPF radar rainfall

1-h ahead flood level forecasting was performed by using as input one step ahead QPF
radar rainfall forecast data in modeling environment. The results obtained were encour-
aging but not satisfactory as shown in Fig. 13 for all selected typhoon events; therefore,
QPF forecasted flood levels were immediately corrected using two techniques (physics-
based correction (PC) and support vector machine (SVM)) after simulation. The compara-
tive analysis of these correction techniques was done using statistical measures. The results
obtained using QPF data are not according to expectation because QPF data itself have
certain errors. During the analysis of QPF radar rainfall data, it was observed that there is a
large over and under prediction fluctuations in all events except Malakas typhoon, as indi-
cated by statistical parameters, CC = 0.93 and RMSE = 1.47 mm. It was also analyzed that
all the other events showed RMSE range 2.16 to 11.07 mm with CC = 0.41–0.75. These
results indicated large error prediction in QPF data may be because QPF needs to use
maximum echo data with 10 min update cycle by grids, while the selected typhoons have
characteristics of thunderstorms with the dramatic change in rainfall intensity within shorts
durations. These characteristics of typhoons make QPESUMS system unable to fully cap-
ture these changes. In the same time, QPF is considered an important variable for river
stage forecast and early warning system of the flood. So, the amount, timing and location
are challenging task for flood warning authorities to forecast; therefore, this study used
QPF data directly to forecast water stage and applied error correction method to improve
the effectiveness of flood warning system.

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Fig. 13.  a Flood level forecasting using radar rainfall (QPF) data and comparison of forecasted flood level correc-
tion using PC and SVM during four typhoon events. b Flood level forecasting using radar rainfall (QPF) data and
comparison of forecasted flood level correction using PC and SVM during two plum rain events

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Table 8  Prediction performance Typhoon event Model Statistical parameters


of QPF, PC and SVM models
CC RMSE (m) ETp(h) EFp(m)

DUJUAN QPF 0.68 0.52 −6 − 1.29


PC 0.99 0.08 0 − 0.35
SVM 0.85 0.42 2 − 1.02
SOUDELOR QPF 0.92 0.94 −7 1.16
PC 0.97 0.74 −1 1.05
SVM 0.94 0.78 −9 1.79
MEGI QPF 0.94 0.39 −1 − 0.82
PC 0.99 0.09 0 − 0.37
SVM 0.91 0.36 1 − 0.29
MALAKAS QPF 0.54 0.33 −2 − 0.48
PC 0.99 0.07 0 − 0.07
SVM 0.93 0.16 −1 − 0.21
0601–0605 QPF 0.89 0.46 2 − 1.03
PC 0.98 0.19 2 − 0.7
SVM 0.85 0.48 3 − 1.15
0701–0709 QPF 0.74 0.87 2 1.82
PC 0.89 0.55 1 1.81
SVM 0.73 0.79 27 1.80

Table 8 evaluates the performance of QPF data input for flood level predication with
other correction approaches. It was analyzed that CC is between 0.54 and 0.94, minimum
for Malakas and maximum for Megi typhoon. The maximum RMSE = 0.94 m is shown
by Soudelor typhoon with ETp = − 7 h, and EFp = 1.16 m. Figure 11 clearly indicates that
the performance of QPF varies greatly in each event and predictions is more effected dur-
ing recession limb of hydrograph in all typhoon events. The error indices related for time
to peak and peak flood level are relatively high compared to observations, but overall
response trend is similar to actual flood level hydrograph. Keeping in view the values of
statistical parameters, it is difficult to use QPF information for flood forecasting and early
flood warning because QPF predictions did not provide correct information related to tim-
ing and amount of flood peak.

4.5 Forecasted flood level correction using physical approach

The under and over predictions of the watershed model using QPF rainfall as input data
are obvious for all hydrographs during rising, peak, and recession limbs because the total
rainfall difference was about 2 to 5 time then observed rainfall. The difference makes QPF
unable to use directly for flood warning; therefore, forecasted flood levels were corrected
by directly integrating physical-based WASH123D model with HEC-HMS hydrological
model and base flow from rating curve inputs. Table 9 compares the performance improve-
ments of 1 h ahead forecasting among data-driven SVR model and physics-based correc-
tion approach. The maximum improvement in a correlation coefficient (CC) and root mean
square error (RMSE) is up to 83% and 85%, respectively. It was analyzed that the forecast-
ing and observations are in excellent agreement (see Table 8); the CC raised above 0.95 and

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Table 9  Assessment of Typhoon event Model Improvement in statistical parameters


prediction improvement using
error correction approaches for CC (%) RMSE (%) ETp(h) EFp(%)
flood level
DUJUAN PC 46 − 85 0 − 73
SVM 25 − 19 2 − 21
SOUDELOR PC 5 − 21 −1 −9
SVM 2 − 17 −9 54
MEGI PC 5 − 77 0 − 55
SVM −3 −8 1 − 65
MALAKAS PC 83 − 79 0 − 85
SVM 72 − 52 −1 − 56
0601–0605 PC 10 − 59 2 − 32
SVM −4 4 3 12
0701–0709 PC 20 − 37 1 −1
SVM −1 −9 – −1

RMSE is below 0.2 m among all typhoon events except Soudelor and July-2017 plum rain
events. In general, the simulation hydrograph patterns are quite close to the observed data,
which reasonably captured the timing of peak flood level as shown in Fig. 13. Although
there is still an acceptable small error in EFp maybe due to QPF prediction itself contains
too much error. The prediction results obtained using physical correction indeed reason-
ably reflect the hydrological changes within the entire events whether in ascending, peak
and receding sections of stage hydrograph and handsomely encountered the errors of fore-
casted flood levels simulated with the help of QPF rainfall as input data. Keeping in view
the high prediction reliability with efficient correction results, a physical approach using
the numerical model (WASH123D) proved as an accurate novel modeling during typhoon
events for a peak as well as low magnitude flood stage forecasting. This approach is helpful
for flood early warning and indicated that physical characteristics of the watershed are very
important for flood level forecasting.

4.6 Forecasted flood level correction using SVM model

The error correction support vector (Eq. 10) from river flood stage and additional typhoon
events as training data are the two key mainstreams used for SVM model to predict 1 h
ahead flood level in Fengshan Creek basin. Table 8 summarizes that the corrected results
are improved compared with QPF original simulation. The results are good as identified by
the values of CC and RMSE but relatively poorer in terms of ETp and EFp. The comparison
between prediction and observations followed the pattern, but the SVM model is unable
to capture the flood peak time and amount. The peak and its time are very important for
flood early warning system. It was analyzed that the response of rising limb for all events
was fast, while receding part is relatively slow. Looking at Fig. 13 for SVM simulation
compared with observations, it can be seen that the model represents the overall hydrol-
ogy and SVM is an attractive way out without requiring knowledge about the complexity
and physical characteristics of the watershed. The drawback associated with SVM includes
peak-value prediction. The peak flood stage forecasting of every event was considerably
overestimated by SVM model, whereas low magnitude water levels were predicted fairly

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

accurate. The reason may be due to the range of training data sets used. The capability of
the model highly depends upon the amount of training data and support vector to provide
higher accuracy in general. So, SVM correction approach may face an insufficient set of
training data and/or the support vector, which is not enough to represent output in this
study.

4.7 Comparison of flood level forecasting

Figure 14(a and b) apparently shows the comparative forecasting performance of physi-
cal correction approach and SVM model, respectively, for all events. According to scatter
plots, both correction methods have positive effects on forecasted flood level using QPF
rainfall as input data. The spread of scattering indicating that both models correspond with
observations is very close for low magnitude water levels up to 42 m. These low values
mostly appeared in rising and recession limbs of flood level hydrographs and prediction
corrections are fairly accurate indicating the strength of integrating physical-based correc-
tion as well as set of training data and support vector for SVM correction. Moreover, it was
analyzed that the forecasting results are overestimated during multi-peak events such as
July-2017 plum rain event. The QPF radar data overestimation itself is high which inher-
ited into flood level forecasting and correction approaches. The peaks are almost overes-
timated in all typhoon events and calculations in terms of linear regressions, PC model
averagely gives a 0.01–6.58% overestimations of peak observed data of all typhoons, while
SVM gives 0.5–7.28% overestimations. The PC approach gives 2.68 m of peak overestima-
tion, while SVM gives 3.07 m during July-2017 plum rain event. The Dujuan, Megi and

Fig. 14  Scatter plots of forecasted and observation flood level for a physical correction approach and
b SVM model, where the line indicates linear regression between corrected forecasted flood level and
observed level (circles)

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

June-2016 plum rain typhoon events indicated underestimation of peak flood level, while
other typhoons represent overestimations and SVM model discrepancies are more obvious
than the physical correction model. These results indicated that flood level forecasting is
purely based on physical equations of river routing and integrating the forecasting model
with physical-based processes provide reasonable accuracy in terms of overall hydrology
rather than data-driven models such as SVM. No doubt, the SVM model improved the
accuracy of forecasted flood levels but over and under predictions of peaks indicating some
deficiency either in support vector or set of training data used. So, hydrologist needs spe-
cial expertise to understand SVM type machine learning models to get better results, while
the physical correction is simply based on hydrology data and its integration provides com-
prehensive predictions.
During the comparison of both models, the coefficient of determination (R2) as shown
in Fig. 14 indicates fairly good agreement between observation and predicted flood lev-
els, but physical correction demonstrates superior prediction capability. Further, Table 9
compares the performance improvement of 1 h ahead forecasting among physical correc-
tion approach and SVM model. The maximum improvements in CC, RMSE and EFp are
83%, 85% and 85%, respectively, for physical correction approach, while for SVM model,
CC = 72%, RMSE = 52% and EFp = 65%. The systematic comparison in terms of error
of time to peak flood level and error of peak flood level reveals the highest reliability of
physical-based correction model and accurately reflects the timing and amount of flood
peak in all selected typhoon except plum rain events. The physical correction model is like
the concept of predictor–corrector algorithm integrated with physics-based WASH123D
and HEC-HMS model in which the model is initially filled with the initial water level,
base flow, QPF-based side flows and flood level simulation using WASH123D, and then,
the prediction is done for next time step. Overall, the analysis and findings indicated that
physical correction approach yields the most accurate flood level predictions for Fengshan
Creek basin.

4.8 Flood alert and warning information

Flood forecasting is a necessary part of the flood management system that provides infor-
mation to the Water Resources Agency for preventative preparation and defense meas-
ure response against extreme typhoon flooding events. Early flood warning may lessen
the overall impact and damage to lives, livestock and property; but that warning system
requires flood forecasting update in real-time like in this study we used two (PC and SVM)
updating approaches for simulated forecast stage hydrographs as shown in Fig. 13. 1 h
ahead updating flood forecasting using physics-based approach is adopted here for flood
warning information to enable individuals and communities to respond appropriately to a
major flood threat to reduce the risk of death, injury and property loss. The selection of the
PC approach is done based on statistical performance and error correction improvement.
According to the Central Weather Bureau (CWB), the accuracy of forecasting reduced sig-
nificantly after 1 h lead time and more than 1 h ahead forecasting is not recommended for
the operation of the early warning system. Flood warning relies on triggers relating to criti-
cal river levels that are indicative of flood states approaching or worsening and are used to
decide when to undertake certain actions during a flood event (WMO 2011). For Fengshan
Creek basin at Hsin-Pu Bridge, CWB established three flood warning stages/levels based
on response actions.

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

1. Warning level 3 = 42.5 m = the river water level is expected to reach the floodplain level
in the next 2 h. The following response actions are recommended
a. Evacuation preparation begins.
b. Disaster prevention and relief mobilization and equipment preparation.
c. Attending that the water level continues to rise.
d. Close or prohibit traffic in river areas.
e. Warning to the evacuation of people in the floodplain.

2. Warning level 2 = 44.4 m = the river water level is expected to reach the designed flood
level (or top of the embankment) in the next 5 h and the follow actions should be
adopted.
a. Warning notification.
b. Watch for rising water levels and rainfall.
c. Disaster prevention and rescue mobilization and equipment standby.
d. Warning level 3 to level 2 alert notification.
e. Evacuation advisory completed and people ready to evacuate.

3. Warning level 1 = 45.2 m = the river water level is expected to reach the designed flood
level (or top of the embankment) in the next 2 h. The following response actions are
recommended.
a. Level 1 flood warning notification.
b. Strengthen dike safety by patrolling.
c. Facilitate embankment and disaster rescue.
d. Mandatory evacuation.

Fig. 15  Schematic of flood warnings, response actions and trigger levels for three flood warning levels at
Hsin-Pu Bridge during Typhoon Soudelor

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Table 10  The flood warning information of selected four typhoon and two plum rain events during this
study
Typhoon event Warning levels Total warn- Rising Release Peak
duration (h) ing time time time time
L3 L2 L1 (h) (h) (h) (h)

Dujuan 6 – – 6 2 3 1
Soudelor 10 11 1 22 15 6 1*
Megi 9 – – 9 4 4 1
Malakas – – – – – – –
0601–0605 8 – – 8 2 5 1+
0701–0709 – – – – – – –
1 3 4 – 7 3 3 1+
2 4 – – 4 1 2 1
3 1 – – 1 – 1
4 7 – – 7 1 5 1
5 1 – – 1 – – 1
6 – – – – – – –
7 4 3 2 9 2 5 2*
8 3 2 – 5 1 2 2+

In peak time column, the number with *represents peak time in warning level 1; the number with + repre-
sents peak time in warning level 2, while the numbers without any symbols represent peak time in warning
level 3

Based on the information collected for flood warning levels, a schematic of flood
warnings and responses along with the arrangement of triggers is illustrated in Fig. 15.
The flood warning levels are illustrated using typhoon Soudelor because the river level
reaches up to warning level 1 and can be used to completely describe the flood sta-
tus and response action. River rise commences flood status indicating alert initiated
response in warning level 3 at 42.5 m water stage and alert initiated after 40 h with the
start of typhoon Soudelor. After flood alert trigger, total duration for warning level 3
was 10 h (6 h from warning level 3 to level 2 in rising part and 4 h from warning level
2 to level 3 in release part). After 6 h of flood alert trigger, the water level hit the flood
warning trigger in warning level 2 and continue to the higher level (severe flood trigger)
within next 9 h in warning level 1 and come back to level 2 from level 1 within next 2 h.
There was only 1 h for the peak that hit warning level 1 (45.2 m). The total warning time
was 22 h from level 3 to level 2 to level 1 in rising part and level 1 to level 2 to level 3
in the release part. The decline of river level below 42.5 m is named as the downgrading
of warning in response actions that happen after 22 h of warning in typhoon Soudelor.
In warning schematic, the observed level indicated warning level 3 while PC approach
provides overestimation of water level within level 3 to level 1 warning that just be on
the safe side in the case of flood warning.
The flood warning details of four typhoons and two plum rain events are explained in
Table 10. According to flood warning levels information, all events hit level 3 warning
except Malakas and plum rain event 0101–0709 has maximum 34 h total warning time
from all the events and for 23 h remain in warning level 3.

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

5 Conclusions

This study proposed a new approach integrated with physics-based WASH123D and HEC-
HMS models to predict 1-h ahead flood level hydrograph in Fengshan Creek basin, north-
ern Taiwan. The comparison of this approach was done with data-driven SVM model and
performances were assessed by using statistical indicators (i.e., CC, RMSE, ETp, EFp).
Four typhoon and two plum rain events (with 620 data sets) were considered for model
calibration and validation process. First the watershed model calibration, validation and
parameterization were performed with rain-gauge rainfall data as input to simulate river
water levels, and then, model application verification was done using QPE radar rainfall
data as input. During the study, the model performance was found better when it used radar
rainfall estimation (QPE) data rather than rain-gauge data. Results of using 1 h ahead radar
rainfall prediction, QPF as input for flood forecasting, were encouraging but not feasible to
use directly for early flood warning system due to errors in peaks flood levels and timing.
Therefore, two correction techniques were suggested and applied to improve the accuracy
of flood forecasting.
Overall, based on thoroughly detailed analysis and comparison, SVM model is con-
sidered as an attractive way out to improve the accuracy of QPF forecasted flood levels
without requiring knowledge about the complexity and physical characteristics of the
watershed. The capability of the model highly depends upon the amount of training data
and support vector to provide higher accuracy in general. The drawbacks associated with
SVM include peak-value predictions and its timing. The peak flood level forecasting of
four events was considerably overestimated, whereas low magnitude water levels were pre-
dicted fairly accurate. The reason may be due to the range of training data sets used. The
SVM model was unable to fully describe the flood level based on its performance during
selected typhoon events and may not be used directly for flood early warning system at this
stage. However, the prediction results could be improved by using different support vectors
and additional training data sets consist of multi-peaks that represent the overall hydrologi-
cal situation of Fengshan Creek basin.
The results obtained by the physics-based approach were much better than the SVM
model. The approach described physics of hydrograph patterns in full manners and outputs
have the same 1 h ahead predictions, in excellent agreement with observations. However,
for multi-peaks July-2017 storm, the peaks were overestimated may be due to QPF original
data errors. This physics-based approach using the numerical model (WASH123D) proved
as an accurate novel modeling during typhoon events and provided high reliability for a
peak as well as low magnitude flood stage forecasting. This approach indicated that physi-
cal characteristics of the watershed are very important for flood and water level forecasting.
The accurate reflections of timing and amount of flood peaks in all selected typhoon by
newly developed physical approach with its operational nature are recommended to use
by the government in future for early warning to reduce the flood impacts during typhoon
events.

Acknowledgements This research is conducted with the financial support of the Ministry of Science and
Technology under Grant no. MOST-108-2621-M-008-009. The authors highly acknowledge the authori-
ties and agencies for the provision of valuable data such as Center for space and remote sensing research
(CSRSR), National Central University; National Land Surveying and Mapping Center, Ministry of Inte-
rior, Taiwan (NLSC); Central Geological Survey (CGS), MOEA; Water Resource Agency (WRA); Central
Weather Bureau (CWB) of Taiwan.

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Author contributions All authors contributed to the study conception and design. The data collection, meth-
odology and analysis were performed by [FH] and [JW] under the supervision and technical support of
[RW]. The first draft of the manuscript was written by [FH]. Review and editing were done by [RW]. All
authors read and approved the final manuscript.

References
Abbot J, Marohasy J (2014) Input selection and optimisation for monthly rainfall forecasting in Queensland,
Australia, using artificial neural networks. Atmos Res 138:166–178. https​://doi.org/10.1016/j.atmos​
res.2013.11.002
Adamowski JF (2008) Development of a short-term river flood forecasting method for snowmelt driven
floods based on wavelet and cross-wavelet analysis. J Hydrol 353:247–266. https​://doi.org/10.1016/j.
jhydr​ol.2008.02.013
BWDB (2015) Annual flood report 2015. Flood forecasting and warning center (FFWC), Bangladesh water
development board (BWDB), Dhaka, Bangladesh
Bellos V, Tsakiris G (2016) A hybrid method for flood simulation in small catchments combining hydro-
dynamic and hydrological techniques. J Hydrol 540:331–339. https​://doi.org/10.1016/j.jhydr​
ol.2016.06.040
Chang CC, Lin CJ (2011) LIBSVM: a library for support vector machines. ACM Trans Intell Syst Technol
2(3):1–27. https​://doi.org/10.1145/19611​89.19611​99
Chang MJ, Chang HK, Chen YC, Lin GF, Chen PA, Lai JS, Tan YC (2018) A support vector machine fore-
casting model for typhoon flood inundation mapping and early flood warning systems. Water 10:1734
Chiang YM, Chang FC (2009) Integrating hydrometeorological information for rainfall-runoff modelling by
artificial neural networks. Hydrol Process 23(11):1650–1659. https​://doi.org/10.1002/hyp.7299
Chu PS, Zhang H, Chang HL, Chen TL, Tofte K (2018) Trends in return levels of 24-hr precipita-
tion extremes during the typhoon season in Taiwan. Int J Climatol 38(14):5107–5124. https​://doi.
org/10.1002/joc.5715
Costabile P, Macchione F (2015) Enhancing river model set-up for 2-D dynamic flood modelling. Environ
Model Softw 67:89–107. https​://doi.org/10.1016/j.envso​ft.2015.01.009
Dehghani M, Saghafian B, Nasiri Saleh F, Farokhnia A, Noori R (2014) Uncertainty analysis of streamflow
drought forecast using artificial neural networks and Monte-Carlo simulation. Int J Climatol 34:1169–
1180. https​://doi.org/10.1002/joc.3754
Roo De et al (2003) Development of a European flood forecasting system. Int J River Basin Manag 1(1):49–
59. https​://doi.org/10.1080/15715​124.2003.96351​92
Feldman AD (2000) Hydrologic modeling system HEC-HMS. Technical reference manual, U.S. Army
corps of engineers. Hydrologic Engineering Center, HEC, Davis
Government of Nepal (2018) Standard operating procedure (SOP) for flood early warning system in Nepal.
Ministry of energy, water resources and irrigation, department of hydrology and meteorology, Kath-
mandu, Nepal
Ghosh S (2010) SVM-PGSL coupled approach for statistical downscaling to predict rainfall from GCM out-
put. J Geophys Res 115:D22102. https​://doi.org/10.1029/2009J​D0135​48
Gumindoga W, Rwasoka DT, Nhapi I, Dube T (2016) Ungauged runoff simulation in Upper Manyame
Catchment, Zimbabwe: applications of the HEC-HMS model. Phys Chem Earth 100:371–382. https​://
doi.org/10.1016/j.pce.2016.05.002
He X, Sonnenborg TO, Refsgaard JC, Vejen F, Jensen KH (2013) Evaluation of the value of radar QPE data
and rain gauge data for hydrological modeling. Water Resour Res 49(2013):5989–6005. https​://doi.
org/10.1002/wrcr.20471​
Hsiao LF, Yang MJ, Lee CS, Kuo HC, Shih DS, Tsai CC, Wang CJ, Chang LY, Chen DYC, Feng L, Hong
JS, Fong CT, Chen DS, Yeh TC, Huang CY, Guo WD, Lin GF (2013) Ensemble forecasting of
typhoon rainfall and floods over a mountainous watershed in Taiwan. J Hydrol 506:55–68. https​://doi.
org/10.1016/j.jhydr​ol.2013.08.046
Hsu TW, Shih DS, Chen WJ (2015) Destructive flooding induced by broken embankments along Linbian
Creek, Taiwan, during typhoon morakot. J Hydrol Eng 20(7):1–9. https​://doi.org/10.1061/(ASCE)
HE.1943-5584.00010​69
Hsu TW, Shih DS, Li CY, Lan YJ, Lin YC (2017) A study on coastal flooding and risk assessment under
climate change in the Mid-Western Coast of Taiwan. Water 9:390. https​://doi.org/10.3390/w9060​390

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Jain SK, Mani P, Jain SK, Prakash P, Singh VP, Tullos D, Kumar S, Agarwal SP, Dimri APA (2018) Brief
review of flood forecasting techniques and their applications. Int J River Basin Manag 16(3):329–344.
https​://doi.org/10.1080/15715​124.2017.14119​20
Kang F, Li J (2016) Artificial bee colony algorithm optimized support vector regression for system reliabil-
ity analysis of slopes. J Comput Civ Eng 30(3):04015040. https​://doi.org/10.1061/(ASCE)CP.1943-
5487.00005​14
Kim S, Matsumi Y, Pan S, Mase H (2016) A real-time forecast model using artificial neural network for
after-runner storm surges on the Tottori Coast. Japan Ocean Eng 122:44–53. https​://doi.org/10.1016/j.
ocean​eng.2016.06.017
Lagmay AMFA, Racoma BA, Aracan KA, Alconis-Ayco J, Saddi IL (2017) Disseminating near real-time
hazards information and flood maps in the Philippines through Web-GIS. J Environ Sci 59:13–23.
https​://doi.org/10.1016/j.jes.2017.03.014
Leahy P, Kiely G, Corcoran G (2008) Structural optimisation and input selection of an artificial neural net-
work for river level prediction. J Hydrol 355:192–201. https​://doi.org/10.1016/j.jhydr​ol.2008.03.017
Li S, Ma K, Jin Z, Zhu Y (2016) A new flood forecasting model based on SVM and boosting learning algo-
rithms. In 2016 IEEE congress on evolutionary computation (CEC), Vancouver, BC, Canada, 1343–
1348. https​://doi.org/10.1109/CEC.2016.77439​44
Lin GF, Chou YC, Wu MC (2013) Typhoon flood forecasting using integrated two-stage support vector
machine approach. J Hydrol 486:334–342. https​://doi.org/10.1016/j.jhydr​ol.2013.02.012
Liong SY, Sivapragasam C (2002) Flood stage forecasting with support vector machines. JAWRA J Am
Water Resour As 38:173–186. https​://doi.org/10.1111/j.1752-1688.2002.tb015​44.x
Lohani AK, Goel N, Bhatia K (2014) Improving real time flood forecasting using fuzzy inference system. J
Hydrol 509:25–41. https​://doi.org/10.1016/j.jhydr​ol.2013.11.021
Markar MS, Clark SQ, Gooda M, Yaowu M, Yali C (2005). Improved flood forecasting for the Yangtze
River in China. 8th international river symposium, Brisbane, 5th–9th September 2005, Australia
Mosavi A, Ozturk P, Chau K (2018) Flood prediction using machine learning models: literature review.
Water 10(11):1536. https​://doi.org/10.3390/w1011​1536
Nayak P, Sudheer K, Rangan D, Ramasastri K (2005) Short-term flood forecasting with a neurofuzzy model.
Water Resour Res 41:W04004. https​://doi.org/10.1029/2004W​R0035​62
Ortiz-García E, Salcedo-Sanz S, Casanova-Mateo C (2014) Accurate precipitation prediction with sup-
port vector classifiers: a study including novel predictive variables and observational data. Atmos Res
139:128–136. https​://doi.org/10.1016/j.atmos​res.2014.01.012
Panda RK, Pramanik N, Bala B (2010) Simulation of river stage using artificial neural network and MIKE
11 hydrodynamic model. Comput Geosci 36:735–745. https​://doi.org/10.1016/j.cageo​.2009.07.012
Pereira Filho AJ, dos Santos CC (2006) Modeling a densely urbanized watershed with an artificial neu-
ral network, weather radar and telemetric data. J Hydrol 317:31–48. https​://doi.org/10.1016/j.jhydr​
ol.2005.05.007
Raje D, Mujumdar PP (2011) A comparison of three methods for downscaling daily precipitation in the
Punjab region. Hydrol Process 25:3575–3589. https​://doi.org/10.1002/hyp.8083
Rezaeianzadeh M, Tabari H, Yazdi AA, Isik S, Kalin L (2014) Flood flow forecasting using ANN, ANFIS
and regression models. Neural Comput Appl 25(1):25–37. https​://doi.org/10.1007/s0052​1-013-1443-6
SCS (1964) National engineering handbook, section 4: hydrology, soil conservation service. US Depart-
ment of Agriculture (USDA), Washington, DC
Shih DS, Chen CJ, Li MH, Jang CS, Chang CM, Liao YY (2019) Statistical and numerical assessments of
groundwater resource subject to excessive pumping: case study in Southwest Taiwan. Water 11(2):360.
https​://doi.org/10.3390/w1102​0360
Shih DS, Chen CH, Yeh GT (2014) Improving our understanding of flood forecasting using earlier hydro-
meteorological intelligence. J Hydrol 512:470–481
Shih DS, Liau JM, Yeh GT (2012a) Model assessments of precipitation with a unified regional circulation
rainfall and hydrological watershed model. J Hydrol Eng 17(1):43–54. https​://doi.org/10.1061/(ASCE)
HE.1943-5584.00004​14
Shih DS, Hsu TW, Chang KC, Juan HL (2012b) Implementing coastal inundation data with an integrated
wind wave model and hydrological watershed simulations. Terr Atmos Ocean Sci 23:513–525. https​://
doi.org/10.3319/TAO.2012.05.03.01(WMH)
Shih DS, Yeh GT (2011) Identified model parameterization, calibration, and validation of the physically
distributed hydrological model WASH123D in Taiwan. J Hydrol Eng 16(2):126–136. https​://doi.
org/10.1061/(ASCE)HE.1943-5584.00002​93
Shrestha DL, Robertson DE, Wang QJ, Pagano TC, Hapuarachchi HAP (2013) Evaluation of numerical
weather prediction model precipitation forecasts for short-term streamflow forecasting purpose. Hydrol
Earth Syst Sci 17:1913–1931. https​://doi.org/10.5194/hess-17-1913-2013

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Smola AJ, Schölkopf B (2004) A tutorial on support vector regression. Stat Comput 14(3):199–222. https​://
doi.org/10.1023/B:STCO.00000​35301​.49549​.88
Su JQ, Wang X, Liang Y, Chen B (2014) GA-based support vector machine model for the prediction of
monthly reservoir storage. J Hydrol Eng 19(7):1430–1437. https​://doi.org/10.1061/(ASCE)HE.1943-
5584.00009​15
Suykens JA, Vandewalle J (1999) Least squares support vector machine classifiers. Neural Process Lett
9(3):293–300. https​://doi.org/10.1023/A:10186​28609​742
USACE (2017) Hydrologic modeling system (HEC-HMS) Application guide: version 4.3, Institute for
Water Resources, Hydrologic Engineering Center, Davis, CA
USACE (2018) Hydrologic modeling system (HEC-HMS) User’s manual: version 4.3, Institute for Water
Resources, Hydrologic Engineering Center (CEIWR-HEC), Davis, CA.
Van den Honert RC, McAneney J (2011) The 2011 Brisbane floods: causes, impacts and implications.
Water 3(4):1149–1173. https​://doi.org/10.3390/w3041​149
Vapnik VN (1999) An overview of statistical learning theory”. IEEE Trans Neural Netw 10(5):988–999.
https​://doi.org/10.1109/72.78864​0
Werner M, Dijk M (2005) Developing flood forecasting systems: examples from the UK, Europe, and Paki-
stan. International conference on innovation advances and implementation of flood forecasting tech-
nology, 17 to 19 October 2005, Tromsø, Norway
WMO (2010) Mission report, WMO Fact-Finding and Needs-Assessment Mission to Pakistan. Available
from: https​://www.wmo.int/pages​/prog/dra/rap/docum​ents/Pakis​tanMi​ssion​Repor​t.pdf. Accessed 5
October 2020
WMO (2011) Manual on flood forecasting and warning, WMO No. 1072. World Meteorological Organiza-
tion, Geneva
WRA (1990) Handbook for hydrological design. Water resources agency (WRA), Ministry of Economic
Affairs (MOEA), Taiwan
Wu MC, Lin GF (2017) The very short-term rainfall forecasting for a mountainous watershed by means of
an ensemble numerical weather prediction system in Taiwan. J Hydrol 546:60–70
Wu RS, Shih DS (2018) Modeling hydrological impacts of groundwater level in the context of climate and
land cover change. Terr Atmos Ocean Sci 29:341–353. https​://doi.org/10.3319/TAO.2017.10.26.01
Wu SJ, Lien HC, Chang CH, Shen JC (2012) Real-time correction of water stage forecast during rainstorm
events using combination of forecast errors. Stoch Environ Res Risk Assess 26:519–531. https​://doi.
org/10.1007/s0047​7-011-0514-4
Xu Z, Li J (2002) Short-term inflow forecasting using an artificial neural network model. Hydrol Process
16(12):2423–2439. https​://doi.org/10.1002/hyp.1013
Yeh GT, Cheng HP, Cheng JR, Lin JHC, Martin WD (1998) A numerical model simulating water flow and
contaminant and sediment transport in WAterSHed systems of 1-D stream-river network, 2-D overland
regime, and 3-D subsurface media (WASH123D: Version 1.0). Prepared for U.S. Environmental Pro-
tection Agency, Environmental Research Laboratory, 1998, technical report CHL-98–19, pp-369
Yeh GT, Shih DS, Cheng JRC (2011) An integrated media, integrated processes watershed model. Comput
Fluids 45(1):2–13. https​://doi.org/10.1016/j.compf​l uid.2010.11.018
Young CC, Liu WC, Wu MC (2017) A physically based and machine learning hybrid approach for accurate
rainfall-runoff modeling during extreme typhoon events. Appl Soft Comput 53:205–216. https​://doi.
org/10.1016/j.asoc.2016.12.052
Yu PS, Chen ST, Chang IF (2006) Support vector regression for real-time flood stage forecasting. J Hydrol
328:704–716. https​://doi.org/10.1016/j.jhydr​ol.2006.01.021
Yu PS, Yang TC, Chen SY, Kuo CM, Tseng HW (2017) Comparison of random forests and support vec-
tor machine for real-time radar-derived rainfall forecasting. J Hydrol 552:92–104. https​://doi.
org/10.1016/j.jhydr​ol.2017.06.020
Yu X, Liong SY, Babovic V (2004) EC-SVM approach for real time hydrologic forecasting. J Hydroinf
6(3):209–223. https​://doi.org/10.2166/hydro​.2004.0016
Zhao M, Hendon HH (2009) Representation and prediction of the Indian Ocean dipole in the POAMA sea-
sonal forecast model. Q J R Meteorol Soc 135(639):337–352. https​://doi.org/10.1002/qj.370

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Natural Hazards

Authors and Affiliations

Fiaz Hussain1,2 · Ray‑Shyan Wu1 · Jing‑Xue Wang1


Fiaz Hussain
fiazhussain@g.ncu.edu.tw; engr.fiaz@uaar.edu.pk
Jing‑Xue Wang
super5307847@gmail.com
1
Department of Civil Engineering, National Central University, Taoyuan City, Taiwan
2
Department of Agricultural Engineering, PMAS-Arid Agriculture University Rawalpindi,
Rawalpindi, Pakistan

13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center
GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers
and authorised users (“Users”), for small-scale personal, non-commercial use provided that all
copyright, trade and service marks and other proprietary notices are maintained. By accessing,
sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of
use (“Terms”). For these purposes, Springer Nature considers academic use (by researchers and
students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and
conditions, a relevant site licence or a personal subscription. These Terms will prevail over any
conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription (to
the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of
the Creative Commons license used will apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may
also use these personal data internally within ResearchGate and Springer Nature and as agreed share
it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not otherwise
disclose your personal data outside the ResearchGate or the Springer Nature group of companies
unless we have your permission as detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial
use, it is important to note that Users may not:

1. use such content for the purpose of providing other users with access on a regular or large scale
basis or as a means to circumvent access control;
2. use such content where to do so would be considered a criminal or statutory offence in any
jurisdiction, or gives rise to civil liability, or is otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association
unless explicitly agreed to by Springer Nature in writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a
systematic database of Springer Nature journal content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a
product or service that creates revenue, royalties, rent or income from our content or its inclusion as
part of a paid for service or for other commercial gain. Springer Nature journal content cannot be
used for inter-library loans and librarians may not upload Springer Nature journal content on a large
scale into their, or any other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not
obligated to publish any information or content on this website and may remove it or features or
functionality at our sole discretion, at any time with or without notice. Springer Nature may revoke
this licence to you at any time and remove access to any copies of the Springer Nature journal content
which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or
guarantees to Users, either express or implied with respect to the Springer nature journal content and
all parties disclaim and waive any implied warranties or warranties imposed by law, including
merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published
by Springer Nature that may be licensed from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a
regular basis or in any other manner not expressly permitted by these Terms, please contact Springer
Nature at

onlineservice@springernature.com

You might also like