You are on page 1of 17

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/297646680

Application of an Automated LSPIV System in a Mountainous Stream for


Continuous Flood Flow Measurements

Article in Hydrological Processes · March 2016


DOI: 10.1002/hyp.10836

CITATIONS READS

38 430

5 authors, including:

Qian Liao Honglei Tang


University of Wisconsin - Milwaukee Kunming University of Science and Technology
43 PUBLICATIONS 671 CITATIONS 6 PUBLICATIONS 104 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

WHRP - WisDOT 0092-13-03: Understanding and Complying with StormWater Mitigation Guidelines from the EPA View project

All content following this page was uploaded by Honglei Tang on 25 November 2020.

The user has requested enhancement of the downloaded file.


HYDROLOGICAL PROCESSES
Hydrol. Process. 30, 3014–3029 (2016)
Published online 19 April 2016 in Wiley Online Library
(wileyonlinelibrary.com) DOI: 10.1002/hyp.10836

Application of an automated LSPIV system in a mountainous


stream for continuous flood flow measurements
Qi-hua Ran,1 Wei Li,1 Qian Liao,2* Hong-lei Tang1 and Meng-yao Wang1
1
Institute of Hydrology and Water Resources, Zhejiang University, Hangzhou 310058, China
2
Department of Civil and Environmental Engineering, University of Wisconsin–Milwaukee, Milwaukee, WI, USA

Abstract:
An imaging-based automated large-scale particle image velocimetry (LSPIV) system for flash flood monitoring is developed and
deployed in a mountainous stream in the Longchi Catchment, Chengdu, China. This system is built from a low-cost Raspberry Pi
board-level computer with a camera module, which can acquire continuous images/videos automatically at programmed
intervals. The minimum quadratic difference algorithm tracks surface patterns as flow tracers to estimate the distribution of
surface velocities. Meanwhile, a stereo imaging-based ‘virtual pole’ method has been developed to reconstruct the three-
dimensional topography with a stereo digital camera, and a cross-sectional bathymetry has been generated without manual
surveying. The varying water stage and water surface gradient, which are critical parameters that affect image rectification and
surface velocity measurements, can also be directly resolved by applying the two imaging modules together. Discharge can then
be estimated with the velocity–area method through selected cross sections.
A flash flood that occurred between 24 July 2014 and 25 July 2014 is selected for analysis. The water surface level reconstructed
from image processing was validated with marked water levels, and a good agreement was found with a root mean square error
of 3.7 cm. The discharge recorded during the flood recession process ranged from approximately 3.5 to 27 m3/s. The rating curve
obtained can be well described by a power function, and the linear regression suggested a Manning’s n roughness coefficient of
0.18 of one specific cross section. Some limitations of the presented large-scale particle image velocimetry system are also put
forward, and possible solutions are provided for future improvements. With these proposed upgrades, the system can provide
valuable datasets of flash floods in steep mountainous streams, which are critically needed for improving our understanding and
modelling of many hydrological processes associated with flood generation, propagation and erosion, as well as for real-time
forecasting. Copyright © 2016 John Wiley & Sons, Ltd.

KEY WORDS LSPIV; flash flood; mountainous stream; discharge estimation; bathymetry reconstruction

Received 21 July 2015; Accepted 1 March 2016

INTRODUCTION In gauged mountainous streams and rivers, traditional


Flash floods are the most common natural disasters in equipment may not perform well because they could be
mountainous catchments where heavy rainfalls are destroyed by the extreme events or the high flow may
quickly transformed into run-off, which may also be exceed the recordable level (Ruiz-Villanueva et al.,
accompanied by landslides and debris flows in high 2013). In addition, rating curves established under normal
gradient streams and cause devastating damage to the flow conditions must be extrapolated to estimate higher
infrastructure and human society (Borga et al., 2007; flow rates, which significantly increases the uncertainty of
Hapuarachchi et al., 2011). A flash flood is generally discharge measurements (Le Coz et al., 2010). A network
defined as a flow event that responses within less than 6 h of gauging stations in mountainous catchments rarely
of an intense or excessive rainstorm (Sweeney, 1992). exists, and rating curves can hardly be dependable, even
Essential hydrologic data, including the water stage, flow under normal conditions. The dominant run-off response
velocity and discharge during the rapid process are may change because of the increase of storm severity, and
critically needed for improving our understanding and the lack of reliable field data for flash floods has
modelling of flood generation and propagation, and for hampered our understanding of these processes (Marchi
real-time forecasting. et al., 2009).
Conventionally, discharge measurements of rivers
without an established gauging station can be conducted
*Correspondence to: Qian Liao, Department of Civil and Environmental using current metres (i.e. mechanical devices such as
Engineering, University of Wisconsin–Milwaukee, EMS Building, impellers or rotating cups; Rantz, 1982), acoustic Doppler
Milwaukee, WI, USA.
E-mail: liao@uwm.edu velocimeters (Muste et al., 2007), acoustic Doppler

Copyright © 2016 John Wiley & Sons, Ltd.


LSPIV FOR MOUNTAINOUS FLOOD MONITORING 3015

current profilers (Gordon, 1989; Oberg and Mueller, velocity measurement to estimate the discharge. The
2007; Simpson, 2001; Yorke and Oberg, 2002), remote automated feature of the LSPIV system makes it widely
sensing methods (Birkinshaw et al., 2014; Costa et al., applicable and ideally suited for remote sites for
2006; Getirana and Peters-Lidard, 2013) or estimation by measuring rapid and extreme flows. The measurement
postevent dendrogeomorphic evidence analysis results indicate that the system is reliable and has the
(Ballesteros Cánovas et al., 2011; Ruiz-Villanueva potential to provide valuable hydrologic data that cannot
et al., 2010). It is, however, not only time consuming be easily obtained through conventional methods.
and labour intensive, but also unsafe and technically
difficult to extend intrusive techniques to conduct field
measurements during high flow events. For hydraulic EXPERIMENT SETTINGS
applications, remote sensing based on satellite is costly,
and covers large target areas, whose resolution is much Experimental site
greater than the width of a typical mountainous stream Field experiments were conducted in the Longchi
(Fujita and Kunita, 2011), and repeat observation Catchment, which is approximately 100 km northwest of
frequency usually depends on the orbital period. More- the City of Chengdu, Sichuan Province, China. It is a
over, the investigation of flash floods (such as survey or typical mountainous watershed with the altitude ranging
direct observation) is event-based and opportunistic from 793 to 3290 m above the sea level, and a catchment
which can hardly be prepared in advance (Borga et al., area of 66 km2. The average annual rainfall is 800–
2008). Unfortunately, many mountainous catchments in 900 mm, with about 80% of which occurs in the monsoon
the world that suffer from recurring flash floods are season (from May to September). The Longxi River is the
ungauged. For example, mountainous streams in South- main stream across the watershed. About 18 km in length,
western China are generally situated in underdeveloped it flows into the Zipingpu Reservoir, which serves as one
regions where resources such as electricity and commu- of the primary water supplies for the city of Chengdu with
nication infrastructures are very limited. It is of a great a population of 14 million. Because of its steep
challenge to measure and monitor flood flows in these topography and quick stream flow response to heavy
areas. precipitation events, the Longchi Catchment has a history
The advent of a nonintrusive, short-range remote of some most severe flash flood-related disasters. This
monitoring technique, named large-scale particle image region is very close to the epicentre of the infamous 2008
velocimetry (LSPIV), has solved some of these issues Wenchuan earthquake which measured 8.0 Ms in magni-
successfully, and can serve as a promising candidate for tude, and caused nearly 70 000 deaths. The earthquake is
flood measurements in mountainous streams. A typical also believed to be the major cause of the loose soil and
LSPIV system adopts video cameras to acquire images of rock materials distributed on steep slopes (Zhou and
a considerably large water surface area under natural light Tang, 2014). A devastating basin-scale debris flow
conditions. LSPIV relies on natural tracers and surface occurred here during a heavy storm event on 13 August
patterns (e.g. foam, floating debris, leaves, turbulence 2010. Landslides and debris flows were triggered
patterns and ripples) for velocity measurements. The simultaneously in 37 gullies, and the total volume of
acquired images are divided into small ‘interrogation debris material reached 778 × 104 m3 in that event (Xu
areas’ on which statistical methods (e.g. cross-correlation) et al., 2012). Almost every house along the Longxi River
are applied repeatedly to estimate tracer displacements on valley was buried in the debris deposition, and many
the surface; hence, the velocity distribution on the water structures (bridges and concrete-line channel) were
surface can be derived. Although LSPIV has been a great severely damaged (refer to Figure 1a). These recent
progress for stream flow measurements in field or catastrophic events have brought some urgent research
laboratory settings, it has rarely been employed for needs in the watershed, especially those related to floods
continuous flood monitoring in mountainous catchments and debris flows. A better understanding of the local
with steep slopes and complex topographies. hydrological and hydraulic properties is a key factor for
In this article, a low-cost LSPIV system is developed to future risk assessments (Chu et al., 2011). As part of this
measure the surface velocity distribution and discharge research, we have deployed a number of hydrological
continuously in an ungauged mountainous stream in sensors including water-level gauges, soil moisture
Southwestern China. Meanwhile, a stereo-imaging sensors and rain gauges along the Longxi River. We
method is applied to reconstruct the channel bathymetry. were, however, not able to deploy conventional flow
Parameters obtained through the 3D topography infor- metres, such as acoustic Doppler velocimeters and
mation are also used for water surface profile measure- acoustic Doppler current profilers, to measure the flow
ments. The reconstructed channel bathymetry and water as a result of the safety concern during dangerous flood
surface elevation are then combined with the surface events. Many instruments were damaged or washed away

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
3016 Q. RAN ET AL.

Figure 1. General information about the experimental site and automated system components. (a) Bulk of rocks and stones from the side slope flushed
into the main river channel; (b) components of the automated large-scale particle image velocimetry (LSPIV) system, including a Raspberry Pi board
computer, a camera module, a timer switch and a lithium battery; (c) location of the system deployment (black dot) in the Longchi Catchment; (d)
enclosure of the LSPIV system installed on a steel bridge over the Longxi River

by torrential flows since the field experimental campaign be combined with soil water content, rainfall process and
started. As a result, fundamental hydrological parameters vegetation for a more comprehensive hydrological
such as the flow stage, velocity and discharge, are still analysis.
missing despite our past efforts.
The primary objective of this study is to test and
calibrate an LSPIV system as a reliable short-range The automated LSPIV system
remote sensing instrument to measure flow speed and Our primary objective is to monitor flash flood
discharge of mountainous rivers under extreme flow automatically and continuously via simple yet robust
conditions. The system was installed on a steel bridge in components. The automated LSPIV system developed for
the middle reach of the Longxi River (refer to Figure 1c), this study comprises four components, including a
with a wide nonobstructive view on the entire river Raspberry Pi board computer, a camera module, a
surface. Bulks of bed rocks and gravels were washed out, 12-volt lithium battery and a timer switch, as illustrated in
and the concrete-line channel bed had been destroyed by Figure 1b. Raspberry Pi (http://www.raspberrypi.org) is a
a number of flood events since it was constructed in popular credit card-sized board computer implemented on
2011. The hydraulic analysis on this site is challenging a Raspbian platform with extended data acquisition
because of the varying channel roughness and bathy- capabilities. As shown in Figure 1b, a digital camera
metry. The hydrologic response here can be attributed to module is connected to the Raspberry Pi for capturing
integrated run-off generation mechanisms, i.e. the Dunne videos or still images of the flow surface. A DC power
overland flow and subsurface flow (Ran et al., 2015). timer switch that controls the discharge of a lithium
Run-off often occurs abruptly after continuous rainfall battery has been programmed to switch on and off for 17
with unpredictable rising and recessing trends, bringing times daily from 8 am to 8 pm, with a 2-min ON time for
great difficulties to conduct direct field measurements every power cycle. The Raspberry Pi is set to run a
during a flood event. An automated LSPIV system, if Python program automatically at every start-up, which
well designed and calibrated, can be desirable for instructs the camera to take one video clip for 5 s, and
acquiring valuable hydraulic data including the velocity, save it as a video file into a SD memory card. With a
water level and discharge. These critical parameters can 20 Ah battery capacity, this configuration allows the

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
LSPIV FOR MOUNTAINOUS FLOOD MONITORING 3017

system to operate continuously and unattended for about sections. Modern survey tools, such as the ground-based
40 days. The whole system can be assembled in a or airborne 3D terrestrial laser scanning, can offer high
compact space, and it only costs approximately 160 US resolution and tremendously accurate bathymetry map
dollars in total, excluding installation and labour fees. (Brasington et al., 2012; Lejot et al., 2007). However, it is
In this project, the automated system was enclosed in a expensive and resource demanding. Because our prima-
weatherproof box, which was approximately rily objective is to develop a low-cost automated flood
60 × 30 × 25 cm in size. The enclosure box was mounted monitoring system, a binocular or trinocular stereo vision
on the top of a 3-metre-high iron post which was bolted method is considered as an alternative for bathymetry
down on a concrete block near the bridge abutment at the measurement. Such method has been applied for
left bank (refer to Figure 1d). In the present study, videos measuring surface wave motions in the field (Wanek
were recorded with the resolution of 1080 pixels and Wu, 2006), which can be extended to our application.
(1920 × 1080 pixels), and the acquisition rate was set to The stereo imaging method presented here adopted a
25 frames per second. The camera has a 3.6 mm fixed- set of direct linear transformation (DLT) equations
focus lens which can image a ground surface area of (Abdel-Aziz and Karara, 1971) to represent the relation-
about 40 m (length) × 30 m (width), with an oblique angle ship between image and physical coordinates illustrated
(with respect to the zenith axis) of 65 degrees, as follows:
approximately. The camera module outputs an H264
L1 X þ L2 Y þ L3 Z þ L4
format videos which are transferred to the storage x¼ ; (1)
memory card (8G) through a CSI ribbon cable. Image L9 X þ L10 Y þ L11 Z þ 1
acquisition parameters, including the frame rate, resolu- L5 X þ L6 Y þ L7 Z þ L8

tion and compression, can be adjusted flexibly to meet L9 X þ L10 Y þ L11 Z þ 1
other specific requirements that optimize the particle
image velocimetry (PIV) interrogation analysis under where (X, Y, Z) represents the physical coordinate, and (x,
various flow conditions. y) is the image coordinate. Lj (j = 1,2,…,11) are linear
The automated LSPIV system described in the preceding transformation parameters based on the pinhole camera
texts has been applied from 10 June 2014, which has model.
captured several floods in situ. We have chosen the largest The DLT equations of a given imaging system can be
one which started at midnight on 24 July 2014 for analysis. obtained through calibration experiment which captures
Because illumination light source was not included for the an image of objects with known physical coordinates. A
current configuration, the flow measurement at nights was least-square approach is then applied to determine the
not available for the case presented here. coefficients of DLT equation uniquely. After calibrating
each camera, reconstruction of the channel bathymetry
can be achieved by matching the image coordinates of a
MEASUREMENT PROCEDURES physical object captured on the stereo image pair.
Denoting the matched image points as (x1, y1) and (x2,
Stream channel bathymetry reconstruction
y2), which correspond to the same physical point, its
The bathymetry of the stream channel is a critical physical coordinates (X, Y, Z) can then be determined by
parameter that characterizes the hydraulics of the river the DLT relations of the two cameras, i.e.

  8 9
 L9ð1Þ x1  L1ð1Þ L10ð1Þ x1  L2ð1Þ L11ð1Þ x1  L3ð1Þ ( ) >  x1 >
 < L4ð1Þ =
 L9ð1Þ y1  L5ð1Þ L10ð1Þ y1  L6ð1Þ L11ð1Þ y1  L7ð1Þ  X L8ð1Þ  y1
 L ð2Þ x  L ð2Þ
 9 2 L10ð2Þ x2  L2ð2Þ L11ð2Þ x2  L3ð2Þ  Y ¼ >
: L4ð2Þ  x2 >
;
(2)
 L ð2Þ y  L1ð2Þ L10ð2Þ y2  L6ð2Þ L11ð2Þ y2  L7ð2Þ  Z L8ð2Þ  y2
9 2 5

flow, as well as the energy dissipation and channel bed where the superscripts (1) and (2) on the DLT parameters
erosion under flooding conditions. Like many mountai- represent the two cameras. Because Equation 2 is an over-
nous streams, the channel of the Longxi River is steep, determined systems with four equations and three
rough and highly irregular. It is subject to frequent change unknowns, a least-square approach can be applied to
because of successive flood erosion and deposition. This estimate (X, Y, Z), instead of finding the unique solution.
brings significant challenges to the proposed studies, as The primary technical challenge is to develop an accurate
frequent field surveys would be needed after every major and efficient algorithm to match the image coordinates of
flood event to update the change of channel cross the same object on the image pair. Usually, a small

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
3018 Q. RAN ET AL.

subimage window frame centred on an image point is search. In this study, however, the reconstruction of the
used for image pattern matching. A searching window 3D channel geometry with the ‘virtual wave gauge’
frame with equal dimension is moved around in the method was not quite successful as it often failed to
pairing image for the best match. Similar to the PIV converge when searching the paring image patterns along
interrogation algorithm, the degree of ‘matching’ can be ‘virtual’ gauges, possibly a result of the complexity of the
quantified by cross-correlation. channel topography. A modified approach (here denoted
In this study, a commercial grade stereo camera as the ‘virtual pole’ in our study) with more reliable
(Fujifilm real 3D W3) was deployed for the 3D channel performance was developed, and its procedures are
topography reconstruction. The bathymetry mapping was summarized as the following:
conducted in June 2014, when the river flow rate was low
and the most part of the riverbed was exposed. The 1 The ‘virtual pole’ method starts by choosing specific
physical axes were defined such that the river streamwise river cross sections with Xi coordinates (i = 1, 2, …, m),
direction was X, the spanwise direction was Y and the and each cross section is divided to a number of equal
vertical direction was Z. The origin of 2D planar was segments with Yj coordinates (j = 1, 2 , …, n).
unseen in the image, whilst the zero of the Z-axis was set 2 For each segment on a cross section with the known
in the lowest ground control points (GCPs). Stereo image horizontal position (Xi, Yj), a vertical ‘virtual pole’ can
pairs were first corrected for lens radial distortion be defined in a vertical range [Zmin, Zmax], which
following the procedures described in previous studies represents the estimated searching range of the channel
(Holland et al., 1997; Wanek and Wu, 2006). Then, the bed elevation at (Xi, Yj).
DLT parameters for each of the stereo pair were 3 This ‘virtual pole’ is then projected on the stereo image
determined through 18 equations comprised nine GCPs. pairs following Equation 1. For every selected point on
All GCPs were chosen to be on large rocks or on the the ‘pole’, the image coordinates are denoted as (x(1),
concrete channel sidewalls, spreading over the field of y(1)) and (x(2), y(2)) for the two images respectively.
view with varying elevations (i.e. the Z coordinate). 4 Particle image velocimetry interrogations are then
These points were marked with bright yellow spray paint, applied to the subimages centred on (x(1), y(1)) and
the physical coordinates of which were measured using a (x(2), y(2)). Because the matching subimages of the
land surveying total station (South NTS-302, refer to channel bed at (Xi, Yj) must be centred at a certain point
Figure 2). along both ‘pole’ images, the point pair with the
Bechle and Wu (2011) applied an Eulerian ‘virtual minimum subimage displacement is assumed to be the
wave gauge’ method for wave surface reconstruction. matching points, and its Z-value is considered as the
Such method aims to find the vertical coordinate (Z) of estimated channel bed elevation at location (Xi, Yj).
the water surface at any given horizontal location (X, Y).
The search for matched image point is conducted along a Specifically for the experiment presented here, 20 cross
virtually place wave gauge at (X, Y). They also applied an sections, with an equal distance of 1.2 m between two
iterative, weighted bisection method to accelerate the consecutive ones, were selected. Each cross section was

Figure 2. Selected ground control points for stereo-imaging calibration were marked by spray paint in the area of interest. X and Y denote the streamwise
and spanwise directions of the main flow, and Z points upward. A total station (South NTS-302) with a high resolution of 3 mm was used for determining
the physical coordinates of the selected ground control points

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
LSPIV FOR MOUNTAINOUS FLOOD MONITORING 3019

divided into 220 segments (X from 3 to 25 m) with an


interval of 0.1 m. For any given horizontal position (Xi,
Yj), we calculated the image displacements on 600 equally
spaced vertical points (from 5 to 7 m) along the two
‘virtual poles’ in order to locate the best-matched point.
Figure 3a shows the example of the reconstructed channel
bed profile at the cross section (X = 8.4 m) and the image
(left camera vision) of a ‘virtual pole’ at (X, Y) = (8.4,
15 m). Figure 3b presents the profile of subimage
Figure 4. Bathymetry profiles of three specified cross sections (X = 9.6,
displacement computed along this ‘virtual pole’, and the 12.0 and 14.4 m) with the highest and lowest water stages during the flash-
minimum displacement is located at Z = 0.7 m, which is flood event presented in this paper
then considered as the bed elevation at (X, Y) = (8.4,
15 m). The ‘virtual pole’ method is validated by
comparing the reconstructed bed elevation at the nine narrowing cross sections downstream. The riverbed slope
GCPs where the world coordinates are known. The root of the field view is quite steep and irregular, as the river
mean square error (RMSE) is 2 cm, and the maximum centreline drops more than 2 m in a short reach of 12 m.
error is 4 cm. This suggests that the water surface will likely have a
Three representative cross-sectional profiles denoted as strong downslope gradient, and it should not be
B, D and F (XB = 9.6 m, XD = 12.0 m and XF = 14.4 m) are considered as horizontal in the following LSPIV analysis.
plotted in Figure 4, which shows a stream reach with
Water surface-level determination and image
orthorectification
Ideally, a stereo imaging method that can measure the
varying water surface profile is desirable for long-term
unattended applications. However, a stereo imaging
system built with Raspberry Pi cameras is still in
progress. The challenge is to synchronize two or multiple
Raspberry Pi cameras, which is the key for stereo imaging
on the moving water surface. In this project, we have
designed an alternative method to obtain the water surface
stage by projecting bathymetry profiles on to flow
images, as described in the succeeding texts.
The transformative DLT parameters of the Raspberry
Pi camera are first obtained through image calibration
with the same GCPs used in the bathymetry mapping.
Although recorded flow videos showed that strong water
surface fluctuation existed during high flow events, we
were only interested in the averaged water surface profile
and discharge for the duration of each video clip which
lasts for 5 s. The synthetic image of the ‘mean’ flow
surface was then created by arithmetically averaging over
all 125 frames of instantaneous flow images. The
bathymetry lines of stream cross sections were projected
on to the ‘mean’ high flow image (refer to Figure 5a). In
order to estimate the water surface level, we have
assumed that the surface is a function of the streamwise
direction only, whilst it is approximately a constant value
along the spanwise direction. In this study, the nominal
streamwise direction is denoted as X in the world
Figure 3. Principle of 3D bathymetry reconstruction method. (a) Scattered
points (black circles) indicate the ‘virtual pole’ at (X, Y) = (8.4, 15 m) with coordinate, although the stream channel meanders slightly
Z varying between [5, 7 m]. (b) Calculating image distances of virtual in the observed reach. Based on this assumption, the
pole points between stereo pair images, the actual Z value can be water surface level measured on the river/bank borderline
determined (0.7 m) when the distance reaches minimum (0.18 pixel). After
repeating this process, cross-sectional bathymetry could be achieved [as represents the averaged surface level over the cross
plotted by bold black line in (a)] section of the same X coordinate. Theoretically, the

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
3020 Q. RAN ET AL.

Figure 5. Water surface slope estimation method. (a) Reconstructed bathymetric lines are projected on the flow image, intersections of these lines and the
water/bank borderlines are used to estimate the water surface elevation at every cross section; blue lines represent the three cross sections B, D and F
shown in Figure 4; (b) profiles of the water surface and the streambed centreline. Scattered square and circle points are measurements; solid and dashed
lines are linear regression of the data, accordingly

river/bank borderline could be determined from the levels along the streamwise direction (circles, and the
synthetic image through image segmentation and edge estimated slope is 12.63 degrees) and the representative
detection method; hence, the intercept between the channel centreline profile (squares, Y = 18 m, with an
borderline and the bed cross section can be calculated estimated slope of 13.39 degrees). It appears that the
from the image. In this study, however, the image slope of the surface profile did not vary significantly (9.54
coordinates of these intercept points were selected degrees to 13.08 degrees) during the process, and it was
manually, e.g. white circles shown in Figure 5a. It should approximately parallel to the bottom although some
also be noted that most of the selected points were on the irregularities existed.
right bank, as part of the borderline on the left bank may For the field application of LSPIV, the camera usually
not be available because of the image occlusion and bank captures images at an oblique angle with perspective
collapse. The Y and Z coordinates of these points can then distortion that means image pixels have nonuniform
be determined with their image coordinates (x, y) and the physical resolutions of the water surface. Therefore, a
known X position following the calibrated DLT relation geometrical transformation (so-called ‘orthorectification’)
as Equation 1. is required to correct the projective distortion. Usually,
To simplify the analysis, we have assumed that the the transformation assumes that the water surface is a
water surface was a two-dimensional plane, and a linear two-dimensional plane with the elevation considered as a
regression was applied to estimate the water surface known constant, which is a reasonable approximation for
profile. Figure 5b shows the measured water surface natural river reach with very mild surface slope. However,

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
LSPIV FOR MOUNTAINOUS FLOOD MONITORING 3021

in the case presented here, the steep slope of the water rise if the water surface slope is ignored in mountainous
surface should be considered as a variable and should be streams, as a result of the nonuniform spatial resolution
taken into consideration for the orthorectification proce- and inaccurate flow direction. As shown in Figure 6a, the
dure. The changes of water stage determined in the boundary of the river surface was irregular, and the river
preceding texts are critical information not only for the width varied frequently. The unstructured triangular mesh
orthorectification, but also for the calculated wetted area with a nominal size of 20 image pixels (or 40 cm
in discharge estimation. physically) were generated with PDE toolbox functions
Figure 6a shows an example of the orthorectified water in Matlab, which conforms to the complex boundary
surface image after projection on the sloped surface, and better.
Figure 6b is the transformed image by projection onto a
horizontal plane. The notable difference between the two Surface velocity calculation
transformed images indicated that significant errors might
Outline of LSPIV algorithm. The LSPIV, stemming
from the PIV methodology, is a promising technique used
in field measurement. Under favourable conditions
including proper light illumination and adequate surface
‘natural tracers’, it can provide instantaneous or averaged
velocity vector fields on a large-scale free surface. Kim
(2006) proposed that multiple factors might affect the
accuracy of LSPIV results via an uncertainty analysis,
including illumination, tracers, hardware, algorithm, etc.
More details of this method have been discussed in a
number of previous publications (Bradley et al., 2002;
Creutin et al., 2003; Dramais et al., 2011; Fujita et al.,
1998; Hauet et al., 2008; Kim et al., 2008; Le Coz et al.,
2010; Jodeau et al., 2008; Muste et al., 2008, 2014).
Natural patterns floating at the free surface are much
more preferred in this application, whilst spreading
artificial ‘tracers’ is not a good option for continuous
unattended monitoring. Fujita and Kunita (2011) con-
cluded that river water surfaces exhibit three primary
natural trackable features in flood flow conditions,
including ripples, floating objects and water colour
differences. As a steep mountainous stream, the Longxi
River is rich of all surface patterns listed in the preceding
texts under a variety of flow conditions. Bubbles and
foams are the dominant surface features under high flow
conditions. Large surface ‘eddies’ are also clearly visible
in the format of ‘ripples’. However, we also observed
some unfavourable features such as water jumps at high
flows. It is not clear yet how this can affect the actual flow
speed measurement.
Cross-correlation-based evaluation is extensively used
for LSPIV interrogation analysis. In the study, we
adopted the minimum quadratic difference (MQD)
algorithm which has been approved to be more accurate
for tracking of large and uneven image patterns (Gui and
Merzkirch, 1996, 2000; Suh, 2003). Instead of finding the
correlation peak, the minimum of the grey-value
difference squared is searched to estimate the average
Figure 6. Comparison of orthorectification results with or without
accounting of water surface slope. (a) Orthorectified surface image pattern displacement:
according to the estimated water surface profile, which is assumed to be a
plane with a constant slope; (b) orthorectified surface image assuming a 1 M 1 N 1
Dðm; nÞ ¼ ∑ ∑ ðg ði; jÞ  g2 ði þ m; j þ nÞÞ2 (3)
horizontal surface plane. PIV interrogation is performed on all nodal
points of a triangular mesh, which conforms to the complex topography
M N i¼0 j¼0 1

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
3022 Q. RAN ET AL.

where g1 and g2 are the distributions of pixel grey values


in the interrogation subwindow frame, and M and N are
the sizes of the interrogation frame. As stated in Gui and
Merzkirch (1996), the MQD method shows its superiority
to cross-correlation as a result of an extra tracking term in
the expanded form of Equation 3 which removes the error
as a result of nonuniform-illuminated conditions.
Images recorded in laboratory PIV experiments are
normally featured with disperse ‘bright’ particles illumi-
nated by a laser sheet on a dark or uniform background,
the error, as a result of the nonuniform illumination, is
usually very small, and the difference between MQD and
cross-correlation is negligible. However, ‘tracers’ in
LSPIV images are more likely continuous patterns. In
the present study, we found that in some surface
locations, particularly at low flow conditions, cross-
correlation often failed as a result of a very low peak-to-
background ratio or reported zero displacement, whilst
the MQD could still produce a good result. We also
noticed that at high flows with rich surface patterns, the
average velocities from MQD and cross-correlation were
nearly identical.
The surface velocity vectors were calculated with each
LSPIV interrogation subwindow frame centred on every
nodal point of these triangles. Fast Fourier Transform is
applied in MQD to accelerate the interrogation process,
and the subpixel precision is achieved through a Fourier
space window shifting technique (Liao and Cowen,
2005). A sample result which shows the vector map of
surface velocities averaged in 5 s is illustrated in Figure 7
a, and the corresponding velocity magnitude is presented
as contour maps as shown in Figure 7b. The surface
velocity magnitude mostly ranges from 1 to 5 m/s. The
velocity distribution on a structured mesh or any given
cross section can then be obtained through interpolation Figure 7. Typical results of the LSPIV system. (a) Mean surface velocity
afterwards. This sample result is obtained during a peak field averaged over a five-second video; streamwise velocity distribution
flow period. The streamwise velocity distributions on on three cross sections B, D and F as shown in Figure 4; (b) contour maps
of surface velocity magnitude
three cross sections (B, D and F, as shown in Figure 4) are
also presented in Figure 7a, which indicate that the
velocity distribution is highly nonuniform across the current LSPIV framework. Improvements can be made by
water surface. Flow reversal can be observed on all cross using multiple synchronized video cameras to reconstruct
sections, which suggests that strong horizontal shear and the instantaneous 3D water surface, which is subject to
swirling exist as a result of the complex topography and our future research.
torrential flow condition. Some of the observed surface
flow patterns can be attributed to the 3D surface Validation of LSPIV algorithm. In this study, the
deformation as a result of local hydraulic jumps and LSPIV results have been validated by comparing with
standing waves. The complex 3D surface structure at high the data acquired by a handheld impeller flow metre.
flow conditions can cause velocity and discharge Because of safety concerns, the test was conducted under
measurement errors, as the surface is forced to map to a a low flow condition. Six validation points were selected
two-dimensional plane. For example, positive or negative where impeller flow metre was positioned at about 5 cm
vertical velocity component as a result of the wave effect below the water surface with the holding rod inserted
can be artificially projected as horizontal velocity into the river to keep the flow metre stationary (refer to
components (Dramais et al., 2011). Uncertainties, as a Figure 8). Instantaneous flow speed can be read out from
result of these errors, are not able to be resolved in the a LCD screen of the flow metre. The data displayed on

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
LSPIV FOR MOUNTAINOUS FLOOD MONITORING 3023

should also be noted that the impeller metre, as a


conventional stream flow measurement tool, also has
some uncertainties, which could affect the result of the
comparison. It is expected that the accuracy of LSPIV
could be higher under high flow conditions, with
abundant surface patterns (primarily foams) as natural
tracers.

Discharge computation
To date, a reliable method to estimate the discharge of
flash flood has yet to be developed, largely a result of the
lack of accurate field measurements (Fenton and Keller,
2001; Moyeed and Clarke, 2005). In conjunction with the
average surface velocity measured from LSPIV and water
Figure 8. Average velocity field calculated by LSPIV under base flow depth profile determined by reconstructed bathymetry
condition, and six validation points selected for comparison with an lines, we are able to calculate the discharge via the
impeller flow metre standard velocity–area method (Bradley et al., 2002;
Turnipseed and Sauer, 2010). In this method, a stream
cross section is divided into segments; the total discharge
the screen fluctuated because of turbulence. Because the is then calculated by integrating the flow rate of all
flow metre does not have a data logger to record the segments, which is estimated as the product of the area of
measurements, the measurement results were represented each subsection and the respective depth-averaged
by manually recording the maximum and minimum velocity along the direction perpendicular to the cross
readings during 1 min. The Raspberry Pi camera was section, i.e.
programmed to record the video continuously through
the entire period of the flow metre measurement. LSPIV n
calculations were applied immediately before and after Q ¼ ∑ ai vi (4)
i¼1
metre deployment at every validation point, and the
mean velocity averaged over the two intervals (15 s) was where Q is the total discharge across a designated cross
selected for comparison with the average of the flow section, ai is the area of subsection i, and vi is the depth-
metre data. The comparison of the results was shown in averaged velocity of each subsection.
Table I. Relative error was calculated as the ratio of the Generally, surface velocity can be an estimated
difference to the flow metre results (average of the surrogate to the depth-averaged velocity with a modifying
maximum and minimum reading). Overall, good agree- coefficient ranging from 0.76 to 0.90, depending on the
ments were found between the two devices with errors specific sites and flow conditions. Muste et al. (2008)
generally less than 8% except at validation Point 3. This pointed out that the velocity coefficient should be
was possibly a result of the lack of surface patterns at determined carefully as it may cause a multiplicative,
that point, where the water surface was relatively systematic discharge error. Theoretically, based on the
smooth, and the river bottom was clearly visible. It assumption of a logarithmic velocity profile, a default
value of 0.85 is widely accepted for river flow studies
(Costa et al., 2000, 2006; Creutin et al., 2003; Rantz,
1982). Uncertainties always exist for the case of steep
Table I. LSPIV and impeller flow metre results at six validation
points mountain flows with complex and dynamic bedforms and
larger bottom roughness. Without a comprehensive study
Point no. Impeller flow metre a
LSPIV b
Relative error c
on this issue with a wide range of field experiments, it
remains to be difficult to assess the variation of the
1 1.3–1.5 1.3458 3.87% velocity coefficient under varying hydraulic conditions,
2 1.0–1.4 1.2357 2.98%
3 0.9–1.1 1.1645 16.45% especially in the complex flow conditions (Gunawan
4 1.0–1.3 1.0618 7.67% et al., 2012). According to Jarrett (1991), for steep rivers,
5 0.9–1.0 0.9080 4.42% a logarithmic velocity profile can hardly develop because
6 1.3–1.4 1.2843 4.87% of a drag from rough riverbed and higher velocity on the
water surface. In this study, a more reasonable coefficient
a
‘Real’ velocity values: Vi = (Vmax + Vmin) / 2, m/s
b
LSPIV results: Vl, m/s of 0.80 in the lower range is chosen uniformly for all
c
Relative error = |(Vl  Vi) / Vi| * 100% cases to estimate the stream discharge.

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
3024 Q. RAN ET AL.

RESULTS indicates that it is of great reliability to determine the


water stage via reconstructed bathymetry lines.
Validation of water surface level measurement
An estimation of the water surface level is critical to Flood hydrograph
determine the river stage variation during the flood. The
measurement with the reconstructed bathymetry method A flood flow that occurred from 8 am on 24 July 2014
is validated by comparing with water level marks which to 2 pm on 25 July 2014 is selected for hydrograph
were spray painted with equal intervals (15 cm) on an analysis. As explained earlier, the data of flows at nights
almost vertical rock in the field of view, as shown in were missing, and our measurements were not able to
Figure 9a. The vertical elevations of these marks were reveal the flood rising process which occurred during the
checked with the total station measurement. After night prior to sunrise on July 24. Six cross sections (from
identifying the intersecting line of water surface and A to F) which occupied the central part of the flow image
rock, the depth could be obtained by accounting the are chosen to calculate the discharge during this process.
number of pixels submerged. On the other side, the Figure 10a shows the time series of the mean, maximum
reconstructed bathymetry line near the marked rock is and minimum of discharge values calculated from the six
selected to obtain the water level following the method cross sections. The recorded hydrograph represents the
illustrated in the ‘Water surface-level determination and recessing part of the flood event. The highest value of
image orthorectification’ section. The comparison discharge recorded was approximately 27 m3/s; it then
between the marked (h-mark) and the reconstructed (h- decreased back to the basic flow of about 3.5 m3/s. The
calculated) during the flood event recorded in this study is peak discharge did not sustain for a long time, as the
presented in Figure 9b, which shows a good one-on-one rainstorm lasted for several hours only.
agreement with a RMSE of 0.037 m. The validation The discharge was calculated over six cross sections
which allowed reduction of the site-specific errors. The
observed reproducibility of the method was actually
excellent, whilst a few measurements still showed
variations. The variations may arise from errors of
velocity and bathymetry measurement, the uncertainty
of the velocity coefficient, as well as the effects of
turbulence and short-term unsteadiness of the stream
flow. The mean velocity is an average of flow in 5 s,
which is indeed too short compared with the timescale of
the large ‘eddies’ in the stream, and the timescale of water
surface fluctuation that causes the short-term transient
flow. Increasing the sampling time may help to reduce the
variability of measured discharge at different cross
sections, but will not reduce the uncertainty associated
with the velocity–area method. In the presented case, the
‘error’ of discharge measurement is represented by the
coefficient of variation, which is the ratio of standard
deviation to the mean of discharges calculated at the six
cross sections. The relative error of each case is also
plotted in Figure 10a. In general, errors are acceptable
with an average of about 14%, and the maximum is about
23%. The time series of the deepest cross-sectional water
depth is shown in Figure 10b, indicating that it changed
from the peak value of 1.8 m to a low level of 0.8 m in a
period of 30 h.
The Froude number (Fr) of the stream flow at a given
cross section is defined as
u
Fr ¼ pffiffiffiffiffi (5)
gh
Figure 9. Water surface level measured through imaging compared with
marked levels. (a) Water level marks on a large rock; (b) measured water
where u is the cross-sectional average velocity, and h ¼
level versus marked ones A=T is the hydraulic depth. A is the area of the cross

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
LSPIV FOR MOUNTAINOUS FLOOD MONITORING 3025

Figure 10. Hydraulic processes of the flood occurred on 24 July 2014. (a) Mean discharge averaged over six cross sections with error bars representing
the range. Crosses represent the data dispersion defined by coefficient of variation; (b) time series of averaged cross-sectional water depth and Fr number

section, and T is the top width of the water surface. even under flood conditions. This suggests that the
Despite the steep slope, the Froude numbers at all the six friction slope Sf can be well approximated by the
cross sections are less than one during the entire water surface slope following the momentum balance
observation period (refer to Figure 10b), suggesting a analysis. The major challenge to apply the Manning’s
subcritical condition for the entire flood event. However, equation in natural streams is the uncertainty associ-
it is noticeable that small hydraulic jumps occurred ated with the estimation of the Manning’s n roughness
sporadically at some areas of the water surface, which coefficient, the value of which depends on a number
suggests a local supercritical flow condition even though of factors, e.g. flow depth, turbulence, bed forms,
the overall flow is subcritical averaged through the cross vegetation, sediment and bed load, cross-sectional
section. These observations agree with recent findings shape, geomorphological processes, floodplain–main
that the mean flow averaged over long reaches remains channel interaction, unsteady flow, etc., as suggested
subcritical for most natural high-gradient channels (Grant, by Trieste and Jarrett (1987). Moreover, the Man-
1997; Lumbroso and Gaume, 2012; Tinkler, 1997). ning’s roughness coefficient in mountainous streams
may not be a constant, depending on the varying
Manning’s n roughness coefficient discharge, friction slope and bed material which is
Manning’s equation has been well accepted as the subject to frequent erosion and sedimentation.
governing law for open channel flow discharge estima- The discharges at cross section B are plotted against
tion. The general format of the Manning’s equation is as AR2/3Sf1/2 calculated from the known parameters, e.g.
follows, cross-sectional area, hydraulic radius and water surface
slope, as they varied during the recorded event, where we
1 have assumed that the friction slope can be approximated
Q ¼ AR2=3 S f 1=2 (6)
n by the water surface slope (refer to Figure 11). A good
linear relation is found, and the least-square regression
where Q is the discharge (m3/s); R represents the suggests that Q = 5.56AR2/3Sf1/2 with the coefficient of
hydraulic radius (m), which is the ratio of the cross- variation R2 = 91%. The linear regression showed that the
sectional area A (m2) over the wetted perimeter P (m); Sf Manning’s roughness coefficient can be determined as
is the friction slope, and under a uniform flow assumption n = 0.18 for cross section B. Good linear relations were
that the downslope force of gravity is balanced entirely by found for the other five cross sections as well, and all the
bottom friction, it is equal to the channel bottom slope; n values were in the interval of [0.14, 0.23]. The
and n is the Manning’s roughness coefficient. estimated n values agreed well with that proposed by
In practice, it is relatively straightforward to determine Zhang et al. (2011) (i.e. n = 0.1~0.3), which was obtained
geometric parameters, i.e. A and R. In addition, the through the sampling of bed materials at the same
Froude number for most river and stream flows is small, experiment site and the calculation using the model

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
3026 Q. RAN ET AL.

valuable tool for future research that investigate the


application and uncertainty of the Manning’s equation for
extreme flood flows, where data are rare and critically
needed.

Rating curve
The flow rate at a gauged stream is traditionally
estimated through a stage–discharge relationship (i.e.
rating curve), which can be established by making
repeated measurements of stage and discharge over a
certain range of flow conditions (Kean and Smith, 2010;
Rantz, 1982; Schmidt, 2002). In gauged catchments,
rating curve usually has a higher accuracy under normal
conditions, but it must be extrapolated for high-flow
Figure 11. Measured discharge, Q, plotted against the Manning’s equation
scaling, AR2/3S1/2. The slope of the linear regression is 5.56, which
conditions. The accuracy of the extrapolation is always
suggests a Manning’s roughness of n = 0.18 questionable despite the possible improvement through
postevent data analysis and numerical simulation (Lang
et al., 2010).
presented by Rickenmann and Recking (2011) that
In the present study, discharge time series at cross
estimates the flow resistance for gravel-bed rivers.
section B from the recorded flood event is plotted against
Despite the good agreement, we noticed that the n
the water stage in Figure 12. A power function is selected
roughness is extremely high, which is possibly a result of
to fit the measured discharge-stage relation, i.e.
the strong turbulence, energy dissipation and form drags
of partially submerged large rocks. It should be noted that Q ¼ aðh  bÞc (7)
a flow like this is far from uniform and channel-
controlled. Successions of waterfall-like features and where h is the measured stage (m), and b is the stage
local hydraulic jumps can be visually observed, which corresponding to zero flow rate (m). The three coeffi-
suggests that the observed stream reach is more likely cients a, b and c are determined based on a nonlinear least
cross section-controlled. It is questionable if the Man- square fit function in Matlab, and they are a = 16.79,
ning’s equation is applicable for such a complex flow. b = 0.50 (m) and c = 1.64. These best fitted parameters are
The observed linear relation between Q and AR2/3Sf1/2 reasonable, as the water stage at a very low flow rate
may not be extrapolated because the effective Manning’s (Q = 1.0 m3/s) was about 0.5 m (c.f., Figure 10 on 23 July
roughness may change as the stage and discharge increase 2014), and the exponent c = 1.64 is close to the theoretical
further (Lumbroso and Gaume, 2012). Nevertheless, the value of 5/3 for a wide rectangular channel-type of
results presented here demonstrate that LSPIV can be a hydraulic control (Le Coz et al., 2014).

DISCUSSIONS AND CONCLUSIONS


Whilst several recent studies have shown LSPIV
applications in mountainous streams (Le Boursicaud
et al., 2016; Le Coz et al., 2010), this paper demonstrates
that an unattended LSPIV system can be applied for long-
term measurements of discharge and other hydrological
processes in ungauged mountainous catchments under a
variety of flow conditions. The system is built with
flexible, robust and off-the-shelf electronic components,
which can grab continuous images/videos automatically.
The computer board (Raspberry Pi) can be easily
programmed with simple commands to control the
camera module to operate at desired intervals. The
overall power consumption is less than 3W. In the
Figure 12. Discharge-stage relation with data obtained between 8 am–8
pm, 24 July 2014 and 8 am–2 pm, 25 July 2014 (dots). Data are fitted with presented application, a 20 Ah battery allows the system
a power-function rating curve to operate for more than a month with a two-minute duty

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
LSPIV FOR MOUNTAINOUS FLOOD MONITORING 3027

cycle and 17 intervals per day. Image processing can be preliminary study. Many storm and flow events started
applied to extract water surface velocity distribution with at nights, including the event presented in this study. As a
the LSPIV algorithm. In addition, a modified stereo- result, the rising process of the flood was not recorded,
imaging approach, which is referred to as the ‘virtual and about one third of the data of the process were
pole’ method, has been presented in this paper. With this missing. The problem can be solved by applying a
method, the stream channel bathymetry can be remotely continuous light source or pulsed flashlight to illuminate
reconstructed without manual surveying. The reconstruct- the water surface at nights (Chaves, 2012). This will,
ed stream cross sections can also be ‘mapped’ onto flow however, significantly increase the power consumption.
images to estimate the elevation and slope of the dynamic A solar panel or other source of energy will be needed to
water surface. Variations of stage and surface slope recharge the battery for long-term deployment. Using an
during a flood flow process are important parameters for infrared LED array as the light source and an infrared
surface image orthorectification, and their errors can version of the Raspberry Pi camera could be another
significantly affect the accuracy of surface velocity and solution for night imaging. The infrared night vision has
discharge measurements, as illustrated through the LSPIV the advantage of being more power efficient and having
sensitivity analysis by Le Boursicaud et al. (2016). fewer disturbances to the wildlife.
Combining the surface velocity distribution and the In the preliminary study, direct stereo imaging is only
channel bathymetry reconstruction, we can estimate the applied for bathymetry reconstruction. The reconstruction
discharge following the standard velocity–area method. of the water surface was indirectly derived by identifying
As a proof-of-concept, the automated LSPIV system is the water-bank interface, with the assumption that surface
deployed in a mountainous stream at the Longchi elevation is constant along the cross-stream direction. The
Catchment from June 2014. A flood event that occurred wave deformation, hardly being accounted in quantity
on 24 July 2014 is selected for analysis. The surface currently, would affect orthorectification accuracy and the
velocity measured with the LSPIV system is validated surface velocity direction which will induce errors in
with an impeller flow metre, and good agreement is found discharge estimation. These limitations can be removed
with acceptable relative errors (lower than 10%). Water by applying two synchronized LSPIV cameras with
surface level reconstructed from stereo-imaging analysis flexible baseline configuration depending on the viewing
is also validated with the marked water depth, and the distance. The geometry of the rapidly moving surface can
result indicates that the accuracy of the imaging method is be directly resolved with the same ‘virtual pole’ method.
less than 4 cm (in terms of the RMSE). The time series of However, the low-cost Raspberry Pi camera employed in
hydraulic processes is calculated at six cross sections this study does not support the hardware external
focussing on the central part of the channel. The triggering; thus, it might not be ideal for the stereo
discharge for this flood event ranged from about 3.5 to imaging of the moving water surface. Although some
almost 27 m3/s, which covers a good portion of the rating degree of synchronization can be achieved through
curve interval. Although a stationary rating curve should software triggering between two computer boards, the
not be expected for the Longxi River because of the temporal accuracy of this method is not documented, and
varying bathymetry induced by frequent erosion and some calibration tests are needed to explore this
deposition, the rating curve obtained from this one event possibility.
does show a good power-law relation. The Manning’s n The most significant limitation lies in the discharge
roughness coefficients can be estimated to be between estimation. Firstly, the geometry and area of the stream
0.14 and 0.23, which agrees well with previously reported cross section can only be measured through imaging
values at the same site. before and after a flood event, when most parts of the
The experimental results suggest that this system can channel bed are exposed. It is, however, not possible to
serve as an applicable and reliable method for flash flood detect the change of bathymetry during the event, as the
monitoring in the mountainous catchment. It can provide bed is subject to continuous erosion and deposition. The
a large set of continuous flow data for an ungauged site unknown variation of the cross section can introduce
with extremely low cost. Hydrological data obtained can errors for discharge estimation if we assume that the
be applied for flood analysis and future model calibration bathymetry is stationary. Secondly, the velocity–area
and prediction. approach depends on an assumed uniform coefficient that
Despite the advantages presented here, the current relates the surface velocity to the depth-averaged velocity.
configuration of the LSPIV system has some limitations. It is almost certain that this coefficient is nonuniform, and
In addition, some uncertainties for flow measurements it can change with water stages. Although some
may not be solved with the LSPIV approach. theoretical and experimental analyses on this coefficient
During night time, imaging technique requires light have been conducted in recent studies (Le Coz et al.,
illumination, which has not been included in the 2010), further investigation and experiments are needed

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
3028 Q. RAN ET AL.

in our future research, given the complexity of the methods. Water Resources Research 42(7): DOI:10.1029/
2005WR004430
bathymetry of this steep mountainous stream. For the Creutin JD, Muste M, Bradley AA, Kim SC, Kruger A. 2003. River
presented study, it remains difficult to evaluate and gauging using PIV techniques: a proof of concept experiment on the
improve the accuracy of discharge measurements that are Iowa River. Journal of Hydrology 277(3): 182–194.
Dramais G, Le Coz J, Camenen B, Hauet A. 2011. Advantages of a mobile
based on surface velocity distribution. LSPIV method for measuring flood discharges and improving stage–
discharge curves. Journal of Hydro-Environment Research 5(4):
301–312.
Fenton JD, Keller RJ. 2001. The Calculation of Streamflow from
ACKNOWLEDGEMENTS Measurements of Stage. Technical Report 01/6, CRC for Catchment
Hydrology, Melbourne, Australia.
The research task was financially supported by the Fujita I, Muste M, Kruger A. 1998. Large-scale particle image velocimetry
National Natural Science Foundation of China (grant for flow analysis in hydraulic engineering applications. Journal of
nos. 51328901 and 51379184) and the China Scholarship Hydraulic Research 36(3): 397–414.
Fujita I, Kunita Y. 2011. Application of aerial LSPIV to the 2002 flood of
Council (no. 201406320048). The authors would like to the Yodo River using a helicopter mounted high density video camera.
express their sincere gratitude to their colleagues who Journal of Hydro-Environment Research 5(4): 323–331.
made great assistances to equipment developments and Getirana ACV, Peters-Lidard C. 2013. Estimating water discharge from
large radar altimetry datasets. Hydrology and Earth System Sciences 17
field measurements in this project, especially Yan Liu, (3): 923–933.
Liguo Zhang, Dawei Guo, Chenge An, Tianheng Feng Gordon RL. 1989. Acoustic measurement of river discharge. Journal of
and Yuxin Ma, and the reviewers who proposed valuable Hydraulic Engineering 115(7): 925–936.
Grant GE. 1997. Critical flow constrains flow hydraulics in mobile-bed
comments. The collaborations of Tsinghua University streams: a new hypothesis. Water Resources Research 33(2): 349–358.
(China) and residents of Longchi county are also Gunawan B, Sun X, Sterling M, Shiono K, Tsubaki R, Rameshwaran P,
appreciated. Knight DW, Chandler JH, Tang X, Fujita I. 2012. The application of
LS-PIV to a small irregular river for inbank and overbank flows. Flow
Measurement and Instrumentation 24: 1–12.
Gui L, Merzkirch W. 1996. A method of tracking ensembles of particle
REFERENCES images. Experiments in Fluids 21(6): 465–468.
Gui L, Merzkirch W. 2000. A comparative study of the MQD method and
Abdel-Aziz YI, Karara HM. 1971. Direct linear transformation from several correlation-based PIV evaluation algorithms. Experiments in
comparator coordinates into object space coordinates in close range Fluids 28(1): 36–44.
photogrammetry. In Proceedings of the ASP/UI Symposium. Close- Hapuarachchi HAP, Wang QJ, Pagano TC. 2011. A review of advances in
Range Photogrammetry: Urbana, IL; 1–18. flash flood forecasting. Hydrological Processes 25(18): 2771–2784.
Ballesteros Cánovas JA, Eguibar M, Bodoque JM, Díez-Herrero A, Stoffel Hauet A, Kruger A, Krajewski WF, Bradley A, Muste M, Creutin JD,
M, Gutiérrez-Pérez I. 2011. Estimating flash flood discharge in an Wilson M. 2008. Experimental system for real-time discharge
ungauged mountain catchment with 2D hydraulic models and estimation using an image-based method. Journal of Hydrologic
dendrogeomorphic palaeostage indicators. Hydrological Processes 25 Engineering 13(2): 105–110.
(6): 970–979. Holland TK, Holman RA, Lippmann TC, Stanley J, Plant N. 1997.
Bechle AJ, Wu CH. 2011. Virtual wave gauges based upon stereo imaging Practical use of video imagery in nearshore oceanographic field studies.
for measuring surface wave characteristics. Coastal Engineering 58(4): IEEE Journal of Oceanic Engineering 22(1): 81–92.
305–316. Jarrett RD. 1991. Wading measurements of vertical velocity profiles.
Birkinshaw SJ, Moore P, Kilsby CG, O’Donnell GM, Hardy AJ, Berry Geomorphology 4(3): 243–247.
PAM. 2014. Daily discharge estimation at ungauged river sites using Jodeau M, Hauet A, Paquier A, Le Coz J, Dramais G. 2008. Application
remote sensing. Hydrological Processes 28(3): 1043–1054. and evaluation of LS-PIV technique for the monitoring of river surface
Borga M, Boscolo P, Zanon F, Sangati M. 2007. Hydrometeorological velocities in high flow conditions. Flow Measurement and Instrumen-
analysis of the 29 August 2003 flash flood in the Eastern Italian Alps. tation 19(2): 117–127.
Journal of Hydrometeorology 8(5): 1049–1067. Kean JW, Smith JD. 2010. Calculation of stage–discharge relations for
Borga M, Gaume E, Creutin JD, Marchi L. 2008. Surveying flash floods: gravel bedded channels. Journal of Geophysical Research, Earth
gauging the ungauged extremes. Hydrological Processes 22(18): 3883. Surface (2003–2012) 115(F3). DOI: 10.1029/2009JF001398
Bradley AA, Kruger A, Meselhe EA, Muste MV. 2002. Flow Kim Y. 2006. Uncertainty Analysis for Non-Intrusive Measurement of
measurement in streams using video imagery. Water Resources River Discharge Using Image Velocimetry. Ph.D. Thesis, Univ. of Iowa,
Research 38(12): 51–1. Iowa City, USA.
Brasington J, Vericat D, Rychkov I. 2012. Modeling river bed Kim Y, Muste M, Hauet A, Krajewski WF, Kruger A, Bradley A. 2008.
morphology, roughness, and surface sedimentology using high Stream discharge using mobile large-scale particle image velocimetry: a
resolution terrestrial laser scanning. Water Resources Research 48 proof of concept. Water Resources Research 44(9): W09502.
(11): W11519. DOI:10.1029/2012WR012223 DOI:10.1029/2006WR005441
Chaves, H. 2012. A weather independent illumination for field LSPIV. Lang M, Pobanz K, Renard B, Renouf E, Sauquet E. 2010. Extrapolation
16th Int Symp on Applications of Laser Techniques to Fluid Mechanics, of rating curves by hydraulic modelling, with application to flood
Lisbon, Portugal, 09–12 July, 2012. frequency analysis. Hydrological Sciences Journal–Journal des Sci-
Chu S, Yu B, Li L, Ma Y, Wu Y, Qi X. 2011. Forming mechanism and ences Hydrologiques 55(6): 883–898.
characteristics of debris flow happened on August 13, 2010 in Jianping Le Coz J, Hauet A, Pierrefeu G, Dramais G, Camenen B. 2010.
Gully. Soil and Water Conservation in China 8: 52–54. Performance of image-based velocimetry (LSPIV) applied to flash-flood
Costa JE, Spicer KR, Cheng RT, Haeni FP, Melcher NB, Thurman EM, discharge measurements in Mediterranean rivers. Journal of Hydrology
Keller WC. 2000. Measuring stream discharge by non-contact methods: 394(1): 42–52.
a proof-of-concept experiment. Geophysical Research Letters 27(4): Le Coz J, Renard B, Bonnifait L, Branger F, Le Boursicaud R. 2014.
553–556. Combining hydraulic knowledge and uncertain gaugings in the
Costa JE, Cheng RT, Haeni FP, Melcher N, Spicer KR, Hayes E, Barrick estimation of hydrometric rating curves: a Bayesian approach. Journal
D. 2006. Use of radars to monitor stream discharge by noncontact of Hydrology 509: 573–587.

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)
LSPIV FOR MOUNTAINOUS FLOOD MONITORING 3029

Lejot J, Delacourt C, Piégay H, Fournier T, Trémélo ML, Allemand P. flash floods in a small ungauged mountain catchment (Central Spain).
2007. Very high spatial resolution imagery for channel bathymetry and Geomorphology 118(3): 383–392.
topography from an unmanned mapping controlled platform. Earth Ruiz-Villanueva V, Bodoque JM, Díez-Herrero A, Eguibar MA, Pardo-
Surface Processes and Landforms 32(11): 1705–1725. Igúzquiza E. 2013. Reconstruction of a flash flood with large wood
Le Boursicaud R, Pénard L, Hauet A, Thollet F, Le Coz J. 2016. Gauging transport and its influence on hazard patterns in an ungauged mountain
extreme floods on YouTube: application of LSPIV to home movies for basin. Hydrological Processes 27(24): 3424–3437.
the post-event determination of stream discharges. Hydrological Schmidt, AR, 2002. Analysis of stage–discharge relations for open-
Processes 30(1): 90–105. channel flows and their associated uncertainties (Doctoral dissertation,
Liao Q, Cowen EA. 2005. An efficient anti-aliasing spectral continuous University of Illinois at Urbana-Champaign).
window shifting technique for PIV. Experiments in Fluids 38(2): 197–208. Simpson MR. 2001. Discharge Measurements Using a Broad-Band
Lumbroso D, Gaume E. 2012. Reducing the uncertainty in indirect estimates Acoustic Doppler Current Profiler (pp. 01–01). US Department of the
of extreme flash flood discharges. Journal of Hydrology 414: 16–30. Interior, US Geological Survey, Sacramento, CA USA.
Marchi L, Borga M, Preciso E, et al. 2009. Comprehensive post-event Suh YK. 2003. Multi-Frame MQD-PIV. KSME international journal 17
survey of a flash flood in Western Slovenia: observation strategy and (10): 1552–1562.
lessons learned. Hydrological Processes 23(26): 3761–3770. Sweeney TL. 1992. Modernized Areal Flash Flood Guid-ance. In NOAA
Moyeed RA, Clarke RT. 2005. The use of Bayesian methods for fitting Technical Memorandum NWS Hydro 44. National Weather Service:
rating curves, with case studies. Advances in Water Resources 28(8): Silver Spring, MD.
807–818. Tinkler KJ. 1997. Critical flow in rockbed streams with estimated values
Muste M, Vermeyen T, Hotchkiss R, Oberg K. 2007. Acoustic velocimetry for for Manning’s n. Geomorphology 20(1): 147–164.
riverine environments. Journal of Hydraulic Engineering 133(12): 1297–1298. Trieste, DJ, Jarrett, RD, 1987. Roughness coefficients of large floods. In
Muste M, Fujita I, Hauet A. 2008. Large-scale particle image velocimetry Irrigation and Drainage Speciality Conference, ‘Irrigation Systems for
for measurements in riverine environments. Water Resources Research the 21st Century’, James LG, English MJ (eds). American Society of
44(4): W00D19. DOI:10.1029/2008WR006950 Civil Engineers: Portland, OR; 32–40.
Muste M, Hauet A, Fujita I, Legout C, Ho HC. 2014. Capabilities of large- Turnipseed DP, Sauer VB. 2010. Discharge Measurements at Gaging
scale particle image velocimetry to characterize shallow free-surface Stations. US Department of the Interior, US Geological Survey,
flows. Advances in Water Resources 70: 160–171. Washington, DC, USA: US Government Printing Office.
Oberg K, Mueller DS. 2007. Validation of streamflow measurements Wanek JM, Wu CH. 2006. Automated trinocular stereo imaging system
made with acoustic Doppler current profilers. Journal of Hydraulic for three-dimensional surface wave measurements. Ocean Engineering
Engineering 133(12): 1421–1432. 33(5): 723–747.
Ran Q, Qian Q, Li W, Fu X, Yu X, Xu Y. 2015. Impact of earthquake- Xu Q, Zhang S, Li W, van Asch TW. 2012. The 13 August 2010
induced-landslides on hydrologic response of a steep mountainous catastrophic debris flows after the 2008 Wenchuan earthquake, China.
catchment: a case study of the Wenchuan earthquake zone. Journal of Natural Hazards and Earth System Sciences 12: 201–216.
Zhejiang University SCIENCE A 16: 131–142. Yorke TH, Oberg KA. 2002. Measuring river velocity and discharge with
Rantz SE. 1982. Measurement and computation of streamflow. In acoustic Doppler profilers. Flow Measurement and Instrumentation 13
Measurement of Stage and Discharge, Vol. 1, Water-Supply Paper (5): 191–195.
2175. US Geological Survey: Washington. Zhang L, Fu X, Guo D, Li T. 2011. Flow resistance in gravel- and
Rickenmann D, Recking A. 2011. Evaluation of flow resistance in gravel- boulder-bed streams: a case study of the Longxi River. Journal of
bed rivers through a large field data set. Water Resources Research 47: Hydraulic Engineering (in Chinese) 44(6): 680–686.
W07538. DOI:10.1029/2010WR009793 Zhou W, Tang C. 2014. Rainfall thresholds for debris flow initiation in the
Ruiz-Villanueva V, Díez-Herrero A, Stoffel M, Bollschweiler M, Wenchuan earthquake-stricken area, Southwestern China. Landslides
Bodoque JM, Ballesteros JA. 2010. Dendrogeomorphic analysis of 11(5): 877–887.

Copyright © 2016 John Wiley & Sons, Ltd. Hydrol. Process. 30, 3014–3029 (2016)

View publication stats

You might also like