You are on page 1of 16

RIVER RESEARCH AND APPLICATIONS

River Res. Applic. 28: 118–133 (2012)


Published online 23 August 2010 in Wiley Online Library
(wileyonlinelibrary.com) DOI: 10.1002/rra.1441

COMPARISON OF EMPIRICAL AND THEORETICAL REMOTE SENSING BASED


BATHYMETRY MODELS IN RIVER ENVIRONMENTS

C. FLENER,* E. LOTSARI, P. ALHO and J. KÄYHKÖ


Department of Geography, University of Turku, 20014 Turku, Finland

ABSTRACT
Knowledge of underwater morphology is an essential component of many hydrological and environmental applications such as flood
modelling and lotic habitat mapping. Remote sensing allows modelling of bathymetry at spatial scales that are impossible to achieve
with traditional methods. However, the use of passive remote sensing for modelling water depth in fluvial environments remains a
challenge.
Different methods of computing bathymetry models based on remotely sensed imagery combined with ground measurements
for calibration were investigated in order to produce a digital bathymetry model of a reach of the river Tana in Lapland. An empirical
deep water correction model was evaluated together with theoretical hydraulically assisted bathymetry (HAB) models.
The empirical model produced good results, correlating to known depths at 0.98 (R2 ¼ 0.96) with a mean error of 12.0 cm. It was
demonstrated that usable levels of accuracy can be achieved with data that had previously been considered unsuitable for
bathymetry modelling. Some issues related to channel substrate were addressed. The models based on hydraulic theory were tested
for the first time outside the area they were developed in. Both models were found to be rather sensitive to certain assumptions, such as
the channel friction parameter. The HAB models are able to produce relative depth estimates that can under certain conditions
approach actual depths at accuracies similar to the empirical model.
Extensive accuracy assessment was performed in order to evaluate the vertical as well as the spatial accuracy of the three models.
Copyright # 2010 John Wiley & Sons, Ltd.
key words: remote sensing; bathymetry modelling; hydraulically assisted bathymetry; deep water correction; Tana river; Finland

Received 4 November 2009; Revised 1 July 2010; Accepted 7 July 2010

INTRODUCTION limitations in shallow waters and the horizontal resolution


is quite poor and the extent is often only on a local or reach
Bathymetry, the measurement of the depth of water bodies,
scale (Milne and Sear, 1997; Gilvear et al., 2004; Marcus
can be seen as an extension to topographic mapping, which
and Fonstad, 2008). The availability of fine resolution aerial
represents land elevation but traditionally renders water
photography allows for a different approach to bathymetric
bodies as flat surfaces. While having a continuous
mapping, which is best suited for shallow water environ-
representation of relief can be a goal in its own right, it
ments. Different methods of deriving bathymetry from the
should be most useful for studying geomorphological
image brightness values of remotely sensed data were
processes such as sediment transportation in the landscape,
pioneered in coastal research (Lyzenga, 1981; Gould and
catchment-scale hydrology (e.g. Bryant and Gilvear, 1999;
Arnone, 1997; Kohler and Philpot, 1998; Sandidge and
Dankers, 2002; Alho and Mäkinen, in press) and flood
Holyer, 1998; Lee et al., 1999; Roberts and Anderson, 1999)
modelling (e.g. Horritt and Bates, 2002; Bates et al., 2003;
and have more recently also been applied to rivers and
Néelz et al., 2006; Schäppi et al., 2008; Lotsari et al., 2010).
streams (e.g. Winterbottom and Gilvear, 1997; Bryant
In flood modelling, the accuracy of the model geometry used
and Gilvear, 1999; Wright et al., 2000; Westaway et al.,
has a remarkable impact upon flood mapping (cf. Bates et
2003; Marcus et al., 2003; Gilvear et al., 2007). Research
al., 2003; Alho et al., 2009a; Koivumäki et al., 2010).
focusing on the optical characteristics of water (Gould and
Therefore, topo-graphic (e.g. Alho et al., 2009b) and
Arnone, 1997; Holden and LeDrew, 2002; Lyzenga et al.,
bathymetric (e.g. Gao, 2009) data collection procedures
2006; Legleiter and Roberts, 2009) helped to establish the
should be developed.
physical basis for the techniques used in remotely sensed
Traditionally bathymetric charts are produced by inter-
bathymetry mapping in rivers as well.
polating data collected by echo-sounding (e.g. Kiss and
Remote sensing based methods make it possible to
Sipos, 2007; Sagawa et al., 2007), but this technique has its
generate accurate maps of shallow water streams by
*Correspondence to: C. Flener, Department of Geography, University of coupling remote sensing data with ground-based measure-
Turku, 20014 Turku, Finland. E-mail: claude.flener@utu.fi ments (e.g. Winterbottom and Gilvear, 1997) while reducing

Copyright # 2010 John Wiley & Sons, Ltd.


COMPARISON OF REMOTE SENSING BASED BATHYMETRY MODELS 119

or eliminating the time-consuming and costly gathering of (Bailly et al., 2008). It is therefore of great interest to further
ground-based measurements. Those models that do require develop methods that use more conventional remotely
some empirical depth data for calibration, need much less sensed imagery to model stream depths in areas where
data than an interpolation would require, while allowing a such data are available and active sensor data are lacking.
depth model to be generated at a much finer spatial Remotely sensed bathymetry modelling methods have so
resolution. far not been tested on sub-arctic rivers. In particular, the
A variety of sensors are available for fluvial research, HAB models have not been tested outside the rivers they
ranging from panchromatic and colour photo- and video- were developed on (Fonstad and Marcus, 2005). Hence, the
graphy to multispectral and hyperspectral scanners (Bryant aims of this paper are
and Gilvear, 1999; Marcus and Fonstad, 2008). Generally
 to evaluate the usability of the Lyzenga model for calcu-
airborne imagery is found to be more useful than satellite
lating bathymetry based on ground data that do not cover
data for river monitoring, due to its higher resolution and
the entire depth spectrum modelled.
potentially flexible timing, which allows for easy compari-
 to evaluate the usability of hydraulically assisted bathy-
son with more traditional field data (Roberts and Anderson,
metry (HAB) models for calculating bathymetry of sub-
1999). Multispectral airborne data have been widely used to
arctic rivers.
map river bathymetry and bottom sediment type classifi-
 to compare the usefulness of the theoretical HAB models
cations (e.g. Lyon et al., 1992; Wright et al., 2000). Marcus
to that of empirical and interpolation models.
(2002) achieved higher accuracy in mapping in-stream
habitats using 1 m resolution 128 band hyperspectral
imagery than with ground-based surveys. The accuracy of
modelled depths was found to decrease in smaller steams PHYSICAL BASIS OF MODELS USED
(Marcus et al., 2003), while Gilvear et al. (2007) concluded
Lyzenga algorithm
that optically remotely sensed depth mapping is somewhat
limited to shallow waters. The empirical model is based on the connection between
Legleiter et al. (2004) concluded that the physical basis water depth and reflectance values measured by a remote
for remote sensing of rivers is sound, providing a solid sensor. Lyzenga (1981) developed the following algorithm
foundation for large-scale, long-term mapping and monitor- to linearise the relationship between pixel values or digital
ing of fluvial systems. Aerial photography related depth numbers (DN) in each band and depth:
calibration issues that occur over catchment-scale study
areas have been addressed with an illumination correction Xi ¼ lnðLi  Lsi Þ (1)
algorithm by Carbonneau et al. (2006). The relationship
where Xi is the variable that is linearly related to water
between the spatial resolution of the sensor and the channel
depth in band i, Li the observed brightness and Lsi is the
morphology was found to be of importance as well
deep water reflectance in the same band.
(Legleiter et al., 2009). Novel approaches to modelling
The algorithm is based on the fact that the attenuation of
river depth (Fonstad and Marcus, 2005), and stream power
light in water is exponential and that the reflected signal
(Jordan and Fonstad, 2005) from remotely sensed images
is composed of other factors than just depth, the most
based on theoretical hydraulic equations rather than
important of these being river bed substrate. The algorithm
extensive field data have been explored with promising
uses deep water reflectance, i.e. the reflectance of water
initial results. The increased areal coverage of remote
deeper than where bottom substrate can influence the signal,
sensing based models compared to non-remote sensing
to isolate the depth factor in the observed reflectance signal.
approaches of computing river bathymetry has been found
Winterbottom and Gilvear (1997) were the first to apply this
to more than make up for the loss in vertical accuracy
to a river, and the method used here is based on theirs.
(Carbonneau et al., 2006; Marcus and Fonstad, 2008). The
However, it is worth noting that the present method differs
developments in aerial imagery based bathymetric model-
in some points from that of Winterbottom and Gilvear
ling methods present great opportunities for practical
(1997), partly due to differences in the data sources
applications such as hydraulic modelling for flood research
available, but also due to some methodological omissions
or habitat modelling on clear water rivers and streams. While
in their paper.
active sensors such as green-wavelength LIDAR can be
used to some extent for accurate depth measurements in
HAB models
clear streams (Feurer et al., 2008; McKean et al., 2008), and
ground penetrating radar can be used in turbid rivers as well The HAB models (Fonstad and Marcus, 2005) do not
(Costa et al., 2000), these methods are expensive and not require any empirically measured depth data. They do
without their own problems, including large footprint size however rely on the same basic principle of attenuation of

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
120 C. FLENER ET AL.

light in water, and hence the covariance of reflectance values While several studies used methods based on the Beer-
and water depth. The aim of these models was to widen the Lambert law (Lyon et al., 1992; Fonstad and Marcus, 2005),
field of applicability by not requiring empirical depth it should be noted that Newman and LeDrew (2001) found
measurements for their calibration. Both models require that the attenuation coefficient may be affected by the
discharge data (the only ground-based input variable), bottom type and depth, and therefore may cause compli-
channel width (derived from the remotely sensed image) and cations in shallow water environments in some circum-
river gradient (derived from elevation data) in addition to the stances.
actual remotely sensed images. Moreover, the hydraulic
equations used in the models require Manning’s roughness
parameter (n) to be determined. Factors affecting modelling
HAB 1 is regression based, like the empirical model,
Depth modelling requires an unobstructed view of the
but instead of regressing against known depths, the depths
area being analysed, implying that in addition to a clear
are derived using hydraulic theory to approximate channel
atmosphere in between the sensor and the water body in
shape by calculating the average and maximum depths for
question, there must not be any overhanging trees, shadows
several river cross-sections based on hydraulic theory and
or ice cover on the river (Fonstad and Marcus, 2005).
relating the minimum, average and maximum pixel values
Furthermore, it is absolutely imperative that the water be as
for each band to these values. The model is established
clear as possible (Marcus et al., 2003). While severely turbid
by regressing all these values over all cross-sections.
water would make it impossible to view the river bed, lighter
This regression-derived model is then applied to the entire
levels of turbidity still influence both the reflectance, the
area of interest to establish depth estimates for each image
extinction depth of light and the diffuse attenuation
pixel (Fonstad and Marcus, 2005).
coefficient of water (Lafon et al., 2002). Studies dealing
HAB 2 was developed for those rivers where some of
with optical remote sensing in fluvial settings have focused
the assumptions that HAB 1 is based on are inappropriate
on clear rivers so far.
(Fonstad and Marcus, 2005). Rather than regressing
In addition to the above constraints it is important that the
against known or calculated depth points, it attempts to
water surface of the area to be analysed be transparent.
compute the rate of attenuation and apply that to the
Surface roughness increases the reflectance of the surface
remotely sensed data by relying on the Beer-Lambert law of
and can cause a severe obstruction to the viewing of the river
logarithmic decay, which describes the logarithmic depen-
bed. Hence, rapids need to be generally excluded when
dence between the distance of light travelled through a
remotely sensing water depth, but even in less turbulent
transmissive medium with minimal scattering, and the
waters, the surface scattering and sun glint may be an
attenuation coefficient of this medium. This can be
obstruction (Marcus and Fonstad, 2008).
employed in a clear water setting (Carbonneau et al.,
Slight surface roughness, such as small ripples on the
2006), where the water column is the transmissive medium
water surface that are caused by wind may be troublesome in
and the attenuation coefficient can be used to derive depth:
some cases, but negligible in others. More than any land
surface, water surfaces tend to accentuate bidirectional
lnðDN=DN0 Þ
D¼ (2) reflectance, which causes some areas of a remotely sensed
b image to be of different quality than others. In some cases,
surface roughness may cause such dominating reflectance
(Fonstad and Marcus, 2005) where D is water depth, that it becomes impossible to remotely sense water depth
DN0 should be the reflectance value of a pixel of river bed from images. One method that has been used to deal with
substrate not covered by any significant amount of water, i.e. this problem is to classify such areas as a separate category
wet river bed of a depth near zero, and b is the attenuation when producing categorised depth maps (Hardy et al.,
coefficient. Since the b value can not be determined without 1994).
field measurements, the HAB 2 model derives it by inserting As the optical remote sensing of water depth is based on
the DN values of a cross-section of the river into Equation 2 the correlation of brightness values and depth, differences in
and using a seed value for b to estimate the depths at substrate can also become problematic. Since different
multiple points along the transect. Using these depth substrates have different reflectance values, the depth
estimates the discharge Q can be computed based on estimates may be influenced in a positive or negative
hydraulic equations (Fonstad and Marcus, 2005). This way, so that depths are either over- or underestimated.
process is iterated with adjusted b values until the outcome The Lyzenga algorithm (Lyzenga, 1981) assumes uniform
for Q equals the measured discharge value, thus yielding substrate for modelling bathymetry, but it can also be
best estimates for D. employed to determine substrate variability. The method

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
COMPARISON OF REMOTE SENSING BASED BATHYMETRY MODELS 121

relies on the fact that the radiance reflected by the substrate


is a linear function of bottom reflectance and an exponential
function of depth. When plotting the deep water corrected X
variables computed from two different bands against one
another, they are predicted to fall on one line if depth varies
while substrate is uniform. If the substrate varies, the points
are expected to fall on two parallel lines. This makes the
creation of substrate variability maps possible.

STUDY AREA
The Tana river basin (Tenojoki in Finnish, Tanaelva in
Norwegian) is located in the northernmost part of
Fennoscandia. The catchment area is approximately
16 000 km2 in size, of which 68% is located in Finnmark
in Norway and 32% in the province of Lapland in Finland
(Mansikkaniemi, 1972). It drains an extensive upland area
and belongs to the sub-arctic zone of Fennoscandia.
The average discharge of the river over the year measured
at Polmak is 168 m3 s1 with the lowest discharge in March
(41 m3 s1), the highest discharge in June (544 m3 s1) and
a yearly average high discharge of 1600 m3 s1 (Alarau-
danjoki et al., 2001).
The study area is located at Granjarga between 27860 2400
and 27890 4700 E and 698550 100 and 698560 600 N (Figure 1).
The area of interest (AOI) is 2.7 km long starting 3 km
downstream of the confluence of the river Utsjoki. This area
is located on a relatively short stretch of river where accurate
discharge data are available together with excellent quality
aerial imagery. The channel in this section is rocky and
rather stable, compared to the sandy river bed further Figure 1. Map of the study area. The Tana catchment is located in the
downstream. Furthermore, the section contains a range of border region between Norway and Finland and drains to the Arctic Ocean.
depths, and reasonable ground data are available as well. The study area (AOI) is located immediately downstream of the confluence
of the main channel of Tana (in Finnish, Tenojoki) and a tributary Utsjoki.

accuracy using the RTK-GPS. Hydrological data from the


METHODS Finnish Environment Institute (SYKE) were utilised in order
to adjust the depth values for each echo-sounded line to the
Input data
water level of the date of the imagery acquisition flight.
The aerial images used in this study were produced on Water level and discharge data were obtained from the
4 July 2005 by TerraTec AS and georectified by Norwegian Onnelansuvanto and Patoniva measuring stations located a
Mapping Authority. The orthophotos were acquired from short distance upstream from the study area. Discharge data
Norsk Eiendomsinformasjon AS in digital form, in GeoTIFF were required for the remote sensing flight date, whereas
format. The images have a spatial resolution of 0.5 m. They water level data were required additionally for the field work
are true colour RGB images at 8 bits per band. days, in order to allow for the depth adjustment calculations
Field data in the form of depth measurements were to be performed. The river surface gradient was calculated
gathered during August 2007 using a sigle-channel echo- based on RTK-GPS points collected within the AOI.
sounder with an accuracy and resolution of 10 cm within its The river was digitised based on the orthophoto data. The
range, deployed from a zodiac and coupled with real time echo-sounding data were extracted and combined with the
kinematic GPS (RTK-GPS). The system suffers from the shoreline data after filtering out measurement errors. The
limitation of not being able to measure depths shallower than water levels were calculated based on the measured depth
0.6 m. Depths were sampled at 5 m intervals, and their values combined with the water level information from
location recorded precisely at 5 cm horizontal and vertical SYKE. The echo-sounding work stretched over a period of

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
122 C. FLENER ET AL.

several weeks, so different data lines were collected on


different days. The water levels of these days were adjusted
to the water level of the aerial photograph flight 2 years
earlier. The digitised shoreline points were assigned a depth
value of 0, since they correspond to the water level of the
flight day.

Lyzenga model
The combined and adjusted depth point data were used
to extract the DN values for each band separately from the
image files. The Lyzenga formula calls for the deep water
reflectance to be determined. This rather crucial step is not
described in Winterbottom and Gilvear (1997). Westaway
et al. (2003), who also used Lyzenga’s algorithm, do not
explain their method for determining this parameter either,
but simply refer to Lyzenga (1981). However, Lyzenga
(1981) never applied his method to rivers, and in the coastal
environment that it was designed for, gathering deep water
reflectance values is a simple matter. A short distance
offshore, the waters become deep enough that the bottom no
longer has an influence on the reflectance values.
Determining this value in a river is a different proposition
though. Gilvear et al. (2007) used in situ spectroscopy to
determine the depth where the river bed no longer influenced
reflectance values. The field spectroscopy option was not
available in the case of the present study, so Lyzenga’s
method of choosing the site from the remotely sensed data
was transferred to the river. Since data were available for a
larger section of the Tana, it was possible to gather Lsi values
outside the study area in deeper water than could be found
within the study reach by choosing the darkest, and hence
deepest area with good image quality and averaging a series
of the darkest pixels of that area to determine the deep water
reflectance values of each band. The river substrate in this
area is virtually identical to that of the AOI and the depth at
the pixel locations likely exceeds the closest measured point of
5.2 m. It was determined that this was the best possible
approximation of this variable using the data available.
The first step of the data analysis involves computing
Lyzenga’s X values for each band, using the formula in
Equation 1. The Lyzenga algorithm isolates the depth signal
from the reflectance measured in the image, turning the
exponential relationship (Figure 2) of reflectance and depth
into a linearised one (Figure 3). The gaps in these (and
subsequent) graphs between depth values of 0 and 0.5 m are
due to the limitation of the echo-sounding process, which
makes data gathering in water shallower than 0.6 m
impossible. The abundance of points of zero depth stems
from the digitising process.
The correlations between Lyzenga’s X values and
measured depth are good (Figure 3), especially in the red Figure 2. Reflectance values for bands 1–3 (RGB) versus depth. The
and green bands. Spectral analysis performed by Winter- abundance of points of 0 depth stem from the shoreline digitisation.

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
COMPARISON OF REMOTE SENSING BASED BATHYMETRY MODELS 123

bottom and Gilvear (1997) revealed that water depth can be


most readily distinguished between 0.5 and 0.8 mm, which
includes bands 1 and 2 of RGB data. In light of this property,
as well as the relatively poor correlation between the blue
band and water depth and the correspondingly poor relation
between X3 and depth (Figures 2 and 3), the model was
built using the red and green bands.

HAB 1 model
The discharge (Q) of the river Tana at the AOI for the
day of image acquisition was 194.5 m3 s1. Manning’s n of
0.033 used in this study was derived from hydraulic data
and had previously been used successfully in hydraulic
modelling studies on the river Tana (Lotsari et al., 2008).
In order to test the sensitivity of the HAB models to the
Manning roughness value, the models were additionally
computed with n ¼ 0.04 and n ¼ 0.05.
A selection of 10 evenly spaced cross-sections was chosen
for this model within the sections of the AOI that consist of
a single channel. Fonstad and Marcus (2005) advise against
using cross-sections in braided sections of a river, since
this leads to complications with the assumptions that the
hydraulic equations are based on. The cross-sections were
sliced and the vertices converted into point data with a distance
of 0.5 m in between each, resulting in an inter-point distance
equal to the resolution of the raster image. The point data
were used to sample the reflectance values for bands 1–3.
The equation to estimate average depth (Da) from
discharge measurements (Q), slope measurements from
map or elevation data (S) and channel width measurements
(W) from remotely sensed imagery was adapted from
Fonstad and Marcus (2005) to be used with a Manning’s n
estimate value, rather than an estimation formula for n:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 
5 Qn 3
Da ¼ pffiffiffi (3)
W S

Now, for each cross-section, the average depth Da can


be related to the average brightness value, the maximum
depth Dmax, which is assumed to be 2Da, to the minimum
brightness value, and the minimum depth Dmin, which was
assumed to be 5 cm (cf. Fonstad and Marcus, 2005), to the
maximum brightness value for each band. Since the relation
of depth to DN values is exponential, a non-linear regression
of the type y¼aebx is performed on these data in order to
derive the depth model to be applied to the image raster.

HAB 2 model
The DN values were sampled at five cross-sections at 1 m
intervals. The input values for the discharge, river gradient
Figure 3. Lyzenga’s X for bands 1–3 versus depth. and roughness coefficients were used in the same way as
for the HAB 1 model. Additionally, the values for DN0

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
124 C. FLENER ET AL.

needed to be determined. Fonstad and Marcus (2005) suggest In order to assess the fit of the model, the modelled depth
two options for selecting DN0: either one uses the brightest values were plotted against the measured values (Figure 4).
pixel of the cross-section, or a pixel adjacent to the river This visual appraisal revealed an interesting cloud of pixels
consisting of the same substrate as the river bed is selected. with depth estimates that are overestimated at a depth
Both methods were tested here, in order to see if the results ranging from about 0.75 to 1.2 m. The same cloud of pixels
differ depending on the selection method for this variable. can be seen in the reflectance data in Figure 2 as well as the
In order to handle the iterative computations of the Xi data in Figure 3, where these points have too low values
attenuation coefficient b, the model was implemented in as for their depth (i.e. DN of band 1 below 75 at depths
an R package (Flener, 2009). A structured data frame between 50 and 120 cm). In general, points with similar
containing reflectance data of a range of transects can thus reflectance values are at least 1 m deeper. The spatial
be analysed and attenuation coefficients are computed until distribution of these points was analysed and found to be
modelled discharge matches the measured value, imple- autocorrelated, leading to the hypothesis that they may be
menting the theory set out in Fonstad and Marcus (2005). caused by differing bottom substrate. The method provided
The package provides optional automatic computation of by Lyzenga (1981) for distinguishing bottom substrate
DN0 values as well as analytical summary capabilities. was used to analyse substrate variability. Figure 5 shows X1
plotted versus X2. While the Xi vs. Xj plot for the entire AOI
looks like the substrate is rather homogeneous (i.e. points
RESULTS fall on one line), the darker end of the spectrum is more
Lyzenga bathymetry model spread out than the brighter end. Plot B in Figure 5 focuses
on the dark points only, and it becomes apparent that here
In order to guarantee unbiased model verification, the AOI the points tend to converge towards two separate lines
was split into a calibration area and a verification area. The with a slight gap in between. This is an indication that the
accuracy of all three models is tested in the same area (cf. cause of the cloud of dark pixels at shallow depth is bottom
Figure 6). Since the relationship between Lyzenga’s X and substrate variability. A subsequent field check confirmed
water depth is linear, the computation of a multiple linear this. The river bed in the area where the low value points
regression allows the construction of the following depth cluster consists of particularly dark rocks, partially covered
model: with algae.
In order to improve the model, the multiple regression
D ¼ 578:327  270:526 X1 þ 166:479 X2 (4) was calculated once more, this time excluding the dark
shallow points, resulting in the following depth model:
All regression coefficients in the models were significant
D ¼ 613:982  255:212 X1 þ 142:780 X2 (5)
at the 0.001 level. The general fit of the models was
calculated as the root mean square deviation (RMSD), which
describes the mean deviation of the estimated values, in the
unit of measure used on the input data (cm). The RMSD Excluding shallow points with low reflective values
on the verification area was 13.3 with an R2 of 0.95. from the analysis marginally improved the result (Table II).

Figure 4. Lyzenga model depths versus measured depths including (A) and excluding (B) non-uniform substrate points.

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
COMPARISON OF REMOTE SENSING BASED BATHYMETRY MODELS 125

Figure 5. Bottom substrate variability plot. X1 vs. X2 (A) and for the subset of the darker points only (B). A pattern of points converging to a single line indicates
homogeneous substrate whereas points converging to parallel lines indicate different substrates.

The new fit can be assessed in Figure 4B which shows the using Lyzenga’s Xi vs. Xj method by dividing X1 by X2 and
new depth estimates plotted vs. measured depth. The RMSD isolating the areas with variability values diverging from 1
was now 12.0 with an R2 of 0.96. Thus the established by more than two standard deviations (Figure 6B). The area
model equation was applied to the image data, creating a with the low reflective shallow points clearly has different
raster consisting of depth estimates (Figure 6A). One should substrate from the rest of the AOI. The model verification
keep in mind though that depth will be overestimated in the data are located in an area with uniform substrate, thus
area of the channel where the shallow points with low fulfilling that premise of the Lyzenga model. However, when
reflectance are located. The bottom variability was modelled comparing the overall accuracy of the AOI, including the

Figure 6. Lyzenga based depth model map (A) and bottom variability map X1 =X2 (B). The white lines in map B indicate the location of the downstream (I) and
upstream (II) reference transects for the assessment of spatial accuracy. The black underlay indicates the verification area. The spline interpolation in map C is
computed from the indicated points. This figure is available in colour online at wileyonlinelibrary.com/journal/rra

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
126 C. FLENER ET AL.

calibration area and hence the variable substrate, the mean


error is 31.51 cm with an R2 of 0.87 when including the dark
non-uniform substrate points and 23.67 cm (R2 ¼ 0.93) when
excluding these points from calibration.

HAB 1 bathymetry model


HAB 1 depth-reflectance equations were computed for
each band using the range of Manning’s n values set out
above. An example of the resulting depth maps produced by
this model can be seen in Figure 8, showing that the overall
shape of the river bed was reproduced rather well by HAB 1.
Visual analysis of the graphs in Figure 7 shows that the best
fit among the three bands is in band 1. The residual standard
error in this case is 10.6 (R2 ¼ 0.97) in the verification area,
whereas on the entire AOI the RMSD is 24.4. The fit of band
2 was not as good as that of band 1, but better than band 3.
The decreasing fit is reflected in an increase of noise in the
depth maps produced as the wavelength decreases resulting
in a rather ‘grainy’ depth map for band 3. The gradient maps
of Figure 8 depict the continuous nature of the bathymetry
raster very well, in particular the detail in the shallow areas,
since this model, unlike the empirical one, does not suffer
from a data gap in this part of the depth spectrum. In order to
facilitate comparison to the results of the other models,
classified depth maps were produced (Figure 8), using an n
value of 0.033, applying the same depth classification and
scale as in Figure 6. These maps clearly show that the HAB 1
model produces a much shallower version of the river bed
than the Lyzenga model. The relative depths are reproduced
in a very similar fashion, but the absolute depth values are
quite different. Furthermore, the decreasing depth signal
with shorter wavelengths is even more clearly illustrated
here, as the map based on band 3 barely reproduces the river
bed at all. The maps produced using HAB 1 with different
roughness values look very similar, differing only in the
scale of the depths. The complete results of this model are
listed in Table II. The most interesting bands here are red and
green, as was the case in the empirical model, which is
corroborated by the spectral analysis findings of Winter-
bottom and Gilvear (1997).
Figure 7. Depth–reflectance graphs for bands 1–3 (RGB) including the best
HAB 2 bathymetry model fit regression lines plotted by the equations shown. The given examples are
the results using a Manning’s roughness value n of 0.033.
HAB 2 yielded results for each band at all tested
Manning’s n values and for manual as well as automatic DN0 of this parameter produces positive minima close to 0. This
selection methods. The resulting bathymetric maps were highlights the importance of the choice of selection method
very similar to those of HAB 1, differing only in the depth for DN0, questioning their simple interchangeability.
scale, not the spatial distribution of depths. Table II The average and maximum depths varied with the DN0
summarises the depth models built with HAB 2. A clear selection method as well as with the roughness coefficient
pattern emerges, differentiating the manual and automatic used, being positively related to n and to wavelength. The
selection methods for the DN0 values. While the models with automatic DN0 selection leads to a larger range of depths
automatically computed DN0 produce negative minimum modelled independent of n. This behaviour is explained by
depths (i.e. pixels above water level), the manual selection comparing the average of the automatically selected DN0

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
COMPARISON OF REMOTE SENSING BASED BATHYMETRY MODELS 127

Figure 8. HAB 1 bathymetry map based on all three bands (1–3) with a Manning’s n of 0.033 with a continuous gradient (A) and a classified depth chart (B) for
each band. Note that the depth scale used is the same for all three bands. To facilitate visual comparison, the classification of the classified depth charts (B) is also
the same as the one used on the map of the empirical model (Figure 6A).

values to the manually selected values (Table I). The model that was computed, the depth values were extracted
automatically selected values are consistently lower (i.e. from the resulting raster at the points’ locations within the
darker/deeper pixels) than the values selected manually from verification area (Figure 6) and correlated to the measured
the shore. depths (Table II).
The correlations of modelled to actual depths are all rather
respectable. The first model in the table is a discretised thin
Assessment of bathymetric accuracy
plate spline interpolation (Wahba, 1990) (Figure 6) of
All models were analysed with respect to the measured measured depth points including the digitised shoreline, and
depth values that were collected by echo-sounding. For each is therefore expected to correlate at a very high level to the

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
128 C. FLENER ET AL.

Table I. Wet river shore substrate reflectance values (DN0) used in The theoretical models, which were computed one band at
HAB 2 based on manual or automatic selection. a time and with a range of Manning’s roughness values as
input, also correlate well with the ground data. A clear
DN0 selection method Red Green Blue
pattern can be seen between the band used in the model and
Manual 246 247 241 the level of correlation. Throughout the HAB 1 and HAB 2
Automatic 199 193 185 model results, the models based on the blue band always
produce worse results than the ones based on the red and
green bands. This is in line with previous findings. In fact,
this is logical as the red part of the visible spectrum is most
very points it is based on, which, at a correlation coefficient readily absorbed by water, thereby allowing its use for the
of 0.98 and an R2 of 0.95, it does. The interpolation was distinction of differences in depths in shallow water. The
performed using a subset of points spaced in as regular a blue part of the visible spectrum on the other hand is
pattern as was feasible with the available data. transmitted through water with much less absorption,
The Lyzenga model also uses the empirical points as making depth modelling in shallow water based on the
input, but the resulting depth values are not based on these blue band rather more difficult (Legleiter and Roberts,
empirical points as directly as is the case with an 2009). Neither the correlation coefficient nor the RMSD
interpolation. Hence a near perfect correlation of the change with the use of different n values.
modelled depths to the measured points is not as readily However, looking at the minimum, average and maximum
implied with an interpolation, yet in the case of the Lyzenga modelled values gives a different perspective on the
model calibrated excluding the non-uniform substrate points accuracy of the fit of a model. Comparing the measured
(–dp), it is identical. minimum (0 cm), average (77.3 cm) and maximum

Table II. Accuracy assessment results for all models in relation to measured depth points as well as the minimum, mean and maximum
modelled depth values. Band is the radiometric band used, n is the roughness value used (for the theoretical models only) and D is the
modelled depth in cm. All statistics are significant at the 0.01 level.

Model Band n Corr. R2 RMSD Dmin Davg Dmax

Spline — — 0.98 0.95 12.5 4.4 86 151


Lyzenga –dp 1,2 — 0.98 0.96 12.0 8.7 87.7 182.9
Lyzenga þdp 1,2 — 0.97 0.95 13.3 4.3 80.8 173.2
HAB 1 1 0.033 0.98 0.97 10.6 7 65.7 125.7
HAB 1 2 0.033 0.97 0.95 13.2 4.1 64.2 129.9
HAB 1 3 0.033 0.95 0.89 19.2 1.2 62.6 133.2
HAB 1 1 0.04 0.98 0.97 10.6 7.9 74 141.6
HAB 1 2 0.04 0.97 0.95 13.2 4.5 71.7 145.5
HAB 1 3 0.04 0.95 0.89 19.3 1.3 70.2 149.7
HAB 1 1 0.05 0.98 0.97 10.6 8.9 84.1 161.4
HAB 1 2 0.05 0.97 0.95 13.2 5.1 81.7 166.1
HAB 1 3 0.05 0.95 0.89 19.3 1.5 79.4 169.7
HAB 2 auto 1 0.033 0.98 0.96 12.5 20.1 58.2 118
HAB 2 auto 2 0.033 0.96 0.93 15.6 27.4 56.8 115
HAB 2 auto 3 0.033 0.94 0.88 20.8 40.2 55.2 109.9
HAB 2 manual 1 0.033 0.98 0.96 12.5 1.3 63.1 110.4
HAB 2 manual 2 0.033 0.96 0.93 15.6 3.5 64.4 106.5
HAB 2 manual 3 0.033 0.94 0.88 20.8 1.8 65.3 101.8
HAB 2 auto 1 0.04 0.98 0.96 12.5 22.5 65.3 132.5
HAB 2 auto 2 0.04 0.96 0.93 15.6 30.7 63.8 129.1
HAB 2 auto 3 0.04 0.94 0.88 20.8 45.1 61.9 123.3
HAB 2 manual 1 0.04 0.98 0.96 12.5 1.5 70.9 123.9
HAB 2 manual 2 0.04 0.96 0.93 15.6 3.9 72.3 119.6
HAB 2 manual 3 0.04 0.94 0.88 20.8 2 73.3 114.3
HAB 2 auto 1 0.05 0.98 0.96 12.5 25.8 74.7 151.5
HAB 2 auto 2 0.05 0.96 0.93 15.6 35.2 73 147.8
HAB 2 auto 3 0.05 0.94 0.88 20.8 51.6 70.8 141.1
HAB 2 manual 1 0.05 0.98 0.96 12.5 1.7 81.1 141.7
HAB 2 manual 2 0.05 0.96 0.93 15.6 4.5 82.8 136.8
HAB 2 manual 3 0.05 0.94 0.88 20.8 2.3 83.8 130.6

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
COMPARISON OF REMOTE SENSING BASED BATHYMETRY MODELS 129

(163.0 cm) depths to the equivalent modelled values in the this is handled by the different models. Since there were
verification area gives an indication of how well the model no complete transects available as ground data, they were
represents the actual dimensions of the river. The Lyzenga computed at locations where an echo-sounding line reached
model gets close to these values, albeit with negative near the shores. The depth values from the transect
minima. While all HAB 2 models (and indeed the HAB 1 interpolations as well as of all previously analysed models
models as well) are much shallower than the measured were subsequently sampled at 1 m intervals and super-
values, those with automatically selected DN0 values are less imposed in Figure 9. The data gap between the digitised
so. This is a result of the fact that the manually selected DN0 shoreline and the echo sounded data are responsible for the
were brighter than the automatically chosen ones. The smoothly curved interpolation leading to the shore on both
manually chosen DN0 values represent a pixel at the shore of sides of the reference transects. This analysis focuses on the
the river, consisting of the same substrate as the river bed section deeper than 60 cm, since the interpolation is more
selected along the entire stretch of the AOI in a process that is accurate there as regular empirical data are available below
necessarily subjective. The automatically selected DN0 values this depth.
are less likely to be of the same levels of brightness, since these All models create higher variability of depths compared
values are selected as the brightest and hence shallowest point to the interpolations. The spatial resolution of the reference
within a cross-section, consisting of a much smaller sample transect and the lack of input points near the shore does
size than in the manual process. Even though they are meant to not allow for small scale variations to be represented. The
represent the reflectance at the surface of water they most resolution of the remotely sensed models on the other hand
likely represent the reflectance just below the surface. allows for much smaller scale variations to be visible. This
However, the selected points are averaged by the HAB 2 implies that the covariance values on their own may be
algorithm and the very nature of an automated process makes misleading, and that visual appraisal of resemblance
the result more objective, which, combined with the greater between transect curves may be a better indication of the
depth range and unchanging RMSD favours the automatic ability of a model to represent the river bed at a given
DN0 selection method as the method of choice for further transect.
applications of HAB 2. The Lyzenga model fits the reference transect relatively
In contrast to the manual HAB 2 models, whose minima well in both transects where the substrate is uniform but it
approach 0, the automatically chosen DN0 values result in overestimates the depths as the substrate changes to the dark
larger negative minima. This is not necessarily incorrect rocks after 100 m on the upstream transect (cf. Figure 6B).
though, as the possibility of miss-classified dry pixels near In the upstream section between 25 and 40 m the model
the shore leaves the possibility for negative values. Indeed, underestimates depth by about half a metre. This section
the Lyzenga model and even the spline interpolation result does not show on the substrate variability map at the current
in negative minima. threshold set for the substrate classification. These findings
Studying the effect of the roughness coefficient on the combined with the decreasing RMSD when including non-
depth range, the sensitivity of both HAB models to this uniform substrate areas in the modelling proves that
parameter is clearly visible again, as the range increases disregarding the assumptions of the model affects the result
markedly with increasing n values, particularly driven by the in a negative manner. In a larger section of river it may be
positive correlation of n to maximum depth. This shows feasible to create the variability map and calibrate the model
that the model is rather sensitive to the estimated Manning’s separately for varying substrates.
roughness variable. The overall accuracy of the models Both HAB models behave very similarly, HAB 1 being
represented by the RMSD is approaching the combined marginally deeper than HAB 2. Using the primary n value of
uncertainty of the sonar–RTK-GPS rig for all but the band 3 0.033, both theoretical models produce river beds that are
based HAB models. Considering the difficulty of defining clearly shallower than the real river bed on both transects.
what an accurate depth measurement actually is in a rocky Most of the river bed is modelled to be around average
river, this is a considerable result. depth. The deeper areas are barely noticeable in these
HAB transects. The cutoff is around a depth of 1 m. The
substrate variability in the upstream transect causes similar
Assessment of spatial accuracy
problems for the HAB models as it did for the Lyzenga
In order to assess how the various models reproduce model.
depths away from the measured points and to gain a useful The transects also illustrate the sensitivity of the HAB
visual appraisal of the relative performance of the various models to the roughness coefficient. The depth range can be
models, a transect was cut through the verification area seen to increase with increasing Manning’s n. The model
(Figure 6I). A second transect was cut through the area using an n of 0.05 comes closest to the depths measured in
with variable substrate (Figure 6II) in order to analyse how the field. However, this is a rather high value for this

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
130 C. FLENER ET AL.

Figure 9. Downstream (A) and upstream (B) river cross-sections (cf. Figure 6) showing the transect, the Lyzenga model calibrated excluding the dark
point area, the HAB 1 and HAB 2 models based on band 1 with an n of 0.033 and HAB 1 based on band 1 with an n of 0.05. All curves except the transect
itself are smoothed with a 2.5 m filter to enhance the readability of the plot. All transect plots are looking downstream, so the distance on the x-axis starts on
the left bank of the river.

particular river setting (cf. USGS, 2008). Since both models DISCUSSION
show the same pattern causing the same weaknesses, while
When assessing the usability of remote sensing based
using completely different methods for deriving depth from
bathymetry models, the context that the resulting terrain
reflectance values, the problem is most likely found in the
models are to be used in is important. The advantages and
theory of how the hydraulic equations are used to derive the
limitations of these models need to be evaluated relative to
reference values. The inability to adjust the roughness
the corresponding properties of other methods of producing
coefficient to discharge and depth as is common practice in
depth models.
hydrological modelling, due to depth being unknown
without field data, is causing a problem to the methodology
Aerial photography based models vs. sonar models
of the theoretically based models.
A general problem exists on the shore of the river, where The overall accuracy of the traditional method for
a pixel with a depth of zero may be darker than the model producing sub-aqueous terrain models based on interp-
would expect, due to it not consisting of the same material olation is directly dependent on the amount of empirical
as the river bed, but instead it may contain sub- input points. The remote sensing based models on the other
aerial vegetation or a shadow of an adjacent tree or hand may not be quite as precise at the measured points as
overhang. Legleiter and Roberts (2005) determined mixed an interpolation raster, but their accuracy does not vary with
radiance on the shore to be one of the major limitations to distance from the ground points, but, assuming similar
remotely sensed bathymetry models for use in fluvial substrate and optical properties of the water column and
geomorphology. Due to the way the reference transect water surface throughout the image, a more even level of
was constructed, based on the widest available straight accuracy can be achieved across the entire area that is
sampling line, very steep shore drop-offs were to be modelled. A sonar interpolation based bathymetry model
expected. This is due to the fact that the echo-sounder never has an irregular level of accuracy, which differs for each
got close to the shore in the shallows. It is clearly not a point, depending on that point’s distance to a measured
feature of a specific model, since all remotely sensed models point. It is possible to produce very accurate terrain models
behave very similarly. It should be noted that the lines on with this method, but a vast amount of ground data are
Figures 9 show a larger gap between the end of the lines required.
and the shore than the model actually produce, due to the This study suggests that models based on optical remote
2.5 m filter used in these plots. sensing can produce reasonably good representations of the

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
COMPARISON OF REMOTE SENSING BASED BATHYMETRY MODELS 131

shape of a river bed. The extent to which this holds true actual depths in the river without ground data. This means
depends on the model as well as on the input data. Visual that while the overall relative depths are well represented,
analysis of the various depth maps reveals that, in fact, all the the quality of the absolute values is not clear. Furthermore,
remote sensing based models produce more realistic river the HAB models do not address substrate variability.
bed shapes than a sonar interpolation model (cf. Figures 6 Fonstad and Marcus (2005) experimented with classifying
and 8). Fonstad and Marcus (2005) came to the conclusion the river according to features that can be distinguished from
that the unfavourable statistical results they achieved when aerial photographs, such as riffles and pools, but the different
comparing the HAB model results to measured depths may reflective properties of varying substrate are not addressed in
be a poor indicator of the utility of the technique. They refer their method.
to the ability of the models to reproduce a plausible looking
river bed that is in line with the classic model of an
asymmetric river and does not produce any unexpected
CONCLUSIONS
abrupt jumps in depth. Compared to an interpolation, the
error of the remote sensing based models is spatially even in It was demonstrated that all the tested remote sensing based
the sense that it is not tied to empirical measurement points. models produce a more complete and contiguous overall
Furthermore, it has been noted that the inaccuracy of depth shape of the river bed than an interpolation model would.
of an individual pixel is offset by an enhanced spatial The empirical bathymetric model produced results with
resolution (Bryant and Gilvear, 1999; Carbonneau et al., a level of error that allows its use for practical applications.
2006; Marcus and Fonstad, 2008). In this study, the vertical accuracy of this model was found
to be 12 cm. It was determined that the accuracy of this
Empirical model limitations and improvements model may still be improved by combining it with substrate
analysis. Moreover, it was demonstrated that despite
The discrepancy between the data requirements and the
inadequacies in the empirical data set, including a 2-year
available data for this study has to be noted, especially in
time gap between the collection of the remotely sensed
the case of the empirical model. Even without eliminating
images and the ground data, an underwater terrain model
the time gap between the acquisition of the remotely sensed
of usable accuracy could still be produced.
images and the ground data, an improvement should be
While both HAB models also produced good overall
possible by gathering ground data in contiguous transects
accuracies, they were shown to be sensitive to estimated
without gaps in any depth ranges, thereby decreasing the
input values such as Manning’s n and the river shore
uncertainty of the model results.
reference pixel value selection method in the case of HAB 2.
Given the current limitations of the data available, the
Given the findings of this study, we can not confidently
results achieved based on Lyzenga’s algorithm are quite
answer the question of whether it is possible to produce a
respectable and, at an accuracy of up to 12 cm, useful for
bathymetric model of a river with virtually no ground data,
hydraulic modelling applications. This study suggests that,
as Fonstad and Marcus (2005) set out to achieve. While it
in the case of a relatively stable channel, the previously
is certainly possible to determine relative depths with this
stressed importance of minimising the time gap between
method, an empirically based model has to be favoured if
aerial image and ground data acquisition (Bryant and
absolute depths are required. On the other hand, the HAB
Gilvear, 1999) may be less crucial than the avoidance of
models may very well have their place in change analysis
gaps in the depth spectrum of the empirical data.
studies, where the absolute depths of a stream may be of less
importance than the relative difference in depth between
Theoretical model limitations and improvements
points. This would permit taking advantage of the very low
The principle of computing the attenuation coefficient data requirements of these models and their usability with
such as is implemented in the HAB 2 model seems like a historic data, for which empirical depth data is impossible
promising avenue to follow. The weakness of both HAB to obtain. With a small amount of field data to establish the
models lies not in the method of computing the relation required roughness parameter, the HAB models could get
between the reflectance values and the water depth, but remarkably more reliable results, while still maintaining
rather in the theoretical computation of the river properties lower data requirements than empirical models.
based on hydrological equations and a rather large set of While some active remote sensing applications, such as
assumptions. This is evidenced by the resulting models swathe beam sonar or, in certain conditions, green LiDAR,
following the same pattern, differing merely in overall depth can produce very high vertical and horizontal resolution
by a few centimetres. The sensitivity of the HAB models to river bed geometries, their cost is considerably higher,
the estimated Manning’s roughness variable makes it especially as the area to be modelled increases. Optical
impossible to determine which model is closest to the remote sensing based river bathymetry, while requiring

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
132 C. FLENER ET AL.

some field data for calibration, potentially allows for the Fonstad M, Marcus W. 2005. Remote sensing of stream depths with
mapping of entire river courses at relatively low cost. hydraulically assisted bathymetry (HAB) models. Geomorphology, 72:
320–339. DOI: 10.1016/j.geomorph.2005.06.005.
Gao J. 2009. Bathymetric mapping by means of remote sensing: methods,
accuracy and limitations. Progress In Physical Geography, 33: 103–116.
ACKNOWLEDGEMENTS DOI: 10.1177/0309133309105657.
Gilvear D, Hunter P, Higgins T. 2007. An experimental approach to the
This study was funded by the Academy of Finland (TULe- measurement of the effects of water depth and substrate on optical and
VAT and FINFLOODS research projects), the Maj and Tor near infra-red reflectance: a field-based assessment of the feasibility of
mapping submerged instream habitat. International Journal of Remote
Nessling Foundation (FLOODAWARE research project)
Sensing, 28: 2241–2256. DOI: 10.1080/01431160600976079.
and TEKES (GIFLOOD research project). The aerial pho- Gilvear DJ, Davids C, Tyler AN. 2004. The use of remotely sensed data to
tography data sets were supplied at an institutional price detect channel hydromorphology; River Tummel, Scotland. River
level by Norsk Eiendomsinformasjon AS. Fieldwork assist- Research and Applications, 20: 795–811.
ance was kindly offered by Mr. Lauri Harilainen, Mrs. Laura Gould R, Arnone R. 1997. Remote sensing estimates of inherent optical
properties in a coastal environment. Remote Sensing of Environment, 61:
Koivumäki and Mr. Sampsa Hämäläinen.
290–301.
Hardy TB, Anderson PC, Neale MU, Stevens DK. 1994. Application of
multispectral videography for the delineation of riverine depths and
mesoscale hydraulic features. In: Effects of Human-Induced Change
REFERENCES on Hydrologic Systems, R. Marston, V. Hasfurther (eds). Proceedings
Annual Water Resources Association: Jackson Hole, WY; 445–454.
Alaraudanjoki T, Elster M, Fergus T, Hoset KA, Moen K, Rönkä E, Holden H, LeDrew E. 2002. Measuring and modeling water column effects
Smith-Meyer S. 2001. Tenojoen eroosio, Tenojoen säilyttäminen on hyperspectral reflectance in a coral reef environment. Remote Sensing
luonnontilaisena lohijokena, Raportti A Eroosion ja sedimentin of Environment, 81: 300–308. PII: S0034-4257(02)00007-X.
kulkeutuminen, Tech. Rep. Tnro 1399R0004, Lapin ympäristökeskus. Horritt M, Bates P. 2002. Evaluation of 1D and 2D numerical models for
Alho P, Mäkinen J. in press: Hydraulic parameter estimations of a 2-D predicting river flood inundation. Journal of Hydrology, 268: 87–99.
model validated with sedimentological findings in the point-bar environ- Jordan DC, Fonstad MA. 2005. Two dimensional mapping of river Bathy-
ment, Hydrological Processes. DOI: 10.1002/hyp.7671 metry and power using aerial photography and GIS on the Brazos River.
Alho P, Hyyppä H, Hyyppä J. 2009a. Consequence of DTM precision for Texas, Geocarto International, 20: 13–20.
flood hazard mapping: a case study in SW Finland. Nordic Journal of Kiss T, Sipos G. 2007. Braid-scale channel geometry changes in a
Surveying and Real Estate Research 6: 21–39. sand-bedded river: Significance of low stages. Geomorphology, 84:
Alho P, Kukko A, Hyyppä H, Kaartinen H, Hyyppä J, Jaakkola A. 2009b. 209–221.
Application of boat-based laser scanning for river survey. Earth Surface Kohler DDR, Philpot WD. 1998. Derivative Based Hyperspectral Algor-
Processes and Landforms, 34: 1831–1838. DOI: 10.1002/esp.1879. ithms for Bathymetric Mapping, in Ocean Optics XIV Conference
Bailly J, Le Coarer Y, Allouis T, Stigermark C, Languille P, Adermus J. Kailua-Kona, Hawaii.
2008. Bathymetry with LiDAR on gravel bed-rivers: quality and limits, in Koivumäki L, Alho P, Lotsari E, Käyhkö J, Saari A, Hyyppä H. 2010.
European Geosciences Union (EGU) General Assembly, 14–18 April, Uncertainties in flood risk mapping: a case study on estimating building
Vienna. damages for a river flood in Finland. Journal of Flood Risk Management.
Bates P, Marks K, Horritt M. 2003. Optimal use of high-resolution topo- DOI: 10.1111/j.1753-318X.2010.01064.x.
graphic data in flood inundation models. Hydrological Processes, 17: Lafon V, Froidefond J, Lahet F, Castaing P. 2002. SPOT shallow water
537–557. DOI: 10.1002/hyp.1113. bathymetry of a moderately turbid tidal inlet based on field measure-
Bryant RG, Gilvear DJ. 1999. Quantifying geomorphic and riparian land ments. Remote Sensing of Environment, 81: 136–148. DOI: PII S0034-
cover changes either side of a large flood event using airborne remote 4257(01) 00340-6.
sensing: River Tay, Scotland. Geomorphology, 29: 307–321. Lee Z, Carder KL, Mobley CD, Steward RG, Patch JS. 1999. Hyperspectral
Carbonneau PE, Lane SN, Bergeron N. 2006. Feature based image proces- remote sensing for shallow waters. 2. Deriving bottom depths and water
sing methods applied to bathymetric measurements from airborne remote properties by optimization. Applied Optics, 38: 3831–3843.
sensing in fluvial environments. Earth Surface Processes and Landforms, Legleiter C, Roberts D. 2005. Effects of channel morphology and sensor
31: 1413–1423. DOI: 10.1002/esp.1341. spatial resolution on image-derived depth estimates. Remote Sensing of
Costa J, Spicer K, Cheng R, Haeni P, Melcher N, Thurman E, Plant W, Environment, 95: 231–247. DOI: 10.1016/j.rse.2004.12.013.
Keller W. 2000. Measuring stream discharge by non-contact methods: A Legleiter C, Roberts DA. 2009. A forward image model for passive optical
proof-of-concept experiment. Geophysical Research Letters, 27: 553– remote sensing of river bathymetry. Remote Sensing of Environment, 113:
556. 1025–1045. DOI: 10.1016/j.rse.2009.01.018.
Dankers R. 2002. Sub-arctic hydrology and climate change: a case study of Legleiter C, Roberts D, Marcus W, Fonstad M. 2004. Passive optical remote
the Tana River Basin in Northern Fennoscandia., Ph.D. thesis, Universi- sensing of river channel morphology and in-stream habitat: Physical basis
teit Utrecht. and feasibility. Remote Sensing of Environment, 93: 493–510. DOI:
Feurer D, Bailly J-S, Puech C, Le Coarer Y, Viau AA. 2008. Very-high- 10.1016/j.rse.2004.07.019.
resolution mapping of river-immersed topography by remote sensing. Legleiter C, Roberts D, Lawrence R. 2009. Spectrally based remote sensing
Progress In Physical Geography, 32: 403–419. DOI: 10.1177/ of river bathymetry. Earth Surface Processes and Landforms, 34: 1039–
0309133308096030. 1059.
Flener C. 2009. Rimplementation of HAB bathymerty models, online: Lotsari E, Alho P, Käyhkö J. 2008. The role of datasets in assessing fluvial-
R-forge HAB bathymetry models project, http://hab-bathy.r-forge.r- geomorphic dynamics of a subarctic catchment under changing climatic
project.org. conditions - case studies at Tana river and Pulmankijoki sub-watershed,

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra
COMPARISON OF REMOTE SENSING BASED BATHYMETRY MODELS 133

Northern Finland and Norway, in Geophysical Research Abstracts, vol. Institution of Civil Engineers. Proceedings. Water Management, 159: 35–
10, EGU2008-A-00619. 43.
Lotsari E, Veijalainen N, Alho P, Käyhkö J. 2010. Impact of climate change Newman CM, LeDrew EF. 2001. Assessment of Beer’s Law of logarithmic
on future discharges and flow characteristics of the Tana River, Sub- attenuation for remote sensing of shallow tropical waters. International
Arctic northern Fennoscandia, Geografiska Annaler Series A-Physical Geoscience and Remote Sensing Symposium, 3: 1539–1541.
Geography. 92A: 263–284. DOI: 10.1111/j.1468-0459.2010.00394.x Roberts A, Anderson J. 1999. Shallow water bathymetry using integrated
Lyon J, Lunetta R, Williams D. 1992. Airborne multispectral scanner data airborne multi-spectral remote sensing. International Journal of Remote
for evaluating bottom sediment types and water depths of the St Marys Sensing, 20: 497–510.
river, Michigan. Photogrammetric Engineering and Remote Sensing, 58: Sagawa T, Komatsu T, Boisnier E, Mustapha KB, Hattour A, Kosaka N,
951–956. Miyazaki S. 2007. A New Applicaton Method for Lyzenga’s Optical
Lyzenga DR. 1981. Remote sensing of bottom reflectance and water Model, in Proceedings of the 2007 Annual Conference of the Remote
attenuation parameters in shallow water using aircraft and Landsat data. Sensing & Photogrammetry Society (RSPSoc2007), Newcastle upon
International Journal of Remote Sensing, 2: 71–82. Tyne.
Lyzenga DR, Malinas NR, Tanis FJ. 2006. Multispectral bathymetry using a Sandidge J, Holyer R. 1998. Coastal bathymetry from hyperspectral obser-
simple physically based algorithm. IEEE Transactions On Geoscience vations of water radiance. Remote Sensing of Environment, 65: 341–352.
and Remote Sensing, 44: 2251–2259. DOI: 10.1109/TGRS.2006.872909. Schäppi B, Perona P, Schneider P, Burlando P. 2008. Integrating river
Mansikkaniemi H. 1972. Flood deposits, transport distancies and roundness bathymetry data with Digital Terrain Models for improved river flow
of loose material in the Tana river valley, Lapland, Turun yliopiston simulations, in EGU General Assembly 2008, edited by A. Richter, vol.
julkaisuja. Sarja A. II. Biologica - Geographica - Geologica, 49: 15–23. 10 of Geophysical Research Abstracts, European Geosciences Union.
Marcus W, Legleiter C, Aspinall R, Boardman J, Crabtree R. 2003. High USGS. 2008. United states Geological Survey: Verified Roughness Charac-
spatial resolution hyperspectral mapping of in-stream habitats, depths, teristics of Natural Channels, [online], http://wwwrcamnl.wr.usgs.gov/
and woody debris in mountain streams. Geomorphology, 55: 363–380. sws/fieldmethods/Indirects/nvalues/.
DOI: 10.1016/S0169-555X(03) 00150-8. Wahba G. 1990. Spline models for observational data, in CBMS-NSF
Marcus WA. 2002. Mapping of stream microhabitats with high spatial Regional Conference Series in Applied Mathematics, vol. 59, SIAM,
resolution hyperspectral imagery. Journal of Geographical Systems, 4: Philadelphia.
113–126. Westaway R, Lane S, Hicks D. 2003. Remote survey of large-scale braided,
Marcus WA, Fonstad MA. 2008. Optical remote mapping of rivers at sub- gravel-bed rivers using digital photogrammetry and image analysis.
meter resolutions and watershed extents. Earth Surface Processes and International Journal of Remote Sensing, 24: 795–815. DOI: 10.1080/
Landforms, 33: 4–24. 01431160110113070.
McKean J, Tonina D, Bohn C, Wright C. 2008. Effects of Bathymetric Lidar Winterbottom SJ, Gilvear DJ. 1997. Quantification of channel bed
Errors on Hydraulic Models, AGU Fall Meeting Abstracts, pp. H867þ. morphology in gravel-bed rivers using airborne multispectral imagery
Milne JA, Sear DA. 1997. Modelling river channel topography using GIS. and aerial photography. Regulated Rivers-Research & Management, 13:
International Journal of Geographical Information Science, 11: 499– 489–499.
519. Wright A, Marcus W, Aspinall R. 2000. Evaluation of multispectral, fine
Néelz S, Pender G, Villanueva I, Wilson M, Wright NG, Bates P, Mason D, scale digital imagery as a tool for mapping stream morphology. Geo-
Whitlow C. 2006. Using remotely sensed data to support flood modelling, morphology, 33: 107–120.

Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. 28: 118–133 (2012)
DOI: 10.1002/rra

You might also like