WATER RESOURCES RESEARCH, VOL. 49, 3165–3179, doi:10.1002/wrcr.

20176, 2013

Estimation of river depth from remotely sensed hydraulic

relationships
Matthew K. Mersel,1,2 Laurence C. Smith,1,3 Konstantinos M. Andreadis,4 and Michael T. Durand5,6
Received 13 October 2012; revised 26 February 2013; accepted 2 March 2013; published 4 June 2013.

[1] The Surface Water and Ocean Topography (SWOT) radar interferometer satellite mission
will provide unprecedented global measurements of water surface elevation (h) for inland water
bodies. However, like most remote sensing technologies SWOT will not observe river channel
bathymetry below the lowest observed water surface, thus limiting its value for estimating river
depth and/or discharge. This study explores if remotely sensed observations of river inundation
width and h alone, when accumulated over time, may be used to estimate this unmeasurable
flow depth. To test this possibility, synthetic values of h and either cross-sectional flow width
(w) or effective width (We, inundation area divided by reach length) are extracted from 1495
previously surveyed channel cross-sections for the Upper Mississippi, Illinois, Rio Grande, and
Ganges-Brahmaputra river systems, and from 62 km of continuously acquired sonar data for
the Upper Mississippi. Two proposed methods (called ‘‘Linear’’ and ‘‘Slope-Break’’) are tested
that seek to identify a small subset of geomorphically ‘‘optimal’’ locations where w or We
covary strongly with h, such that they may be usefully extrapolated to estimate mean cross-
sectional flow depth (d). While the simplest Linear Method is found to have considerable
uncertainty, the Slope-Break Method, identifying locations where two distinct hydraulic
relationships are identified (one for moderate to high flows and one for low flows), holds
promise. Useful slope breaks were discovered in all four river systems, ranging from 6 (0.04%)
to 242 (16%) of the 1495 studied cross-sections, assuming channel bathymetric exposures
ranging from 20% to 95% of bankfull conditions, respectively. For all four rivers, the derived
depth estimates from the Slope-Break Method have root mean squared errors (RMSEs) of
<20% (relative to bankfull mean depth) assuming at least one channel bathymetry exposure of
25% or greater. Based on historic discharge records and HEC-RAS hydraulic modeling, the
Upper Mississippi and Rio Grande rivers experience adequate channel exposures at least 60%
and 42% of the time, respectively. For the Upper Mississippi, so-called ‘‘reach-averaging’’
(spatial averaging along some predetermined river length) of native-resolution h and We values
reduces both RMSE and longitudinal variability in the derived depth estimates, especially at
reach-averaging lengths of 1000–2000 m. These findings have positive implications for
SWOT and other sensors attempting to estimate river flow depth and/or discharge solely from
incomplete, remotely sensed hydraulic variables, and suggest that useful depth retrievals can be
obtained within the spatial and temporal constraints of satellite observations.
Citation: Mersel, M. K., L. C. Smith, K. M. Andreadis, and M. T. Durand (2013), Estimation of river depth from remotely sensed
hydraulic relationships, Water Resour. Res., 49, 3165–3179, doi:10.1002/wrcr.20176.

1
Department of Geography, University of California, Los Angeles, Cali- 1. Introduction
fornia, USA.
2
Now at U.S. Army Cold Regions Research and Engineering Laboratory [2] Terrestrial runoff to rivers is a significant term in the
(CRREL), Hanover, New Hampshire, USA.
3
global water balance and a principle source of fresh water
Department of Earth and Space Sciences, University of California, Los for human and ecosystem use [Vörösmarty et al., 2010], yet
Angeles, California, USA.
4
Jet Propulsion Laboratory, California Institute of Technology, Pasa-
global knowledge of the spatial and temporal dynamics of
dena, California, USA. river flow is surprisingly poor [Alsdorf et al., 2007b;
5
School of Earth Sciences, The Ohio State University, Columbus, Ohio, Durand et al., 2010a]. Stream gages, the traditional method
USA. for measuring discharge, are in decline globally [Stokstad,
6
The Byrd Polar Research Center, The Ohio State University, Colum- 1999; Shiklomanov et al., 2002], and where gages do exist,
bus, Ohio, USA.
these data are not always shared among countries or agen-
Corresponding author: M. K. Mersel, Cold Regions Research and Engi- cies. Furthermore, gages are inherently limited to providing
neering Laboratory, RS/GIS 103, 72 Lyme Rd., Hanover, NH 03755,
USA. ([email protected])
information only at single points along a river and fail to
capture three-dimensional dynamics of fluvial systems
©2013. American Geophysical Union. All Rights Reserved. including overbank flow, flood waves, and multichannel
0043-1397/13/10.1002/wrcr.20176 flow. These limitations, combined with rising worldwide

3165
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

stress on river systems owing to industrialization, population measurements of terrestrial and ocean water surface eleva-
growth, and climate change [United Nations Educational, tions globally with high spatial resolution. Through instanta-
Scientific and Cultural Organization, 2012], motivate devel- neous detection of both h and A, SWOT will acquire repeated
opment of new approaches for understanding river dynamics DEMs of terrestrial water surfaces wider than 100 m.
from satellite observations that would compliment ground- Repeat-pass measurements of h and A will essentially map
based measurements. any exposed portion of a river channel’s bathymetry between
[3] Remote sensing of rivers is a relatively immature but the highest and lowest water surfaces encountered over time.
rapidly emerging subdiscipline within hydrology that is Additionally, topographic mapping of fully exposed river ba-
advancing new approaches to the study of fluvial systems thymetry (i.e. any portion of the river channel that happens to
[Smith, 1997; Alsdorf et al., 2007b; Durand et al., 2010a; be fully dry at the time of one or more satellite overpasses)
Marcus and Fonstad, 2010]. Furthermore, the unique spa- and adjacent floodplains will be achieved using standard
tial perspective afforded from satellite and aircraft sensors interferometric techniques (in the case of SWOT, multiple
allows for observation and understanding of rivers in ways overpasses would likely be required due to lower signal-to-
both infeasible and fundamentally different from traditional noise ratio over land). Owing to the very bright returns of
ground-based methods. One approach is the use of profiling near-nadir Ka band radar over water, instantaneous measure-
oceanographic radar altimeters to retrieve measurements ments of A, h, and water surface slope will be obtained glob-
of water surface elevation (h) along transects where orbit ally at least once every 11 days, and near-daily at high
paths cross water bodies [e.g., Koblinsky et al., 1993; latitudes.
Birkett, 1998; Birkett et al., 2002; Frappart et al., 2005; [6] While global satellite measurements of A, h and
Cretaux and Birkett, 2006; Calmant et al., 2008; Birkett water surface slope (dh/dx) will have numerous scientific
and Beckley, 2010; Lee et al., 2011]. However, such tech- and practical applications, they do not permit direct calcu-
niques are generally limited to lakes, reservoirs, and very lation of river discharge (m3/s). Global knowledge of river
large rivers and, like stream gages, are inherently point discharge would, for instance, greatly improve knowledge
based. Other studies have mapped spatial variations in river of the availability and fluxes of surface water, especially in
inundation area (A) as a proxy for changing stage or dis- remote or developing regions where few river gages exist.
charge, using visible/near-infrared or synthetic aperture ra- However, unless the full channel bathymetry is either inde-
dar (SAR) backscatter imagery [e.g., Smith et al., 1995, pendently known or is adequately observed over the SWOT
1996; Prigent et al., 2001; Brakenridge et al., 2005; Papa mission lifetime (i.e., the river is repeatedly imaged while
et al., 2006; Smith and Pavelsky, 2008; Khan et al., 2011]. completely dry), the depth of river flow below the lowest
However, these methods typically require ancillary data (e.g., observed h will remain unknown. This unknown residual
from stream gages or digital elevation models (DEMs)) and flow depth is a critical obstacle to estimating river discharge
are most effective for rivers where discharge fluctuations are using SWOT and other satellite technologies.
accommodated largely by width adjustments (e.g., braided riv- [7] Alternate approaches for remote sensing of river
ers and overbank flows in floodplain systems). Furthermore, depth are few in number. Several studies have exploited the
cloud and vegetation cover limits the use of optical sensors, attenuation of bottom reflectance in the water column using
while SAR backscatter techniques are limited by difficulties optical imagery [e.g., Legleiter et al., 2004, 2009; Marcus
related to wind roughening of the water surface [Smith, 1997]. and Fonstad, 2008; Legleiter and Roberts, 2009], but spec-
[4] The preceding approaches measure either point- tral scattering from suspended solids limits this approach to
based h, or spatially varying A, but not both. An exciting clear, shallow streams where the channel bottom can be
development in satellite remote sensing of river hydraulics seen. Others have used data assimilation techniques to
is 3-D imaging, first advanced using repeat-pass interfero- combine simulated observations of h with a hydrodynamic
metric SAR (InSAR) to measure relative changes in h over model to solve for depth and discharge simultaneously
time and space [e.g., Alsdorf et al., 2000, 2001, 2007a; Lu [e.g., Andreadis et al., 2007; Durand et al., 2008; Bianca-
et al., 2005; Jung and Alsdorf, 2010; Jung et al., 2010]. maria et al., 2011], while another method [Durand et al.,
This particular method can detect temporal changes in h to 2010b] estimates stream depth using an algorithm based on
a vertical precision of centimeters, but requires inundated the Manning equation. Although these last two approaches
woody vegetation for signal return and only measures rela- show considerable promise, limitations of such model-
tive changes (i.e., height anomalies, dh/dt) over time, not h. based approaches include their computational expense and
Hydraulic surface slopes, therefore, cannot be mapped with reliance on parameterization data that are not always
this method. Furthermore, while repeat-pass InSAR dem- available.
onstrates the vast potential of remote sensing for studying [8] Few studies have explored the river depth-estimation
3-D river dynamics, no sensor currently exists to quantify problem from a purely empirical standpoint. Using a large
terrestrial surface water elevations, storages, and fluxes data set compiled from U.S. Geological Survey cross-sec-
globally over time and space [Alsdorf et al., 2007b; tions, Bjerklie [2007] developed a simple regression equa-
Durand et al., 2010b]. tion to estimate bankfull mean depth (dbf, the mean depth at
[5] The Surface Water and Ocean Topography (SWOT) bankfull discharge) from observed values of bankfull width
satellite, a joint project between NASA and the French and channel slope. This equation produced a large standard
space agency, the Centre National d’Etudes Spatiales error of 58%, leading the authors to call for ‘‘improved
(CNES), is planned for launch in 2020 and has strong poten- methods to estimate bankfull depth from observed varia-
tial to overcome many of the aforementioned limitations bles’’ [Bjerklie, 2007].
(http://swot.jpl.nasa.gov). SWOT will use a Ka-band inter- [9] One such method could be to exploit empirical rela-
ferometric wide-swath altimeter, to provide unprecedented tionships between covarying, interrelated hydraulic variables

3166
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

(e.g., width w, depth d, and velocity v), where stable rela- [12] Even at a single cross-section, temporal variability
tionships exist, to estimate one variable (i.e., d) using in HG relationships is common, particularly when compar-
another. Identification of these empirical relationships has ing high- versus low-flow conditions [Lewis, 1966; Rich-
been a central theme in fluvial geomorphology for decades, ards, 1976; Phillips, 1990; Jowett, 1998]. Indeed, Lewis
both at single locations (called ‘‘at-a-station hydraulic geom- [1966] highlights substantial ‘‘slope breaks’’ in HG power-
etry (HG)’’) and longitudinally along a river (‘‘downstream law relationships that often occur at low flows, suggesting
HG’’). Classic HG theory formulates the empirical relation- that remotely sensed HG relationships calibrated at moder-
ships of w, d, and v with discharge (Q), expressed as the ate to high flow levels would thus lead to large errors if ex-
simple power functions w¼aQb, d¼cQf, and v¼kQm, where a, trapolated to low flows. Instead of a single empirical
b, c, f, k, and m are empirical constants unique to a particular function existing relating each hydraulic variable (w, d, and
river or cross-section and determined through accumulation v) to discharge, his study found that in-channel HG relation-
of field measurements [Leopold and Maddock, 1953]. ships are often best described by two functions, one for
Because Q¼wdv, the exponents b, f, and m describe the moderate to high flows and one for low flows [Lewis, 1966].
‘‘trade-offs’’ between flow width, depth, and velocity (i.e., Similarly, from a remote sensing perspective, a We-h rela-
bþfþm¼1), controlled mainly by the unique bathymetric tionship derived from satellite observations during moderate
shape of a stream channel at a given location (for at-a-station to high flows might be expected to lead to large errors if ex-
HG). Indeed, where stable at-a-station HG relationships exist, trapolated to estimate river depth d. If a low-flow We-h rela-
they form the basis for traditional stream gage discharge esti- tionship could be detected, however (again, from many
mates, through construction of an empirical h-Q relationship observations of We and h accumulated over time), its extrap-
(rating curve) correlating continuous measurements of h to olation might be better suited for estimating d.
occasional in situ measurements of Q. For locations where [13] The objective of this study is to explore the feasibil-
discharge adjustments are depth sensitive (i.e., changes in Q ity of estimating unobservable mean depth d (either cross-
are significantly accommodated by adjustments in depth) and sectional mean depth or reach-averaged mean depth) from
for flows confined within the channel banks, h thus becomes a remotely sensed measurements of water surface elevation
reliable proxy for discharge. h, cross-sectional flow width w, or reach-scale effective
[10] Stable at-a-station HG relationships have long been width We alone. To do this, we extract synthetic values of
observed in field measurements, and it now appears possible h, and w from 1495 field surveyed river cross-sections com-
to also detect them from space [Smith and Pavelsky, 2008]. piled for the Ganges-Brahmaputra, Rio Grande, Illinois,
This suggests that remotely sensed A (or more properly and Upper Mississippi river systems; and values of h and
remotely sensed ‘‘effective width’’ We, which is A divided by We from a sonar-derived, continuously gridded (5 m resolu-
some defined reach length, Smith et al. 1995, 1996) and h, tion) bathymetric data set for an overlapping 62 km reach
such as will be acquired by SWOT, will be useful for estimat- of the Upper Mississippi. For the Rio Grande and Upper
ing certain hydraulic properties of river channels. In particu- Mississippi rivers only, synthetic values of h and w are also
lar, for those locations along a river where a stable generated using the U.S. Army Corps of Engineers (USACE)
relationship between h and We is detected in satellite observa- Hydraulic Engineering Center––River Analysis System 1-D
tions accumulated over a range of flow conditions, this rela- hydraulic model (HEC-RAS) version 4.1.0 (http://www.
tionship might then be extrapolated to the unobserved portion hec.usace.army.mil/software/hec-ras/), thus allowing the
of the river channel to estimate the lowest channel bed eleva- results for these two rivers to also be presented in terms of
tion (zmin), from which mean flow depth d is easily computed. temporal discharge exceedance probability [Dingman, 1994].
[11] In principle, even discharge Q could be estimated if While an essentially unlimited number of channel geome-
all of the necessary HG coefficients can be determined. A tries exist in nature, preliminary assessment of the acquired
key uncertainty in adapting at-a-station HG to the remote cross-section and bathymetry data sets revealed two simple
sensing context, however, is that numerous studies have recurring relationships between w (or We) and h. These two
shown that HG coefficients are highly variable along natu- w-h (or We-h) relationships were found in all five data sets to
ral river courses [e.g., Leopold and Maddock, 1953; Rich- varying degrees despite substantial differences with respect
ards, 1973; Knighton, 1975; Park, 1977] meaning that to channel geometry, location, and river type. Based on this
such functions are site specific and not transferable to other preliminary assessment of the data, the generated data sets of
locations. Furthermore, this heterogeneity introduces the w versus h (or We versus h) are then used to test two methods
key question of how empirical HG relationships, tradition- for estimating d by exploiting these two simple satellite-
ally obtained only at surveyed river cross-sections, are observable w-h relationships. The ‘‘Linear Method’’ extrapo-
influenced by ‘‘reach-averaging’’ (spatial averaging over lates w-h or We-h relationships to estimate d only for those
some defined reach length) that is inherently necessary locations where a single, strongly linear correlation between
using remote-sensing technologies [Smith and Pavelsky, these two variables is observed, across a range of different
2008]. SWOT radar echoes, in particular, must be spatially channel exposures (i.e., down to some minimum low-flow h).
averaged over reach lengths of hundreds to thousands of The ‘‘Slope-Break Method’’, motivated by Lewis [1966],
meters (depending on the width of the river or water body) extrapolates w versus h (or We versus h) relationships to esti-
in order to improve their associated measurement preci- mate d only for those locations where two strongly linear cor-
sions of h [Durand et al., 2010a]. The implications of this relations are observed, one for moderate to high flows and
level of reach-averaging on empirical HG relationships are one for low flows. Both methods are then assessed as to the
not well understood, with a bare handful of studies examin- abundance of locations identified and the corresponding qual-
ing this question to date [Stewardson, 2005; Smith and ity of their extrapolated depth estimates. Finally, the influ-
Pavelsky, 2008; Fonstad and Marcus, 2010]. ence of reach-averaging is examined using the continuously

3167
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

gridded sonar data set for the upper Mississippi River, to shifting islands. For the Rio Grande River, a total of 150
examine a range of spatially averaged We-h relationships that cross-sections were surveyed by Tetra Tech, Inc. along
could realistically be sampled by SWOT or other imaging 172 km from the Caballo Dam in southern New Mexico to
remote sensing technologies. the American Dam near the U.S./Mexico border (Figure 1).
This part of the Rio Grande is a semiarid environment, sur-
2. Methods rounded primarily by farmland, and regulated by the
upstream Caballo Dam and reservoir. Mean discharge is
2.1. Data and Study Areas 20 to 30 m3/s with channel widths ranging from 30 to
[14] A total of 1495 previously acquired field-survey 130 m. For the Illinois River, 482 cross-sections were sur-
cross-sections were compiled from various sources for six veyed along 338 km by the USACE, in an area of temper-
tributary rivers of the Ganges-Brahmaputra system, Ban- ate climate, surrounded mostly by farmland, and regulated
gladesh (http://www.iwmbd.org), and reaches of the Rio by a series of locks and dams. Mean discharge down this
Grande River, USA [Tetra Tech Inc., 2005], Illinois River, reach increases from approximately 280 to 1000 m3/s,
USA [U.S. Army Corps of Engineers, 2004], and Upper with channel widths ranging from 70 to 3300 m. For the
Mississippi River, USA [U.S. Army Corps of Engineers, Upper Mississippi, a total of 639 cross-sections were sur-
2004]. Additionally, continuous gridded bathymetric data veyed by the USACE over 315 km, upstream of its conflu-
overlapping 62 km of the Upper Mississippi cross-section ence with the Ohio River. The Upper Mississippi lies in an
data set were obtained (http://www.umesc.usgs.gov/aquatic/ area of temperate climate, is surrounded largely by farmland,
bathymetry/download.html). Viewed collectively, these four and is also regulated by a series of locks and dams. Dis-
study areas represent a wide range of rivers in terms of their charge along this reach averages between 6000 and 7000
size (50–18,000 m wide) and discharge (1–50,000 m/s3). m3/s and channel widths range from 200 to 2300 m. An
The locations of these data sets and study areas are shown in overlapping, continuously gridded (5 m  5 m) bathymetric
Figure 1. The total number of cross-sections, as well as the data set, obtained primarily from depth soundings and sup-
minimum, mean, and maximum values of bankfull width plemented with manual measurements using a calibrated
(wbf) and bankfull mean depth (dbf) for each of the five data sounding pole was also obtained for a 62 km river reach of
sets acquired are listed in Table 1. this site. This continuously gridded bathymetric data set was
[15] Each cross-section in the compiled database consists adjusted to a constant reference water surface by the devel-
of a single transect of x and z values (distance perpendicu- opers of the data set, thus removing water surface slope
lar to the direction of flow and distance above sea level, while preserving the shape of the channel [Rogala, 1999].
respectively) surveyed at a particular location along the While a comprehensive assessment of the accuracies of
river. For the Ganges-Brahmaputra system, a total of 224 these data sets and of the data collection techniques used to
cross-sections were surveyed by the Bangladesh Institute of derive them is beyond the scope of this study, accuracies are
Water Modeling (IWM) along seven tributary rivers (Brah- those typically associated with ADCP, LiDAR, and sonar
maputra, Ganges, Jamuna, Padma, Surma, Upper Meghna, measurements.
and Lower Meghna rivers) (Figure 1). This river system
constitutes one of the largest in the world, with a mean dis- 2.2. Extraction of Synthetic Width and h Values
charge of 40,000 m3/s and channel widths occasionally [16] Synthetic values of water surface width w (for
exceeding 10 km. Much of it is anastomosing, with wide, cross-sections) and effective width We (for the gridded
multithreaded channels interspersed with permanent and Upper Mississippi data) and h were extracted to test two

Figure 1. Location map of study areas and field data sets. 1495 intermittent surveyed cross-sections
were obtained from various sources along those river reaches highlighted in red. Blue box highlights a
62km reach of the Upper Mississippi for which gridded bathymetric data were also acquired. Black
symbols mark the locations of USGS stream gages used for calculation of discharge statistics.

3168
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

Table 1. Total Number of Cross-Sections, and Minimum, Mean, and Maximum Values of Bankfull Width (wbf) and Bankfull Mean
Depth (dbf) From Each of Four River Cross-Section Data Sets and One Gridded Bathymetric Data Set
Number of
River cross-sections Min. wbf Mean wbf Max. wbf Min. dbf Mean dbf Max. dbf

Upper Mississippi 639 240.89 798.37 2313.01 2.20 7.30 12.69

Illinois 482 73.81 414.99 3293.86 0.59 4.59 12.56
Rio Grande 150 37.19 71.68 130.87 0.60 1.71 3.60
Ganges-Brahmaputra 224 45.04 6945.60 17844.50 0.43 4.26 11.25
Upper Mississippi (gridded) 12788a 515.00 1410.37 3094.03 1.01 3.72 12.74
a
When extracted every 5 m from gridded bathymetric data set.

methods for estimating d. We assume that h refers to geo- section, and range from nearly 100% channel exposure in
detic water elevation above sea level. Cross-sections near some locations (at 99% exceedance probability) to nearly
bridges or dams were excluded from subsequent analysis. 0% channel exposure (at 40% exceedance probability).
These values of w, We, and h, treated as error-free synthetic [19] To study the effects of reach-averaging on remotely
SWOT retrievals for the purposes of this analysis, were sensed hydraulic relationships, synthetic values of We and
generated as follows. reach-averaged h were extracted from the gridded Upper
[17] For each of the 1495 cross-sections, values of w and Mississippi bathymetric data set as follows. First, these
h were extracted at different channel depths, corresponding data were discretized into 12,788 5 m sections using the
to percentages of bankfull mean depth dbf ranging from USACE HEC-GeoRAS extension for ArcGISV R . Each sec-

100% to 5% in increments of 5%. dbf was visually deter- tion thus represents the average bathymetry for the shortest
mined at each cross-section by selecting the highest sur- possible reach length (5 m) permitted by this high-resolu-
veyed elevation that could be confidently determined as in- tion data set. Synthetic pairs of We and h were then
channel. At locations where multiple channels were extracted from all 12,788 sections in much the same man-
encountered along the same surveyed transect (primarily ner as for the cross-section database, except the reference
the anastomosing Ganges-Brahmaputra river system) all dbf was defined as the highest recorded elevation for each
channels were treated as a single cross-section. These section instead of from visual inspection (unlike the sur-
derived pairs of w and h allow for exploration of w-h rela- veyed cross-sections, the bathymetric data do not extend
tionships as a function of percentage ‘‘channel exposure’’, onto the river floodplain). Synthetic pairs of We and h were
i.e., the percentage of the channel area lying above the extracted from each section at water levels corresponding
waterline for a given h. to percentages of dbf ranging from 100% to 5%, then spa-
[18] To estimate the probability that a required quantity tially averaged for reach-lengths of 50, 100, 500, 1000,
of channel exposure might actually occur at least once dur- 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, and
ing a nominal satellite mission lifetime (e.g., 3 years for 10,000 m. The mathematical mean of h (for a given per-
SWOT), a second data set of synthetic w-h pairs was gener- centage of dbf) was taken from all cross-sections within a
ated using HEC-RAS, to simulate river discharges along given spatial-averaging length. The spatial-averaging
the Rio Grande and Upper Mississippi rivers. The neces- ‘‘window’’ was then shifted 5 m and the mean taken again.
sary inputs and calibration parameters for these particular Thus, fewer reach-averaged sections were extracted as
river reaches were previously defined by their respective reach-averaging length increased. Unlike traditional sur-
developers [Tetra Tech Inc., 2005; U.S. Army Corps of veyed cross-sections, these synthetic, reach-averaged We
Engineers, 2004]. For the Upper Mississippi River, this and h values derived from a continuous bathymetric data
modeling was restricted to the downstream 430 of 639 set enable assessment of how spatial averaging of remotely
cross-sections, owing to excessive influence of man-made sensed observations may influence hydraulic We-h relation-
structures on the HEC-RAS simulations. For each river, ships in natural river systems.
steady-state discharges ranging from 99% to 40% exceed-
ance probability (the percentage of time that a given dis- 2.3. Linear Method for Depth Estimation
charge is exceeded based on historic streamflow records) [20] The simplest approach to estimating mean flow
were simulated and the corresponding values of w, h, and d depth below some lowest observed water level (hmin) seeks
extracted from the surveyed cross-sections. These exceed- river locations with a strongly linear relationship between
ance probabilities were determined using 10 years (2001– all pairs of w (or We) and h lying above hmin, and assumes
2010) of daily discharge records from four USGS gaging this relationship may be extrapolated below that level down
stations along the Upper Mississippi River and one along to w¼0 to obtain zmin (e.g., Figure 2, top row).
the Rio Grande River (Figure 1). A 99% exceedance proba- [21] To implement this, the derivatives (approximated by
bility roughly corresponds with the lowest magnitude dis- finite forward differences) of h with respect to w (dh/dw)
charge, and thus the greatest percentage of channel were first calculated between all simulated in-channel
exposure, that could potentially be encountered over a mis- water surfaces (i.e., between each two adjacent water surfa-
sion lifetime, while a 40% exceedance probability roughly ces) above a given hmin for each cross-section in the com-
corresponds with the highest magnitude discharge that piled database. Those cross-sections where all observed
remains in-channel at all cross-sections along each river. values of dh/dw were within 60.015 of each other were
Thus, the channel exposure percentages corresponding with flagged. This value was determined from trial-and-error so
these exceedance probabilities are unique to each cross- as to select for locations with strong w-h relationships

3169
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

Figure 2. (a and c) Two conceptual, idealized river bathymetric cross-sections and (b and d) their
respective plots of w versus h for a range of hypothetical flow levels. A single linear w-h relationship
(Figure 2b) exists for cross-section (Figure 2a), whereas two linear relationships, one for moderate to
high flows and one for low flows (Figure 2d) exist for cross-section (Figure 2c).

while still retaining a sufficient number of cross-sections be discerned, may improve the derived estimates of zmin
from the sample pool for which to estimate depth. For all and dest. To test this idea, a second depth-estimation algo-
locations satisfying this criterion (called ‘‘optimal’’ loca- rithm was developed that selects for slope-break locations
tions), the mean of the derivatives of h with respect to only where a second, distinct low-flow width-h relationship
w ðdh=dwÞ for all water surfaces above hmin was extrapo- can also be detected, which is then extrapolated as before
lated to compute the minimum channel elevation (zmin), the to estimate zmin and dest. As such, this method seeks partic-
elevation for w¼0. Next, zmin was subtracted from hmin to ular locations where two stable width-h relationships are
compute the maximum channel depth (dmax) then halved to found instead of one (e.g., Figure 2, bottom). The term
compute an estimate of d (dest) (given the perfectly linear ‘‘slope break’’ refers to the break in slope of a line fit
relationship between width and h that this method assumes, through a scatterplot of h versus w (or We) at locations
dest ¼ dmax /2 for any given h). An example of an optimal where this requirement is met (e.g., Figure 3, bottom). This
location detected by the Linear Method is displayed in slope break defines the h at which the set of hydraulic rela-
Figure 3 (top). tionships defined for moderate to high flows is replaced by
[22] The Linear Method was tested for all cross-sections a new set of hydraulic relationships defined for low flows.
in the compiled database. For those cross-sections satisfy- [24] For every cross-section (or reach-averaged section,
ing the algorithm’s requirements (optimal locations), the for continuously gridded We) for each river, dh/dw was first
resultant depth estimates dest were compared with true d calculated for all synthetic water surfaces above a given
(for a given hmin) to assess the accuracy of the approach. hmin (i.e., between each two adjacent water surfaces), as for
This process was repeated for a range of possible hmin val- the Linear Method. dh=dw was calculated from the four
ues (i.e., from 80% to 5% of dbf), to assess performance of highest water surfaces at each cross-section (the number of
the Linear Method from low- to high-channel bathymetry observations for calculating dh=dw was arbitrarily chosen ;
exposure. this value had little effect on the results) and was then used
to compare each subsequent (i.e., at lower elevation) value
2.4. Slope-Break Method for Depth Estimation of dh/dw. If a subsequent value of dh/dw deviated suffi-
[23] An obvious limitation of the Linear Method is that ciently from dh=dw (a dh/dw value <0.3 dh=dw), the
hydraulic relationships determined at moderate to high cross-section was flagged as having a slope break. Other-
flows are often inappropriate for low flows, owing to the wise, dh=dw was recalculated to include the subsequent
presence of HG ‘‘slope breaks’’ as described earlier [Lewis, value of dh/dw and the process continued until all values
1966]. This suggests that restricting the described linear were compared. Note that the value 0.3, which defines the
extrapolation to a low-flow width-h relationship, if one can threshold for what constitutes a slope break, was chosen

3170
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

Figure 3. Examples of two field-surveyed cross-sections (a and c) identified as ‘‘optimal’’ by the Lin-
ear and Slope-Break methods, respectively. For cross-section (Figure 3a) the in-channel w-h relationship
(Figure 3b) remains strongly linear with depth, and would thus be suitable for remote estimation of zmin
and d based on remotely sensed observations of w and h alone using the Linear Method. In contrast, the
channel geometry of cross-section (Figure 3c) yields a w-h plot (Figure 3d) with two distinct linear rela-
tionships, one for moderate to high flows and one for low flows. If water levels drop sufficiently such
that the slope-break becomes observable by satellite (>70% channel exposure required, for this particu-
lar cross-section), the low-flow w-h relationship may be usefully extrapolated using the Slope-break
Method to estimate zmin and d based on remotely sensed observations of w and h alone.

through simple trial and error, but the value of this thresh- the Rio Grande and Upper Mississippi rivers. Finally, the
old did not have much effect on results. For cross-sections Slope-Break Method was tested on synthetic measurements
where a slope break was detected, dest was only estimated extracted from the Upper Mississippi gridded bathymetric
at those locations where all values of dh/dw below the slope data set for a range of reach-averaging length scales.
break and above hmin remained consistent with each other
(i.e., within 60.015, as for the Linear Method). Extrapola- 3. Results
tion of dh=dw to w ¼ 0 below this break was then used to
estimate zmin and dest as before. An example of an optimal 3.1. Depth Estimates at Channel Cross-Sections Using
location detected by the Linear Method is displayed in Fig- the Linear Method
ure 3 (bottom). [26] Standard and root mean squared errors (RMSEs) for
[25] The Slope-Break Method was tested for all cross- mean depth retrievals estimated using the Linear Method
sections in the compiled database. For those cross-sections are shown in Figure 4, varying as a function of percent
satisfying the algorithm’s criteria (optimal locations) result- channel exposure. Each blue symbol represents the stand-
ant values of dest were compared with true d (for a given ard error of the derived dest (percent error relative to dbf) at
hmin) to assess the accuracy of this approach. As with the any location deemed optimal (i.e., having a single strongly
Linear Method, this process was repeated for a range of linear w-h relationship) for a given channel exposure. Algo-
possible hmin values (i.e., from 80% to 5% of dbf), to assess rithm performance is highly site specific, with standard errors
how width-h relationships might vary as a function of river in depth retrieval varying several orders of magnitude regard-
channel exposure. Results were compared with those of the less of channel exposure (blue symbols, Figure 4). Viewed
Linear Method for each river (Figure 6). This initial com- collectively, the RMSE values (red symbols) for cross-sec-
parison of the two methods made clear that the Slope-Break tion exposures ranging from 95% (high exposure) to 20%
Method outperformed the Linear Method for all rivers, thus (low exposure) for all optimal locations detected range from
further testing was limited to the Slope-Break Method 11% to 302%, 18% to 377%, 760% to 1288%, and 7% to
alone. The Slope-Break Method was further tested on 109% for the Upper Mississippi, Illinois, Rio Grande, and
HEC-RAS-generated water surface profiles for reaches of Ganges-Brahmaputra river systems, respectively. Both the

3171
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

Figure 4. Standard and RMS errors for ‘‘optimal’’ locations for remote depth estimation, identified by
applying the Linear Method to synthetic values of w and h extracted from 1495 cross-sections from four
river systems. Errors are generally high (note log-normal scale). Each blue plus symbol represents the
standard error (percent error in dest relative to dbf) for a single optimal location detected for a given per-
centage of channel exposure. The RMSE for all estimates for a given percentage of channel exposure is
plotted in red. In all four river systems, errors tend to improve but the total number of optimal locations
detected tends to decrease with increasing channel exposure. Note that for the Rio Grande, no optimal
locations are detected when channel exposure exceeds 70%.

RMSE values and the number of optimal locations detected by the Slope-Break Method (Figure 5) as compared with
tend to decrease with increasing channel exposure for all the Linear Method (Figure 4). As with the Linear Method,
four river systems (the reasons for this are discussed in sec- RMSE values tend to improve with increasing channel ex-
tion 4). For example, at 20% channel exposure 942 (63%) of posure for all four river systems, but in contrast to the Lin-
the total 1495 cross-sections were detected as optimal loca- ear Method, the number of optimal locations detected tends
tions, while only 166 (11%) were detected as optimal at 95% to increase with increasing channel exposure for the Slope-
channel exposure. At least some optimal locations were Break Method (Figure 6; the reasons for this are discussed
detected for all levels of channel exposure (20%–95%) in all in section 4). For example, at 20% channel exposure only 6
rivers except the Rio Grande, where no optimal locations (0.4%) of the 1495 cross-sections were detected as optimal
were detected at levels of channel exposure >70%. locations, whereas 242 (16%) were detected as optimal
locations at 95% channel exposure. Optimal locations were
3.2. Depth Estimates at Channel Cross-Sections Using detected at all levels of channel exposure (20%–95%) for
the Slope-Break Method both the Upper Mississippi and Ganges-Brahmaputra rivers
[27] Standard and RMS errors for mean depth retrievals and at levels of channel exposure 30% for the Illinois
estimated using the Slope-Break Method are shown in Fig- River, but 80% channel exposure was required to detect
ure 5, varying as a function of percent channel exposure. optimal locations along the Rio Grande River.
Viewed collectively, the RMSE values (red symbols) for
cross-section exposures ranging from 95% (high exposure) 3.3. Depth Estimates at Channel Cross-Sections Using
to 20% (low exposure) for all optimal locations detected the Slope-Break Method and Water Surface Profiles
range from 3% to 20%, 4% to 35%, 2% to 8%, and 4% to Simulated From HEC-RAS
65%, for the Upper Mississippi, Illinois, Rio Grande, and [28] To exploit the described slope breaks in w-h relation-
Ganges-Brahmaputra river systems, respectively. There is a ships, water levels must fall to the required level of channel
sharp reduction in both RMSE values and the range of dest exposure and be imaged at least once during a satellite mis-
errors for a given percentage of channel exposure produced sion lifetime. To assess how probable this occurrence might

3172
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

Figure 5. Standard and RMS errors for ‘‘optimal’’ locations for remote depth estimation, identified by
applying the Slope-break Method to synthetic values of w and h extracted from 1495 cross-sections from
four river systems. Note use of normal scale and significantly reduced errors relative to Figure 4. Each
blue plus symbol represents the standard error (percent error in dest relative to dbf) at a single optimal
location detected for a given percentage of channel exposure. The RMSE for all estimates for a given
percentage of channel exposure is plotted in red. In all four river systems, errors tend to improve and the
total number of optimal locations detected also tends to increase with increasing channel exposure. Note
that for the Illinois and Rio Grande, no optimal locations are detected when channel exposure is less
than 30% and 80%, respectively.

be, HEC-RAS simulations of water surface profiles were ance probability is found for the Upper Mississippi. Optimal
produced for the Upper Mississippi and Rio Grande rivers locations were detected along the Upper Mississippi for all
given their historic discharge statistics from USGS gaging tested exceedance probabilities (40–99%), while only for
stations. Standard and RMSE for dest retrievals estimated exceedance probabilities 58% for the Rio Grande River.
using the Slope-Break Method on HEC-RAS water surface As with previous tests using the Slope-Break Method, the
profile simulations are shown in Figure 7, varying as a func- number of optimal locations detected tends to increase with
tion of discharge exceedance probability. Using the Upper increasing channel exposure for both rivers. For the Rio
Mississippi as an example, at 40% exceedance probability Grande, while only 2 (1.3%) of the 150 cross-sections were
(meaning that this magnitude of discharge is exceeded 40% detected at 58% exceedance probability, 6 (4.0%) were
of the time, and thus this level of channel exposure (or detected at 90% exceedance probability. For the Upper Mis-
greater), occurs roughly 60% of the time), 14 of the 430 sissippi, 14 (3.3%) of 430 cross-sections were detected as
cross-sections tested meet the criteria as optimal slope- optimal locations at 40% exceedance probability, while 42
break locations, with a RMSE of 22.95% relative to dbf. (9.8%) were detected at 99% exceedance probability.
Viewed collectively, the standard errors in depth retrievals
range from 0.08% to 86% and 0.01% to 3.60% for the 3.4. Depth Estimates for River Reaches Using the
Upper Mississippi and Rio Grande rivers, respectively. The Slope-Break Method and Reach-Averaged Effective
RMSE values for exceedance probabilities ranging from Width We and h
99% (high exposure) to 40% (low exposure) for all optimal [29] Unlike intermittent cross-sections, the continuously
locations detected range from 20% to 31% and 0.75% to gridded bathymetric data set for the Upper Mississippi
2.25% for the Upper Mississippi and Rio Grande rivers, River enables assessment of how reach-averaging (spatial-
respectively. While RMSE values tend to decrease with averaging) of continuous measurements of We and h may
increasing exceedance probability (corresponding to influence the quality of the obtained dest retrievals. Standard
increasing channel exposure) for the Rio Grande, a slight and RMS errors for dest retrievals estimated using the
trend of increasing RMSE values with increasing exceed- Slope-Break Method on reach-averaged We and h are shown

3173
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

Figure 6. Comparison of the percentage of cross-sections detected as optimal (left column), and their
corresponding RMSE values (calculated from the percent error in dest relative to dbf, for all estimates for
a given percentage of channel exposure) (right column) for the Linear Method (red) and Slope-Break
Method (blue) for each of the four river systems studied. Note the opposing trends in the number of opti-
mal locations detected by each method, and the superior performance of the Slope-Break Method across
all levels of channel exposure.

3174
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

Figure 7. Standard and RMS errors for optimal locations identified by applying the Slope-break
Method to simulated water surface elevations (h) modeled in HEC-RAS for a range of discharge exceed-
ance probabilities (the percentage of time that a given discharge is exceeded, based on historic stream-
flow records) along the Upper Mississippi and Rio Grande Rivers. Each blue plus symbol represents the
standard error (percent error in dest relative to dbf) for a single optimal location detected for a given dis-
charge exceedance probability. The RMSE for all estimates for a given exceedance probability is plotted
in red. Useful slope-breaks are detected at flow levels which are exceeded 40% and 58% of the time
for the Upper Mississippi and Rio Grande rivers, respectively (thus indicating that adequate channel
exposures occur roughly 60% and 42% of the time, respectively.

in Figure 8, varying as a function of percent channel expo- levels) in order to identify hydraulic relationships that may
sure as before. A summary comparison of how these errors usefully be extrapolated to estimate depth. The minimum
vary with reach-averaging length is plotted in Figure 9. As amount of channel exposure necessary for the method to
for the slope-break analysis based on cross-sections, both work varies between rivers and for different locations along
RMSE values and the range of errors for reach-averaged the same river. While the variability in required channel ex-
sections tend to decrease with increasing channel exposure posure is likely due to in part to the different geometries
for those reaches where slope breaks were detected. Like- and hydrologic regimes of the rivers studied, this finding is
wise, the number of optimal locations for a given reach- also potentially related to the limited sample sizes of the
averaging length tends to increase with increasing channel cross-sectional and bathymetric data sets used.
exposure, as shown in Figure 9. However, as reach-averag- [31] Testing of the Linear Method, for example, shows
ing length increases, the number of optimal locations for a that a simple linear extrapolation of w-h relationships
given percentage of channel exposure decreases, and a observed at moderate to high flow levels can lead to wildly
greater percentage of channel exposure is required in order inaccurate depth estimates (Figure 4). In contrast, testing of
for optimal slope breaks to be detected. RMSE values for a the Slope-Break Method suggests that extrapolation of low-
given percentage of channel exposure remain relatively con- flow w-h relationships at those locations where a second lin-
sistent at all reach-averaging lengths up to 2000 m, but ear relationship is found can substantially reduce this error.
tend to increase somewhat at longer lengths. Optimal slope- Indeed, as compared to the Linear Method the Slope-Break
breaks were not detected at any level of channel exposure Method demonstrates superior performance in flow depth
for reach-averaging lengths exceeding 7000 m. However, estimation for all four rivers at all levels of channel exposure
between reach-averaging lengths of 5 m and 7000 m the (Figure 6). Although the Linear Method technically detects a
range of errors for all optimal locations detected for a given greater number of ‘‘optimal’’ locations, with less channel ex-
percentage of channel exposure tends to decrease as reach- posure, as compared with the Slope-Break Method, this
averaging length scale is increased. The reasons for this are seeming benefit is offset by considerably larger errors in the
discussed next. derived depth estimates. Indeed, as channel exposure
increases, the Linear Method detects fewer optimal locations,
as many locations displaying ‘‘stable’’ w-h relationships at
4. Discussion and Conclusions moderate to high flow levels are later revealed to have unsta-
[30] The findings of this study suggest that remotely ble w-h relationships if more of the channel is exposed (e.g.,
sensed measurements of river cross-sectional or reach-aver- Figure 3, bottom row). While a small number of purely linear
aged flow width (w or We) and water surface elevation h w-h relationships were preserved throughout the entire range
alone may be useful for estimating mean river depth d, at of possible h values for the cross-sections studied here, such
select locations, if a sufficient number of observations are locations could not in practice be detected from space unless
accumulated over time so as to identify stable empirical they became fully exposed (i.e., the river dries up) during at
relationships between the two variables. Because moderate- least one satellite overpass. Therefore, despite the Linear
to high-flow hydraulic relationships often do not extend to Method’s apparent appeal of identifying numerous candidate
low flows, a sufficient portion of a river channel’s bathyme- locations with low amounts of channel exposure, its overly
try must be observed (i.e., h must fall to sufficiently low simplistic assumptions yield poor depth-estimation results.

3175
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

Figure 8. Standard and RMS errors for optimal locations for remote depth estimation, identified by
applying the Slope-break Method to synthetic values of reach-averaged effective width We and h
extracted from 62 km of continuously-gridded bathymetry along the Upper Mississippi. The level of spa-
tial-averaging performed ranges from 5 m (native resolution) to 7 km river reach lengths. Each blue plus
symbol represents the standard error (percent error in dest relative to dbf) at a single optimal location
detected for a given percentage of channel exposure The RMSE for all estimates for a given percentage
of channel exposure is plotted in red. For a given percentage of channel exposure, both depth retrieval
errors and the total number of optimal locations identified decrease substantially with increasing reach-
averaged length. For this particular river, a reach-averaged length of approximately 1000–2000 m strikes
an optimal balance between improved quality of the depth-retrieval estimates and total number of opti-
mal locations found.

[32] The Slope-Break Method mitigates this weakness systems, Figure 5), is somewhat surprising given that slope
by selecting for locations where two linear trends in the breaks were initially assumed to occur only at low-flow h val-
width-h relationship are detected instead of one. While still ues. Nonetheless, the overall conclusion to be drawn is that
a simple model, this more closely reflects real-world chan- the Slope-Break Method detects fewer optimal locations than
nel geometries associated with cut banks and alluvial bar the Linear Method, but offers substantially smaller errors.
formation, as well as field-based geomorphic observations [33] Even for the Slope-Break Method, substantial, site-
of high- versus low-flow hydraulic geometries [e.g., Lewis, specific errors remain if the method is applied to the scale
1966; Richards, 1976; Jowett, 1998]. This also plausibly of traditional surveyed cross-sections, long known to vary
explains why the Slope-Break Method, in contrast to the Lin- greatly from one place to the next along natural river sys-
ear Method, tends to detect more optimal locations as channel tems [e.g., Leopold and Maddock, 1953; Richards, 1973;
exposure increases, because as h falls and submerged slope Knighton, 1975; Park, 1977]. In contrast, longitudinal
breaks are revealed such locations become qualified for the reach-averaging of We and h substantially reduces uncer-
former and disqualified for the latter. The detection of numer- tainty in the obtained depth estimates (Figure 8). As such,
ous slope breaks even with channel exposures as low as 20% reach-averaging does not appear to diminish the quality of
(e.g., the Upper Mississippi and Ganges-Brahmaputra river observed empirical hydraulic relationships, and may actually

3176
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

Figure 9. Comparison of (a) the percentage of reaches detected as optimal locations by the Slope-
Break Method; and (b) their corresponding RMSE values (calculated from the percent error in dest rela-
tive to dbf, for all estimates for a given percentage of channel exposure), both varying with reach-aver-
aged length and percentage of channel exposure. The number of optimal cross-sections identified
generally increases with channel exposure but decreases with increasing reach-averaging length. RMSE
values for dest tend to decrease with increasing channel exposure and remain fairly consistent for reach-
averaged lengths between 5 m and 2000 m, but degrade at longer reach-averaged lengths.

improve their robustness through spatial-averaging of local like sensor planned for first flights over the Sacramento
geomorphic heterogeneities. Therefore, while increased reach- River in 2013 (http://swot.jpl.nasa.gov/Airswot/)). While
averaging tends to reduce the total number of optimal loca- our approach requires no bathymetric or floodplain DEM, a
tions detected, the quality of the derived depth estimate is thorough assessment of how SWOT (or other sensors)
improved, at least for the Upper Mississippi data set tested width and h measurement errors would propagate further
here (Figure 8). For this particular river, a high-quality depth uncertainty to the derived depth estimates is warranted.
retrieval is ensured using reach lengths of 1000 to 2000 m, Finally, it is important to reiterate that neither the Linear
which is roughly 2–3 times mean bankfull channel width. nor Slope-Break method can possibly produce continuous
This finding agrees with Smith and Pavelsky [2008], who depth estimates everywhere along a river. Instead, they
found that remotely sensed HG b-exponents reached stable seek out a small subset of ideal channel locations where
values once reach-averaging lengths are increased to 2–3 simple, linear correlations exist between a river’s flow
times river floodplain width. width and mean flow depth.
[34] This finding of improved depth estimation through [36] Despite these limitations, this study puts forth a first,
greater reach-averaging of observed hydraulic variables is purely empirical method for estimating river flow depth
especially promising in the SWOT context, given its tech- based solely on remotely sensed measurements of river water
nological requirement to spatially average raw native-reso- surface elevation h and effective width We using no ancillary
lution radar echoes to improve the measurement precision data whatsoever. Even intermittent depth estimates at a sub-
of its h measurement. Although examining the impact of set of ideal channel locations along a given river reach could
SWOT measurement precision on the conceptual depth esti- potentially be used to back-out depth estimates throughout
mation technique presented here lies beyond the scope of the remaining portion of the reach (e.g., using hydraulic rela-
this study, like all radars SWOT measurement precision is tionships developed at optimal locations to estimate depth
improved through increased spatial-averaging, an approach elsewhere). Successfully applied, intermittent retrievals of
compatible with river reach-averaging. Furthermore, the mean flow depth could aid remote estimation of river dis-
exceedance probability simulations presented in Figure 7 charge, either as inputs to the Manning equation-based slope-
suggest that the temporal sampling of SWOT and other sat- area method [Dingman, 1994] or data assimilation into
ellites is more than adequate to encounter the required expo- hydrodynamic models. The simplicity of the Slope-Break
sures of channel bathymetry, with sufficiently low flows in Method is much of its appeal, offering potential value as a
the Upper Mississippi and Rio Grande rivers occurring at first-order depth estimator for global river studies from space.
least 60% and 42% of the time, respectively.
[35] This study was limited by the availability and limi-
tations of field-based data sets, and the predominant use of [37] Acknowledgments. This research was funded by the NASA
Physical Oceanography Program (grant NNX10AE96G), managed by Eric
traditional surveyed cross-sections. Unlike imaging tech- Lindstrom. In situ cross-section data for the Ganges-Brahmaputra river
nologies, cross-sections provide only 2-D information, and system was provided by Faissal Hossain (Tennessee Technical University)
only at intermittent locations. Even the continuous sonar as part of a Memorandum of Understanding between the Institute of Water
Modeling––Bangladesh and Tennessee Technical University. In situ cross-
bathymetric data set for the Upper Mississippi was some- section data for the Rio Grande (surveyed by Tetra Tech Inc.) and for the
what removed from real-world conditions, owing to its re- Upper Mississippi and Illinois rivers (surveyed by the U.S. Army Corps of
moval of surface water slope dh/dx and lack of bankfull Engineers) was provided to the authors by Edward Beighley (FM Global).
The continuous in situ bathymetric data set for the Upper Mississippi was
coverage. Future work should apply the methods developed provided by the USGS, as part of the Long Term Resource Monitoring
herein to more realistic riverine measurements, e.g., meas- Program. The authors thank Doug Alsdorf, Paul Bates, and one anonymous
urements from AirSWOT, an airplane-mounted, SWOT- reviewer for their constructive feedback on this paper.

3177
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

References Jung, H. C., et al. (2010), Characterization of complex fluvial systems using
remote sensing of spatial and temporal water level variations in the Ama-
Alsdorf, D. E., J. M. Melack, T. Dunne, L. A. K. Mertes, L. L. Hess, and zon, Congo, and Brahmaputra Rivers, Earth Surf. Proc. Landforms,
L. C. Smith (2000), Interferometric radar measurements of water level 35(3), 294–304, doi:10.1002/esp.1914.
changes on the Amazon flood plain, Nature, 404(6774), 174–177, Khan, S. I., et al. (2011), Satellite remote sensing and hydrologic modeling
doi:10.1038/35004560. for flood inundation mapping in Lake Victoria basin: Implications for
Alsdorf, D. E., L. C. Smith, and J. M. Melack (2001), Amazon floodplain hydrologic prediction in ungauged basins, IEEE Trans. Geosci. Remote
water level changes measured with interferometric SIR-C radar, IEEE Sens., 49(1), 85–95, doi:10.1109/TGRS.2010.2057513.
Trans. Geosci. Remote Sens., 39(2), 423–431, doi:10.1109/36.905250. Knighton, A. D. (1975), Variations in at-a-station hydraulic geometry, Am.
Alsdorf, D. E., E. Rodrıguez, and D. P. Lettenmaier (2007a), Measuring J. Sci., 275(2), 186–218, doi:10.2475/ajs.275.2.186.
surface water from space, Rev. Geophys., 45, RG2002, doi:10.1029/ Koblinsky, C. J., R. T. Clarke, A. C. Brenner, and H. Frey (1993), Measure-
2006RG000197. ment of river level variations with satellite altimetry, Water Resour. Res.,
Alsdorf, D. E., P. Bates, J. Melack, M. Wilson, and T. Dunne (2007b), Spa- 29(6), 1839–1848, doi:10.1029/93WR00542.
tial and temporal complexity of the Amazon flood measured from space, Lee, H., R. E. Beighley, D. Alsdorf, H. C. Jung, C. K. Shum, J. Duan,
Geophys. Res. Lett., 34, L08402, doi:10.1029/2007GL029447. J. Guo, D. Yamazaki, and K. Andreadis (2011), Characterization of terres-
Andreadis, K. M., E. A. Clark, D. P. Lettenmaier, and D. E. Alsdorf (2007), trial water dynamics in the Congo Basin using GRACE and satellite radar
Prospects for river discharge and depth estimation through assimilation altimetry, Remote Sens. Environ., 115(12), 3530–3538, doi:10.1016/
of swath-altimetry into a raster-based hydrodynamics model, Geophys. j.rse.2011.08.015.
Res. Lett., 34, L10403, doi:10.1029/2007GL029721. Legleiter, C. J., and D. A. Roberts (2009), A forward image model for pas-
Biancamaria, S., et al. (2011), Assimilation of virtual wide swath altimetry sive optical remote sensing of river bathymetry, Remote Sens. Environ.,
to improve Arctic river modeling, Remote Sens. Environ., 115(2), 373– 113(5), 1025–1045, doi:10.1016/j.rse.2009.01.018.
381, doi:10.1016/j.rse.2010.09.008. Legleiter, C. J., D. A. Roberts, W. A. Marcus, and M. A. Fonstad (2004),
Birkett, C. M. (1998), Contribution of the TOPEX NASA radar altimeter to Passive optical remote sensing of river channel morphology and in-
the global monitoring of large rivers and wetlands, Water Resour. Res., stream habitat: Physical basis and feasibility, Remote Sens. Environ.,
34(5), 1223–1239. 93(4), 493–510, doi:10.1016/j.rse.2004.07.019.
Birkett, C. M., and B. Beckley (2010), Investigating the performance of the Legleiter, C. J., D. A. Roberts, and R. L. Lawrence (2009), Spectrally based
Jason-2/OSTM radar altimeter over lakes and reservoirs, Mar. Geod., remote sensing of river bathymetry, Earth Surf. Proc. Landforms, 34(8),
33(S1), 204–238, doi:10.1080/01490419.2010.488983. 1039–1059, doi:10.1002/esp.1787.
Birkett, C. M., L. A. K. Mertes, T. Dunne, M. H. Costa, and M. J. Jasinski Leopold, L., and T. Maddock (1953), The hydraulic geometry of stream
(2002), Surface water dynamics in the Amazon Basin: Application of satel- channels and some physiographic implications, Geol. Surv. Prof. Paper.
lite radar altimetry, J. Geophys. Res., 107(D20), 8059–8080, doi:10.1029/ [Available at: http://scholarworks.umass.edu/fishpassage_reports/252.].
2001JD000609. Lewis, L. A. (1966), The adjustment of some hydraulic variables at dis-
Bjerklie, D. M. (2007), Estimating the bankfull velocity and discharge for charges less than one cfs, Prof. Geogr., 18(4), 230–234, doi:10.1111/
rivers using remotely-sensed river morphology information, J. Hydrol., j.0033-0124.1966.00230.x.
341(3–4), 144–155, doi:10.1016/j.jhydrol.2007.04.011. Lu, Z., M. Crane, O.-I. Kwoun, C. Wells, C. Swarzenski, and R. Rykhus
Brakenridge, G. R., S. V. Nghiem, E. Anderson, and S. Chien (2005), (2005), C-band radar observes water level change in swamp forests, Eos
Space-based measurement of river runoff, Eos Trans. AGU, 86(19), 185– Trans. AGU, 86(14), 141, doi:10.1029/2005EO140002.
188. Marcus, W. A., and M. A. Fonstad (2008), Optical remote mapping of riv-
Calmant, S., F. Seyler, and J. Cretaux (2008), Monitoring continental sur- ers at sub-meter resolutions and watershed extents, Earth Surf. Proc.
face waters by satellite altimetry, Surv. Geophys., 29(4), 247–269, Landforms, 33(1), 4–24, doi:10.1002/esp.1637.
doi:10.1007/s10712-008-9051-1. Marcus, W. A., and M. A. Fonstad (2010), Remote sensing of rivers: The
Cretaux, J.-F., and C. Birkett (2006), Lake studies from satellite radar altim- emergence of a subdiscipline in the river sciences, Earth Surf. Proc.
etry, C. R. Geosci., 338(14–15), 1098–1112, doi:10.1016/ Landforms, 35(15), 1867–1872, doi:10.1002/esp.2094.
j.crte.2006.08.002. Papa, F., C. Prigent, F. Durand, and W. B. Rossow (2006), Wetland dynam-
Dingman, S. L. (1994), Physical Hydrology, 2nd ed., Prentice Hall, Upper ics using a suite of satellite observations: A case study of application and
Saddle River, N. J. [Available at: http://www.cce.uri.edu/cels/nrs/whl/ evaluation for the Indian Subcontinent, Geophys. Res. Lett., 33, L08401,
Teaching/nrs592/2009/Class%203%20Wetland%20Riparian%20Biogeo doi:10.1029/2006GL025767.
chemistry/Soil%20Hydrology%20Bckground%20Dingman.pdf, accessed Park, C. C. (1977), World-wide variations in hydraulic geometry exponents
19 Sept. 2012.]. of stream channels: An analysis and some observations, J. Hydrol.,
Durand, M., K. M. Andreadis, D. E. Alsdorf, D. P. Lettenmaier, D. Moller, 33(1–2), 133–146, doi:10.1016/0022–1694(77)90103-2.
and M. Wilson (2008), Estimation of bathymetric depth and slope from Phillips, J. D. (1990), The instability of hydraulic geometry, Water Resour.
data assimilation of swath altimetry into a hydrodynamic model, Geo- Res., 26(4), 739–744, doi:10.1029/WR026i004p00739.
phys. Res. Lett., 35, L20401, doi:10.1029/2008GL034150. Prigent, C., E. Matthews, F. Aires, and W. B. Rossow (2001), Remote sens-
Durand, M., E. Rodriguez, D. E. Alsdorf, and M. Trigg (2010a), Estimating ing of global wetland dynamics with multiple satellite data sets, Geo-
river depth from remote sensing swath interferometry measurements of phys. Res. Lett., 28(24), 4631–4634, doi:10.1029/2001GL013263.
river height, slope, and width, IEEE J. Selec. Top. Appl. Earth Obs. Richards, K. S. (1973), Hydraulic geometry and channel roughness; a non-
Remote Sens., 3(1), 20–31, doi:10.1109/JSTARS.2009.2033453. linear system, Am. J. Sci., 273(10), 877–896, doi:10.2475/ajs.273.10.877.
Durand, M., L.-L. Fu, D. P. Lettenmaier, D. E. Alsdorf, E. Rodriguez, and Richards, K. S. (1976), Complex width-discharge relations in natural river
D. Esteban-Fernandez (2010b), The surface water and ocean topography sections, Geol. Soc. Am. Bull., 87(2), 199–206, doi:10.1130/0016-
mission: Observing terrestrial surface water and oceanic submesoscale 7606(1976)87<199:CWRINR>2.0.CO;2.
eddies, Proc. IEEE, 98(5), 766–779, doi:10.1109/JPROC.2010.2043031. Rogala, J. T. (1999), Bathymetry - Chapter 2 extracted from James
Fonstad, M. A., and W. A. Marcus (2010), High resolution, basin extent T. Rogala, 1999. Methodologies employed for bathymetric mapping and
observations and implications for understanding river form and process, sediment characterization as part of the Upper Mississippi River System
Earth Surf. Process. Landforms, 35(6), 680–698, doi:10.1002/esp.1969. Navigation Feasibility Study, ENV Rep. 13, Interim Report for the Upper
Frappart, F., F. Seyler, J.-M. Martinez, J. G. Leon, and A. Cazenave (2005), Mississippi River––Illinois Waterway System Navigation Study; pre-
Floodplain water storage in the Negro River basin estimated from micro- pared for U.S. Geol. Surv., U.S. Army Eng. Dist., Rock Isl., St. Louis.
wave remote sensing of inundation area and water levels, Remote Sens. Shiklomanov, A. I., R. B. Lammers, and C. J. Vörösmarty (2002), Wide-
Environ., 99(4), 387–399, doi:10.1016/j.rse.2005.08.016. spread decline in hydrological monitoring threatens Pan-Arctic Research,
Jowett, I. G. (1998), Hydraulic geometry of New Zealand rivers and its use Eos Trans. AGU, 83(2), 13, doi:10.1029/2002EO000007.
as a preliminary method of habitat assessment, Reg. Rivers Res. Manage., Smith, L. C. (1997), Satellite remote sensing of river inundation area, stage,
14(5), 451–466, doi:10.1002/(SICI)1099-1646(1998090)14:5<451: : and discharge: A review, Hydrol. Process., 11, 1427–1439, doi:10.1002/
AID-RRR512>3.0.CO;2-1. (SICI)1099-1085(199708)11:10<1427: :AID-HYP473>3.3.CO;2-J.
Jung, H. C., and D. Alsdorf (2010), Repeat-pass multi-temporal interfero- Smith, L. C., and T. M. Pavelsky (2008), Estimation of river discharge,
metric SAR coherence variations with Amazon floodplain and lake habi- propagation speed, and hydraulic geometry from space: Lena River,
tats, Int. J. Remote Sens., 31(4), 881–901, doi:10.1080/014311609029 Siberia, Water Resour. Res., 44, 03427, doi:10.1029/2007WR006133.
02609.

3178
 MERSEL ET AL.: REMOTELY SENSED ESTIMATION OF RIVER DEPTH

Smith, L. C., B. L. Isacks, R. R. Forster, A. L. Bloom, and I. Preuss (1995), Tetra Tech, Inc. (2005), FLO-2D Model Development Below Caballo Dam
Estimation of discharge from braided glacial rivers using ERS 1 synthetic URGWOM: Final Report on FLO-2D Model Development, Tetra Tech
aperture radar: First results, Water Resour. Res., 31, 1325–1330, Inc., Albuquerque, N. M.
doi:10.1029/95WR00145. U.S. Army Corps of Engineers (2004), Upper Mississippi River System
Smith, L. C., B. L. Isacks, A. L. Bloom, and A. B. Murray (1996), Estima- Flow Frequency Study: Final Report, U. S. Army Corps of Eng., Rock
tion of discharge from three braided rivers using synthetic aperture radar Isl. Dist.
satellite imagery: Potential application to ungaged basins, Water Resour. United Nations Educational, Scientific and Cultural Organization (2012),
Res., 32(7), 2021–2034, doi:10.1029/96WR00752. The United Nations World Water Development report 4, United Nations
Stewardson, M. (2005), Hydraulic geometry of stream reaches, J. Hydrol., Educ., Sci. and Cult. Organ., Paris, France.
306(1–4), 97–111, doi:10.1016/j.jhydrol.2004.09.004. Vörösmarty, C. J. et al. (2010), Global threats to human water security
Stokstad, E. (1999), Scarcity of rain, stream gages threatens forecasts, Sci- and river biodiversity, Nature, 467(7315), 555–561, doi:10.1038/
ence, 285(5431), 1199–1200, doi:10.1126/science.285.5431.1199. nature09440.

3179