ArticlePDF Available

Modeling multidecadal surface water inundation dynamics and key drivers on large river basin scale using multiple time series of Earth observation and river flow data

Article

Modeling multidecadal surface water inundation dynamics and key drivers on large river basin scale using multiple time series of Earth observation and river flow data

Abstract and Figures

Periodically inundated floodplain areas are hot spots of biodiversity and provide a broad range of ecosystem services but have suffered alarming declines in recent history. Despite their importance, their long-term surface water (SW) dynamics and hydroclimatic drivers remain poorly quantified on continental scales. In this study, we used a 26 year time series of Landsat-derived SW maps in combination with river flow data from 68 gauges and spatial time series of rainfall, evapotranspiration and soil moisture to statistically model SW dynamics as a function of key drivers across Australia's Murray-Darling Basin (∼1 million km²). We fitted generalized additive models for 18,521 individual modeling units made up of 10 × 10 km grid cells, each split into floodplain, floodplain-lake, and nonfloodplain area. Average goodness of fit of models was high across floodplains and floodplain-lakes (r²>0.65), which were primarily driven by river flow, and was lower for nonfloodplain areas (r²>0.24), which were primarily driven by rainfall. Local climate conditions were more relevant for SW dynamics in the northern compared to the southern basin and had the highest influence in the least regulated and most extended floodplains. We further applied the models of two contrasting floodplain areas to predict SW extents of cloud-affected time steps in the Landsat series during the large 2010 floods with high validated accuracy (r²>0.97). Our framework is applicable to other complex river basins across the world and enables a more detailed quantification of large floods and drivers of SW dynamics compared to existing methods.
Content may be subject to copyright.
RESEARCH ARTICLE
10.1002/2016WR019858
Modeling multidecadal surface water inundation dynamics and
key drivers on large river basin scale using multiple time series
of Earth-observation and river flow data
V. Heimhuber
1
, M. G. Tulbure
1
, and M. Broich
1
1
School of Biological, Earth and Environmental Science, University of New South Wales, Sydney, New South Wales,
Australia
Abstract Periodically inundated floodplain areas are hot spots of biodiversity and provide a broad range
of ecosystem services but have suffered alarming declines in recent history. Despite their importance, their
long-term surface water (SW) dynamics and hydroclimatic drivers remain poorly quantified on continental
scales. In this study, we used a 26 year time series of Landsat-derived SW maps in combination with river
flow data from 68 gauges and spatial time series of rainfall, evapotranspiration and soil moisture to statisti-
cally model SW dynamics as a function of key drivers across Australia’s Murray-Darling Basin (1 million
km
2
). We fitted generalized additive models for 18,521 individual modeling units made up of 10 310 km
grid cells, each split into floodplain, floodplain-lake, and nonfloodplain area. Average goodness of fit of
models was high across floodplains and floodplain-lakes (r
2
>0.65), which were primarily driven by river
flow, and was lower for nonfloodplain areas (r
2
>0.24), which were primarily driven by rainfall. Local climate
conditions were more relevant for SW dynamics in the northern compared to the southern basin and had
the highest influence in the least regulated and most extended floodplains. We further applied the models
of two contrasting floodplain areas to predict SW extents of cloud-affected time steps in the Landsat series
during the large 2010 floods with high validated accuracy (r
2
>0.97). Our framework is applicable to other
complex river basins across the world and enables a more detailed quantification of large floods and drivers
of SW dynamics compared to existing methods.
1. Introduction
Wetlands and other periodically inundated surface water (SW) areas are hot spots of biodiversity [Gopal,
2009; Pearson et al., 2013; Arthington et al., 2014] play a crucial role in global climatic, hydrologic, and bio-
geochemical cycles [Hamilton, 2010] and provide numerous ecosystem services of value to people [Maltby
and Acreman, 2011; Costanza et al., 2014]. In recent history however, increasing development of water
resources, land use transformations, and agricultural intensification have led to an alarming disappearance
and decline of floodplains and other types of wetlands around the world [Finlayson et al., 1999; Lemly et al.,
2000; Jones et al., 2009; Costanza et al., 2014], with the most comprehensive recent review estimating a loss
of 69–75% of inland wetland area present in 1900 AD [Davidson, 2014]. Our study area, the semi-arid Mur-
ray-Darling Basin (MDB) in Australia, is a good example of this global trend, where intensive agricultural and
water resources development in the second half of the 20th century has led to deterioration of many of the
basins floodplain wetlands [Kingsford and Thomas, 2004]. This problem was further intensified by the recent
Millennium Drought (from mid-1990s to 2009), the most severe recorded hydrological drought in the MDB,
which was followed by widespread extreme flooding in 2010/2011 commonly referred to as the La Nina
Floods [Leblanc et al., 2012].
As a result of the dramatic global loss in wetland area and the paramount value of terrestrial SW resources
in the face of widely proclaimed global freshwater crisis [V
or
osmarty et al., 2010; Bakker, 2012; World
Economic Forum, 2015], the usage of remote sensing for quantifying the distribution and temporal variabili-
ty of SW over large areas and periods of time has become an important field of research in recent years
[Alsdorf and Lettenmaier, 2003; Alsdorf et al., 2007; Papa et al., 2010; Tulbure and Broich, 2013; Bates et al.,
2014; Bierkens, 2015; Hu et al., 2015; Kuenzer et al., 2015; Lettenmaier et al., 2015; Pekel et al., 2016; Tulbure
et al., 2016; World Bank, 2016]. A better understanding of SW extent dynamics, that is the distribution of
Key Points:
We quantified the role of river flow
and hydroclimatic drivers in surface
water dynamics across a large and
complex river basin
Integrating a 26 year Landsat time
series of surface water with flow data
produced accurate statistical flood
models for 18,521 areas
The role of hydroclimatic drivers in
flooding dynamics differed across the
system and between different types
of water bodies
Correspondence to:
V. Heimhuber,
valentin.heimhuber@unsw.edu.au
Citation:
Heimhuber, V., M. G. Tulbure, and
M. Broich (2017), Modeling
multidecadal surface water inundation
dynamics and key drivers on large
river basin scale using multiple time
series of Earth-observation and river
flow data, Water Resour. Res.,53,
doi:10.1002/2016WR019858.
Received 28 SEP 2016
Accepted 10 JAN 2017
Accepted article online 13 JAN 2017
V
C2017. American Geophysical Union.
All Rights Reserved.
HEIMHUBER ET AL. MODELING SURFACE WATER 1
Water Resources Research
PUBLICATIONS
inland water bodies through time, at continental scale can enable better management of water resources
[Alsdorf et al., 2003; Overton, 2005; Powell et al., 2008; Viney et al., 2014], assist in managing floods and
droughts [Sakamoto et al., 2007; Chen et al., 2013; Huang et al., 2014; Hu et al., 2015] or help to identify prior-
ity conservation areas by studying ecological connectivity across space and time [Tulbure et al., 2014;
Bishop-taylor et al., 2015]. Remotely sensed series of SW maps have further been used to quantify reservoir
storage [Gao et al., 2012; Zhang et al., 2014b] and to estimate wetland area in the context of malaria man-
agement [Midekisa et al., 2014].
Large-scale flooding events are a critical component of the terrestrial water cycle [Alsdorf et al., 2007] but
their propagation through river systems and the corresponding dynamics remain poorly quantified on con-
tinental or global scales [Bates et al., 2014]. The traditional approach for quantifying the extent of floods is
through hydrodynamic modeling and increases in computation power and data availability have led to
major advances in continental and even global scale flood modeling in recent years [Yamazaki et al., 2011;
Getirana et al., 2012; Neal et al., 2012, 2015; Cauduro et al., 2013; Sampson et al., 2015]. Schumann et al.
(2016) applied a hydrodynamic model to simulate floodplain inundation across Australia continuously over
a 40 year period. Nevertheless, parameterization of these models remains difficult and accurate representa-
tion of complex river and floodplain topographies and the corresponding storage effects is still limited by
the quality of suitable data sets such as digital elevation models (DEMs) with global coverage [Bates et al.,
2014]. A promising alternative for quantifying and monitoring SW dynamics over large areas and long peri-
ods of time is the use of time series of satellite data [Alsdorf et al., 2007; World Bank, 2016], with the Landsat
and MODIS satellites currently representing the most commonly used data source [Musa et al., 2015;
Karpatne et al., 2016]. The unique spectral signature of SW in the infrared and visible bands allows for sepa-
ration between water and dry land surface by applying a suitable classification technique to satellite imag-
ery, enabling the generation of time series of SW maps on large river basin [Tulbure and Broich, 2013; Huang
et al., 2014; Kuenzer et al., 2015; Tulbure et al., 2016], continental [Klein et al., 2014; Mueller et al., 2016], and
even global scales [Papa et al., 2010; Klein et al., 2015; Pekel et al., 2016].
Apart from providing a promising data source for calibrating large-scale hydrodynamic models [Karim et al.,
2011; Ticehurst et al., 2013], such series of SW maps derived from satellite imagery have previously been
used for developing statistical inundation models by empirically linking remotely sensed SW extents to river
flow or stage [Costelloe et al., 2003; Gumbricht et al., 2004; Powell et al., 2008; Westra and De Wulf, 2009;
Frazier and Page, 2009; Ren et al., 2010; Jung et al., 2011; Leauthaud et al., 2013; Sims et al., 2014; Huang et al.,
2014; Ogilvie et al., 2015; Azman et al., 2016]. In addition to their high cost-benefit ratio, satellite-based inun-
dation models are useful for a variety of applications including the prediction of wetland inundation extent
[Gumbricht et al., 2004; Overton, 2005; Westra and De Wulf, 2009; Huang et al., 2014] or the improvement of
environmental flow strategies [Shaikh et al., 2001; Powell et al., 2008], a water management technique for
recovering floodplain ecosystem health by releasing flushes of water from reservoirs. Most existing models,
however, are based on a single cohesive floodplain site, used only a limited number of manually selected
satellite images and did not account for lag times between river gauges and the area of inundation
response. Heimhuber et al. [2016] addressed these limitations by developing a prototype framework for
modeling SW extent dynamics, depicted in a Landsat-derived time series of SW maps (1986–2011) [Tulbure
et al., 2016], continuously as a function of river flow (Q), local rainfall (P), soil moisture (SM), and evapotrans-
piration (ET), yet to be tested and applied on large river basin scale. In this study, we built on and adapted
this framework to model SW extent dynamics and drivers across the entire MDB (1 million km
2
).
Like many large river basins, the MDB contains a broad variety of smaller river systems with unique climate
and hydromorphological characteristics, flow and flooding regimes as well as highly variable levels of river
regulation and water abstraction for irrigated agriculture and other uses [Murray-Darling Basin Authority
(MDBA), 2010; Leblanc et al., 2012]. All these aspects impose difficulties for accurate quantification of the dis-
tribution and movement of SW across river and floodplain systems [Rudorff et al., 2014], which is key knowl-
edge for managing these resources more sustainably [World Bank, 2016]. Satellite-based time series of SW
extent, which are rapidly becoming more available [Karpatne et al., 2016], hold good potential for overcom-
ing these difficulties, but their application for modeling large-scale inundation processes remains poorly
explored [Bates et al., 2014]. Our study addresses this gap by integrating multiple novel time series of Earth-
observation and ground data for quantifying and modeling SW extent dynamics and drivers within a large
and complex river system. Specific objectives were to (1) statistically model inundation extent dynamics
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 2
across the MDB, (2) quantify the role of river flow and local climate drivers (i.e., P,ET, and SM) in SW dynam-
ics, and (3) apply the models developed in (1) to predict SW extents for highly cloud-affected Landsat obser-
vations during the 2010/2011 La Nina Floods for two floodplains with contrasting flooding regimes.
2. Materials and Methods
2.1. Study Area
The study area of this research is one of Australia’s largest river systems and the country’s bread basket, the
predominantly arid MDB (Figure 1). The MDB is often divided into a northern part covering the entire Dar-
ling River catchment until its confluence with the Murray River and a southern part covering the remaining
area to the south [Leblanc et al., 2012; Huang et al., 2014]. The basin hosts almost 30,000 wetlands [MDBA,
2010] with 16 listed as ‘‘wetlands of international importance’’ [Ramsar Convention Secretariat, 2014] and
around 200 specified in the ‘‘Directory of Important Wetlands in Australia’’ [Environment Australia, 2001].
Together, all wetlands cover around 6% of the total area of the MDB and 89% of all wetland areas are flood-
plains [Kingsford et al., 2004]. The system’s long-term average annual rainfall is 469 mm of which around
90% is lost due to evapotranspiration [Leblanc et al., 2012]. The MDB has a pronounced climate gradient
with average annual rainfall decreasing and climate variability and evapotranspiration increasing from S-E
to N-W [MDBA, 2010]. Due to this climate gradient and the diverse terrain, flooding regimes differ
Figure 1. Location, major rivers [BOM, 2012], ecohydrological-zonation [Huang et al., 2013], and topography [CSIRO and Geoscience Australia, 2011] of the Murray-Darling Basin along
with the spatial framework used for modeling SW dynamics. Colored 10 310 km grid cells contained SW areas that are hydraulically connected (i.e., floodplain and floodplain-lake cate-
gory) to one of the 68 model gauges (colors of grid cells correspond to the color of their respective modeling gauge). For better readability of Figure 1a, irrigation areas [BRS, 2008] and
the floodplain and floodplain-lake SW categories (derived from Kingsford et al. [2004]) are only shown in Figures 1b and 1c. The nonfloodplain category comprised all remaining areas
outside the two floodplain categories (colored grid cells and uncolored areas). Blue areas outside of grid cells in Figure 1a illustrate the location of mapped water bodies within the non-
floodplain SW category [Kingsford et al., 2004].
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 3
substantially across the MDB, with many of the ephemeral subcatchments in the northern basin having a
dry-land river flow regime characterized by extreme variability with long dry spells, punctuated by large
flood events [Bunn et al., 2006]. In the rivers of the southern basin, floods typically occur in winter and
spring as a result of reliable rainfall and snowmelt in the runoff generating catchment areas and can last for
several months at a time in the lower Murray River [Penton and Overton, 2007].
The very different hydrology of the northern compared to the southern basin is reflected in the fact that
despite the large size of its catchment, the Darling River only contributes about 14% of the total Murray Riv-
er inflows under natural conditions [Leblanc et al., 2012]. The MDB is a hot spot for agricultural activity with
irrigated agriculture accounting for about 60% of Australia’s total agricultural water use and the overall SW
use is about half of the MDB’s average annual water availability [Leblanc et al., 2012]. This very high level of
water usage and related water resources development has led to the deterioration of many of the basin riv-
er and floodplain ecosystems [Kingsford, 2000; Kingsford and Thomas, 2004; Leblanc et al., 2012].
2.2. Model Development and Validation
The following paragraphs provide a detailed description of the input data, the spatial modeling framework,
and the statistical techniques used for modeling SW extent as a function of key driver variables across the
MDB. In this study, we built on and adapted a prototype modeling framework that was developed for three
local river and floodplain systems in the MDB and is explained in detail in Heimhuber et al. [2016]. This
modeling framework is based on a segmentation of the river basin into spatial units that are suitable for
establishing meaningful statistical variable relationships. This segmentation accounts for the hydrological
structure of the river basin, the type of water body that is being modeled as well as for the hydraulic con-
nection of water bodies to a river with available gauge data. For each spatial unit, a statistical model
between a numerical time series of SW extent as the dependent variable and Q,P,ET, and SM as potential
driver variables was then fitted and used to analyze the variable relationships.
2.2.1. Time Series of Surface Water Extent
For modeling SW extent holistically across the study area, we used an existing, currently unprecedented
Landsat-based time series of SW extent for the MDB [Tulbure et al., 2016]. This time series was derived from
all available Landsat 5 and 7 scenes between 1986 and 2011 (25,000) of the USGS/EROS L1T standard
product with less than 50% overall cloud cover by using random forest classification models for water and
clouds. For each satellite observation time step, the series consisted of a raster file in which each pixel was
classified into water, dry-land, cloud or no-data. A rigorous accuracy assessment revealed overall classifica-
tion accuracy of 99%. The median number of classified Landsat scenes per year per path/row was 13 images
as opposed to the potentially possible number of 22 scenes as a result of cloud cover [Tulbure et al., 2016].
2.2.2. Spatial Modeling Framework
Modeling SW extent dynamics consistently across the MDB required the development of a spatial modeling
framework. Here we built on a previously developed and tested prototype framework [Heimhuber et al.,
2016] to model SW dynamics across the entire river basin (Figure 1). To account for the hydrological struc-
ture of the study area, we used an existing ecohydrological (EH) zonation of the MDB into 89 zones with
uniform ecological and hydrological characteristics [Chen et al., 2012; Huang et al., 2013] as the major units
for analysis and modeling. We then combined this zonation with a regular 10 310 km grid, which enabled
the application of a consistent modeling approach across the river system. The MDB contains a large variety
of SW areas and even though the majority of wetlands are floodplains, not all floodplains have a direct
hydraulic connection to a river with available gauge data. To account for the different dynamics of different
type of SW areas, we categorized them into floodplains, floodplain-lakes, and nonfloodplain areas (hereafter
referred to as SW categories) based on the type of wetland and its connection to gauged rivers and mod-
eled them separately using a tailored modeling approach (see Figures 1b and 1c). It is important to note
here that this categorization did not intend to accurately define floodplain areas across the MDB. Instead, it
is intended to differentiate between SW areas (floodplains and lakes) that are primarily driven by the
dynamics of larger rivers with available gauge data and all remaining SW areas which are driven by more
local rainfall runoff processes. Hence, the nonfloodplain SW category contained a large variety of SW areas
including isolated wetland systems, lakes and areas of irrigated agriculture.
For generating the SW categorization, we first selected 68 river gauges with long-term records to establish
homogenous gauge coverage across the MDB (Figure 1). We then used the Australian Geofabric [BOM,
2012], a fully connected and directed river network, to perform network routing operations and define
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 4
hydraulic connectivity between river gauges and grid cells that contained floodplain or lake areas as
defined by an existing static wetland layer for the MDB [Kingsford et al., 2004] (see colored grid-cells in Figure
1). If the river network indicated that there was a hydraulic connection between a grid cell and the river
gauge, all floodplain areas in that cell were assigned to the floodplain category, and all lakes to the
floodplain-lake category. The river network did not always represent the hydraulic connectivity that may occur
on vast floodplains during times of flooding. To address this limitation, we used a Landsat-based relative inun-
dation frequency layer [Tulbure et al., 2016] as an additional indicator for hydraulic connectivity to assign those
SW areas to the floodplain category that are only connected to gauged rivers during major flooding events. In
some areas, this methodology led to numerous modeling grid cells with hydraulic connectivity to a gauge
that did not have any wetlands according to the static wetland layer but showed significant SW dynamics in
the Landsat-based inundation frequency layer. For these grid cells that predominantly occurred in the very
N-E of the MDB, we considered all areas that were inundated at least once during the analysis period as flood-
plain area, in order to capture and model the corresponding SW extent dynamics accordingly.
For generating the final modeling units, we applied a buffer of 300 m to both the floodplain and the
floodplain-lake areas to ensure full coverage of floodplain areas that might inundate during the largest
floods. The resulting areas were then clipped to individual modeling units based on the regular grid and
the EH-zonation so that each grid cell potentially consisted of a combination of floodplain, floodplain-lake,
and nonfloodplain area. The resulting framework comprised a total of 4120 individual floodplain and
447 floodplain-lake modeling units across the MDB which were each linked to one of the 68 river gauges
(Figure 1). In addition, there were a total of 13,954 grid cells that contained nonfloodplain area.
2.2.3. River Flow Data
The river gauges that were used to model SW dynamics across the MDB were selected based on the follow-
ing considerations. The flow data should cover the entire analysis period from 1986 to 2011, should be with-
out major gaps and there should roughly be at least one gauge per EH-zone. Out of the 68 selected gauges,
50 had full data coverage for the analysis period, nine covered between 20 and 25 years and nine less than
20 years, with the two most problematic ones having less than 10 years of data. Since the majority of
gauges had full or nearly full data coverage, we did not interpolate any flow data and instead, only used
data for the exact period of coverage of each gauge in the modeling process. All flow data were down-
loaded from respective government repositories [NSW Government, 2014; Government of SA, 2016; QLD
Government, 2016; State Government VIC, 2016].
2.2.4. Driver Variables and Data Preparation
For each modeling unit within each of the three SW categories, we derived a numerical time series of SW
extent in the form of the fraction of the percentage of SW area of the total area of the cloud-free modeling
unit. We dropped SW extents of Landsat observations, where more than 40% of the area of a modeling unit
was classified as cloud. The resulting time series of SW extent served as the dependent variable for each SW
category and was modeled as a function of all driver variables, using a dynamic multivariate regression
framework. In order to analyze the role of local climate conditions in SW dynamics, we used spatially explicit
daily time series of P,ET, and SM as additional driver variables (hereafter referred to as local climate drivers)
besides river flow, which are all known to influence the magnitude and duration of floods [S
anchez-Carrillo
et al., 2004; Lamontagne and Herczeg, 2009]. The selection criteria for the corresponding data sets were that
they had to be spatially explicit and cover the entire analysis period. The Ptime series was based on interpo-
lation of gauge records throughout Australia [BOM, 2015] and the ET time series was modeled output of the
landscape component of the Australian Water Resource Assessment System (AWRA-L 4.5) continental scale
water balance model [Vaze et al., 2013; Viney et al., 2014]. The SM time series was based on a combination
of active and passive microwave data from multiple satellites [Liu et al., 2012; Wagner et al., 2012]. All three
local climate drivers were converted into numerical daily time series by computing spatial averages on the
level of the 10 310 km grid cells.
2.2.5. Statistical Modeling Approach
During the development of the basin-wide SW extent model, the original linear regression approach [Heim-
huber et al., 2016] was found not to accurately capture the large variety of SW extent to Qrelationships that
occurred on modeling units across the MDB. Therefore, we used generalized additive models (GAM) [Wood,
2011], which allowed us to model the nonlinear relationships between SW extent and Qusing a smooth
function, while the remaining explanatory variables were modeled additively based on linear regression.
The resulting model equation was:
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 5
SWextentt5b01b1SWextentðt21Þ1sðLagðQÞÞ1b2P1b3ET1b4SM1e(1)
Where SWextent
t
is the SW extent (%) at time t,SWextent
(t21)
is the SW extent of the previous available
Landsat observation (t21), sis the GAM smooth function, Lag(Q) is the discharge at the related gauge
lagged by the time it takes for a flood to travel from the gauge to the respective modeling cell, Pis local
rainfall, ET is evapotranspiration, SM is soil moisture, and eis the error term. The equation used for modeling
SW extent dynamics on nonfloodplain areas is the same as equation (1) but without discharge as a predictor
variable, given that those areas are not hydraulically connected to a river. For each SW category, SW extent
was also modeled as a function of SW extent on the previous available satellite observation as a fixed vari-
able, which is hereafter referred to as the lagged dependent variable. Lagged dependent variables are often
used for modeling time series data in economics [Keele and Kelly, 2006; Shumway and Stoffer, 2006] and
help to improve predictions when a process is known to depend on the state of the process as observed on
a previous time step, which is the case for SW extent dynamics. The smooth function in the GAM models is
essentially a combination of two or more polynomial curves, which are joined together at ‘‘knots,’’ with the
smoothness of the curve depending on a single smoothing parameter [Wood, 2011]. Due to the large num-
ber of modeling units, it was not feasible to define this smoothing parameter individually for each model.
Instead, we used a global smoothing parameter of four for all GAM models. Based on visual examination of
model fits, this value led to good approximations of the variable relationship in a variety of different areas
of the MDB.
The local climate drivers were modeled linearly and in the form of sums for the 16 day time period before
each observation for Pand ET (mm/16 days) and a moving average for the same time period for SM (%).
While some variables (Qfor floodplain and floodplain-lakes and Pfor nonfloodplain models) and the lagged
dependent variable were always included in the models (hereafter referred to as fixed variables), the local
climate drivers P,ET, and SM, respectively, were subject to an automated stepwise variable selection process
and only included in the models if adding them led to a reduction in root mean squared error (RMSE) in
fivefold cross validation (CV). The variables were added to the model in the order given in equation (1). For
the nonfloodplain models, only ET and SM were subject to the automated variable selection process since P
was the fixed driver for this model category, with Qdropped from these nonfloodplain models. To ensure
that only meaningful empirical relationships were modeled, we only used SW extent observations >0 and
set a minimum of 20 observations >0 as a threshold for fitting a model. In addition, we also masked out res-
ervoirs with surface area of more than 5 km
2
for analysis. We quantified the performance of the statistical
models based on the adjusted r
2
and RMSE in fivefold CV. The absolute and relative improvement in the
model’s RMSE in fivefold CV after accounting for the selected local climate drivers was used to analyze the
role of these drivers in SW extent dynamics. The absolute improvement in RMSE was calculated as the dif-
ference in fivefold CV RMSE before and after accounting for the local climate drivers and by dividing this dif-
ference by the initial model RMSE (before accounting for the additional drivers), we obtained the relative
improvement in RMSE. In addition, we also quantified the percentage of models in each zone, for which
each local climate driver was selected. The management of time series data and all statistical modeling was
implemented in R[R Development Core Team, 2008].
2.2.6. Quantifying Flood Travel Times
To account for the time that it takes for a flood to travel from the gauge to the modeling unit, Qwas lagged
by a previously quantified lag time before including it into the models. This lag time for discharge (hereafter
referred to as Qlag) is the modeled timing between discharges recorded at the gauge and the correspond-
ing remotely sensed inundation response of a downstream or upstream floodplain or floodplain-lake
modeling unit. Qlags were quantified by fitting GAM spline models between SW extent on the floodplain
and floodplain-lake units and Qlagged by each possible number of days between a lower limit of 220 and
an upper limit of 40 days, and selecting the number of days that led to the highest r
2
. For a detailed analysis
of the effect of Qlags as well as the suitability of GAM for capturing different Qto SW extent relationships,
we exemplified four floodplain modeling units with different SW extent dynamics and distances to the
respective modeling gauges (see Ex-A to Ex-D in Figure 1).
2.2.7. Predicting Surface Water Extent
One promising application of statistical SW inundation models is that they can be used to predict SW extent
time steps that are not captured by the Landsat time series as a result of cloud cover. Here we used two of
the four previously selected example floodplain modeling units that have contrasting inundation regimes
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 6
(see Ex-A and Ex-B in Figure 1) and applied the statistical models to predict SW extents at a regular 8 day
time step during the 2010/2011 La Nina Floods. These predictions were implemented using the previously
fitted GAM models and the predict function in R. We validated these predictions by calculating r
2
and RMSE
between predictions and Landsat observations of SW extent that were available during the selected time
periods.
3. Results
3.1. Example Floodplain Modeling Units
In many areas across the basin, the automated quantification of Qlags was a key step for establishing a
meaningful Qto SW extent relationship, to which the GAM models were then fitted along with the other
driver variables. The relationship between Qand SW extent is highly dependent on the floodplain topog-
raphy [Frazier and Page, 2009] and was found to vary substantially across different floodplains in the
study area (Figure 2). The four example floodplain modeling units (Ex) were located 140 (Ex-A), 85 (Ex-B),
130 (Ex-C), and 40 km (Ex-D) from their respective modeling gauges and had Qlagsof11,5,12,and4
days, respectively, which equals to an average flood propagation speed of 13 km/d. It can be seen that
the application of Qlags reduces the scatter in the data (comparing data points before (green) and after
(black) applying Qlags in Figure 2) and reveal distinctive curves, which indicate the flows that are need-
ed to achieve a certain SW extent on the corresponding floodplain area. In comparison to the two Paroo
floodplain units (Ex-B and Ex-C), the Lower Murray floodplain unit (Ex-A) does not respond to increased
river flows until a threshold of about 500 m
3
/s is reached and the floodplain starts to fill up. In the Paroo
floodplain units, increasing river flow immediately leads to increasing floodplain inundation until a satu-
ration point, above which only the highest peak flows of the 26 year analysis period leads to further
increase in SW extent. It is also evident that the Qlag has a more pronounced effect on the variable rela-
tionship in the Paroo examples (Ex-B and Ex-C) as compared to the lower Murray (Ex-A) example, where
there is already little scatter in the relationship before applying a lag time. The Barwon River example
(Ex-D) shows a linear relationship for the entire range of recorded discharge values which was also cap-
tured well by the GAM model.
3.2. Basin-Wide Modeling
Only 2953 out of all 4120 individual floodplain modeling units exceeded the threshold of more than 20 >0
observations of SW extent that was set for fitting statistical models. For the floodplain-lake category, 332
out of 447 units were modeled and 8926 out of all 13,954 nonfloodplain units (Table 1). For illustration pur-
poses, we averaged model metrics of individual modeling units spatially for the floodplain and floodplain-
lake categories based on zones that result from combining all grid cells that were modeled using the same
gauge (Figures 3 and 4). For the nonfloodplain models, we summarized the individual modeling results on
the basis of EH-zones (Figure 5). Despite using a single smoothing parameter for the GAM spline function,
the models were able to capture the large variety of Qto SW extent relationships occurring across the basin.
The average r
2
for all floodplain models within gauge connected cells was 0.65, with better average good-
ness of fit in the northern (0.68) as compared to the southern (0.58) MDB (Figure 3a). Floodplain-lakes
showed a similar pattern in model performance with an overall average r
2
of 0.66 for the entire MDB and r
2
of 0.7 and 0.59 for the northern and southern basin, respectively (Figure 4a). In contrast to this, the non-
floodplain models had a low total-basin r
2
of 0.25 and slightly better goodness of fit in the southern (0.28)
than in the northern (0.24) basin (Figure 5a). In this category, there was a noticeable pattern with the high-
est r
2
predominantly occurring in the S-W part of the basin, where 27 adjacent EH-zones achieved an aver-
age r
2
of 0.33 (see highlighted EH-zones in Figure 5a).
The relative improvement in model RMSE (fivefold CV) resulting from the lagged dependent variable was
by far the highest in the floodplain-lake category, with 47% improvement as compared to 22% for flood-
plains and 27% for nonfloodplain areas. The lagged dependent variable also led to the highest absolute
RMSE improvement in the floodplain-lake category with 0.15 compared to 0.05 for floodplains and 0.005 for
nonfloodplains and there was no noticeable difference between the northern and southern basin. In com-
parison to the improvements resulting from the lagged dependent variable, the improvements in RMSE
after adding the local climate drivers were small for all three model categories. The average absolute RMSE
reduction of the floodplain models across the entire basin was 0.002, with 0.0005 in the southern and
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 7
0.0026 in the northern basin (compare Figure 3b). In this category, there was also a distinctive patch of 14
adjacent zones in the N-W of the basin with a particularly high average RMSE improvement of 0.0036 for
1160 individual floodplain units (see highlighted areas in Figure 3b). It is important to note here that RMSEs
of floodplain models also differed across the basin before accounting for the local climate drivers (initial
RMSE), with an average initial RMSE of 0.048 in the northern compared to 0.025 in the southern basin
(Table 1).
To account for the different magnitude in RMSE of models prior to the variable selection process, we also
calculated the relative improvement in RMSE that was achieved through the local climate drivers (Figures
3c, 4c, and 5c). We found that the local climate drivers helped to improve predictive performance of flood-
plain models by 5.5% in the northern compared to 2.0% in the southern basin with an overall average
improvement of 4.9% (Table 1). Furthermore, there were 280 individual floodplain models where the local
climate drivers led to RMSE improvement of more than 10% of which 251 were in the northern and 29 were
in the southern basin. The average initial RMSE of these 280 floodplain models prior to accounting for the
Figure 2. River flow to SW extent relationship (black dots) and GAM model fits (blue lines) for four example floodplain modeling units (Ex)
located within the Lower Murray (Ex-A), Paroo (Ex-B, Ex-C), and Barwon River (Ex-D) (location is given in Figure 1). Green dots illustrate the
same relationship before applying Qlag times.
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 8
Table 1. Summary Table of the Statistical Modeling Results
a
Model Category
and Part of Basin
Number of
Modeling Units
Number of
Units Modeled r
2
RMSE
Improvement
Lagged
Dependent
Variable RMSE Without
Local Climate
Drivers
RMSE With
Local Climate
Drivers
RMSE
Improvement
Climate Drivers
Percentage of
Models With
(abs.) (rel.) (abs.) (rel.)
Number of
Models With
>10% RMSE
Improvement PETSM
Floodplains All 4120 2953 0.65 0.0419 22% 0.0413 0.0393 0.0020 4.86% 280 9% 58% 36% 40%
Floodplains South 1242 874 0.58 0.0260 25% 0.0247 0.0243 0.0005 1.97% 29 3% 48% 32% 34%
Floodplains North 2878 2079 0.68 0.0510 21% 0.0483 0.0456 0.0026 5.48% 251 12% 63% 37% 43%
Floodplain-Lakes All 447 332 0.66 0.1467 47% 0.0809 0.0784 0.0025 3.04% 29 9% 54% 30% 41%
Floodplain-Lakes S 190 136 0.59 0.1166 41% 0.0734 0.0716 0.0019 2.54% 4 3% 53% 34% 32%
Floodplain-Lakes N 257 196 0.70 0.1675 50% 0.0888 0.0853 0.0035 3.89% 25 13% 54% 28% 48%
Nonfloodplains All 13954 8926 0.25 0.0053 27% 0.0752 0.0750 0.00015 0.20% 935 10% 48% 51%
Nonfloodplains S 5779 3281 0.28 0.0048 36% 0.0802 0.0801 0.00007 0.09% 251 8% 41% 41%
Nonfloodplains N 8175 5345 0.24 0.0056 22% 0.0541 0.0539 0.00016 0.29% 684 13% 49% 56%
a
Results are given separately for the three model categories (i.e., floodplains, floodplain-lakes, and nonfloodplain areas) as well as for the entire, northern and southern
Murray-Darling Basin.
Figure 3. Model statistics for the floodplain category across the Murray-Darling Basin, averaged over grid cells that used the same modeling gauge. Shown are zonal averages of (a) r
2
,
(b) the absolute, and (c) relative improvement in RMSE after adding the local climate driver variables and the percentage of models, for which adding (d) P, (e) ET, and (f) SM led to RMSE
improvement. For better illustration of patterns in the modeling results, the symbology was adapted individually for each result metric.
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 9
local climate drivers was high with a RMSE of 0.070 compared to an overall average initial RMSE of 0.041 for
all models. This illustrates that large improvement in RMSE after accounting for the local climate drivers
may partially be linked to high initial model RMSEs. To understand this connection better, we regressed the
absolute RMSE improvement resulting from the local climate drivers against the initial RMSEs for all individ-
ual models. These regressions revealed an average r
2
of 0.2 for the floodplain and 0.3 for the nonfloodplain
category which indicates that despite a limited connection between these two metrics, high absolute
improvements after accounting for the local climate drivers can be achieved for low initial model RMSEs
and vice versa.
For floodplain-lakes, the average improvement in RMSE after accounting for the local climate drivers was
0.0025 across the basin and twice as high in the northern (0.0035) than in the southern (0.0019) basin. For
nonfloodplains, where Pwas a fixed variable and only ET and SM were accounted for in the variable selec-
tion process, the average improvement was much lower as compared to the other two model categories
with a reduction in RMSE of 0.00015 across the basin and 0.00007 and 0.00016 for the southern and north-
ern basin respectively. While the distribution of r
2
-based model performance between the northern and
southern basin was the opposite between the floodplain and floodplain-lake categories and the nonflood-
plain category, the RMSE improvement after accounting for the local climate drivers was higher in the
northern basin for all three model categories.
Figure 4. Model statistics for the floodplain-lake category across the Murray-Darling Basin, averaged over grid cells that used the same modeling gauge. Shown are zonal averages of
(a) r
2
, (b) the absolute, and (c) relative improvement in RMSE after adding the local climate driver variables and the percentage of models, for which adding (d) P, (e) ET, and (f) SM led to
RMSE improvement. For better illustration of patterns in the modeling results, the symbology was adapted individually for each result metric.
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 10
Based on the agreement between absolute and relative improvements in RMSE, we found the north-
ern and N-W part of the MDB to represent a hot spot for the role of the local climate drivers in
explaining SW dynamics in both the floodplain (see Figures 3b and 3c) and floodplain-lake (see
Figures 4b and 4c) category. In comparison, many of the nonfloodplain areas that had the highest
absolute RMSE improvements showed the lowest relative improvements after accounting for the local
climate drivers (see Figures 5b and 5c). The corresponding zones contain large areas of irrigated agri-
culture, which inhibit the establishment of meaningful variable relationships due to the high level of
artificial SW areas that are subject to irrigation management strategies. The artificially increased inun-
dation frequency of the corresponding areas (see irrigation areas in Figures 5b and 5c) can be seen
when looking at the number of >0 observations in the nonfloodplain areas (Figure 5f), which is higher
in these areas compared to the surrounding areas. The resulting dynamics of these SW areas led to
poor model performance and consequently high model RMSEs, which make larger absolute improve-
mentsinRMSEmorelikelywhenaccounting for the local climate drivers. At the same time, the very
low relative improvements in RMSE indicate that the local climate drivers did not play significant roles
in the dynamics of these artificial SW bodies. Instead, the relative RMSE improvements indicate that
similar to the other two model categories, the local climate drivers were most important for explaining
SW dynamics on nonfloodplain areas in the N-W part of the MDB.
Figure 5. Model statistics for the nonfloodplain SW category across the Murray-Darling Basin, averaged over the ecohydrological zones [Huang et al., 2013]. Shown are zonal averages of
(a) r
2
, (b) the absolute, and (c) relative improvement in RMSE after adding the local climate driver variables along with irrigation areas [BRS, 2008] and the percentage of models, for
which adding (d) ET and (e) SM led to RMSE improvement. (f) The number of SW extent observations >0 of each modeling grid cell for the 26 year analysis period. For bett er illustration
of patterns in the modeling results, the symbology was adapted individually for each result metric.
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 11
Both the absolute and relative improvements in the model’s predictive performance (RMSE improvement)
were based on each model’s individual combination of local climate drivers as determined in the variable
selection process. To get a better understanding of the role of each local climate driver, we also analyzed
their individual importance in general and across different zones in the MDB (d, e, f in Figures 3 and 4 and
d, e, in Figure 5). For both floodplain categories, Pwas the most important local climate driver and was
selected for 58% of all individual floodplain models as compared to 36% for ET and 40% for SM. There was a
noticeable difference in the role of Pfor explaining SW extent dynamics across the basin with 63% of all
floodplain model units in the northern basin accounting for Pcompared to 48% in the southern basin. P
helped to improve the predictive power of more than 70% of all floodplain models in many of the zones in
the N-W of the MDB (Figure 3d) with a maximum percentage of 77% within a zone with more than 50 indi-
vidual models in the zone that contains example cell Ex-A in Figure 1. The highest percentage within a zone
with more than 50 individual models where SM helped to explain SW extent dynamics on floodplains was
67% which was also achieved in the same zone.
Pwas selected for 54% of all floodplain-lake models and ET and SM for 30% and 41%, respectively. Similar
to the floodplain category, the highest percentages of models accounting for these drivers occurred in the
N-W of the MDB, which represented a hot spot for these local climate drivers. For the nonfloodplain catego-
ry, ET and SM were selected for about half of all models across the basin with a noticeable difference of
49% compared to 41% (ET) and 56% compared to 41% (SM) of models accounting for these drivers in the
northern and southern part of the basin, respectively. These findings indicate that Pplayed the most impor-
tant role for explaining SW dynamics on floodplains and floodplain-lakes followed by SM and ET. Additional-
ly, Pwas the fixed driver variable for the nonfloodplain models and the RMSE improvements resulting from
the local climate drivers (ET and SM) were by far the smallest in this category which indicates that Palso
played the most important role in explaining SW dynamics across nonfloodplain areas. In summary, our
results suggest that the local climate drivers played a more important role in SW dynamics in the northern
compared to the southern basin for all three model categories, with the N-W of the MDB representing a hot
spot.
3.3. Modeling the 2010/2011 La Nina Floods
The extreme dynamics during the large 2010/2011 La Nina Floods revealed the strengths and limitations of
Landsat data and the statistical inundation models for quantifying SW extent dynamics during major flood-
ing events. The very different flooding dynamics of the two example floodplain modeling units can be seen
clearly with several short but intense flood bursts in the Paroo site (Ex-C in Figure 6), which is located close
to the runoff generating catchment, as compared to the Lower Murray site (Ex-A in Figure 6), where the
large flood lasted for over half a year (see Figure 1, for location of Ex-A and Ex-C). The Paroo example illus-
trates the limitation of Landsat for capturing the short and intense flooding as a result of the comparatively
Figure 6. Comparison of the ability of statistical inundation models for quantifying SW extent dynamics during the 2010/2011 La Nina
floods based on two example floodplain modeling units with contrasting flooding regimes (see Ex-A and Ex-C in Figure 1). Shown are the
hydrographs of the modeling gauge along with observed (green bars) and statistically modeled (black bars) SW extents. For each example
floodplain unit, r
2
and RMSE of the validation against observed Landsat-based SW extents during the analysis period are given.
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 12
long satellite revisit interval and missing observations resulting from cloud cover. Even though this area in
the Paroo is located within overlapping satellite paths resulting in a potential time step of 8 days, there was
not a single observation with less than 50% cloud cover which would allow to quantify the flooding dynam-
ics during the 1 month period of the largest flood peak in March 2010. In both examples, the statistical inun-
dation models were able to overcome this limitation by predicting SW extents of the missing time steps
accurately.
Validation of predicted SW extents (black bars in Figure 6) against the observed Landsat-based SW extents
(green bars in Figure 6) revealed an r
2
of 0.97 and a RMSE of 2% for the Paroo floodplain unit and an r
2
of
0.97 and RMSE of 4% for the Murray example. It can be seen that in the Murray example, the prediction of
the maximum SW extent during the 9 month period of the flood (see 2011-03 in Figure 6a) is slightly bigger
than the highest available Landsat observation in the same period. This means that there was a SW extent
observation in a similar range during a different flood before the illustrated time period to which the model
was trained as no predictions above the maximum satellite-observed SW extent were made. The statistical
inundation models captured the contrasting flooding dynamics of the two example cells well and provided
predictions of SW extent at a regular interval of 8 days. Based on the error statistics of the validation time
steps, these predictions were a good approximation of the corresponding temporal flooding dynamics for
both examples. Both of the example floodplain modeling units had high r
2
for their respective GAM model
fits with an r
2
of 0.9 in the Paroo and 0.88 in the Murray example, which partly explains the high accuracy of
the predicted validation time steps. Considering the differences in goodness of fit across the basin for all
three model categories (see Figures 3a, 4a, and 5a), the accuracy of predicting missing time steps is likely to
vary correspondingly for different modeling units and would likely lead to the least accurate predictions for
the nonfloodplain category.
4. Discussion
4.1. Modeling Framework
Despite the growing need for better understanding the role of periodically inundated SW areas in the ter-
restrial water cycle and the Earth system in general, the propagation of large floods through river systems
and the corresponding SW dynamics remain poorly quantified on continental or global scales [Bates et al.,
2014]. Here we used an unprecedented Landsat time series of SW maps to quantify inundation dynamics
and the role of local climate drivers across a large and highly regulated river system. We quantified these
complex dynamics by building on a transferable grid-based framework [Heimhuber et al., 2016] and devel-
oped a model that comprised 68 river gauges and 18,521 individual spatial modeling units among three
SW categories (i.e., floodplains, floodplain-lakes, and nonfloodplains), out of which 12,211 exceeded the
threshold of 20 >0 observations of SW extent. The development of this framework required a variety of
ambiguous decisions in some areas such as selecting the most suitable gauge out of multiple candidates or
defining hydraulic connectivity between model gauges and grid cells on the edge of large and scattered
floodplains (i.e., top right corner in Figure 1b). Additionally, despite using comparatively small modeling
units, the framework was still a simplified approximation of the complex hydrological and hydraulic struc-
ture of the river system. Nevertheless, it enabled us to capture and model the spatiotemporal dynamics of
large floods and their propagation through the river system at an unprecedented level of detail.
The majority of all SW areas in the MDB are floodplains and our modeling framework achieved high average
goodness of fit (r
2
0.65) for the floodplain and floodplain-lake model categories. A key step for achieving
good model performance across the large variety of river and floodplain locations was the quantification of
Qlags. In many areas of the basin, meaningful variable relationships were revealed only after quantifying
and applying these Qlag times to the river flow data prior to fitting the models (i.e., Figure 2). Although this
automated quantification can lead to implausible Qlags for some individual modeling units [Heimhuber
et al., 2016], the basin-wide estimation of flow travel times that this study provides is an important step
toward quantifying the propagation of floods across large river systems. Furthermore, the application of Q
lags allowed us to provide statistical models of SW extent for all major river and floodplain systems across
the basin, which are essentially rating curves that relate flow in the river to the corresponding inundation
extent response on the modeling unit (i.e., Figure 2). In an unprecedented effort to restore the health of the
MDB’s water-dependent ecosystems, the Australian Government is currently undertaking large investments
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 13
in buying back water licenses and improving irrigation efficiencies to increase the share of water available
for environmental flows in the MDB [Burke, 2007; Banks and Docker, 2014]. In the context of this effort, our
whole-of-basin inundation response models can provide useful knowledge for managing the distribution of
environmental flow water across the MDB holistically and more effectively [Powell et al., 2008; Chen et al.,
2011; Sims et al., 2014].
4.2. Surface Water Categories
The unique dynamics of different type of SW bodies have previously been addressed by either focusing
study sites entirely on lakes [Gao et al., 2012; Wang et al., 2014; Zhang et al., 2014b; Hu et al., 2015]or flood-
plains [Frazier and Page, 2009; Jung et al., 2011; Sims et al., 2014], or by using large spatial modeling units
and modeling all SW areas simultaneously [Huang et al., 2014]. We found that nonfloodplain areas had by
far the lowest model performance (average r
2
of 0.24), which means that modeling all SW areas within a giv-
en EH-zone simultaneously would have resulted in lower average goodness of fit across the basin. The poor
performance of this category may partly be owed to the fact that the corresponding SW areas were mod-
eled as a function of Pover the grid cell as the fixed driver variable instead of Qso that potentially occurring
lateral inflows of SW through the ungauged local drainage network into the modeling unit were not
accounted for. This category also covered all areas that were not part of the two floodplain categories and
hence, comprised a wide range of natural and artificial SW bodies such as farm dams, irrigation paddies,
and isolated wetland systems. Additionally, many nonfloodplain areas did not exhibit sufficient SW dynam-
ics for establishing meaningful empirical relationships. In contrast to this, the S-W of the MDB contains
numerous isolated ephemeral river and wetland systems (see blue areas outside of colored grid cells in
Figure 1a) that seem to have a pronounced relationship with Pand had comparatively good model perfor-
mance (i.e., Figure 5a). Another major difference that we found between the three model categories was
that the lagged dependent variable was clearly most important for explaining SW extent dynamics in the
floodplain-lake category compared to the other two categories as a result of the slow changes in SW area
that occur on lakes. This implies that modeling applications focusing on more static SW areas such as lakes
and reservoirs can particularly benefit from statistical methods designed for dynamic processes such as the
regression-based lagged dependent variable approach used in this study or Auto Regressive Integrated
Moving Average (ARIMA) models. All these differences between the three model categories illustrate that
future studies should take into account that SW areas of large and complex rivers systems can have highly
variable dynamics that may require tailored modeling approaches.
4.3. Local Climate Drivers
We found that Pwas the most important local climate driver for explaining SW extent dynamics in all three
model categories, followed by SM and ET, which may partly be the result of the different nature of the data
sets (i.e., observed versus modeled) that we used for representing these drivers [Heimhuber et al., 2016]. In
addition, Pis arguably the main input of water to a modeling unit despite river flow, which might be anoth-
er reason for the higher importance of this variable. We also quantified the combined role of the local cli-
mate drivers for SW dynamics and found that they played a more relevant role in the northern as compared
to the southern MDB with a noticeable hot spot in the N-W. It is important to note that the improvements
in model RMSE achieved through the local climate drivers were generally small and hardly ever exceeded
10%. Nevertheless, the model results showed distinctive spatial patterns that were in line with key site char-
acteristics of the MDB, which are likely to have an effect on the role of local climate drivers in large-scale
inundation processes.
The highest levels of river regulation are found in the mountainous areas along the S-E and eastern bound-
ary of the basin [CSIRO, 2008; MDBA, 2010] and the adjacent zones to the west have large areas of irrigated
agriculture (i.e., Figure 5b). These areas are affected by high levels of SW abstraction and diversion, which
highly alters river flow and floodplain inundation regimes and may consequently inhibit the establishment
of meaningful statistical relationships between the model variables. In comparison, the Paroo and Warregoo
river systems in the very N-W of the MDB have little to no intervention to their natural flow regimes, are of
ephemeral and, in the case of the Paroo, also of terminal nature [MDBA, 2010], which means that the drying
of inundated floodplains is solely the result of infiltration and evapotranspiration. Additionally, the flood-
plains of these river systems are extensive, shallow, and sparsely vegetated as compared to many of the
southern basin’s floodplains, which are often more confined [Tulbure et al., 2016] and covered with
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 14
floodplain forest. It is likely that the increased importance of the local climate drivers in the N-W of the basin
is the result of a combination of these characteristics. However, we applied a simplified data-driven empiri-
cal modeling approach for quantifying complex land surface processes and further analyses are needed to
reinforce these findings on the role of hydroclimatic site characteristics in SW extent dynamics. It is well
known that the local climate characteristics before and during flooding such as P,ET, and SM can influence
the extent and duration of large-scale inundation processes [Overton et al., 2006; World Bank, 2016] but to
our best knowledge, their role in long-term SW extent dynamics has not been quantified across a large river
basin to date. Future work on SW dynamics should take into account that local Pand other local climate
drivers can play an important role in the flooding and drying dynamics of certain SW areas in addition to
lateral inflows and outflows of SW.
4.4. Quantifying Surface Water Inundation Dynamics
Traditionally, the inundation characteristics of an area have been quantified based on the statistical return
period of floods. The corresponding maximum flood extents are typically obtained through hydrodynamic
modeling for a range of relevant return periods (i.e., 5, 25, and 100 years) and incorporated into flood haz-
ard maps, which provide useful and well known tools for water resources and disaster management [Heim-
huber et al., 2015]. This physically based approach has been upscaled to generate return period flood
hazard maps at 90 m resolution for the whole terrestrial land surface between 568S and 608N[Sampson
et al., 2015]. There have been recent advances in improving the representation of complex floodplain top-
ographies and corresponding interactions between rivers and their adjacent floodplains in large-scale
hydrodynamic models [Yamazaki et al., 2011; Neal et al., 2012, 2015; Sampson et al., 2015]. In addition, a
recent study [Schumann et al., 2016] has applied a hydrodynamic model to generate a 40 year climatology
of floodplain inundation dynamics across Australia and found good agreement with total inundated areas
derived from a Landsat-based time series of SW extent. Nevertheless, these large-scale hydrodynamic mod-
els are still limited for accurately quantifying the complex and fine-scaled dynamics of periodically inundat-
ed and potentially shifting SW areas across large regulated river basins and over long periods of time
[Alsdorf and Lettenmaier, 2003].
In comparison to hydrodynamic modeling, satellite-based approaches typically do not provide information
about the return periods of floods. Instead, SW dynamics are characterized based on composite flooding
frequency maps which are derived directly from a series of classified images. Landsat-based studies have
expressed inundation frequency as the number of times a given pixel was flagged as water, divided by the
total number of satellite observations during a given year or the entire analysis period [Mueller et al., 2016;
Tulbure et al., 2016]. One of the limitations with this approach is that flooding may be underestimated, since
there are typically less cloud-free observations available during floods and peak flood extents are often
missed as a result of the 16 day revisit cycle of the Landsat satellite. As an alternative, MODIS imagery has
been used to create daily series of SW maps which were combined into temporally consistent composite
maps, showing the number of days that every pixel was inundated during a given year [Klein et al., 2015;
Kuenzer et al., 2015]. CSIRO has combined MODIS-based SW maps with river flow data, on which frequency
analysis was performed, to generate return period inundation maps for the MDB [Huang et al., 2014].
Despite their potentially achievable daily time step, MODIS-based SW time series are limited by their coarse
250 or 500 m resolution as well as their, in comparison to Landsat, lower SW classification accuracies and
limitations for capturing fine-scaled water bodies and inundation patterns [Chen et al., 2013]. It is evident,
that there is currently no validated method available for quantifying SW inundation dynamics continuously
at a high Landsat-like resolution on subcontinental or continental scale.
Our study partly addresses this limitation by using a Landsat-based time series of SW maps to provide mod-
els that statistically relate SW extent to river flow and other hydrological key drivers, for which long-term
daily data is available. The resulting statistical models can predict SW extent for each individual floodplain
and floodplain-lake modeling unit (max. 10 310 km) at a regular 8 or 16 day time step, which can lead to
substantially improved quantification of the temporal dynamics of large floods (i.e., Figure 6) as compared
to using only cloud-free satellite observations. One limitation of this approach was that estimation of predic-
tive accuracy was based on validation against a limited number of available SW extent observations. In
addition, the accuracy of predictions partly depends on the goodness of fit of each individual model, which
was variable across the MDB but predominantly high for all gauge connected SW areas.
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 15
Assuming that a certain numerical SW extent always corresponds to the same distribution of SW across the
modeling unit, statistical predictions of SW extent could be converted into SW maps, to generate a spatial
time series with regular time step. Our modeling units, however, are still large compared to a Landsat or
MODIS pixel so that conversion of numerical SW extents back into the spatial domain would be subject to a
variety of uncertainties. To address this limitation, future work should evaluate the suitability of image
fusion algorithms for quantifying SW dynamics [Gao et al., 2006; Zhu et al., 2010]. Image fusion takes advan-
tage of the complementary data characteristics of different satellite sensors such as Landsat and MODIS
and has previously been used to generate image time series with Landsat-like resolution and a regular time
step of MODIS composite images of eight days [Hilker et al., 2009; Walker et al., 2012; Emelyanova et al.,
2013; Knauer et al., 2016]. Although a few applications have applied image fusion to areas that are subject
to flooding [Emelyanova et al., 2013; Zhang et al., 2014a], a comprehensive analysis of the suitability for
quantifying SW extent dynamics on large river basin scale is currently missing. Therefore, future research
should assess the potential of data-fusion for quantifying long-term and large-scale SW extent dynamics
and compare this approach with the statistical modeling approach presented here.
5. Conclusions
We integrated 26 year long time series of Landsat-derived SW maps, P,ET, and SM with flow data from 68
river gauges to model SW extent dynamics and the role of drivers holistically across the MDB, and quanti-
fied flood propagation times for all of the basin’s key river and floodplain sites. GAM models captured the
large variety of nonlinear empirical Qto SW extent relationships across the basin well and the lagged
dependent variable led to large improvements in model performance, particularly for the floodplain-lake
category, which exhibits the slowest changes in SW extents. Categorizing SW areas was another crucial step
for achieving high goodness of fit of statistical inundation models, which was much higher for gauge-
connected SW areas (i.e., floodplains and floodplain-lakes) than for the remaining SW areas (nonfloodplain
areas) and better in the northern compared to the southern MDB. The local climate drivers were also more
important for SW dynamics in the northern basin with the N-W representing a hot spot. These N-S gradients
that our models revealed were in good accordance with key basin characteristics such as lower levels of riv-
er regulation, vaster and shallower floodplains and higher climate variability in the northern as compared to
the southern MDB. Our study illustrates that integrating multidecadal time series of earth observation data
with statistical modeling techniques can provide cost-effective tools for improving the management of lim-
ited SW resources and better understanding of long-term inundation dynamics and respective drivers. The
data-driven modeling framework is applicable to other large and complex river basins across the world and
provides statistical models that can predict SW extent for cloud-affected Landsat observations or during the
peak of floods and hence, allows a more detailed quantification of the spatiotemporal inundation dynamics
of large floods as compared to existing approaches.
References
Alsdorf, D., D. Lettenmaier, and C. V
or
osmarty (2003), The need for global, satellite-based observations of terrestrial surface waters, Eos
Trans. AGU,84(29), 269–280.
Alsdorf, D. E., and D. P. Lettenmaier (2003), Tracking fresh water from space, Science,301, 1491–1494.
Alsdorf, D. E., E. Rodr
ıguez, and D. P. Lettenmaier (2007), Measuring surface water from space, Rev. Geophys.,45, RG2002, doi:10.1029/
2006RG000197.
Arthington, A. H., P. C. Godfrey, R. G. Pearson, F. Karim, and J. Wallace (2014), Biodiversity values of remnant freshwater floodplain lagoons
in agricultural catchments: Evidence for fish of the Wet Tropics bioregion, northern Australia, Aquat. Conserv. Mar. Freshwater Ecosyst.,
25, 336–352, doi:10.1002/aqc.2489.
Azman, E. B., G. E. Y. Ying, L. Y. Y. Gladys, S. Yiting, B. S. Wu, and K. N. Irvine (2016), An exploratory application of remote sensing technolo-
gies and statistical analysis to provide rapid and cost effective inundation predictions for the Tonle Sap Lake Floodplain System,
J. Geogr. Environ. Earth Sci. Int., 5(3), 1–13, doi:10.9734/JGEESI/2016/23531.
Bakker, K. (2012), Water security: Research challenges and opportunities, Science,337, 914–915, doi:10.1126/science.1226337.
Banks, S. A., and B. B. Docker (2014), Delivering environmental flows in the Murray-Darling Basin (Australia)-legal and governance aspects,
Hydrol. Sci. J.,59(3–4), 688–699, doi:10.1080/02626667.2013.825723.
Bates, P. D., J. C. Neal, and D. Alsdorf (2014), Observing global surface water flood dynamics, Surv. Geophys.,35, 839– 852, doi:10.1007/
s10712-013-9269-4.
Bierkens, M. F. P. (2015), Global hydrology 2015: State, trends, and directions, Water Resour. Res.,51, 5198–5216, doi:10.1016/0022-
1694(68)90080-2.
Bishop-taylor, R., M. G. Tulbure, and M. Broich (2015), Surface water network structure, landscape resistance to movement and flooding
vital for maintaining ecological connectivity across Australia’ s largest river basin, Landscape Ecol.,30, 2045–2065, doi:10.1007/s10980-
015-0230-4.
Acknowledgments
The Landsat-based time series of SW
maps will be made freely available as
part of the Australian Research Data
Storage Infrastructure in accordance
with the rules of the funding agency
and embargo regulations of UNSW.
The Australian Geofabric river network
[BOM, 2012], the gridded Pdata, and
the AWRA-L ET data are available
through the Australian Bureau of
Meteorology (http://www.bom.gov.au/
water/landscape/). The SM time series
is provided by the European Space
Agency’s CCI (Climate Change
Initiative) project. The static wetland
layer for the MDB can be requested
from the Murray-Darling Basin
Authority. All river flow data used in
this analysis can be obtained from
Australian state government
repositories (see section 2.2.3). This
work was supported by the Australian
Research Council Linkage grant
(LP130100408) with co-funding from
the Murray-Darling Basin Authority
entitled ‘‘A novel approach for
assessing environmental flows using
satellite data.’’ The ecohydrological
zonation of the MDB used in this study
was provided by Chang Huang. The
static wetland layer for the MDB was
provided by Richard Kingsford and the
ET data were provided by the Bureau
of Meteorology with support from
Albert Van Dijk and Andrew Frost.
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 16
BOM (2012), Australian Geofabric, Bureau of Meteorology (Australian Government), ACT, Australia. [Available at http://www.bom.gov.au/
water/geofabric/, last access: 12 May 2016.]
BOM (2015), Australian Daily Rainfall Gridded Data, Bureau of Meteorolog y (Australian Government), ACT, Australia. [Available at http://
www.bom.gov. au/climate/how/newproducts/IDCdrgrids.shtml, last access: 12 May 2016.]
BRS (2008), Combined Irrigation Areas of Australia Dataset, Bur. of Rural Sci., Canberra, Australia.
Bunn, S. E., M. C. Thoms, S. K. Hamilton, and S. J. Capon (2006), Flow variability in dryland rivers: Boom, bust and the bits in between, River
Res. Appl.,22, 179–186, doi:10.1002/rra.904.
Burke, T. (2007), Water Act 2007—Basin Plan, Commonwealth of Australia.
Cauduro, R., D. De Paiva, D. C. Buarque, W. Collischonn, and M. Bonnet (2013), Large-scale hydrologic and hydrodynamic modeling of the
Amazon River basin, Water Resour. Res.,49, 1226–1243, doi:10.1002/wrcr.20067.
Chen, Y., S. M. Cuddy, B. Wang, L. E. Merrin, D. Pollock, and N. Sims (2011), Linking inundation timing and extent to ecological response
models using the Murray-Darling Basin Floodplain Inundation Model (MDB-FIM), in MODSIM2013, 20th International Congress on Model-
ling and Simulation, edited by J. Piantadosi, R. S. Anderssen, and J. Boland, Modell. and Simul. Soc. of Aust. and N. Z., 1–6 Dec. 2013.
Chen, Y., S. M. Cuddy, L. E. Merrin, C. Huang, D. Pollock, N. Sims, B. Wang, and Q. Bai (2012), Murray-Darling Basin Floodplain Inundation
Model Version 2.0 (MDB-FIM2), technical report, CSIRO Water for a Healthy Country Flagship, Australia.
Chen, Y., C. Huang, C. Ticehurst, L. Merrin, and P. Thew (2013), An evaluation of MODIS daily and 8-day composite products for floodplain
and wetland inundation mapping, Wetlands,33, 823–835, doi:10.1007/s13157-013-0439-4.
Costanza, R., R. de Groot, P. Sutton, S. van der Ploeg, S. J. Anderson, I. Kubiszewski, S. Farber, and R. K. Turner (2014), Changes in the global
value of ecosystem services, Glob. Environ. Change,26, 152–158, doi:10.1016/j.gloenvcha.2014.04.002.
Costelloe, J. F., R. B. Grayson, R. M. Argent, and T. A. McMahon (2003), Modelling the flow regime of an arid zone floodplain river, Diamanti-
na River, Australia, Environ. Model. Software,18, 693–703, doi:10.1016/S1364-8152(03)00071-9.
CSIRO (2008), Water Availability in the Murray-Darling Basin. A Report From CSIRO to the Australian Government From the CSIRO Murray-
Darling Basin Sustainable Yields Project, CSIRO, Australia.
CSIRO and Geoscience Australia (2011), 1 Second SRTM Derived Digital Elevation Models User Guide, Version 1.0.4, Geosci . Australia, Canberra,
Australia.
Davidson, N. C. (2014), How much wetland has the world lost? Long-term and recent trends in global wetland area, Mar. Freshwater Res.,
64, 934–941, doi:10.1071/MF14173.
Emelyanova, I. V., T. R. McVicar, T. G. Van Niel, L. T. Li, and A. I. J. M. van Dijk (2013), Assessing the accuracy of blending Landsat-MODIS sur-
face reflectances in two landscapes with contrasting spatial and temporal dynamics: A framework for algorithm selection, Remote Sens.
Environ.,133, 193–209, doi:10.1016/j.rse.2013.02.007.
Environment Australia (2001), A Directory of Important Wetlands in Australia, 3rd ed., Environ. Australia, Canberra, Australia.
Finlayson, C. M., N. C. Davidson, A. G. Spiers, and N. J. Stevenson (1999), Global wetland inventory-current status and future priorities, Mar.
Freshwater Res.,50, 717–727, doi:10.1071/MF99098.
Frazier, P., and K. Page (2009), A reach-scale remote sensing technique to relate wetland inundation to river flow, River Res. Appl.,25, 836–
849, doi:10.1002/rra.1183.
Gao, F., J. Masek, M. Schwaller, and F. Hall (2006), On the blending of the Landsat and MODIS surface reflectance: Predicting daily Landsat
surface reflectance, IEEE Trans. Geosci. Remote Sens.,44(8), 2207–2218, doi:10.1109/TGRS.2006.872081.
Gao, H., C. Birkett, and D. P. Lettenmaier (2012), Global monitoring of large reservoir storage from satellite remote sensing, Water Resour.
Res.,48, W09504, doi:10.1029/2012WR012063.
Getirana, A. C. V., A. Boone, D. Yamazaki, B. Decharme, N. Mognard, and F. Papa (2012), The Hydrologica l Modeling and Analysis Platform
(HyMAP): Evaluation in the Amazon Basin, J. Hydrometeorol.,13, 1641–1665, doi:10.1175/JHM-D-12-021.1.
Gopal, B. (2009), Biodiversity in wetlands, in The Wetlands Handbook, edited by E. Maltby and T. Barker, Wiley-Blackwell, Oxford, U. K.
Government of SA (2016), WaterConnect, Gov. of South Australia, Australia. [Available at http://www.waterconnect.sa.gov.au, last access:
12 April 2016.]
Gumbricht, T., P. Wolski, P. Frost, and T. S. McCarthy (2004), Forecasting the spatial extent of the annual flood in the Okavango delta,
Botswana, J. Hydrol.,290, 178–191, doi:10.1016/j.jhydrol.2003.11.010.
Hamilton, S. K. (2010), Wetlands of large rivers: Flood plains, in River Ecosystem Ecology: A Global Perspective, edited by G. E. Likens, pp.
607–610, Elsevier, Oxford, U. K.
Heimhuber, V., J. Hannemann, and W. Rieger (2015), Flood risk management in remote and impoverished areas—A case study of Onaville,
Haiti, Water,7, 3832–3860, doi:10.3390/w7073832.
Heimhuber, V., M. G. Tulbure, and M. Broich (2016), Modeling 25 years of spatio-temporal surface water and inundation dynamics on large
river basin scale using time series of Earth observation data, Hydrol. Earth Syst. Sci.,20, 2227–2250, doi:10.5194/hess-20-2227-2016.
Hilker, T., M. A. Wulder, N. C. Coops, N. Seitz, J. C. White, F. Gao, J. G. Masek, and G. Stenhouse (2009), Generation of dense time series syn-
thetic Landsat data through data blending with MODIS using a spatial and temporal adaptive reflectance fusion model, Remote Sens.
Environ.,113, 1988–1999, doi:10.1016/j.rse.2009.05.011.
Hu, Y., J. Huang, Y. Du, P. Han, and W. Huang (2015), Monitoring spatial and temporal dynamics of flood regimes and their relation to wet-
land landscape patterns in Dongting Lake from MODIS time-series imagery, Remote Sens.,7, 7494–7520, doi:10.3390/rs70607494.
Huang, C., Y. Chen, and J. Wu (2013), GIS-based spatial zoning for flood inundation modelling in the Murray-Darling Basin, in MODSIM2013,
20th International Congress on Modelling and Simulation, edited by J. Piantadosi, R. S. Anderssen, and J. Boland, Modell. and Simul. Soc.
of Aust. and N. Z. [Available at www.mssanz.org.au/modsim2013.]
Huang, C., Y. Chen, and J. Wu (2014), Mapping spatio-temporal flood inundation dynamics at large river basin scale using time-series flow
data and MODIS imagery, Int. J. Appl. Earth Observ. Geoinform.,26, 350–362, doi:10.1016/j.jag.2013.09.002.
Jones, K., Y. Lanthier, P. van der Voet, E. van Valkengoed, D. Taylor, and D. Fern
andez-Prieto (2009), Monitoring and assessment of wetlands
using Earth Observation: The GlobWetland project, J. Environ. Manage.,90, 2154–2169, doi:10.1016/j.jenvman.2007 .07.037.
Jung, H. C., D. Alsdorf, M. Moritz, H. Lee, and S. Vassolo (2011), Analysis of the relationship between flooding area and water height in the
Logone floodplain, Phys. Chem. Earth,36, 232–240, doi:10.1016/j.pce.2011.01.010.
Karim, F., C. Petheram, S. Marvanek, C. Ticehurst, J. Wallace, and B. Gouweleeuw (2011), The use of hydrodynamic modelling and remote
sensing to estimate floodplain inundation and flood discharge in a large tropical catchment, in MODSIM201 1, 19th International Con-
gress on Modelling and Simulation, edited by F. Chan, D. Marinova, and R. S. Anderssen, Modell. Simul. Soc. of Aust. and N. Z., 12–16
December, pp. 3796–3802. [Available at http://mssanz.org.au/modsim2011.]
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 17
Karpatne, A., A. Khandelwal, X. Chen, V. Mithal, J. Faghmous, and V. Kumar (2016), Global Monitoring of inland water dynamics: State-of-
the-art, challenges, and opportunities, in Computational Sustainability, Studies in Computational Intelligence , vol. 645, edited by J. L
assig,
K. Kersting, and K. Morik, pp. 121–147, Springer Int. Publ., Switzerland.
Keele, L., and N. J. Kelly (2006), Dynamic models for dynamic theories: The ins and outs of lagged dependent variables, Polit. Anal.,14,
186–205, doi:10.1093/pan/mpj006.
Kingsford, R. (2000), Ecological impacts of dams, water diversions and river management on floodplain wetlands in Australia, Austral Ecol.,
25, 109–127.
Kingsford, R. T., and R. F. Thomas (2004), Destruction of wetlands and waterbird populations by dams and irrigation on the Murrumbidgee
River in Arid Australia, Environ. Manage.,34(3), 383–396, doi:10.1007/s00267-004-0250-3.
Kingsford, R. T., K. Brandis, R. F. Thomas, P.Crighton, E. Knowles, and E. Gale (2004), Classifying landform at broad spatial scales: The distri-
bution and conservation of wetlands in New South Wales, Australia, Mar. Freshwater Res.,55, 17–31, doi:10.1071/MF03075.
Klein, I., A. J. Dietz, U. Gessner, A. Galayeva, A. Myrzakhmetov, and C. Kuenzer (2014), Evaluation of seasonal water body extents in Central
Asia over the past 27 years derived from medium-resolution remote sensing data, Int. J. Appl. Earth Observation Geoinform.,26,
335–349, doi:10.1016/j.jag.2013.08.004.
Klein, I., A. Dietz, U. Gessner, and S. Dech (2015), Results of the Global WaterPack: A novel product to assess inland water body dynamics
on a daily basis, Remote Sens. Lett.,6(1), 78–87, doi:10.1080/2150704X.2014.1002945.
Knauer, K., U. Gessner, F. Rasmus, and K. Claudia (2016), An ESTARFM fusion framework for the generation of large-scale time series in
cloud-prone and heterogeneous landscapes, Remote Sens.,8(5), 425, doi:10.3390/rs8050425.
Kuenzer, C., I. Klein, T. Ullmann, E. Georgiou, R. Baumhauer, and S. Dech (2015), Remote sensing of river delta inundation: Exploiting the
potential of coarse spatial resolution, temporally-dense MODIS time series, Remote Sens.,7, 8516–8542, doi:10.3390/rs70708516.
Lamontagne, S., and A. L. Herczeg (2009), Hydrology and ecohydrology of Australian semi-arid wetlands, Hydrol. Processes,23, 3413–3414,
doi:10.1002/hyp.7462.
Leauthaud, C., G. Belaud, S. Duvail, R. Moussa, O. Gr
unberger, and J. Albergel (2013), Characterizing floods in the poorly gauged wetlands
of the Tana River Delta, Kenya, using a water balance model and satellite data, Hydrol. Earth Syst. Sci.,17, 3059–3075, doi:10.5194/hess-
17-3059-2013.
Leblanc, M., S. Tweed, A. Van Dijk, and B. Timbal (2012), A review of historic and future hydrological changes in the Murray-Darling Basin,
Glob. Planet. Change,80–81, 226–246, doi:10.1016/j.gloplacha.2011.10.012.
Lemly, A., R. Kingsford, and J. Thompson (2000), Irrigated agriculture and wildlife conservation: Conflict on a global scale, Environ. Manage.,
25(5), 485–512, doi:10.1007/s002679910039.
Lettenmaier, D. P., D. Alsdorf, J. Dozier, G. J. Huffman, M. Pan, and E. F. Wood (2015), Inroads of remote sensing into hydrologic science dur-
ing the WRR era, Water Resour. Res.,51, 7309–7342, doi:10.1002/2015WR017616.
Liu, Y. Y., W. A. Dorigo, R. M. Parinussa, R. A. M. de Jeu, W. Wagner, M. F. McCabe, J. P. Evans, and A. I. J. M. van Dijk (2012), Trend-preservi ng
blending of passive and active microwave soil moisture retrievals, Remote Sens. Environ.,123, 280–297, doi:10.1016/j.rse.2012.03.014.
Maltby, E., and M. C. Acreman (2011), Ecosystem services of wetlands: Pathfinder for a new paradigm, Hydrol. Sci. J.,56(8), 1341–1359, doi:
10.1080/02626667.2011.631014.
Midekisa, A., G. B. Senay, and M. C. Wimberly (2014), Multisensor earth observations to characterize wetlands and malaria epidemiology in
Ethiopia, Water Resour. Res.,50, 8791–8806, doi:10.1002/2014WR015634.
Mueller, N. et al. (2016), Water observations from space: Mapping surface water from 25 years of Landsat imagery across Australia, Remote
Sens. Environ.,174, 341–352, doi:10.1016/j.rse.2015.11.003.
Murray-Darling Basin Authority (MDBA) (2010), Guide to the Proposed Basin Plan: Technical Background, Murray-Darling Basin Auth., Canber-
ra, Australia.
Musa, Z. N., I. Popescu, and A. Mynett (2015), A review of applications of satellite SAR, optical, altimetry and DEM data for surface water
modelling, mapping and parameter estimation, Hydrol. Earth Syst. Sci.,19, 3755–3769, doi:10.5194/hess-19-3755-2015.
Neal, J., G. Schumann, and P. Bates (2012), A subgrid channel model for simulating river hydraulics and floodplain inundation over large
and data sparse areas, Water Resour. Res.,48, W11506, doi:10.1029/2012WR012514.
Neal, J. C., N. A. Odoni, M. A. Trigg, J. E. Freer, J. Garcia-pintado, D. C. Mason, M. Wood, and P. D. Bates (2015), Efficient incorporation of
channel cross-section geometry uncertainty into regional and global scale flood inundation models, J. Hydrol.,529, 169–183, doi:
10.1016/j.jhydrol.2015.07.026.
NSW Government (2014), Pinneena 10.2 Continuous Monitoring: Users Guide. New South Wales Surface Water Data Archive, NSW Dep. of Primary
Ind., Off. of Water, Sydney, NSW, Australia. [Available at http://waterinfo.nsw.gov.au/pinneena/cm.shtml, last access: 12 April 2016.]
Ogilvie, A., G. Belaud, C. Delenne, J. S. Bailly, J. C. Bader, A. Oleksiak, L. Ferry, and D. Martin (2015), Decadal monitoring of the Niger Inner
Delta flood dynamics using MODIS optical data, J. Hydrol.,523, 368–383, doi:10.1016/j.jhydrol.2015.01.036.
Overton, I. C. (2005), Modelling floodplain inundation on a regulated river: Integrating GIS, remote sensing and hydrological models, River
Res. Appl.,21, 991–1001, doi:10.1002/rra.867.
Overton, I. C., K. McEwan, and J. R. Sherrah (2006), The River Murray Floodplain Inundation Model (RiM-FIM) Hume Dam to Lower Lakes,
CSIRO Water for a Healthy Country technical report, CSIRO, Canberra, Australia.
Papa, F., C. Prigent, F. Aires, C. Jimenez, W. B. Rossow, and E. Matthews (2010), Interannual variability of surface water extent at the global
scale, 1993–2004, J. Geophys. Res.,115, D12111, doi:10.1029/2009JD012674.
Pearson, R. G. A., P. C. B. Godfrey, A. H. B. Arthington, J. C. Wallace, F. Karim, and M. Ellison (2013), Biophysical status of remnant freshwater
floodplain lagoons in the Great Barrier Reef catchment: A challenge for assessment and monitoring, Mar. Freshwater Res.,64, 208–222,
doi:10.1071/MF12251.
Pekel, J., A. Cottam, N. Gorelick, and A. S. Belward (2016), High-resolution mapping of global surface water and its long-term changes,
Nature,540, 418–422, doi:10.1038/nature20584.
Penton, D. J., and I. C. Overton (2007), Spatial Modelling of Floodplain Inundation Combining Satellite Imagery and Elevation Models, in
MODSIM 2007 International Congress on Modelling and Simulation, edited by L. Oxley and D. Kulasiri, pp. 1464–1470, Modell. and Simul.
Soc. of Aust. and N. Z.
Powell, S. J., R. A. Letcher, and B. F. W. Croke (2008), Modelling floodplain inundation for environmental flows: Gwydir wetlands, Australia,
Ecol. Modell.,211, 350–362, doi:10.1016/j.ecolmodel.2007.09.013.
QLD Government (2016), Water Monitoring Portal, Queensland Gov., Australia. [Available at https://water-monitoring.information.qld.gov.
au/, last accessed 19 Jan 2017.]
R Development Core Team (2008), R: A Language and Environment for Statistical Computing, R Found. for Stat. Comput., Vienna, Austria.
[Available at http://www.R-project.org, last access: 12 May 2016.]
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 18
Ramsar Convention Secretariat (2014), The List of Wetlands of International Importance. [Available at http://www.ramsar.org/library, last
access: 13 Sep 2016.]
Ren, S., R. T. Kingsford, and R. F. Thomas (2010), Modelling flow to and inundation of the Macquarie Marshes in arid Australia, Environmet-
rics,21, 549–561, doi:10.1002/env.1002.
Rudorff, C. M., J. M. Melack, and P. D. Bates (2014), Flooding dynamics on the lower Amazon floodplain: 1. Hydraulic controls on water ele-
vation, inundation extent, and river-floodplain discharge, Water Resour. Res.,50, 619–634, doi:10.1002/2013WR014091.
Sakamoto, T., N. Van Nguyen, A. Kotera, H. Ohno, N. Ishitsuka, and M. Yokozawa (2007), Detecting temporal changes in the extent of annual
flooding within the Cambodia and the Vietnamese Mekong Delta from MODIS time-series imagery, Remote Sens. Environ.,109, 295–313,
doi:10.1016/j.rse.2007.01.011.
Sampson, C. C., A. M. Smith, P. D. Bates, J. C. Neal, L. Alfieri, and J. E. Freer (2015), A high-resolution global flood hazard model, Water Resour.
Res.,51, 7358–7381, doi:10.1002/2015WR016954.
S
anchez-Carrillo, S., D. G. Angeler, R. S
anchez-Andr
es, M. Alvarez-Cobelas, and J. Garatuza-Pay
an (2004), Evapotranspiration in semi-arid
wetlands: Relationships between inundation and the macrophyte-cover: Open-water ratio, Adv. Water Resour.,27, 643–655, doi:
10.1016/j.advwatres.2004.02.018.
Schumann, G. J., D. Stampoulis, A. M. Smith, and C. C. Sampson (2016), Rethinking flood hazard at the global scale, Geophys. Res. Lett.,43,
10,249–10,256, doi:10.1002/2016GL070260.
Shaikh, M., D. Green, and H. Cross (2001), A remote sensing approach to determine environmental flows for wetlands of the Lower Darling
River, New South Wales, Australia, Int. J. Remote Sens.,22(9), 1737–1751.
Shumway, R. H., and D. S. Stoffer (2006), Time Series Analysis and its Applications: With R Examples, Springer, New York.
Sims, N. C. et al. (2014), RiM-FIM Floodplain Inundation Modelling for the Edward-Wakool, Lower Murrumbidgee and Lower Darling River
Systems. Report prepared for the Murray-Darling Basin Authority, CSIRO Water for a Healthy Country Flagship, Canberra, Australia.
State Government VIC (2016), Water Monitoring, State Gov. Victoria, Australia. [Available at http://data.water.vic.gov.au/monitoring.htm,
last access: 12 April 2016.]
Ticehurst, C. J., Y. Chen, F. Karim, D. Dutta, and B. Gouweleeuw (2013), Using MODIS for mapping flood events for use in hydrological
and hydrodynamic models: Experiences so far, in MODSIM2013, 20th International Congress on Modelling and Simulation, edited by
J. Piantadosi, R. S. Anderssen, and J. Boland, pp. 1721–1727, Modell. and Simul. Soc. of Aust. and N. Z. [Available at www.mssanz.org.au/
modsim2013.]
Tulbure, M. G., and M. Broich (2013), Spatiotemporal dynamic of surface water bodies using Landsat time-series data from 1999 to 2011,
ISPRS J. Photogramm. Remote Sens.,79, 44–52, doi:10.1016/j.isprsjprs.2013.01.010.
Tulbure, M. G., S. Kininmonth, and M. Broich (2014), Spatiotemporal dynamics of surface water networks across a global biodiversity
hotspot-implications for conservation, Environ. Res. Lett.,9, 114012, doi:10.1088/1748-9326/9/11/114012.
Tulbure, M. G., M. Broich, S. V Stehman, and A. Kommareddy (2016), Surface water extent dynamics from three decades of seasonally con-
tinuous Landsat time series at subcontinental scale in a semi-arid region, Remote Sens. Environ.,178, 142–157, doi:10.1016/j.rse.2016.
02.034.
Vaze, J. et al. (2013), The Australian Water Resource Assessment Modelling System (AWRA), in MODSIM2013, 20th International Congress on
Modelling and Simulation, edited by J. Piantadosi, R. S. Anderssen, and J. Boland, Modell. and Simul. Soc. of Aust. and N. Z. [Available at
www.mssanz.org.au/modsim2013.]
Viney, N., J. Vaze, R. Crosbie, B. Wang, W. Dawes, and A. Frost (2014), AWRA-L v4.5: Techn ical Description of Model Algorithms and Inputs,
CSIRO, Canberra, Australia.
V
or
osmarty, C. J., et al. (2010), Global threats to human water security and river biodiversity, Nature,467, 555–561, doi:10.1038/nature09549.
Wagner, W., W. Dorigo, R. De Jeu, D. Fernandez, J. Benveniste, E. Haas, and M. Ertl (2012), Fusion of active and passive microwave observa-
tions to create an essential climate variable data record on soil moisture, in ISPRS Annals of the Photogrammetry,Remote Sens. and Spa-
tial Inform. Sci., vol. I-7, 2012 XXII ISPRS Congress, edited by M. Shortis, W. Wagner, and J. Hyypp
a, 25 August–1 September 2012, ISPRS,
Melbourne, Australia.
Walker, J. J., K. M. De Beurs, R. H. Wynne, and F. Gao (2012), Evaluation of Landsat and MODIS data fusion products for analysis of dryland
forest phenology, Remote Sens. Environ.,117, 381–393, doi:10.1016/j.rse.2011.10.014.
Wang, J., Y. Sheng, and T. S. D. Tong (2014), Monitoring decadal lake dynamics across the Yangtze Basin downstream of Three Gorges
Dam, Remote Sens. Environ.,152, 251–269, doi:10.1016/j.rse.2014.06.004.
Westra, T., and R. R. De Wulf (2009), Modelling yearly flooding extent of the Waza-Logone floodplain in northern Cameroon based on
MODIS and rainfall data, Int. J. Remote Sens.,30(21), 5527–5548, doi:10.1080/01431160802672872.
Wood, S. N. (2011), Fast stable restricted maximum likelihood and marginal likelihood estimation of semiparametric generalized linear
models, J. R. Stat. Soc.,73(1), 3–36.
World Bank (2016), Earth observation for water resources management [License: Creative Commons Attribution CC BY 3.0 IGO], in Current
Use and Future Opportunities for the Water Sector, edited by I. P. Garc
ıa et al., World Bank Group, Washington, D. C., doi:10.1596/978-1-
4648-0475-5.
World Economic Forum (2015), Global Risks 2015, 10th ed., World Econ. Forum, Geneva. [Available at www.weforum.org/risks.]
Yamazaki, D., S. Kanae, H. Kim, and T. Oki (2011), A physically based description of floodplain inundation dynamics in a global river routing
model, Water Resour. Res.,47, W04501, doi:10.1029/2010WR009726.
Zhang, F., X. Zhu, and D. Liu (2014a), Blending MODIS and Landsat images for urban flood mapping, Int. J. Remote Sens.,35(9), 3237–3253,
doi:10.1080/01431161.2014.903351.
Zhang, S., H. Gao, and B. S. Naz (2014b), Monitoring reservoir storage in South Asia from multisatelli te remote sensing, Water Resour. Res.,
50, 8927–8943, doi:10.1002/2014WR015829.
Zhu, X., J. Chen, F. Gao, X. Chen, and J. G. Masek (2010), An enhanced spatial and temporal adaptive reflectance fusion model for complex
heterogeneous regions, Remote Sens. Environ.,114, 2610–2623, doi:10.1016/j.rse.2010.05.032.
Water Resources Research 10.1002/2016WR019858
HEIMHUBER ET AL. MODELING SURFACE WATER 19
... org/record/2438110#.XbyctS0ZO3U) that are due to hydroclimatic events and land-use change (Tulbure and Broich, 2019) per season per year. However, the 16-day overpass frequency of Landsat and occasional cloudy days only allowed assessment of surface water and flooding extent at irregular intervals, the shortest being 16 days (Heimhuber et al., 2017), and thus floods were likely missed. Other work addressed this challenge by modeling flooding dynamics with two main approaches. ...
... Other work addressed this challenge by modeling flooding dynamics with two main approaches. One approach is data-driven top-down statistical models, such as generalized additive models or linear regression with river flow and climate data as covariates to predict flooding as well as predicting flooding dynamics in between cloudy Landsat observations (Heimhuber et al., 2017;Walker et al., 2020). The second approach is satellite data fusion such as blending Landsat and MODIS data using STARTFM and ESTARTFM algorithms (Emelyanova et al., 2013;Heimhuber et al., 2018;Ticehurst et al., 2014). ...
... In the northern part of the MDB, including the Paroo, the time of year with highest amount of flooding is summer due to increased rainfall in their corresponding catchments. Given that the Paroo is located close to runoff-generating catchment areas, local rainfall events have a more prominent effect on flooding dynamics (Heimhuber et al., 2016;Heimhuber et al., 2017). However, the flooding picked up by HLS in 2019 occurred at the end of March (Fig. 4), due to the ex-tropical cyclone Trevor delivering a needed relief to drought-affected southwest Queensland and northwest New South Wales in the Paroo and Warrego river catchments, with over 100 mm of rainfall recorded in the area between 26 and 30 March (Australian Government Bureau of Meteorology, 2019). ...
Article
Full-text available
Spatiotemporal quantification of surface water and flooding is essential given that floods are among the largest natural hazards. Effective disaster response management requires near real-time information on flood extent. Satellite remote sensing is the only way of monitoring these dynamics across vast areas and over time. Previous water and flood mapping efforts have relied on optical time series, despite cloud contamination. This reliance on optical data is due to the availability of systematically acquired and easily accessible optical data globally for over 40 years. Prior research used either MODIS or Landsat data, trading either high temporal density but lower spatial resolution or lower cadence but higher spatial resolution. Both MODIS and Landsat pose limitations as Landsat can miss ephemeral floods, whereas MODIS misses small floods and inaccurately delineates flood edges. Leveraging high temporal frequency of 3–4 days of the existing Landsat-8 (L8) and two Sentinel-2 (S2) satellites combined, in this research, we assessed whether the increased temporal frequency of the three sensors improves our ability to detect surface water and flooding extent compared to a single sensor (L8 alone). Our study area was Australia’s Murray-Darling Basin, one of the world’s largest dryland basins that experiences ephemeral floods. We applied machine learning to NASA’s Harmonized Landsat Sentinel-2 (HLS) Surface Reflectance Product, which combines L8 and S2 observations, to map surface water and flooding dynamics. Our overall accuracy, estimated from a stratified random sample, was 99%. Our user’s and producer’s accuracy for the water class was 80% (±3.6%, standard error) and 76% (±5.8%). We focused on 2019, one of the most recent years when all three HLS sensors operated at full capacity. Our results show that water area (permanent and flooding) identified with the HLS was greater than that identified by L8, and some short-lived flooding events were detected only by the HLS. Comparison with high resolution (3 m) PlanetScope data identified extensive mixed pixels at the 30 m HLS resolution, highlighting the need for improved spatial resolution in future work. The HLS has been able to detect floods in cases when one sensor (L8) alone was not, despite 2019 being one of the driest years in the area, with few flooding events. The dense optical time-series offered by the HLS data is thus critical for capturing temporally dynamic phenomena (i.e., ephemeral floods in drylands), highlighting the importance of harmonized data such as the HLS.
... A multi-variate approach could quantify simultaneous drivers, while the removal of managed water bodies, which are more strongly influenced by water management than weather (Fergus et al., 2020), may help to isolate climate responses. For more static water bodies such as lakes and reservoirs, which DSWEmod is more adept at identifying due to pixel size limitations, lagged drivers are shown to be valuable in driving surface water extent dynamics (Heimhuber et al., 2017) and may be worth considering in future analyses. Other changes in LULC can have underlying impacts on surface water dynamics, including changes in urban area (e.g., expansion and intensity), agricultural practices (e.g., irrigation, evapotranspiration), vegetation cover and type (e.g., water use), and could improve our understanding of regional drivers of surface water (Schilling et al., 2008;Stonestrom et al., 2009). ...
Article
Full-text available
Satellite imagery is commonly used to map surface water extents over time, but many approaches yield discontinuous records resulting from cloud obstruction or image archive gaps. We applied the Dynamic Surface Water Extent (DSWE) model to downscaled (250-m) daily Moderate Resolution Imaging Spectroradiometer (MODIS) data in Google Earth Engine to generate monthly surface water maps for the conterminous United States (US) from 2003 through 2019. The aggregation of daily observations to monthly maps of maximum water extent produced records with diminished cloud and cloud shadow effects across most of the country. We used the continuous monthly record to analyze spatiotemporal surface water trends stratified within Environmental Protection Agency Ecoregions. Although not all ecoregion trends were significant (p<0.05), results indicate that much of the western and eastern US underwent a decline in surface water over the 17-year period, while many ecoregions in the Great Plains had positive trends. Trends were also generated from monthly streamgage discharge records and compared to surface water trends from the same ecoregion. These approaches agreed on the directionality of trend detected for 54 of 85 ecoregions, particularly across the Great Plains and portions of the western US, whereas trends were not congruent in select western deserts, the Great Lakes region, and the southeastern US. By describing the geographic distribution of surface water over time and comparing these records to instrumented discharge data across the conterminous US, our findings demonstrate the efficacy of using satellite imagery to monitor surface water dynamics and supplement traditional instrumented monitoring.
... However, spatially aggregated irrigation data does exist for large areas within our study region (Torak & Painter, 2011) and efforts are being made to increase available data on irrigation water use across the U.S. (Painter et al., 2021). In addition to irrigation data, future work consisting of adding explanatory variables (Liang et al., 2019;Tulbure & Broich, 2019) to improve our understanding of the driver categories should consider including soil type, soil moisture, slope, elevation, stream gage data, and timelag factors such as precipitation or surface water area from the proceeding season (Heimhuber et al., 2017;Liang et al., 2019;Tulbure & Broich, 2019;Walker et al., 2020). Additionally, error was likely introduced with the annual application of the ∼5-year NLCD data before 2007. ...
Article
Full-text available
Surface water is the most readily accessible water resource and provides an array of ecosystem services, but its availability and access are stressed by changes in climate, land cover, and population size. Understanding drivers of surface water dynamics in space and time is key to better managing our water resources. However, few studies estimating changes in surface water account for climate and anthropogenic drivers both independently and together. We used 19 years (2000–2018) of the newly developed Dynamic Surface Water Extent Landsat Science Product in concert with time series of precipitation, temperature, land cover, and population size to statistically model maximum seasonal percent surface water area as a function of climate and anthropogenic drivers in the southeastern United States. We fitted three statistical models (linear mixed effects, random forests, and mixed effects random forests) and three groups of explanatory variables (climate, anthropogenic, and their combination) to assess the accuracy of estimating percent surface water area at the watershed scale with different drivers. We found that anthropogenic drivers accounted for approximately 37% more of the variance in the percent surface water area than the climate variables. The combination of variables in the mixed effects random forest model produced the smallest mean percent errors (mean −0.17%) and the highest explained variance (R² 0.99). Our results indicate that anthropogenic drivers have greater influence when estimating percent surface water area than climate drivers, suggesting that water management practices and land‐use policies can be highly effective tools in controlling surface water variations in the Southeast.
... For example, researchers developed maps of surface water and flooding extent dynamics (SWD) using statistically representative measures of accuracy based on the entire Landsat archive across one of the largest agricultural dryland basins in the world . They then used the SWD maps together with time series of vegetation dynamics to quantify space time patterns of vegetation response to flooding (Broich et al., 2018;Shendryk et al., 2016), to assess drivers of flooding dynamics across space and time (Heimhuber et al., 2016(Heimhuber et al., , 2017, and to identify surface water habitats that play central roles as biodiversity hubs for water dependent biota (Bishop-Taylor et al., 2015, 2017. This type of work, while technically possible at global scales, will prove less insightful due to generalizations that will need to be made even if enough budget is available for their global development. ...
Article
Full-text available
Unprecedented amounts of analysis‐ready Earth Observation (EO) data, combined with increasing computational power and new algorithms, offer novel opportunities for analysing ecosystem dynamics across large geographic extents, and to support conservation planning and action. Much research effort has gone into developing global EO‐based land‐cover and land‐use datasets, including tree cover, crop types, and surface water dynamics. Yet there are inherent trade‐offs between regional and global EO products pertaining to class legends, availability of training/validation data, and accuracy. Acknowledging and understanding these trade‐offs is paramount for both developing EO products and for answering science questions relevant for ecology or conservation studies based on these data. Here we provide context on the development of global EO‐based land‐cover and land‐use datasets, and outline advantages and disadvantages of both regional and global datasets. We argue that both types of EO‐derived land‐cover datasets can be preferable, with regional data providing the context‐specificity that is often required for policy making and implementation (e.g., land‐use and management, conservation planning, payment schemes for ecosystem services), making use of regional knowledge, particularly important when moving from land cover to actors. Ensuring that global and regional land‐cover and land‐use products derived based on EO data are compatible and nested, both in terms of class legends and accuracy assessment, should be a key consideration when developing such data. Open access high‐quality training and validation data derived as part of such efforts are of utmost importance. Likewise, global efforts to generate sets of essential variables for climate change, biodiversity, or eventually land use, which often require land‐cover maps as inputs, should consider regionalized, hierarchical approaches to not sacrifice regional context. Global change impacts manifest in regions, and so must the policy and planning responses to these challenges. EO data should embrace that regions matter, perhaps more than ever, in an age of global data availability and processing.
... Secondly, the efficacy of the ESA-CCI SM product is well established by several studies that have validated the ESA-CCI SM product against in-situ soil moisture observations by comparing it against ground-based observations from various sites around the globe (Dorigo et al., 2015;Fang et al., 2016;An et al., 2016;Shen et al., 2016;Mao et al., 2017;Dorigo et al., 2017). Moreover, the widespread applications of the ESA-CCI soil moisture product in different fields, such as climate variability and change (Shrivastava et al., 2017;Unnikrishnan et al., 2017), land atmosphere interactions Park et al., 2017), global biogeochemical cycles and ecology Tang et al., 2017), hydrological and land surface modelling (Asoka et al., 2017;Heimhuber et al., 2017), drought applications (Cammalleri et al., 2017;Das and Maity, 2015;Yan et al., 2017;McNally et al., 2017), and hydrometeorological applications (Zhan et al., 2017), to improve our earth system understanding are well recognized by the climate science community (Dorigo et al., 2017). Hence, these advantages of having the long temporal global coverage and its well alignment with the ground observations make the use of ESA-CCI SM product highly potential over different products available. ...
Article
The Climate Change Initiative (CCI) program of the European Space Agency (ESA) (ESA-CCI) provides a global surface Soil Moisture (SM) product by merging the information from various Active and Passive sensors on board for the period of 1971–2014. This study aims to assimilate this surface SM product into a recently developed model, namely Statistical Soil Moisture Profile (SSMP) model, to investigate the efficacy of this combination in developing a vertical soil moisture profile database over a climatologically vast area, such as Indian mainland. The SSMP model is a spatially varying statistical framework that couples the memory and forcing of SM from the overlying layers and utilizes the information of Hydrologic Soil Groups (HSGs) that makes it spatially transferable. The model estimates SM at four depths i.e. 10, 20, 51 and 102 cm using the surface SM information (~0–5 cm). The developed database is named as ‘ESA-CCI Driven Vertical Soil Moisture Profile (root zone) Database (EDVSMPD)’ and the simulated values are validated with observed SM data at many monitoring stations across India, maintained by India Meteorological Department (IMD). As an exemplary demonstration of the utility of the data, basin scale reconstruction of the historical agricultural droughts and its spatial distribution, using the developed EDVSMPD product, is illustrated for a medium sized drought-stricken river basin, namely Wardha, in the central part of India. The resulting SM product for a considerable length of period (more than 30 years) will be immensely useful for many hydro-climatological studies.
Article
Exploring the factors influencing ecosystem services (ESs) is crucial for protecting them. Habitat fragmentation is a critical factor that has been proved to have positive or negative effects on various ESs. However, the detailed effects of different habitat fragmentation processes have not been thoroughly explored. This study aimed to explore the complicated influences of different habitat fragmentation processes on the variations in ESs, and identify the most significant influencing processes. The Yangtze River Delta region of China was selected as the study area. Three basic habitat fragmentation processes were quantified: a decrease in habitat area, an increase in habitat isolation, and an increase in habitat edge. Generalized additive models and a geographical detector were adopted to explore the influences of different fragmentation processes on the variations in the four ESs. The results suggested that different habitat fragmentation processes had linear/nonlinear/combined effects on different ESs, and the decrease in habitat area was the most significant fragmentation process. Carbon storage and habitat quality were affected significantly by habitat fragmentation. However, habitat fragmentation contributed less to the variations in water yield and soil retention, because they were more closely associated with climate factors. The influences of these fragmentation processes exhibited a positive and nonlinear interaction. These findings suggest that integrating the nonlinear effects of fragmentation in practical ESs management and landscape planning might be essential. Moreover, our study provides a fundamental reference for related research.
Thesis
SMOS and SMAP passive microwave sensors provide soil moisture (SM) data at 40 km resolution every 2-3 days globally, with a 0-5 cm sensing depth relevant for climatic and meteorological applications. However, SM data would be required at a higher (typically 100 m or finer) spatial resolution for many other regional (hydrology) or local (agriculture) applications. Optical/thermal and radar sensors can be used for retrieving SM proxies at such high spatial resolution, but both techniques have limitations. In particular, optical/thermal data are not available under clouds and under plant canopies. Moreover, radar data are sensitive to soil roughness and vegetation structure, which are challenging to characterize from outer space, and have a repeat cycle of at least six days, limiting the observations' temporal frequency. In this context, the leading principle of the thesis is to propose a new approach that combines passive microwave, optical/thermal, and active microwave (radar) sensors for the first time to retrieve SM data at 100 m resolution on a daily temporal scale. Our assumption is first to rely on an existing disaggregation method (DISPATCH) of SMOS/SMAP SM data to meet the target resolution achieved by radars. DISPATCH is originally based on the soil evaporative efficiency (SEE) retrieved over partially vegetated pixels from 1 km resolution optical/thermal (typically MODIS) surface temperature and vegetation cover data. The disaggregated SM data is then combined with a radar-based SM retrieval method to exploit the sensing capabilities of the Sentinel-1 radars. Finally, the efficacy of the assimilation of satellite-based SM data in a soil water balance model is assessed in terms of SM predictions at the 100 m resolution and daily temporal scale. As a first step, the DISPATCH algorithm is improved from its current version by mainly 1) extending its applicability to fully vegetated optical pixels using the temperature vegetation dryness index and an enhanced vegetation cover product, and 2) increasing the targeted downscaling resolution from 1 km to 100 m using Landsat (in addition to MODIS) optical/thermal data. The 100 m resolution disaggregated SM product is validated with in situ measurements collected over irrigated areas in Morocco, showing a daily spatial correlation in the range of 0.5-0.9. As a second step, a new algorithm is built on a synergy between DISPATCH and radar 100 m resolution data. In practice, the DISPATCH SM product available on clear sky days is first used to calibrate a radar radiative transfer model in the direct mode. Then the calibrated radar radia- tive transfer model is used in the inverse mode to estimate SM at the spatio-temporal resolution of Sentinel-1. Results indicate a positive correlation between satellite and in situ measurements in the range of 0.66 to 0.81 for a vegetation index lower than 0.6. As a third and final step, an optimal assimilation method is used to interpolate 100 m resolution SM data in time. The assimilation exercise is undertaken over irrigated crop fields in Spain. The analyzed SM product derived from the assimilation of 100 m resolution DISPATCH SM is consistent with irrigation events. This approach can be readily applied over large areas, given that all the required input (remote sensing and meteorological) data are available globally.
Article
The Darling River system in Australia is under pressure from water extraction and climate change. Management interventions such as environmental flow releases require understanding of water storage dynamics and the connectivity of floodplains and wetlands. Such knowledge can be gleaned from the long observational record of the Landsat series of satellite sensors and high (< 5 m) resolution digital elevation models derived from airborne light detection and ranging (LiDAR). Here, for the first time, we develop and demonstrate an approach to reconstruct 16-day floodplain water dynamics, including extent, depth, and volume for a long Landsat time series (1987 to present). Time series mapping of surface water extent at 5-m resolution was achieved by topographic downscaling of Landsat-derived surface water data. We propose a simple and effective algorithm to restore missing data in the images caused by, e.g., cloud and shadows, swath edges and the Landsat 7 Scan Line Corrector (SLC) failure, thereby increasing the number of useable images five-fold. The 5-m surface water extent maps clearly delineate the narrow river channel and the boundary of floodplain wetlands. They can capture the development, peak and retreat of flood events. By combining Landsat and airborne LiDAR observations, we produced time series of surface water depth mapping at 5-m resolution, accounting for the degree of hydraulic surface water connectivity. Based on these maps, we derived 16-day floodplain volume dynamics for 1987 to present. The correlation coefficient between upstream river flow records and floodplain volume time series was 0.88, indicating that the estimates were robust. The algorithms developed can be used for ongoing very high-resolution mapping to assist in managing human water use and environmental health in the Murray-Darling Basin.
Article
Optical satellite imagery is commonly used for monitoring surface water dynamics, but clouds and cloud shadows present challenges in assembling complete water time series. To test whether the daily revisit rate of Moderate Resolution Imaging Spectroradiometer (MODIS) satellite imagery can reduce cloud obstruction and improve high‐frequency surface water mapping, we compared map results derived from Landsat (30‐m) and MODIS (250‐m) data across the state of California for 2003–2019. We adapted the Dynamic Surface Water Extent (DSWE) model in Google Earth Engine to generate surface water map composites from MODIS imagery every 5, 10, 15, and 30 days, and compared products to monthly Landsat‐based DSWE maps. Results for DSWEmod (DSWE MODIS) in California suggest that more than 5% data loss (cloud obstruction, etc.) was present in only 2% of the 15‐day time series, as compared to 32% of the monthly Landsat DSWE time series. The five‐day DSWEmod composites averaged 8.4% obscuration in the winter months. Area estimates derived from cloud‐filtered MODIS and Landsat monthly products have the highest linear correlations compared to streamgage discharge records, suggesting that monthly scale analyses best explain the relationship between surface water area and general streamflow dynamics. Shorter‐interval DSWEmod products have lower correlations but utility for understanding the timing of surface water peaks and past flood events.
Preprint
Full-text available
Abstract - Surface water is the most readily accessible water resource and provides an array of ecosystem services, but is stressed by changes in climate, land cover, and population size. Understanding drivers of surface water dynamics in space and time is key to better managing our water resources. However, few studies estimating changes in surface water account for climate and anthropogenic drivers both independently and together. We used 19 years (2000-2018) of the newly developed Dynamic Surface Water Extent Landsat Science Product in concert with time series of precipitation, temperature, land cover, and population size to statistically model maximum seasonal percent surface water area as a function of climate and anthropogenic drivers in the Southeastern U.S. We fitted three statistical models (linear mixed effects, random forests, and mixed effects random forests) and three groups of explanatory variables (climate, anthropogenic, and their combination) to assess the accuracy of estimating percent surface water area at the watershed scale with different drivers. We found that anthropogenic drivers accounted for approximately 37% more of the variance in the percent surface water area than the climate variables. The combination of variables in the mixed effects random forest model produced the smallest mean percent errors (mean -0.17%) and the highest explained variance (R2 0.99). Our results indicate that anthropogenic drivers have greater influence when estimating percent surface water area than climate drivers, suggesting that water management practices and land use policies can be highly effective tools in controlling surface water variations in the Southeastern U.S.
Article
Full-text available
The location and persistence of surface water (inland and coastal) is both affected by climate and human activity and affects climate, biological diversity and human wellbeing. Global data sets documenting surface water location and seasonality have been produced from inventories and national descriptions, statistical extrapolation of regional data and satellite imagery, but measuring long-term changes at high resolution remains a challenge. Here, using three million Landsat satellite images, we quantify changes in global surface water over the past 32 years at 30-metre resolution. We record the months and years when water was present, where occurrence changed and what form changes took in terms of seasonality and persistence. Between 1984 and 2015 permanent surface water has disappeared from an area of almost 90,000 square kilometres, roughly equivalent to that of Lake Superior, though new permanent bodies of surface water covering 184,000 square kilometres have formed elsewhere. All continental regions show a net increase in permanent water, except Oceania, which has a fractional (one per cent) net loss. Much of the increase is from reservoir filling, although climate change is also implicated. Loss is more geographically concentrated than gain. Over 70 per cent of global net permanent water loss occurred in the Middle East and Central Asia, linked to drought and human actions including river diversion or damming and unregulated withdrawal. Losses in Australia and the USA linked to long-term droughts are also evident. This globally consistent, validated data set shows that impacts of climate change and climate oscillations on surface water occurrence can be measured and that evidence can be gathered to show how surface water is altered by human activities. We anticipate that this freely available data will improve the modelling of surface forcing, provide evidence of state and change in wetland ecotones (the transition areas between biomes), and inform water-management decision-making.
Article
Full-text available
The usage of time series of Earth observation (EO) data for analyzing and modeling surface water extent (SWE) dynamics across broad geographic regions provides important information for sustainable management and restoration of terrestrial surface water resources, which suffered alarming declines and deterioration globally. The main objective of this research was to model SWE dynamics from a unique, statistically validated Landsat-based time series (1986–2011) continuously through cycles of flooding and drying across a large and heterogeneous river basin, the Murray–Darling Basin (MDB) in Australia. We used dynamic linear regression to model remotely sensed SWE as a function of river flow and spatially explicit time series of soil moisture (SM), evapotranspiration (ET), and rainfall (P). To enable a consistent modeling approach across space, we modeled SWE dynamics separately for hydrologically distinct floodplain, floodplain-lake, and non-floodplain areas within eco-hydrological zones and 10km × 10km grid cells. We applied this spatial modeling framework to three sub-regions of the MDB, for which we quantified independently validated lag times between river gauges and each individual grid cell and identified the local combinations of variables that drive SWE dynamics. Based on these automatically quantified flow lag times and variable combinations, SWE dynamics on 233 (64 %) out of 363 floodplain grid cells were modeled with a coefficient of determination (r2) greater than 0.6. The contribution of P, ET, and SM to the predictive performance of models differed among the three sub-regions, with the highest contributions in the least regulated and most arid sub-region. The spatial modeling framework presented here is suitable for modeling SWE dynamics on finer spatial entities compared to most existing studies and applicable to other large and heterogeneous river basins across the world.
Article
Full-text available
Monitoring the spatio-temporal development of vegetation is a challenging task in heterogeneous and cloud-prone landscapes. No single satellite sensor has thus far been able to provide consistent time series of high temporal and spatial resolution for such areas. In order to overcome this problem, data fusion algorithms such as the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) have been established and frequently used in recent years to generate high-resolution time series. In order to make it applicable to larger scales and to increase the input data availability especially in cloud-prone areas, an ESTARFM framework was developed in this study introducing several enhancements. An automatic filling of cloud gaps was included in the framework to make best use of available, even partly cloud-covered Landsat images. Furthermore, the ESTARFM algorithm was enhanced to automatically account for regional differences in the heterogeneity of the study area. The generation of time series was automated and the processing speed was accelerated significantly by parallelization. To test the performance of the developed ESTARFM framework, MODIS and Landsat-8 data were fused for generating an 8-day NDVI time series for a study area of approximately 98,000 km2 in West Africa. The results show that the ESTARFM framework can accurately produce high temporal resolution time series (average MAE (mean absolute error) of 0.02 for the dry season and 0.05 for the vegetative season) while keeping the spatial detail in such a heterogeneous, cloud-prone region. The developments introduced within the ESTARFM framework establish the basis for large-scale research on various geoscientific questions related to land degradation, changes in land surface phenology or agriculture.
Article
Full-text available
Aims: The objective of this study is twofold: i) explore a simple, empirical relationship based on freely available, remotely sensed data and the water levels recorded at Prek Kdam (one week prior) to predict total inundated area of the Tonle Sap flood plain; and ii) use the relationship to provide a preliminary demarcation of flood risk zones around the Tonle Sap Lake. 2 Methodology: Three approaches were adopted in this study: 1) Classification: to examine flooded regions during wet seasons by Landsat images. 2) Regression model: to explore the relationship between the flooded areas and water level at Prek Kdam, near the mouth of the Tonle Sap Lake. 3) Visualization and estimation: To observe dynamics of inundation and predict the potential flooded areas based on the regression. Results: The adoption of GIS and remote sensing helps the delineation of flood zones. The results of the statistical analysis demonstrated a strong linear relationship between water levels at Prek Kdam and flooded areas at the Tonle Sap Lake. Together with the adoption of GIS and remote sensing technologies, the regression model can be further used to support flood prediction, management and regional planning. Conclusion: This research develops a flood warning tool for the government and the public to intuitively evaluate the potential flooding areas in the Tonle Sap Lake during monsoon seasons. It can further help the government prepare for flood risk management and develop a sustainable environment.
Article
Flooding is governed by the amount and timing of water spilling out of channels and moving across adjacent land, often with little warning. At global scales, flood hazard is typically inferred from streamflow, precipitation or from satellite images, yielding a largely incomplete picture. Thus, at present, the floodplain inundation variables, which define hazard, cannot be accurately predicted nor can they be measured at large scales. Here we present, for the first time, a complete continuous long-term simulation of floodplain water depths at continental scale. Simulations of floodplain inundation were performed with a hydrodynamic model based on gauged streamflow for the Australian continent from 1973 to 2012. We found the magnitude and timing of floodplain storage to differ significantly from streamflow in terms of their distribution. Furthermore, floodplain volume gave a much sharper discrimination of high hazard and low hazard periods than discharge. These discrepancies have implications for characterizing flood hazard at the global scale from precipitation and streamflow records alone, suggesting that simulations and observations of inundation are also needed.
Article
Summary. Recent work by Reiss and Ogden provides a theoretical basis for sometimes preferring restricted maximum likelihood (REML) to generalized cross-validation (GCV) for smoothing parameter selection in semiparametric regression. However, existing REML or marginal likelihood (ML) based methods for semiparametric generalized linear models (GLMs) use iterative REML or ML estimation of the smoothing parameters of working linear approximations to the GLM. Such indirect schemes need not converge and fail to do so in a non-negligible proportion of practical analyses. By contrast, very reliable prediction error criteria smoothing parameter selection methods are available, based on direct optimization of GCV, or related criteria, for the GLM itself. Since such methods directly optimize properly defined functions of the smoothing parameters, they have much more reliable convergence properties. The paper develops the first such method for REML or ML estimation of smoothing parameters. A Laplace approximation is used to obtain an approximate REML or ML for any GLM, which is suitable for efficient direct optimization. This REML or ML criterion requires that Newton–Raphson iteration, rather than Fisher scoring, be used for GLM fitting, and a computationally stable approach to this is proposed. The REML or ML criterion itself is optimized by a Newton method, with the derivatives required obtained by a mixture of implicit differentiation and direct methods. The method will cope with numerical rank deficiency in the fitted model and in fact provides a slight improvement in numerical robustness on the earlier method of Wood for prediction error criteria based smoothness selection. Simulation results suggest that the new REML and ML methods offer some improvement in mean-square error performance relative to GCV or Akaike's information criterion in most cases, without the small number of severe undersmoothing failures to which Akaike's information criterion and GCV are prone. This is achieved at the same computational cost as GCV or Akaike's information criterion. The new approach also eliminates the convergence failures of previous REML- or ML-based approaches for penalized GLMs and usually has lower computational cost than these alternatives. Example applications are presented in adaptive smoothing, scalar on function regression and generalized additive model selection.
Chapter
Inland water is an important natural resource that is critical for sustaining marine and terrestrial ecosystems as well as supporting a variety of human needs. Monitoring the dynamics of inland water bodies at a global scale is important for: (a) devising effective water management strategies, (b) assessing the impact of human actions on water security, (c) understanding the interplay between the spatio-temporal dynamics of surface water and climate change, and (d) near-real time mitigation and management of disaster events such as floods. Remote sensing datasets provide opportunities for global-scale monitoring of the extent or surface area of inland water bodies over time. We present a survey of existing remote sensing based approaches for monitoring the extent of inland water bodies and discuss their strengths and limitations. We further present an outline of the major challenges that need to be addressed for monitoring the extent and dynamics of water bodies at a global scale. Potential opportunities for overcoming some of these challenges are discussed using illustrative examples, laying the foundations for promising directions of future research in global monitoring of water dynamics.