Modeling 25 years of spatio-temporal surface water and inundation dynamics on large river basin scale using time series of Earth observation data

Abstract

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.

Full text

Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016
www.hydrol-earth-syst-sci.net/20/2227/2016/
doi:10.5194/hess-20-2227-2016
© Author(s) 2016. CC Attribution 3.0 License.
Modeling 25 years of spatio-temporal surface water and inundation
dynamics on large river basin scale using time series of Earth
observation data
Valentin Heimhuber, Mirela G. Tulbure, and Mark Broich
School of Biological, Earth & Environmental Sciences, University of New South Wales, Sydney, NSW 2052, Australia
Correspondence to: Valentin Heimhuber (valentin.heimhuber@unsw.edu.au)
Received: 15 October 2015 – Published in Hydrol. Earth Syst. Sci. Discuss.: 13 November 2015
Accepted: 11 April 2016 – Published: 10 June 2016
Abstract. The usage of time series of Earth observation (EO)
data for analyzing and modeling surface water extent (SWE)
dynamics across broad geographic regions provides impor-
tant information for sustainable management and restoration
of terrestrial surface water resources, which suffered alarm-
ing 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 re-
gression 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 flood-
plain, floodplain-lake, and non-floodplain areas within eco-
hydrological zones and 10 km ×10 km grid cells. We ap-
plied this spatial modeling framework to three sub-regions
of the MDB, for which we quantified independently vali-
dated 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 quanti-
fied flow lag times and variable combinations, SWE dynam-
ics on 233 (64 %) out of 363 floodplain grid cells were mod-
eled with a coefficient of determination (r2) greater than 0.6.
The contribution of P, ET, and SM to the predictive perfor-
mance 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 enti-
ties compared to most existing studies and applicable to other
large and heterogeneous river basins across the world.
1 Introduction
Periodically inundated areas such as floodplains play a ma-
jor role in the healthy function of river systems and perform
many ecosystem services of value to people such as the re-
tention of flood water, nutrients, and sediment and the provi-
sion of food, clean water, and groundwater recharge (Hamil-
ton, 2010; Lemly et al., 2000; Maltby and Acreman, 2011;
Robertson et al., 1999; Tockner et al., 1999). Floodplains are
particularly important within water stressed areas with high
rainfall variability and semi-arid climate conditions as they
help to sustain smaller discharges during the dry season, re-
sulting in improved overall availability of water (Teferi et
al., 2010). During the last century, increasing development
of water resources, land use transformations, and agricultural
intensification have led to an alarming disappearance and de-
cline of terrestrial surface water resources (Finlayson and
Spiers, 1999; Jones et al., 2009; Lemly et al., 2000). A recent
study estimates that nearly two-thirds of all terrestrial fresh-
water wetlands were lost between 1997 and 2011, expressed
as a reduction of the global area from 165 to 60 million ha
(Costanza et al., 2014). Consequently, there is a pressing
need for improved management and restoration of terrestrial
surface water resources, which requires cost-effective meth-
ods for mapping and analyzing the distribution and dynam-
ics of surface water across large spatial and temporal scales
Published by Copernicus Publications on behalf of the European Geosciences Union.

2228 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
(Alsdorf et al., 2007; Bakker, 2012; Finlayson et al., 1999;
Vörösmarty et al., 2015).
Recent advances in the availability and spatial and tem-
poral resolution of geospatial data along with improved
processing capabilities enabled the development of various
continental- to global-scale hydrodynamic models with im-
proved representation of channel and floodplain inundation
dynamics (Cauduro et al., 2013; Getirana et al., 2012; Neal et
al., 2015, 2012; Sampson et al., 2015; Yamazaki et al., 2011).
Although these models can provide information about the
distribution of surface water across extended areas and peri-
ods of time, they require complex parameterization, are com-
putationally intensive, and depend on the accuracy of digital
elevation models (DEMs) with global coverage. As an alter-
native, Earth observation (EO) data, combined with statis-
tical modeling techniques, represents a promising and cost-
effective approach for systematic observation and quantifica-
tion of surface water (Alsdorf et al., 2007; Overton, 2005).
New satellite, airborne, and ground-based remote sensing
data with high spatial, temporal, and radiometric resolution
are quickly becoming more accessible (Nativi et al., 2015).
EO data enable analyses of changes in the availability and
distribution of surface water at continental or sub-continental
scales based on comparison of snapshots of the state of
the system at two (Baker et al., 2007; Teferi et al., 2010)
or multiple points in time (Huang et al., 2014b; Zhao et
al., 2011). The opening of archives of continuous optical
satellite data, such as Landsat and MODIS imagery, further
increased the potential of performing time series analysis of
remotely sensed surface water extent (SWE) and inundation
dynamics over large areas and long periods of time (Klein
et al., 2014; Kuenzer et al., 2015; Mccarthy et al., 2003;
Sakamoto et al., 2007; Tulbure and Broich, 2013; Tulbure et
al., 2016). Such EO-based analyses of SWE dynamics are
to be distinguished from EO-based flood mapping, which
focuses on flooding of areas that are not frequently inun-
dated and large-scale damage assessment of floods (Kuen-
zer et al., 2015). In comparison to change analysis based on
multiple observations, time series analysis refers to tempo-
rally dense monitoring of land surface dynamics over a de-
fined period of time (Broich et al., 2011; Wagner et al., 2015).
Synthetic aperture radar (SAR) has the advantage of not be-
ing affected by cloud cover for mapping surface water, but
the availability of long-term SAR time series data for large
areas is still limited (Yan et al., 2015). Accordingly, optical
satellite data currently represent the main choice for time se-
ries analysis of SWE dynamics.
Empirical models of SWE on floodplains derived from
optical satellite data as a function of discharge or water
height in the adjacent river (Table 1) have been previously
developed, with case studies including the Okavango Delta
(∼15 000 km2) (Gumbricht et al., 2004), the Waza-Logone
floodplain in Cameron (∼3000 km2) (Jung et al., 2011; Wes-
tra and De Wulf, 2009), the Tana River delta in Kenya
(∼1300 km2, Leauthaud et al., 2013), and various flood-
plains across the Murray–Darling Basin (MDB) in Australia
(Table 1, study no. 4, 5, 6, 7, 8). Table 1 gives an overview
of studies in which SWE derived from continuous optical
satellite imagery was empirically modeled as a function of
river flow and other driver variables for different floodplain
sites. For example, the RIM-FIM (River Murray Floodplain
Inundation Model) (Sims et al., 2014) used between four and
seven manually selected Landsat images during the rising
side of flood hydrographs in the period from 1984 to 2012
in combination with high-resolution DEMs to create empir-
ical models of floodplain inundation as a function of river
flow for 11 zones in the MDB. Huang et al. (2014a) used
Moderate-resolution Imaging Spectroradiometer (MODIS)
imagery during the biggest annual floods between 2001 and
2010 to develop a model that provides maximum SWEs for
river-flow levels with a range of average return periods for
90 zones covering the entire MDB (∼1 million km2). While
such EO-based inundation models can provide cost-effective
tools for sustainable management of water resources at sub-
continental scales, there is currently still a gap between mod-
els of SWE at local scale and high spatial resolution (Table 1,
study no. 2, 4, 5, 6) and sub-continental scale and coarse
resolution, which cover large areas but lack local detail (Ta-
ble 1, study no. 7). Using a unique Landsat-based time se-
ries (1986–2011) of validated surface water extent (Tulbure
et al., 2016), the overall aim of this research was to model
SWE dynamics at large river basin scale with locally relevant
detail, focusing on the MDB of Australia as a case study. The
Landsat-based SWE time series is unprecedented in its spa-
tial (30 m ×30 m) and temporal (every 16 days) resolution
and provides unique insights into 25 years of SWE dynamics
across the entire MDB, including the Millennium Drought
(1997–2009) (Leblanc et al., 2012), and major floods (e.g.,
2010–2011 La Nina floods).
Despite the great value of large-scale inundation models
for water resources management, there remains potential for
improving the usage of time series of EO data for modeling
SWE at sub-continental scale and multi-decadal time peri-
ods. One of the major limitations of existing approaches is
that most of them are based on a small number of satellite
images, which are typically acquired before, during, and af-
ter the occurrence of peak flow of manually selected floods
(see event-based models in Table 1). The resulting models are
event based, limiting them to forecasting a single maximum
flood extent for a given peak flow. Considering the impor-
tance of flood propagation and duration for biodiversity as
well as the increasing availability of time series of EO data,
an event-based approach has drawbacks and a dynamic mod-
eling approach, where each time step of the SWE time series
is accounted for, is desirable (Chen et al., 2014; Hamilton,
2010; Shaikh et al., 2001). Therefore, this study aimed to
model SWE dynamics continuously through cycles of flood-
ing and drying, using all observations of the SWE time series
along with a modeling approach suited for time series data.
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/

V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water 2229
Table 1. Overview of studies that modeled optical satellite-based observations of surface water as a function of river flow and other driver variables.
Reference Modeling Study site Modeling Dependent variable Time step Predictor variables
technique (size) unit
Inundation extent River flow/ Rainfall Soil Evapotranspiration
(spatial resolution/ height moisture
time period)
[1] Multiple linear Logone Floodplain, Three Annual maximum inundation Annual/ Cumulative – Antecedent –
(Westra and De Wulf, 2009) regression Chad and Cameroon sub-regions extent based on three event-based runoff in (MODIS)
(∼3000 km2) 16-day composite upstream
MODIS Vegetation Indices catchment
images aggregated into at different
a single flood extent time points
(250 m ×250 m/2000–2005)
[2] Second Logone Floodplain, Entire Continuous time series ≥16 days River heights at – – –
(Jung et al., 2011) polynomial Chad and Cameroon study of inundation extent five locations
regression (∼3000 km2) site derived from Landsat (two
and time (30 m ×30 m/33 images ENVISAT based
shifting between Januaru 2006 and three
and November 2008) from gauges)
[3] Multiple linear Okawango Delta, Seven Annual maximum inundation Annual/ Cumulative Cumulative – Cumulative
(Gumbricht et al., 2004) regression Botswana sub-regions extent derived from daily event-based runoff in 10 months (one gauge)
(including (∼15 000 km2) NOAA AVHRR satellite data 10 months (two gauges)
previous year (1 km ×1 km/1985 and 2000) preceding the
inundation yearly flood at
extent) an upstream gauge
[4] Flexible local Macquarie Marshes, Entire Annual maximum inundation Annual/ Cumulative Cumulative – –
(Ren et al., 2010) polynomial Australia study extent from Landsat MSS event-based annual annual
regression (∼2000 km2) site and TM imagery during river flow (two gauges)
(LOESS) times of spring flooding (upstream
(30 m ×30 m/1979–2006) gauge)
[5] Inundation extent Three floodplain sites, 11 Between four and seven Event-based Discharge – – –
(Sims et al., 2014) linked to Murray–Darling Basin, sub-regions Landsat images per zone on day of
corresponding Australia corresponding to a range Landsat image
flow level (no of river-flow values at rising (one gauge
model provided) hydrograph limbs per zone)
(30 m ×30 m/1984–2012)
[6] Inundation extent Murrumbidgee River Six zones Inundation extend Event-based Flood peak – – –
(Frazier and Page, 2009) linked to and floodplains, along for 22 selected floods discharge of
corresponding Australia the reach derived from Landsat all selected
flow level (no (∼640 km) imagery using sliding floods
model provided) maximum wetland extent (one gauge for
(technique (Frazier et al., 2003) each of six
(30 m ×30 m/1989–2008) sub-reaches)
[7] Inundation extent Murray–Darling Basin, 90 Annual maximum inundation Annual/ Annual peak flow – – –
(Huang et al., 2014a) linked to Australia sub-regions extent mapped from seven event-based (one gauge
corresponding (∼1 million km2) MODIS images during per sub-region)
flow level (no maximum annual flood
model provided) (250 m ×250 m/2001–2010)
www.hydrol-earth-syst-sci.net/20/2227/2016/ Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016

2230 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
Table 1. Continued.
Reference Modeling Study site Modeling Dependent Time step Predictor variables
technique (size) unit variable
Inundation extent River flow/ Rainfall Soil Evapotranspiration
(spatial resolution/ height moisture
time period)
[8] Inundation extent Maquire Marches, Entire Maximum flood extent Annual/ Peak flows of – – –
(Chen et al., 2014) linked to Australia study of two 1in10 year floods event-based two 1 in 10 year
corresponding flow (∼2000 km2) site derived from MODIS floods based on
level (no model 8-day composites flood frequency
provided) (250 m ×250 m/2000–2011) analysis (two gauges)
[9] Tailored flood routing Tana River delta, Entire Time series of inundation Daily Daily Daily – Modeled
(Leauthaud et al., 2013) and water balance Kenya study extents based on discharge (one gauge) (monthly)
model calibrated (∼1300 km2) site classification of (one upstream
with inundation 434 MODIS 8-day composite images gauge)
extents (250 m ×250 m/2002–2011)
[10] Gridded conceptual Diamantina River 5 km×5km Inundation extents for Daily Daily Spatial – Spatial
(Costelloe et al., 2003) water balance and and floodplains, grid a variety of flood events discharge time time
flow routing model Australia derived from Landsat and (one upstream series series
calibrated with (∼330 km) NOAA-AVHRR imagery used gauge) (daily) (monthly)
inundation extents for definition of flow
paths between grid cells
(30 m ×30 m and 1 km ×1 km)
[11] Semi-distributed water Gwydir Wetlands, Channels, 15 daily inundation extents Daily Daily discharge Daily – Daily
(Powell et al., 2008) balance and inundation Australia flow paths for one large flood used (two gauges in (one gauge) (one gauge)
model (∼2200 km2) and wetlands for model calibration the study area)
derived from classified NOAA-AVHRR
(1 m ×1 km)
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/

V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water 2231
Even though the extent of floodplain inundation largely
depends on the discharge and water level in the river, the hy-
drologic conditions of the floodplain as well as the local cli-
mate before, during, and after a flood play an important role
in the flooding and drying behavior of floodplains. For many
water bodies that are not connected to rivers, local rainfall
(P) is the main source of inundation (Kingsford et al., 2001).
Increased soil moisture (SM) prior to flooding usually leads
to reduced transmission losses and thus to a larger flood ex-
tent and longer flood duration for a given flow level com-
pared to dry antecedent conditions of the floodplain (Over-
ton, 2005). Additionally, evapotranspiration (ET) is a major
component of the water balance of surface water bodies es-
pecially in semi-arid regions such as the MDB (Lamontagne
and Herczeg, 2009; Sánchez-Carrillo et al., 2004). Besides
river flow as the key driver for floodplain inundation, five
of the studies that developed EO-based inundation models
accounted for Pand four of them also for ET during or be-
fore flooding, whereas only one study accounted for the an-
tecedent SM condition of the floodplain (Table 1). Account-
ing for the local climate before and during flooding is most
commonly done based on a modeling approach that requires
definition of conceptual water balance and flow routing mod-
els (Table 1, study no. 9, 10, 11). In order to understand the
key factors that drive the dynamics of SWE over extended
areas, we modeled SWE as a function of river flow and spa-
tially explicit time series of P, SM, and ET and quantified
the contribution of each of these variables in explaining the
variability of SWE.
Even though some of the larger study sites are divided into
smaller sub-regions for modeling (Table 1), only two studies
accounted for lag times between discharge recorded at the
gauge (Westra and De Wulf, 2009) or water surface eleva-
tion at key points (Jung et al., 2011) and the correlated SWE
in different areas of the sub-region. The overall aim of this
study was to develop a holistic and data-driven methodology
for modeling SWE and its drivers through periods of flood-
ing and drying across a large and heterogeneous river basin.
Specific key objectives of this study were to
1. develop a transferable spatial modeling framework that
allows for the application of a holistic modeling ap-
proach across the study area;
2. model lag times between remotely sensed SWE per
modeling unit and recordings of discharge at available
river gauges;
3. model SWE and quantify the role of drivers (i.e., river
flow, ET, SM, P) in explaining SWE variability across
space and time.
Figure 1. Major rivers and topography of the Murray–Darling
Basin. Source: DEM (Geoscience Australia and CSIRO, 2011).
2 Methods
2.1 Study area
The study area of this research was one of Australia’s largest
river system, the Murray–Darling Basin (more than 1 mil-
lion km2). The eastern and southern border of the MDB is
marked by the Great Dividing Range (Fig. 1), which is also
where most of the MDB’s surface water runoff is generated
(CSIRO, 2008). Outside of these partly humid highlands, the
majority of the MDB is characterized by extensive and flat
low-lying plains with arid and semi-arid climatic conditions
and slowly meandering rivers with vast floodplains (MDBA,
2010). The long-term average annual rainfall of the MDB
is 469 mm of which around 90 % evaporates or transpires
back into the atmosphere. There is also a pronounced cli-
mate gradient across the MDB with average annual rainfall
decreasing and climate variability increasing from southeast
to northwest (Leblanc et al., 2012).
The MDB has almost 30 000 wetlands (MDBA, 2010)
with 16 wetlands listed as “wetlands of international impor-
tance” (Ramsar Convention Secretariat, 2014) and around
200 specified in the Directory of Important Wetlands in Aus-
tralia (Environment Australia, 2001). The majority of wet-
lands are floodplains, which cover around 6 % of the MDB’s
total area (Kingsford et al., 2004). In the second half of the
20th century, the MDB’s water resources have been devel-
oped intensively with agriculture now taking up around 80%
of the MDB’s area (CSIRO, 2008) and accounting for more
than 80 % of the MDB’s average annual surface water use
of 11.3 km3year−1(Leblanc et al., 2012). Reduction in both
frequency and size of flooding events resulting from agricul-
tural development within the MDB has led to deterioration
www.hydrol-earth-syst-sci.net/20/2227/2016/ Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016

2232 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
in the health of many surface water ecosystems (Kingsford,
2000). On top of this human-induced stress, between 1997
and 2009 the MDB experienced the most severe drought
since the records began, during which the floodplains and
wetlands of the lower MDB received very little or no in-
flows with devastating environmental and ecological impacts
(Leblanc et al., 2012). These key features illustrate that wa-
ter requires extensive and well-planned management in the
MDB for which a basin-wide empirical model of surface wa-
ter and inundation dynamics can be a valuable tool.
2.2 Data
The dependent variable used in this study was based on a
remotely sensed time series of validated, open-surface wa-
ter and flooding extent derived from the seasonally continu-
ous archive of over 25000 Landsat TM and ETM+images
available for the entire MDB from 1986 to 2011 (Tulbure
et al., 2016). This SWE time series was developed using ran-
dom forest classification models for surface water and clouds
(Tulbure et al., 2016). The overall classification accuracy was
99 %, with 87 % (±3 % standard error) and 96 % (±2 %) pro-
ducer’s and user’s accuracy of surface water, respectively.
SWE dynamics showed high inter- and intra-annual variabil-
ity across the MDB, highlighting the vulnerability of SWE
to hydroclimatic variability (Tulbure et al., 2016). The SWE
time series used Landsat images with ≤50 % cloud cover
(Tulbure et al., 2016), resulting in times between subsequent
observations of SWE from 16 days (Landsat temporal reso-
lution) to a multiple of 16 days.
In order to model SWE continuously through periods of
flooding and drying, we used spatially explicit time series
of rainfall, evapotranspiration, and near-surface soil mois-
ture along with river flow as predictor variables (Table 2).
The selection criteria for these variables were that the tem-
poral extent of the data sets had to be spatially explicit and
long enough to cover the entire period of the SWE time se-
ries. River-flow data were acquired for gauges that had com-
plete records of daily discharge expanding over the entire
time frame of the SWE time series and downloaded from re-
spective state repositories (State Government Victoria, 2015;
Queensland Government, 2015; Government of South Aus-
tralia, 2015; New South Wales Government, 2015).
2.3 Spatial modeling framework
A schematic overview of the data processing, analysis and
spatial modeling framework used in this analysis is provided
in Fig. 2. Due to the large geographic extent of the study
area and the related heterogeneity of SWE dynamics across
the basin, the study area was split into suitable sub-units for
modeling SWE. Overton et al. (2009) developed a segmenta-
tion of the MDB into zones with uniform ecological and hy-
drological characteristics (EH-zones), which is described as a
trade-off between the finer resolution of river and floodplain
behavior and available river gauges. This zonation was used
for the development of the MDB-FIM (Chen et al., 2012)
and subsequently adapted and improved by accounting for
the hydrologic structure of the MDB (Huang et al., 2013)
(Figs. 2a, 3). The zonation was specifically developed to en-
able hydrological and hydraulic modeling on a whole-of-
basin scale while preserving key ecologic and hydrologic en-
tities, and served as the basic spatial segmentation in this
analysis. For each EH-zone, the most suitable river gauge
for modeling SWE was specified by this zonation and the
size of the resulting 89 zones ranged from a maximum of
59 991 km2to a minimum of 541 km2with an average zone
size of 11 935 km2.
The MDB contains numerous small and ephemeral rivers
and other water bodies that are not connected to major river
systems. As opposed to floodplains, we expected the SWE
of these water bodies to be mainly driven by local rainfall
and evapotranspiration. Furthermore, the SWE dynamics of
floodplain lakes differ greatly from those of shallow flood-
plains, especially with respect to the retention of flood wa-
ter after a flood has passed. Therefore, we used an existing
static wetland layer (Kingsford et al., 2004) to categorize the
entire study area into floodplain, floodplain-lake, and non-
floodplain areas, so that the heterogeneous dynamics of SWE
on these different entities could be accounted for in the mod-
eling process (Fig. 2b, c). The definition of the floodplain and
floodplain-lake category was based on hydraulic connectiv-
ity of surface water bodies to river systems with available
discharge data for modeling. We defined hydraulic connec-
tivity based on the Geofabric Surface Network (Common-
wealth of Australia (Bureau of Meteorology), 2012), a fully
connected and directed spatial river network, which allowed
for performing upstream and downstream routing operations
through the river network based on the location of available
river gauges (Fig. 2). As a result of this approach, several
water bodies that were defined as floodplains by the static
wetland layer were assigned to the non-floodplain instead of
the floodplain category. Accordingly, the floodplain-lake cat-
egory comprised all non-permanent lakes, for which the river
network indicated hydraulic connectivity to a river gauge.
Since SWE dynamics of reservoirs are mainly a function of
respective management strategies, they were masked out for
generating the surface water categories.
To enable similar modeling conditions across the study
area and to identify local spatial patterns in the role of cli-
mate drivers (P, SM, ET) of SWE dynamics, we imposed
a regular grid on top of the EH-zonation (Fig. 2c). Although
the imposition of a regular grid at times led to a less than ideal
spatial segmentation of the river and floodplain structure, it
allowed us to quantify the relationship between SWE and hy-
drologic key parameters at a much finer scale compared to
using only the regional EH-zonation. Another advantage of
using a regular grid was that it was straight forward to im-
plement at sub-continental or even global scales, and did not
require a comprehensive analysis of the river structure. In re-
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/

V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water 2233
Table 2. Overview of spatial time series used for modeling SWE dynamics.
Product Description Period Interval Resolution Sources
Surface water Time series of open- 1986–2011 ≥16 days 30 m Tulbure et al. (2016)
surface water extent
across the entire
Murray–Darling Basin
derived from Landsat imagery
Rainfall Time series of rainfall based 1980–2014 daily 5 km Commonwealth of Australia
on interpolation of rainfall (Bureau of Meteorology), 2015
gauge records throughout
Australia (Australian
Bureau of Meteorology)
Evapotranspiration Time series of actual 1980–2014 daily 5 km Viney et al. (2014),
evapotranspiration Vaze et al. (2013)
(modeled output of
the landscape component
of the Australian Water
Resource Assessment System
(AWRA-L 4.5) continental scale
water balance model)
Soil moisture Time series of near-surface 1986–2011 daily 25 km Liu et al. (2012),
soil moisture derived from Wagner et al. (2012)
active and passive satellite
microwave sensors
(European Space Agency’s CCI
(Climate Change Initiative)
soil moisture project)
gards to defining a suitable grid cell size, a finer grid leads
to a decrease in the fraction of cells that contain any flood-
plain area and consequently also to an increase in cells that
contain very small fractions of floodplain. We chose 10 km
as the most suitable cell size for modeling as a trade-off be-
tween sufficient spatial detail for capturing the variability in
SWE at a local scale and the suitability and spatial resolution
of data for modeling the driver variables.
2.4 Statistical modeling of surface water extent
dynamics
2.4.1 Data pre-processing
In order to prepare geospatial time series data sets for sta-
tistical analysis and modeling, we summarized all data sets
based on the grid cells and surface water categories. For all
spatially explicit driver variables (i.e., ET, SM, P), we de-
veloped numeric daily time series per grid cell by computing
spatial averages for each day. For the dependent variable, we
derived a time series of surface water area for each surface
water category in each grid cell from the SWE time series.
Even though a cloud cover threshold of 50 % per Landsat
scene was already applied for generating the SWE time se-
ries, there could still be excessive cloud cover of up to 100 %
in individual 10 km ×10 km grid cells. To preserve a maxi-
mum number of valid surface water observations while main-
taining acceptable levels of noise and uncertainty in the SWE
time series, we therefore applied another cloud cover thresh-
old of 40 % per individual grid cell. To account for scale ef-
fects resulting from different fractions of each surface water
category in different grid cells, we used the fraction of SWE
per cloud-free area of each category in a cell instead of ac-
tual areas (Table 3). For data management and modeling pur-
poses, all data were stored and handled in time series format
using the zoo infrastructure for regular and irregular time se-
ries (Zeileis and Grothendieck, 2005) in R (R Development
Core Team, 2008).
2.4.2 Model development and specification
To better understand SWE dynamics and its drivers across
the study area, we developed dynamic multiple linear regres-
sion models with surface water as the dependent variable and
four predictor variables (P, SM, ET, and river-flow data; Ta-
ble 3) for each surface water category per grid cell. One of the
key objectives was to take advantage of each time step of the
SWE time series and to model surface water extent contin-
uously through periods of flooding and drying. We achieved
this by including the previous SWE observation as an addi-
www.hydrol-earth-syst-sci.net/20/2227/2016/ Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016

2234 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
Figure 2. Overview of the analysis design and spatial modeling framework. Sources: river network (Commonwealth of Australia (Bureau of
Meteorology), 2012), surface water categories derived from Kingsford et al. (2004), and eco-hydrological zonation (Huang et al., 2013).
Table 3. Definition of model parameters per grid cell.
Parameter Signification Type Unit
SWE Surface water extent Fraction of SWE on cloud-free surface water category area km2km−2
QDischarge 10-day moving average m3s−1
PRainfall Previous 16-day sum mm 16 days−1
ET Evapotranspiration Previous 16-day sum (actual) mm 16 days−1
SM Soil moisture Previous 16-day average m3m−3
tional predictor variable into the model equation. Models in
which a lagged-dependent variable is used as an additional
predictor variable are referred to as dynamic linear regres-
sion models or lagged-dependent variable (LDV) models and
are commonly used for the analysis and forecasting of time
series data in economics (Keele and Kelly, 2006; Shumway
and Stoffer, 2006). The lagged-dependent variable introduces
a temporal component into the model, so that the SWE at a
given time step is also a function of the SWE of the previ-
ous time step in the time series. The equation that includes
all potential predictor variables used for modeling SWE on
floodplains and floodplain lakes is shown in Eq. (1).
SWEt=β0+β1Lag(Q) +β2SWE(t−1)+β3P+β4ET
+β5SM +e, (1)
where SWEtis the surface water extent at time t, SWE(t−1)
is the surface water extent of the previous available Land-
sat observation (t−1), 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 SWE dynamics
on non-floodplain areas is the same as Eq. (1) but without
discharge as a predictor variable, given that those areas are
not connected to a river.
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/

V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water 2235
Figure 3. (a) Eco-hydrological zonation, irrigation areas, and wetlands of the Murray–Darling Basin and surface water categories of the
three sub-regions – (b) Paroo, (c) Murray and (d) Murrumbidgee – used for illustrating the abilities of the spatial modeling framework for
modeling SWE dynamics on a local scale (per grid cell) across large river basins. Sources: surface water categories derived from Kingsford
et al. (2004), eco-hydrological zonation (Huang et al., 2013), and irrigation areas (Bureau of Rural Sciences, 2008).
We used the dynlm (Zeileis, 2014) R package for dynamic
linear modeling and time series regression for implementa-
tion and analysis of dynamic models based on the numerical
input time series. To account for the time that it takes for wa-
ter to travel from the gauge where it is recorded to a modeling
cell, discharge is incorporated into the equation by applying
a lag time. This lag time for discharge (Qlag) is the modeled
timing between daily flows measured at the gauge and the
correlated satellite-observed inundation response of a down-
stream or upstream grid cell. After quantification of Qlags,
we used a 10-day moving average of discharge instead of
daily flows in order to match the inter-daily variability and
dynamics of the discharge time series with the dynamics of
the SWE time series (16-day time step). For local rainfall and
actual ET we used the sum of 16 days before each Landsat
observation (including the day the image was taken) because
Pand ET that occurred more than 16 days before were al-
ready accounted for by the previous SWE observation. Com-
pared to ET and P, a previous 16-day moving average of SM
before each Landsat observation was used, since SM charac-
terizes the condition of the floodplain or land surface rather
than input or output of water to the system such as ET and P.
By transforming daily time series of the predictor variables
into moving averages and sums, we smoothed out the high
inter-daily variation of these variables, which is not reflected
in the SWE time series and thus not suitable for explaining
the variation of SWE in modeling. The final approach for
including the driver variables into the models was based on
our conceptual understanding of SWE dynamics, where the
SWE on a given observation is a function of river flow dur-
ing the observation, the SWE of the previous observation and
changes in P, ET, and SM between the previous and current
observation (Table 3).
2.4.3 Variable selection and model validation
We used the coefficient of determination (r2) as a measure
of how well the variability in SWE was explained by the
models. For quantifying the relative importance of the pre-
dictor variables on SWE, we tested whether accounting for
P, ET, and SM leads to an improvement in the predictive
performance of models. We used the root mean squared error
(RMSE) in 5-fold cross-validation (CV) (hereafter referred
to as CV-RMSE) to quantify predictive performance of mod-
els. The RMSE is the root of the mean of the squared resid-
uals of the prediction and is in the same unit as the depen-
dent variable, which ranges between 0 and 1 (Table 3). For
5-fold CV, we split the data into five equally sized chrono-
www.hydrol-earth-syst-sci.net/20/2227/2016/ Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016

2236 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
logical subsets. We then fitted the model to four subsets of
training data and used it to predict the remaining subset as
test data. We repeated this process for the other four constel-
lations of training and test data and the CV-RMSE was cal-
culated by averaging the RMSE of all five predictions. The
variable selection process was implemented in R by using the
cross-validation tools for regression models (cvTools) pack-
age (Alfons, 2012).
We used step-wise variable selection to find the variable
combination that leads to the best predictive performance as
measured by CV-RMSE. Since we considered Qand Pas
the key drivers of SWE dynamics in floodplains and non-
floodplains, respectively, the initial models included only Q
as a predictor for floodplain and floodplain-lake models and
only Pfor non-floodplain models. The base model is the
initial model after adding the LDV as a predictor variable.
Based on the order of predictor variables as given by Eq. (1),
we then added one variable at a time to the base model and
kept the variable in the model if it led to a reduction in CV-
RMSE. Hereby, a very small improvement in CV-RMSE was
sufficient for including a variable into the model. If adding
a certain variable did not lead to an improvement in CV-
RMSE, this variable was no longer considered in the variable
selection process. The predictor variables that were selected
based on this process are hereafter referred to as additional
predictor variables (APV). The model that is obtained after
adding the APV to the base model is hereafter referred to as
the final model.
2.4.4 Quantification of flow lag times
Preliminary analysis of available gauges and their suitability
for modeling showed that there are often multiple suitable
gauges for grid cells that contain floodplain or floodplain-
lake areas and that large EH-zones with multiple major in-
flows and outflows require more than one gauge to model
all floodplain cells. Therefore, we considered all available
gauges with suitable discharge data and used the river net-
work to select the most suitable gauges for modeling. We
then quantified the most suitable lag time for discharge by
iterating through a variety of positive and negative Qlags at
5-day intervals and selected the one that led to the highest
correlation between discharge at the gauge and surface water
extent on the respective cell. We selected 5-day intervals to
account for the fact that there is no exact lag time for each
cell because flow travel times are a function of discharge and
overbank flow, so that elevated discharge values during times
of flooding are likely to result in different flow travel times
compared to low flows (Overton et al., 2006). For each possi-
ble lag time, we developed a simple linear regression model
with surface water area as the dependent variable and lagged
discharge as the predictor variable and used the adjusted r2
as a measure for correlation. The lag that led to the highest r2
was then assigned to each cell and used for modeling SWE
on floodplains and floodplain lakes. For floodplain lakes, the
estimation of discharge lag times was more difficult, since the
SWE on these surface water bodies is only correlated to dis-
charge on the rising side of the flood hydrograph, whereas the
draining of these water bodies is primarily driven by infiltra-
tion and evapotranspiration. To account for this, we used only
increasing SWE observations (observations of surface water
that were higher than the previous observation) for quantify-
ing Qlags for floodplain lakes.
2.4.5 Case study and experiment design
One of the main objectives of this study was to develop a
spatial modeling framework that enables capturing SWE dy-
namics on a local scale, while being applicable to large (i.e.,
sub-continental) and heterogeneous areas. We selected three
sub-regions (Fig. 3) across the study area for illustrating the
modeling and analysis results. Based on the climate charac-
teristics of the MDB (see Sect. 2.1), we expected SWE dy-
namics and the role of the predictor variables to differ sub-
stantially amongst the three sub-regions.
Each of the three sub-regions contains important flood-
plain wetland systems that are listed under the Directory
of Important Wetlands in Australia (Environment Australia,
2001). The Paroo sub-region comprises large parts of the Pa-
roo river system, which together with the neighboring War-
regoo River is considered the last of 26 major rivers in the
MDB without large dams and diversions and consequently
little or no manipulation of the natural flow and inundation
regimes (Kingsford et al., 2001). These river systems expe-
rience semi-arid to arid climate conditions and have a flow
regime typical for dryland rivers, which is characterized by
extreme variability with long dry spells, punctuated by large
and unpredictable flood events and more frequent flow pulses
that lead to in-channel flows without achieving floodplain in-
undation (Bunn et al., 2006). Another important feature of
the Paroo river system is that it is predominantly a terminal
river system that has only connected with the Darling River
a few times in recorded history through the Paroo overflow
(Fig. 5).
The Murray sub-region covers the lower Murray River and
its adjacent floodplains along a ∼350 km stretch, starting
from the location where the Darling River merges into the
Murray River. This section of the Murray River is highly reg-
ulated by a sequence of locks, weirs and storage facilities.
The Murrumbidgee sub-region covers a ∼300 km stretch
of the Murrumbidgee River and its adjacent floodplains as
well as parts of the mountainous runoff-generating catch-
ment area. Both the Murray and the Murrumbidgee sub-
regions are highly regulated and contain areas of irrigated
agriculture, which is likely to have a pronounced impact on
SWE dynamics.
In the southern part of the MDB including the Murray
and the Murrumbidgee sub-regions, floods naturally occur
in winter and spring as a result of reliable rainfalls and
snowmelt and typically last for several months (Penton and
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/

V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water 2237
Overton, 2007). In the northern part of the MDB including
the Paroo, floods typically occur in the summer months as
the result of increased rainfall activity in the corresponding
catchment areas during this time of the year. Floodplain in-
undation in the Murray sub-region is largely driven by dis-
charge generated by rainfall in distant upstream catchments.
In comparison, the Paroo and the Murrumbidgee sub-regions
are located close to runoff-generating catchment area; thus,
local rainfall will likely have a more pronounced effect on
SWE dynamics as compared to the Murray. The runoff and
climate characteristics differed substantially among the three
sub-regions (Table 4).
3 Results
3.1 Flood propagation and flow lag times
There was more than one available river gauge in all three
sub-regions (Table 5), which allowed us to validate Qlag es-
timates based on a comparison of flow data from two gauges
in the same river reach. For each pair of validation gauges, we
randomly selected four floods of different magnitudes that
had a single pronounced flood peak and calculated the time
difference between the day of occurrence of the flood peak at
the upstream and downstream gauge (Table 5).
Qlags for the Murray sub-region were modeled for flow
data of two different gauges (Fig. 4a, b). We found that de-
spite the long length of the river reach, using one gauge for
all floodplain modeling cells of the sub-region led to realistic
estimates of Qlags. Both the most upstream (2-A) and most
downstream located gauge (2-B) led to a gradual increase
and decrease in Qlags, respectively, along the 350 km reach
in accordance with a validated lag time of 12 days between
gauge 2-A and 2-B (Table 5). Despite the overall realistic
flow propagation pattern for both gauges, there were a few
outlier cells, for which Qlags were different as compared to
the average Qlag of neighboring cells. For floodplain lakes,
we expected Qlags to be in the same range as the Qlags
of the surrounding floodplains, with potential minor differ-
ences resulting from the larger volume and slower filling
behavior of these water bodies compared to shallow flood-
plains. For both gauge 2-A and 2-B, Qlags were in good
accordance with the Qlags of the surrounding floodplains
for some floodplain lakes (e.g., lakes around Lake Limbra)
but deviated substantially for others (e.g., lakes around Lake
Woolpolool and Lake Wallawalla) (Fig. 4a, b).
Due to the complexity of the river and floodplain network
in the Paroo, the usage of a single gauge was found to be in-
sufficient for modeling all floodplains across this sub-region.
The Paroo River receives lateral inflows from the Warregoo
River, which first passes the large Yantabulla swamp before
merging into the Paroo further downstream (Fig. 5). The Ker-
ribree creek is a second lateral inflow into this sub-region
from the Warregoo River but is likely not connected to the
Paroo River (Timms, 2009). To account for these lateral in-
flows, Qlags of the Paroo sub-region were modeled using
two gauges. The main branch of the Paroo River (west of
line L-A in Fig. 5a) was modeled using gauge 1-C (Fig. 5a).
Similar to the Murray, Qlags in the area of the gauge (1-C)
were 0 days and showed a gradual increase and decrease in
downstream and upstream directions, respectively, with the
exception of a patch of connected floodplain area in the east
of the southernmost floodplains. The validated flow lag time
of 9 days between river gauge 1-B and 1-E (Table 5) was cap-
tured well by the model with Qlags consistently approach-
ing −5 days in floodplain areas around gauge 1-B and 5 days
around gauge 1-E. In the most downstream located flood-
plains of this sub-region, Qlags approached the upper limit
of 60 days that was defined for modeling. The lateral inflows
from the Warregoo River to the Yantabulla swamp and Ker-
ribree creek (east of line L-A in Fig. 5a) were modeled using
Gauge 1-B, for which records only date back to 1991. De-
spite the limited temporal coverage of this gauge, Qlags for
the northern inflows represent a realistic pattern, with a no-
ticeable abrupt increase of about 5–10 days along the passage
through the Yantabulla swamp. The increase in Qlags from
0 to 5 days occurs in the area of validation gauge 1-D, which
is in accordance with the validated flow lag time of 4 days
between gauge 1-A and 1-D (Table 5). For floodplains and
floodplain lakes of the Kerribree creek, automated quantifi-
cation of Qlags did not lead to realistic results as indicated
by negative Qlags.
In the Murrumbidgee sub-region, quantification of Qlags
was based on two gauges along the reach (3-A and 3-B) and
led to realistic flow propagation patterns up to a point where
the floodplains divert into two branches (point A in Fig. 6).
Before point A, Q-lags were in good accordance with the val-
idated flow lag time of 5 days between gauges 3-A and 3-B
and between 3-B and 3-C (Table 5). Downstream of point A,
Qlags abruptly approached the pre-defined upper limit of
possible lag times consistently for the remaining floodplain
cells.
3.2 Model performance
In order to quantify the relative importance of the LDV and
APV for predicting SWE, we compared the CV-RMSE of
the initial, base and final models as defined in Sect. 2.4.3
(Table 6). Since the majority of water bodies in all three
sub-regions are floodplains, this section is mainly focused
on this surface water category. For all three sub-regions,
adding the LDV to the initial model of the floodplain cate-
gory yielded large improvements in CV-RMSEs, with an av-
erage improvement of 81 % (1CV-RMSE =0.19) for the
Paroo, 81 % (1CV-RMSE =0.16) for the Murray, and 87 %
for the Murrumbidgee (1CV-RMSE =0.10). In compari-
son to that, adding the APV (i.e., P, ET, SM) to the base
model, only led to further CV-RMSE improvement of 5.2 %
(1CV-RMSE =0.0029) for Paroo, 0.3 % (1CV-RMSE =
www.hydrol-earth-syst-sci.net/20/2227/2016/ Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016

2238 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
Figure 4. Model results of the Murray sub-region: (a) Qlags for river gauge 2-A, (b) Qlags for river gauge 2-B, and (c) r2of floodplain
and floodplain lakes (final models).
Figure 5. Model results of the Paroo sub-region: (a) Qlags and (b) r2for the floodplain, and floodplain-lake surface water category (final
models) based on river gauge 1-A (all cells east of line L-A) and 1-C (all cells west of line L-A), (c) r2of non-floodplain category (final
models).
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/

V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water 2239
Table 4. Long-term average climate and runoff characteristics along with total, floodplain, and floodplain-lake area of the three sub-regions
used for illustration of the spatial modeling framework.
Averages (1986–2011) Rainfall ET Discharge Total area Floodplains Floodplain lakes
Paroo 294 290 12 50 593 7908 (15.6) 479 (0.9)
Murray 270 264 139 9132 3109 (34.1) 784 (8.6)
Murrumbidgee 453 448 90 13 032 907 (7.0) 5 (0.04)
Unit mm year−1mm year−1m3s−1km2km2(% of total area) km2(% of total area)
Table 5. Official names, ID, and start dates of river gauges and data used for this analysis along with estimated average flow travel times
between gauges used for validation of Qlags.
Sub-region EH-zone Name ID Start date Validation gauge (lag time)
Paroo (east) 27 423202C 1-A 1991
Paroo 27 A424201 1-B 1967
Paroo 27 424002 1-C 1975
Paroo (east) 27 423005 1-D 1993 (incomplete) 1-A (4 days)
Paroo 27 424001 1-E 1980 (incomplete) 1-B (9 days)
Murray 75 A4260505 2-A 1957 2-B (12 days)
Murray 79 A4260528 2-B 1985
Murrumbidgee 48 410001 3-A 1980 3-B (5 days)
Murrumbidgee 49 410005 3-B 1920 3-C (5 days)
Murrumbidgee 49 410078 3-C 1995 (incomplete)
0.0001) for the Murray and 0.8 % (1CV-RMSE =0.0003)
for the Murrumbidgee. Since the RMSE is in the same unit
as the dependent variable, it partly depends on the average
magnitude of the SWE ratio on local floodplain units and is
consequently not suitable for comparing model performance
among the three sub-regions but only within each sub-region.
In comparison to the RMSE, r2is independent of the mag-
nitude of the dependent variable and thus suitable for com-
paring model performance between the sub-regions. The r2
of initial floodplain models is a measure of how well SWE
dynamics are explained by river flow as the only predictor
variable after accounting for Qlags. The average r2of ini-
tial models was much higher in the Murray sub-region with
a value of 0.58 as compared to the Paroo (0.33) and Mur-
rumbidgee (0.25). Despite this large difference in r2of ini-
tial models, average r2of floodplain models in the Paroo
increased to the same level as the Murray (0.67) after ac-
counting for the LDV (0.62) and APV (0.67). For the Mur-
rumbidgee, the performance of final floodplain models was
much lower compared to the other two sub-regions with an
average r2of 0.41. The average r2of all 363 floodplain mod-
els across the three sub-regions was 0.63. Similar to the CV-
RMSE, accounting for the APV only led to small further im-
provements in model r2compared to the large improvements
resulting from the LDV. Analysis of the spatial distribution of
r2-based performance of final models showed that there were
no distinct patterns for floodplain and floodplain-lake models
of the Murray (Fig. 4c), Paroo (Fig. 5a), and Murrumbidgee
sub-regions (results not shown).
The average r2of final models of the floodplain-lake cat-
egory was 0.69 for the Paroo, 0.68 for the Murray, and 0.47
for the Murrumbidgee. For the non-floodplain category, av-
erage r2of final models was 0.42 for the Paroo (Fig. 5b),
0.27 for the Murray (results not shown) and 0.32 for the Mur-
rumbidgee (results not shown). In comparison to floodplain
and floodplain-lake models, r2-based model performance of
non-floodplain models showed distinct spatial patterns in
the Paroo with predominantly high r2(>0.6) for all mod-
eling cells that contained surface water bodies that are not
hydraulically connected to the modeling gauges (see non-
floodplain waterbodies in Fig. 5b).
The large differences in performance of initial and final
models of the floodplain category were mainly a result of
the distinct relationships between remotely sensed SWE on
local floodplain units and river flow in each of the three sub-
regions (Fig. 7). In the Paroo, most floodplain areas showed
SWE time series with similar temporal patterns to example
cells Ex-A and Ex-B (locations shown in Fig. 3), which were
characterized by long dry periods and short but intense flood
pulses. Both example cells are located in the area of overlap-
ping Landsat satellite paths resulting in approximately dou-
ble the number of SWE observations and a shorter modeling
time step compared to the Murray and Murrumbidgee exam-
ple cells. Most flood pulses were captured well by the SWE
time series in both cells but example cell Ex-B (Qlag =
25 days) illustrates that the river-flow data recorded at the
gauge becomes less representative for SWE dynamics with
increasing distance to the gauge. For cell Ex-B, the SWE
www.hydrol-earth-syst-sci.net/20/2227/2016/ Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016

2240 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
Figure 6. Model results of the Murrumbidgee sub-region: Qlags for river gauge 3-A (used for modeling the reach between gauge 3-A and
3-B) and 3-B (used for modeling all areas downstream of gauge 3-B).
Table 6. Average CV-RMSE and r2of initial, base, and final models of the floodplain category of the three sub-regions (definition of initial,
base, and final models is given in Sect. 2.4.3).
Sub-region Paroo Murray Murrumbidgee
Number of floodplain models 255 57 51
CV-RMSE of initial models 0.239 0.204 0.122
CV-RMSE of base models (initial model +LDV) 0.046 0.039 0.016
CV-RMSE of final models (initial model +LDV +APV) 0.043 0.039 0.015
r2of initial models 0.33 0.58 0.25
r2of base models (initial model +LDV) 0.62 0.67 0.39
r2of final models (initial model +LDV +APV) 0.67 0.67 0.41
time series showed prolonged inundation of floodplain ar-
eas for several months after river flow at the gauge has re-
turned to dry conditions resulting in a much lower r2of
the initial model of example cell Ex-B (0.45) as compared
to example cell Ex-A (0.83). After accounting for the LDV,
the r2of example cell Ex-B increased to 0.79, which illus-
trates the importance of the LDV in the final models. Due
to the large size of the Paroo sub-region, SWE dynamics
on many of the floodplain areas differed from the dynam-
ics of river flow at the modeling gauge as in example cell
Ex-B, which explains the low average r2of initial flood-
plain models in this sub-region. In the Murray, most of the
floodplain areas showed similar SWE dynamics to the ex-
ample cell for which the SWE time series closely resembles
the dynamics of river flow whenever a certain minimum flow
threshold was exceeded in the river. Inundated area reduces
quickly when river flow recedes to pre-flood levels, which ex-
plains the comparatively high initial r2of the example flood-
plain model (0.74). Out of the three sub-regions, the Mur-
rumbidgee was most affected by cloud cover during times
of flooding due to its proximity to mountainous headwater
catchments. Consequently, SWEs during times of flooding
were not represented well in the SWE time series (i.e., miss-
ing time steps) for the majority of floods resulting in a low
initial r2of 0.39. In comparison to the Paroo, accounting for
the LDV did not lead to large improvements in model perfor-
mance, with r2of the final model of 0.79 (7 %) in the Murray
and 0.43 (10 %) in the Murrumbidgee example cells.
3.3 Predictor variable combinations and spatial
patterns
Based on the results of the variable selection process, we
calculated the percentage of models in each sub-region for
which inclusion of APV led to improvement of CV-RMSE
(Table 7). On average, Pand SM were more important for
explaining SWE dynamics on floodplains in the Paroo as
compared to the Murray and Murrumbidgee sub-regions, for
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/

V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water 2241
Figure 7. Time series of surface water extent ratio on floodplains and discharge in the corresponding river gauge for four example grid cells
for the period from 1992 to 1999. (Vertical green bars indicate time steps of the SWE time series; location of example grid cells is shown in
Fig. 3.)
which Pand ET were selected for about 50 % of all flood-
plain models. In general, local rainfall was the most influ-
ential APV and helped to explain SWE dynamics in 241
(66 %) out of 363 floodplain models and 41 (89 %) out of
46 floodplain-lake models. ET was selected for approxi-
mately half of all floodplain models in the Murrumbidgee
and Murray sub-regions and 36 % of models in the Paroo. For
floodplain-lake models, local rainfall and evapotranspiration
were more important in the Paroo compared to the Murray
sub-region. There were only four modeling cells that con-
tained floodplain-lake area in the Murrumbidgee sub-region
so that a comparison with the other two zones was of lim-
ited value for this category. For non-floodplain models, we
found a similar pattern for actual ET and SM for all three
sub-regions with both variables being selected for one-third
to one-half of all models with an exception in the Murray,
where SM was only selected for 11 % of the models.
In comparison to r2-based model performance, there were
distinct spatial patterns for the contribution of the LDV and
the APV to predictive performance as measured by CV-
RMSE of floodplain and floodplain-lake models. For in-
Table 7. The role of local climate variables for modeling SWE dy-
namics on floodplains, showing the percentage of grid cells in each
sub-region where automated variable selection led to improvements
in CV-RMSE and thus inclusion of each P, ET, and SM into the
final models.
Sub-region Paroo Murray Murrumbidgee
Number of floodplain models 255 57 51
Models including P74 % 46% 51 %
Models including ET 36 % 49 % 47 %
Models including SM 47 % 23 % 35 %
Number of floodplain-lake models 28 15 3
Models including P100 % 73 % 67 %
Models including ET 43 % 33 % 33 %
Models including SM 43 % 47 % 0 %
Number of non-floodplain models 502 68 169
Models including ET 37 % 56 % 46 %
Models including SM 54 % 25 % 33 %
stance in the Paroo, the LDV led to the largest improve-
ment in CV-RMSE in areas with poor predictive performance
(high CV-RMSE) of the initial models such as the area of the
www.hydrol-earth-syst-sci.net/20/2227/2016/ Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016

2242 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
Figure 8. (a) CV-RMSE of the initial models of the floodplain and floodplain-lake surface water category in the Paroo sub-region and
improvement in CV-RMSE after accounting for the (b) LDV (base models) and (c) APV (final models).
Yantabulla swamp, the most downstream-located dead-end
floodplain area and the floodplains of the Kerribree creek
(Fig. 8a, b). The Kerribree creek along with the Yantabulla
swamp also showed the highest contribution of the APV for
explaining SWE dynamics (Fig. 8c). The floodplains of the
Kerribree creek are a good example showing that previous
floodplain conditions (i.e., LDV) and local climate drivers
(i.e., APV) explain the variability in SWE (i.e., models with
r2>0.6, Fig. 5) for floodplains on which SWE dynamics are
poorly correlated with river flow from available gauges as in-
dicated by initially high CV-RMSE (Fig. 8a). In general, the
magnitude of the improvements in CV-RMSE resulting from
the APV was small compared to the improvements resulting
from including the LDV.
The complex nature of the relationship between SWE
and its drivers over time was identified by analyzing cross-
correlations and lag effects of the time series used for mod-
eling. The correlation coefficient indicates how variation of
one variable over time is explained by a second variable. We
computed this coefficient for a range of positive and nega-
tive lag times between different pairs of input time series.
For the floodplain area in the Paroo example cell Ex-A (lo-
cation shown in Fig. 3), SM was partly driven by P(max-
imum correlation coefficient 0.43 at a lag of 3 days), indi-
cated by short-term peaks in the otherwise strongly seasonal
SM signal during larger rainfall events (Fig. 9a). We found
the highest cross-correlation (maximum correlation coeffi-
cient of 0.84 at a lag of 3 days) between river flow and SWE
(Fig. 9b).
This floodplain example illustrates the importance of the
hydrologic condition of the floodplain as well as the local cli-
mate conditions before, during and after flooding. The three
consecutive floods starting in 2000 (referred to as the “2000
floods” in Fig. 9) led to prolonged inundation of parts of the
floodplain for an entire year until another minor flooding oc-
curred shortly before the end of 2000 (referred to as the “end
of 2000 flood” in Fig. 9). In comparison to that, the flood
that occurred after 3 dry years in 2004 (referred to as the
“2004 flood” in Fig. 9) had a higher peak flow in the river
and a larger maximum surface water extent but only caused
short-term flooding of the floodplain, which rapidly returned
to dry conditions after the flood had passed. Comparison of
SM and ET and their respective trend components (Fig. 9c)
showed that both parameters were significantly higher during
the time period of the 2000 floods than in the time period of
the 2004 flood (see trend components in Fig. 9c). ET had a
distinct seasonal component with a peak during the summer
months but it is also a function of water available for evap-
otranspiration, which is mainly governed by Pand SM. The
maximum correlation coefficient between ET and SM was
0.44 at a lag of 24 days and 0.53 at a lag of 26 days between
ET and P.
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/

V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water 2243
Figure 9. Relationship of variables used for modeling the floodplain area of the Paroo example cell Ex-A (location shown in Fig. 3) for
the period from 1999 to 2006: (a) soil moisture and local rainfall, (b) discharge and surface water extent ratio, and (c) soil moisture and
evapotranspiration with respective trend components (dashed lines).
4 Discussion
Globally, terrestrial surface water resources continue to suf-
fer degradation and decline (Costanza et al., 2014; Jones et
al., 2009) and there is currently a new boom in hydro power
development, expected to reduce the number of remaining
free-flowing large rivers by 21 % (Zarfl et al., 2015). As a
result of this trend, there is a need for balancing the socio-
economic demands on water resources with the requirements
of the environment for maintaining ecological health of ter-
restrial freshwater ecosystems. In absence of dense in situ
monitoring networks, time series of Earth observation data
combined with adequate modeling techniques represent a
promising tool for quantifying water resources across large
and heterogeneous river basins, which is needed for balanc-
ing these competing demands. In this study, we used dynamic
linear regression to model SWE through cycles of flooding
and drying as a function of key hydrological drivers. We de-
veloped a spatial modeling framework that allowed us to ac-
count for the unique SWE dynamics of hydrologically dis-
tinct floodplain, floodplain-lake, and non-floodplain surface
water bodies. Our empirical inundation models are season-
ally continuous and suitable for predicting surface water ex-
tent and retention on each modeling cell based on locally
quantified combinations of predictor variables.
4.1 Model selection
The appropriate form of a model depends on its specific ob-
jectives (Bennett et al., 2013). The main objective for devel-
oping SWE models was to identify distinctive patterns in the
role of lagged surface water observations and predictor vari-
ables for SWE dynamics across space and time.
We chose a dynamic modeling approach because includ-
ing the previous SWE observation as an explanatory variable
allowed us to account for the complex and spatially varying
flooding and drying behavior of different water bodies. Slow
flooding and drying behavior of periodically inundated wa-
terbodies leads to an increase in autocorrelation in the SWE
time series, resulting in a higher probability of a large inun-
dated area at time t, if there was a large inundated area on
the previous time step (t−1). By including the LDV into
the models, we were able to overcome the limitations of ex-
isting event-based approaches for modeling SWE, in which
observations during the falling limb of floods, which likely
contain water from the previous observation, were not used
for model development (see event-based models in Table 1).
Additionally, Keele and Kelly (2006) suggested that in most
cases, LDV models are more appropriate than static models
if a process is known to be dynamic, meaning that the pro-
cess at time tis a function of history as modified by new
information, which applies to SWE dynamics. Furthermore,
www.hydrol-earth-syst-sci.net/20/2227/2016/ Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016

2244 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
our findings indicate that the degree of improvement in CV-
RMSE resulting from accounting for the LDV yields infor-
mation about the flooding and especially drying behavior of
waterbodies. Large improvements in predictive performance
as measured by CV-RMSE resulting from accounting for the
LDV were commonly linked with poor predictive perfor-
mance of initial models and surface water bodies that tend
to retain floodwater for prolonged periods of time. In both
cases, including the LDV played an essential role in achiev-
ing good model performance across each sub-region. Based
on these considerations, this study illustrates that a dynamic
modeling framework has a variety of advantages compared to
a static modeling approach for modeling SWE through flood-
ing and drying cycles and over long periods of time.
One of the drawbacks of using a LDV model is that the re-
sulting models are limited to forecasting SWE for one Land-
sat time step (commonly ≥16 days), since the previous SWE
observation needs to be specified as part of the predictive
model equation. Nevertheless, the resulting models are still
of high practical relevance, considering that a 16-day fore-
cast of SWE across the Murray–Darling Basin is valuable
for many water management applications. Additionally, the
models can be used for estimating missing time steps in the
SWE time series resulting from discarding images with ex-
cessive cloud cover.
4.2 Quantification of flow lag times
In previous studies, one gauge was commonly used for
synchronous inundation modeling over comparatively large
floodplain units without applying lag times (Table 1). Our
results illustrate that the continuous Landsat-based SWE
time series yields sufficient information for estimating Q-
lag times and consequently the spatio-temporal dynamics of
SWE on a finer than sub-basin scale. By classifying wa-
ter bodies into three categories and using a regular grid of
10 km ×10 km cells, we were able to quantify the local com-
bination of predictor variables on individual grid cell level
and to apply a tailored modeling approach to surface water
bodies with distinctive SWE dynamics.
Our results show that for floodplain and floodplain-lake
areas, accurate quantification of Qlags is an essential step
for developing SWE models with good (r2>0.6) predictive
performance. For all three sub-regions, automated Q-lag es-
timation led to realistic lag time patterns, as confirmed by
validation at different gauging stations, with gradual increase
or decrease of Qlags in downstream and upstream direc-
tion away from the gauge. There were, however, a number
of scattered outlier cells, for which Qlags deviated from
the average Qlag of neighboring cells. These outliers were
likely a result of the large differences in SWE dynamics from
cell to cell as a result of river regulation and variable flood-
plain geometries. Furthermore, the available data sets did not
allow for an exact distinction between permanent and non-
permanent floodplain lakes. Permanent and semi-permanent
lakes that are connected to a river have a less pronounced
discharge to SWE relationship and slow draining behavior
compared to other floodplain areas. The Q-lag relationship
was therefore more difficult to model as reflected in more
frequent occurrence of outliers for floodplain lake compared
to floodplain Qlags. For future applications, these outliers
could be eliminated by only allowing increasing lag times
for modeling cells along the reach (in downstream direction)
or by averaging the Qlag for each cell based on the Qlags
of the surrounding cells.
In the most southerly located floodplain areas of the Pa-
roo sub-region, Qlags increased abruptly and eventually ap-
proached the upper limit of 60 days that was defined for mod-
eling. This is likely a result of the hydrologic characteris-
tics of this terminal river system. Due to the large extent of
this sub-region and the abundance of vast and shallow flood-
plains, the flow pattern recorded in the Darling River in the
north of the sub-region changes significantly when flowing
downstream and filling up vast floodplains. As a result of this,
flow travel times from gauge 1-C (most southerly gauge with
flow data covering the analysis period) to far downstream lo-
cated floodplain cells are assumed to be strongly dependent
on the magnitude and duration of floods and thus difficult to
quantify from the data. The unrealistic Qlags of the Kerri-
bree creek floodplains (Fig. 5a) as well as a patch of con-
nected floodplain area in the east of the southernmost flood-
plains are most likely a result of the fact that due to the above-
mentioned drastic changes to the flow regime, the river-flow
data of the modeling gauge were not representative of SWE
dynamics on these areas. There were also few surface water
observations of large magnitude in the east of the southern-
most floodplains during the analysis period, which imposed
difficulties for quantifying flow travel times from the gauge
to these areas.
The abrupt increase in Q-lags downstream of point A in
the Murrumbidgee sub-region (Fig. 6) was not reflected in
the flow data (see validation in Table 6), indicating that our
approach for automated estimation of Qlags did not work in
this area. This was likely a result of the high level of river
regulation in this area as illustrated by the large areas of irri-
gated agriculture in the close proximity of this reach and the
resulting drastic changes to the flow regime in downstream
direction from the gauge.
4.3 Advances in modeling surface water extent
dynamics from space
One of the key objectives of this study was to develop a
highly automated, data-driven top-down approach for mod-
eling SWE dynamics. Most existing studies of satellite-based
empirical inundation modeling (Table 1) required extensive
site analysis, manual image and data selection and tailor-
ing of the modeling approach to local site characteristics.
While these approaches are likely to result in improved in-
undation models for the local floodplain system for which
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/

V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water 2245
they were developed, they are not applicable across large
river basins with highly variable SWE dynamics across space
and time. Consequently, they would not be suitable for mod-
eling SWE on sub-continental scales, where complex pa-
rameterization of site characteristics is not feasible. Existing
studies on large-river basin or sub-continental scale provide
empirical inundation models linked to higher levels of un-
certainty because variables are often averaged for compar-
atively large, yet potentially hydrologically complex spatial
entities. Such models may potentially neglect fine-scale vari-
ations in SWE, which could lead to high levels of uncer-
tainty in the results, poor model performance or unexpected
model behavior. Nevertheless, modeling SWE dynamics on
sub-continental scale requires a simplification of the process
to a certain degree and segmentation of the study site into
adequate spatial units. By providing a novel approach for
balancing the local complexity of SWE dynamics with the
requirements and limitations of modeling surface water pro-
cesses on large river basin scale, this study is filling the gap
between existing approaches of large-scale and high level of
uncertainty and finer-scale and complex site analysis. The
key feature of this approach was the spatial modeling frame-
work, which allowed us to model each cell based on auto-
matically quantified flow travel times and to capture SWE
dynamics on a local scale across a large and heterogeneous
river basin.
Additionally, the spatial modeling framework allowed us
to apply a tailored modeling approach to surface water bod-
ies that are not hydraulically connected to river systems. This
surface water category was modeled as a function of local
rainfall as the key driver rather than river flow. Average r2
of final models of this surface water category was much
lower for all three sub-regions compared to the other two
categories. Similar to the floodplain and floodplain-lake cat-
egories, the LDV played an important role in achieving uni-
form model performance in non-floodplain models. One of
the reasons for the low average r2of non-floodplain mod-
els was our definition of this category, which comprised all
remaining land surface areas that were not assigned to the
floodplain and floodplain-lake categories. We used this defi-
nition to capture inundation of land surface areas that are not
defined as water bodies by the static wetland layer, which
could be caused by intense local rainfall events. As a result
of this definition of the non-floodplain category, many of the
non-floodplain modeling cells did not have sufficient surface
water observations for fitting a model equation. These find-
ings indicate that a definition, in which the non-floodplain
category is limited to dynamic surface water areas rather than
the entire remaining land surface, could be more suitable for
modeling this category. The SWE time series used in this
study quantifies the spatial and temporal distribution of sur-
face water over 25 years at 30 m pixel size and can thus be
used to characterize the non-floodplain category.
The large size of the study area introduces some limi-
tations for modeling SWE with Landsat-based time series.
Overlapping satellite paths and discarding of images with ex-
cessive cloud cover led to varying sample densities across
the study area and sometimes across single EH-zones. The
MDB expands across eight Landsat paths from east to west,
which resulted in seven overlapping paths with a width of
approximately 40 km. In these areas of overlapping satellite
paths, the number of observations was approximately twice
as high as in the remaining areas resulting in variable tempo-
ral density of the SWE time series used for modeling across
the study area. Discarding images or parts of images with ex-
cessive cloud cover introduced gaps into the otherwise regu-
lar time series of SWE (every 16 days). Since these missing
time steps in the time series were not explicitly accounted for
during the fitting of the models, these gaps led to inclusion
of SWE observations into the models that were sometimes
several 16-day time steps apart from the current observation.
4.4 Role of the local climate for surface water extent
dynamics
Another key objective of this research was to analyze the role
of APV (i.e., P, ET and SM) in SWE dynamics, based on in-
tegration of spatial time series of rainfall, near-surface soil
moisture and actual evapotranspiration. We used dynamic
linear regression and 5-fold cross-validation for performing
a data-driven analysis of the role of these driver variables
across space and time.
Since we developed a highly automated and data-driven
approach for modeling SWE dynamics, we did not consider
modeling SWE processes based on simplified water balance
models as done in other studies on a smaller-scale (Table 1,
study no. 9, 10, 11). Instead, we performed a variety of ex-
ploratory analysis (e.g., Figs. 7, 9) to get a better understand-
ing of the relationships between variables and to find the
most suitable representation of each variable in the models
(Table 4). Even though analysis of cross-correlations indi-
cated that for some variables the correlation with SWE is the
highest after applying relatively large lag times, we did not
apply lag times for any of the APV for modeling and variable
selection. Instead, the final modeling approach was based on
our conceptual understanding of SWE dynamics, for which
the SWE at time t, should be a function of river flow at time
t, the SWE at time (t −1)and the changes in P, ET, and SM
between (t −1)and t. Using a 10-day moving average for
river flow helped to account for the fact that especially for
modeling cells located far away from a gauge, Qlags are an
approximation and partly depend on the magnitude of floods
so that actual daily values become less suitable for modeling
SWE dynamics.
Cross-validation is commonly used for variable selection
and the high value of predictive modeling for explanatory
purposes and capturing patterns and relationships in rich data
sets is widely recognized (Arlot and Celisse, 2010; Bennett
et al., 2013; Shmueli, 2010). We found that local rainfall was
the most important additional driver variable across all three
www.hydrol-earth-syst-sci.net/20/2227/2016/ Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016

2246 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
sub-regions and that the APV were more important for mod-
eling and predicting SWE in the Paroo sub-region compared
to the other two sub-regions. The Paroo is the largest, most
arid and least regulated sub-region, which, in combination
with its extensive shallow floodplains, might be the reason
for the increased importance of the APV in this area.
Based on our understanding of SWE dynamics, we ex-
pected actual evapotranspiration to help explain the variabil-
ity in SWE particularly on slow draining surface water areas
such as the floodplains of the Paroo sub-region since these
waterbodies mainly drain as a result of infiltration and ET.
Despite these assumptions, ET was the least influential local
climate variable in the Paroo. One of the reasons for this may
be that the ET data are modeled output of a land surface and
water balance model (Table 2). Even though this model is
calibrated against streamflow and partly accounts for surface
water body dynamics, including inflows from runoff and dis-
charge (Viney et al., 2014), model estimates of actual evapo-
transpiration might be of limited accuracy over extended dy-
namic surface water areas such as large floodplains.
In comparison to ET, the Pand SM data sets are of obser-
vational nature. The rainfall data set was generated based on
spatial interpolation of point measurements and the SM time
series was generated based on remotely sensed data sets. The
active and passive microwave data used for generating the
SM time series is expected to be largely influenced by ex-
tended open water bodies, including inundated floodplains
as well as by dense vegetation cover (Liu et al., 2012; Ye
et al., 2015). Consequently, this data set may be biased for
floodplain areas during flooding, where large-scale inunda-
tion is likely to drive soil moisture estimates towards satu-
ration so that the SM data are partly a direct observation of
SWE.
Overall, the improvements in CV-RMSE after accounting
for the APV were small compared to the improvements after
accounting for the LDV in all sub-regions, which is likely a
result of the complex relationship between SWE dynamics
and the local climate drivers. This relationship is non-trivial
to quantify due to the absence of spatially continuous infor-
mation on inundation volumes and infiltration rates (missing
variables) as well as resolution limits of the data sets used
here. A conceptual water balance model (i.e., a dynamic hy-
drologic model), as an alternative modeling framework, ca-
pable of accounting for all fluxes of water into and out of
each spatial modeling unit, would likely be limited by the
same factors. Such a modeling approach might, however, be
more suitable for capturing the variable effect of Pand ET on
SWE over time, resulting from the fact that a certain strong
rainfall event is more likely to influence the SWE during
times of inundation as compared to dry conditions. Alterna-
tively, in an event-based modeling approach, the APV could
be used to characterize the antecedent conditions of flood-
plains for each modeled flood, which are known to have a
strong influence on the magnitude and duration of inundation
on floodplains. Such an event-based modeling approach may
be useful for certain applications but is not suitable for mod-
eling SWE dynamics continuously through cycles of flood-
ing and drying as done in this study.
5 Conclusions
In this study, we statistically modeled SWE over a period
of 25 years across three large and hydrologically distinct
sub-regions of the MDB. We developed a spatial modeling
framework that allowed us to apply a tailored modeling ap-
proach to hydrologically distinct floodplain, floodplain-lake
and non-floodplain areas and to quantify local driver com-
binations on the level of 10km ×10 km grid cells. Based on
this spatial modeling framework, we modeled SWE continu-
ously through cycles of flooding and drying on 946 modeling
cells across the three sub-regions. Automated quantification
of flow lag times was a key step for modeling floodplains
and floodplain lakes on the level of individual grid cells. Ac-
counting for the LDV was important for achieving uniformly
good model performance across the three sub-regions. Freely
available spatial time series of P, ET, and SM allowed us to
analyze the role of local hydrological and climatic conditions
in explaining variability in SWE. Even though the contribu-
tion of local rainfall, evapotranspiration and soil moisture
to the predictive performance of SWE models were small
compared to the large improvements resulting from the LDV,
these variables were important for achieving good model per-
formance in a variety of hydrologically distinctive areas. Ad-
ditionally, local rainfall was the main driver for modeling
SWE on all surface water bodies that are not connected to
river systems with available flow data. The models devel-
oped in this study provide unique insights into the inunda-
tion and retention behavior of surface water bodies and local
driver combinations and can provide a valuable tool for im-
proving water resource management in the MDB. Compared
to the more complex framework of continental- or global-
scale hydrodynamic models our modeling approach provides
a data-driven and highly automated alternative for analyz-
ing SWE dynamics. Future work will focus on applying the
presented methodology to the entire MDB, providing a first
basin wide and seasonally continuous inundation model. The
spatial modeling framework is transferable to other large and
heterogeneous river basins across the world provided that hy-
draulic connectivity of surface water bodies and river sys-
tems can be determined.
Data availability
The Landsat-based time series of surface water extent 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 rainfall data product can be purchased from the Aus-
tralian Bureau of Meteorology. The soil moisture time series
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/

V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water 2247
is provided and distributed free of charge by the European
Space Agency’s CCI (Climate Change Initiative) soil mois-
ture project. The Geofabric Surface Network and the AWRA-
L evapotranspiration data product are available free of charge
through the Australian Bureau of Meteorology. A data portal
for public distribution of the AWRA-L data is currently in de-
velopment. The static wetland layer for the Murray–Darling
Basin can be requested from the Murray–Darling Basin Au-
thority. All river-flow data used in this analysis can be ob-
tained partly free of charge from respective state government
repositories as specified in Sect. 2.2 of this paper.
Author contributions. Valentin Heimhuber, Mirela G. Tulbure, and
Mark Broich designed this research and interpreted the results.
Valentin Heimhuber developed the model code and performed all
data analysis. Valentin Heimhuber prepared the manuscript with
contributions from all co-authors.
Acknowledgements. 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 eco-hydrological 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. The evapotranspiration data
were provided by the Bureau of Meteorology with support from
Albert Van Dijk and Andrew Frost. We would also like to thank
Iurii Shendryk, Robbi Bishop-Taylor, and Miles Davitt for their
valuable feedback on this manuscript.
Edited by: J. Freer
References
Aksoy, H., Unal, N. E., Eris, E., and Yuce, M. I.: Stochas-
tic modeling of Lake Van water level time series with jumps
and multiple trends, Hydrol. Earth Syst. Sci., 17, 2297–2303,
doi:10.5194/hess-17-2297-2013, 2013.
Alfons, A.: cvTools: Cross-validation tools for regression models, R
package version 0.3.2., available at: http://CRAN.R-project.org/
package=cvTools, last access: 12 May 2016, 2012.
Alsdorf, D. E., Rodríguez, E., and Lettenmaier, D. P.: Mea-
suring Surface Water from Space, Rev. Geophys, 45, 1–24,
doi:10.1029/2006RG000197, 2007.
Arlot, S. and Celisse, A.: A survey of cross-validation procedures
for model selection, Stat. Surv., 4, 40–79, doi:10.1214/09-SS054,
2010.
Baker, C., Lawrence, R., Montagne, C., and Patten, D.:
Change detection of Wetland ecosystems using Landsat
Imagery and change vector analysis, Wetlands, 27, 610–
619, doi:10.1672/0277-5212(2007)27[610:CDOWEU]2.0.CO;2,
2007.
Bakker, K.: Water Security: Research Challenges and Opportuni-
ties, Science, 337, 914–915, doi:10.1126/science.1226337, 2012.
Bennett, N. D., Croke, B. F. W., Guariso, G., Guillaume, J. H. A.,
Hamilton, S. H., Jakeman, A. J., Marsili-Libelli, S., Newham,
L. T. H., Norton, J. P., Perrin, C., Pierce, S. A., Robson, B., Sep-
pelt, R., Voinov, A. A., Fath, B. D., and Andreassian, V.: Charac-
terising performance of environmental models, Environ. Modell.
Softw., 40, 1–20, doi:10.1016/j.envsoft.2012.09.011, 2013.
Box, G. E. P. and Jenkins, G. M.: Time series analysis, forecasting
and control, Holden-day series in time series analysis, Holden-
Day, Oakland, CA, 1970.
Broich, M., Hansen, M. C., Potapov, P., Adusei, B., Lindquist, E.,
and Stehman, S. V: Time-series analysis of multi-resolution op-
tical imagery for quantifying forest cover loss in Sumatra and
Kalimantan, Indonesia, Int. J. Appl. Earth Obs., 13, 277–291,
doi:10.1016/j.jag.2010.11.004, 2011.
BRS (Bureau of Rural Sciences): Combined irrigation areas of Aus-
tralia dataset, BRS, Canberra, 2008.
Bunn, S. E., Thoms, M. C., Hamilton, S. K., and Capon, S. J.: Flow
variability in dryland rivers: Boom, bust and the bits in between,
River Res. Appl., 22, 179–186, doi:10.1002/rra.904, 2006.
Cauduro, R., De Paiva, D., Buarque, D. C., Collischonn, W., and
Bonnet, M.: Large-scale hydrologic and hydrodynamic modeling
of the Amazon River basin, Water Resour. Res., 49, 1226–1243,
doi:10.1002/wrcr.20067, 2013.
Chen, Y., Cuddy, S. M., Merrin, L. E., Huang, C., Pollock, D., Sims,
N., Wang, B., and Bai, Q.: Murray–Darling Basin Floodplain
Inundation Model Version 2.0 (MDB-FIM2), Technical report,
CSIRO Water for a Healthy Country Flagship, 2012.
Chen, Y., Wang, B., Pollino, C. A., Cuddy, S. M., Merrin, L. E.,
and Huang, C.: Estimate of flood inundation and retention on
wetlands using remote sensing and GIS, Ecohydrology, 7, 1412–
1420, doi:10.1002/eco.1467, 2014.
Commonwealth of Australia (Bureau of Meteorology): Aus-
tralian Geofabric, available at: http://www.bom.gov.au/water/
geofabric/, last access: 12 May 2016, 2012.
Commonwealth of Australia (Bureau of Meteorology): Australian
Daily Rainfall Gridded Data, available at: http://www.bom.gov.
au/climate/how/newproducts/IDCdrgrids.shtml, last access: 12
April 2016, 2015.
Costanza, R., de Groot, R., Sutton, P., van der Ploeg, S., Anderson,
S. J., Kubiszewski, I., Farber, S., and Turner, R. K.: Changes in
the global value of ecosystem services, Global Environ. Chang.,
26, 152–158, doi:10.1016/j.gloenvcha.2014.04.002, 2014.
Costelloe, J. F., Grayson, R. B., Argent, R. M., and McMahon, T. A.:
Modelling the flow regime of an arid zone floodplain river, Dia-
mantina River, Australia, Environ. Modell. Softw., 18, 693–703,
doi:10.1016/S1364-8152(03)00071-9, 2003.
CSIRO: 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, Aus-
tralia, 2008.
Environment Australia: A directory of important wetlands in Aus-
tralia, 3rd Edn., Environment Australia, Canberra, 2001.
Finlayson, C. M. and Spiers, A. G.: Global Review of Wetland Re-
sources and Priorities for Inventory, Supervising Scientist Report
No. 144, Canberra, 1999.
Finlayson, C. M., Davidson, N. C., Spiers, A. G., and Stevenson,
N. J.: Global wetland inventory – current status and future priori-
ties, Mar. Freshwater Res., 50, 717–727, doi:10.1071/MF99098,
1999.
www.hydrol-earth-syst-sci.net/20/2227/2016/ Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016

2248 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
Frazier, P. and Page, K.: A Reach-Scale Remote Sensing Technique
to Relate Wetland Inundation to River Flow, River Res. Appl.,
25, 836–849, doi:10.1002/rra.1183, 2009.
Geoscience Australia and CSIRO: 1 Second SRTM Derived Digi-
tal Elevation Models User Guide, Version 1.0.4. Canberra, ACT,
Geoscience Australia, 2011.
Getirana, A. C. V., Boone, A., Yamazaki, D., Decharme, B., Mog-
nard, N., and Papa, F.: The Hydrological Modeling and Analy-
sis Platform (HyMAP): Evaluation in the Amazon Basin, J. Hy-
drometeorol., 13, 1641–1665, doi:10.1175/JHM-D-12-021.1,
2012.
Government of South Australia: WaterConnect, available at: http:
//www.waterconnect.sa.gov.au, last access: 12 April 2016, 2015.
Gumbricht, T., Wolski, P., Frost, P., and McCarthy, T.: Fore-
casting 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, 2004.
Hamilton, S. K.: Wetlands of Large Rivers: Flood plains, in: River
ecosystem ecology: a global perspective: a derivative of Ency-
clopedia of inland waters, Elsevier Inc., 607–610, 2010.
Huang, C., Chen, Y., and Wu, J.: GIS-based spatial zoning for flood
inundation modelling in the Murray–Darling Basin, in 20th In-
ternational Congress on Modelling and Simulation, Adelaide,
Australia, 1–6 December 2013, available at: www.mssanz.org.
au/modsim2013, last access: 12 May 2016, 2013.
Huang, C., Chen, Y., and Wu, J.: Mapping spatio-temporal flood
inundation dynamics at large river basin scale using time-series
flow data and MODIS imagery, Int. J. Appl. Earth Obs., 26, 350–
362, doi:10.1016/j.jag.2013.09.002, 2014a.
Huang, C., Peng, Y., Lang, M., Yeo, I.-Y., and McCarty, G.: Wet-
land inundation mapping and change monitoring using Landsat
and airborne LiDAR data, Remote Sens. Environ., 141, 231–242,
doi:10.1016/j.rse.2013.10.020, 2014b.
Jones, K., Lanthier, Y., van der Voet, P., van Valkengoed, E., Taylor,
D., and Fernández-Prieto, D.: Monitoring and assessment of wet-
lands using Earth Observation: the GlobWetland project, J. Envi-
ron. Manage., 90, 2154–69, doi:10.1016/j.jenvman.2007.07.037,
2009.
Jung, H. C., Alsdorf, D., Moritz, M., Lee, H., and Vassolo, S.:
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, 2011.
Keele, L. and Kelly, N. J.: Dynamic models for dynamic theories:
The ins and outs of lagged dependent variables, Polit. Anal., 14,
186–205, doi:10.1093/pan/mpj006, 2006.
Kingsford, R.: Ecological impacts of dams, water diversions and
river management on floodplain wetlands in Australia, Austral.
Ecol., 25, 109–127, 2000.
Kingsford, R. T., Thomas, R. F., and Curtin, A. L.: Conservation
of wetlands in the Paroo and Warrego River catchments in arid
Australia, Pacific Conservation Biology, 7, 21–33, 2001.
Kingsford, R. T., Brandis, K., Thomas, R. F., Crighton, P.,
Knowles, E., and Gale, E.: Classifying landform at broad spa-
tial scales: the distribution and conservation of wetlands in
New South Wales, Australia, Mar. Freshwater Res., 55, 17–31,
doi:10.1071/MF03075, 2004.
Klein, I., Dietz, A. J., Gessner, U., Galayeva, A., Myrzakhmetov,
A., and Kuenzer, C.: 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 Obs., 26, 335–
349, doi:10.1016/j.jag.2013.08.004, 2014.
Kuenzer, C., Klein, I., Ullmann, T., Georgiou, E., Baumhauer, R.,
and Dech, S.: Remote Sensing of River Delta Inundation: Ex-
ploiting the Potential of Coarse Spatial Resolution, Temporally-
Dense MODIS Time Series, Remote Sens., 7, 8516–8542,
doi:10.3390/rs70708516, 2015.
Lamontagne, S. and Herczeg, A. L.: Hydrology and ecohydrology
of Australian semi-arid wetlands, Hydrol. Process., 23, 3413–
3414, doi:10.1002/hyp.7462, 2009.
Leauthaud, C., Belaud, G., Duvail, S., Moussa, R., Grünberger, O.,
and Albergel, J.: Characterizing floods in the poorly gauged wet-
lands 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, 2013.
Leblanc, M., Tweed, S., Van Dijk, A., and Timbal, B.: A
review of historic and future hydrological changes in the
Murray–Darling Basin, Global Planet. Change, 80–81, 226–246,
doi:10.1016/j.gloplacha.2011.10.012, 2012.
Lemly, A., Kingsford, R., and Thompson, J.: Irrigated Agriculture
and Wildlife Conservation: Conflict on a Global Scale, Environ.
Manage., 25, 485–512, doi:10.1007/s002679910039, 2000.
Liu, Y. Y., Dorigo, W. A., Parinussa, R. M., de Jeu, R. A. M., Wag-
ner, W., McCabe, M. F., Evans, J. P., and van Dijk, A. I. J. M.:
Trend-preserving blending of passive and active microwave
soil moisture retrievals, Remote Sens. Environ., 123, 280–297,
doi:10.1016/j.rse.2012.03.014, 2012.
Maltby, E. and Acreman, M. C.: Ecosystem services of wetlands:
pathfinder for a new paradigm, Hydrol. Sci. J., 56, 1341–1359,
doi:10.1080/02626667.2011.631014, 2011.
Mccarthy, J. M., Gumbricht, T., Mccarthy, T., Frost, P., Wessels,
K., and Seidel, F.: Flooding Patterns of the Okavango Wetland in
Botswana between 1972 and 2000, Ambio, 32, 453–457, 2003.
MDBA: Guide to the proposed Basin Plan Technical background,
Murray–Darling Basin Authority, Canberra, 2010.
Nativi, S., Mazzetti, P., Santoro, M., Papeschi, F., Craglia, M., and
Ochiai, O.: Big Data challenges and solutions in building the
Global Earth Observation System of Systems, Environ. Modell.
Softw., 68, 1–26, doi:10.1016/j.envsoft.2015.01.017, 2015.
Neal, J., Schumann, G., and Bates, P.: 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, 2012.
Neal, J. C., Odoni, N. A., Trigg, M. A., Freer, J. E., Garcia-Pintado,
J., Mason, D. C., Wood, M., and Bates, P. D.: Efficient incorpora-
tion 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, 2015.
New South Wales Government: Waterinfo, available at: http://
waterinfo.nsw.gov.au/pinneena/cm.shtml, last access: 12 April
2016, 2015.
Overton, I. C.: 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, 2005.
Overton, I. C., McEwan, K., and Sherrah, J. R.: The River Murray
Floodplain Inundation Model (RiM-FIM) Hume Dam to Lower
Lakes. CSIRO Water for a Healthy Country Technical Report
2006, CSIRO, Canberra, 2006.
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/

V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water 2249
Overton, I. C., Colloff, M. J., Doody, T. M., Henderson, B., and
Cuddy, S. M.: “Ecological Outcomes of Flow Regimes in the
Murray–Darling Basin”, Report prepared for the National Water
Commission by CSIRO Water for a Healthy Country Flagship,
CSIRO, Canberra, 2009.
Penton, D. J. and Overton, I. C.: Spatial Modelling of Floodplain In-
undation Combining Satellite Imagery and Elevation Models, in:
MODSIM 2007 International Congress on Modelling and Simu-
lation, edited by: Oxley, L. and Kulasiri, D., Modelling and Sim-
ulation Society of Australia and New Zealand, 1464–1470, 2007.
Powell, S. J., Letcher, R. A., and Croke, B. F. W.: Mod-
elling floodplain inundation for environmental flows:
Gwydir wetlands, Australia, Ecol. Model., 211, 350–362,
doi:10.1016/j.ecolmodel.2007.09.013, 2008.
Queensland Government: Water Monitoring Portal, available at:
https://www.dnrm.qld.gov.au/water/water-monitoring- and-data/
portal, last access: 12 April 2016, 2015.
R Development Core Team: R: A language and environment for sta-
tistical computing. R Foundation for Statistical Computing, Vi-
enna, Austria, available at: http://www.R- project.org, last access:
12 May 2016, 2008.
Ramsar Convention Secretariat: The List of Wetlands of Interna-
tional Importance, available at: http://www.ramsar.org/library,
last access: 12 May 2016, 2014.
Ren, S., Kingsford, R. T., and Thomas, R. F.: Modelling flow to and
inundation of the Macquarie Marshes in arid Australia, Environ-
metrics, 21, 549–561, doi:10.1002/env.1002, 2010.
Robertson, A. I., Bunn, S. E., Boon, P. I., and Walker, K. F.: Sources,
sinks and transformations of organic carbon in Australian flood-
plain rivers, Mar. Freshwater Res., 50, 813–829, 1999.
Sakamoto, T., Van Nguyen, N., Kotera, A., Ohno, H., Ishitsuka,
N., and Yokozawa, M.: Detecting temporal changes in the ex-
tent 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, 2007.
Sampson, C. C., Smith, A. M., Bates, P. D., Neal, J. C.,
Alfieri, L., and Freer, J. E.: A high-resolution global
flood hazard model, Water Resour. Res., 51, 7358–7381,
doi:10.1002/2015WR016954, 2015.
Sánchez-Carrillo, S., Angeler, D. G., Sánchez-Andrés, R., Alvarez-
Cobelas, M., and Garatuza-Payán, J.: 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, 2004.
Shaikh, M., Green, D., and Cross, H.: A remote sensing approach
to determine environmental flows for wetlands of the Lower Dar-
ling River, New South Wales, Australia, Int. J. Remote Sens., 22,
1737–1751, 2001.
Shmueli, G.: To Explain or to Predict?, Stat. Sci., 25, 289–310,
doi:10.1214/10-STS330, 2010.
Shumway, R. H. and Stoffer, D. S.: Time Series Analysis and Its
Applications: With R Examples, Springer, New York, 2006.
Sims, N. C., Warren, G., Overton, I. C., Austin, J., Gallant, J.,
King, D. J., Merrin, L. E., Donohue, R., McVicar, T. R., Hod-
gen, M. J., Penton, D., Chen, Y., Huang, C., and Cuddy, S.: 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, 2014.
State Government Victoria: Water Monitoring, available at: http://
data.water.vic.gov.au/monitoring.htm, last access: 12 April 2016,
2015.
Teferi, E., Uhlenbrook, S., Bewket, W., Wenninger, J., and
Simane, B.: The use of remote sensing to quantify wetland
loss in the Choke Mountain range, Upper Blue Nile basin,
Ethiopia, Hydrol. Earth Syst. Sci., 14, 2415–2428, doi:10.5194/
hess-14-2415-2010, 2010.
Timms, B. V: Waterbirds of the Saline Lakes of the Paroo, arid-
zone Australia?: A review with special reference to diversity and
conservation, Nat. Resour. Env. Iss., 15, 46, 8 pp., 2009.
Tockner, K., Pennetzdorfer, D., Reiner, N., Schiemer, F., and
Ward, J. V.: Hydrological connectivity, and the exchange of
organic matter and nutrients in a dynamic river-floodplain
system (Danube, Austria), Freshwater Biol., 41, 521–535,
doi:10.1046/j.1365-2427.1999.00399.x, 1999.
Tulbure, M. G. and Broich, M.: Spatiotemporal dynamic
of surface water bodies using Landsat time-series data
from 1999 to 2011, ISPRS J. Photogramm., 79, 44–52,
doi:10.1016/j.isprsjprs.2013.01.010, 2013.
Tulbure, M. G., Broich, M., Stehman, S. V., and Kommareddy,
A.: Surface water extent dynamics from three decades of sea-
sonally continuous 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, 2016.
Vaze, J., Viney, N., Stenson, M., Renzullo, L., Van Dijk, A., Dutta,
D., Crosbie, R., Lerat, J., Penton, D., Vleeshouwer, J., Peeters,
L., Teng, J., Kim, S., Hughes, J., Dawes, W., Zhang, Y., Leighton,
B., Joehnk, K., Yang, A., Wang, B., Frost, A., Elmahdi, A.,
Smith, A., and Daamen, C.: The Australian Water Resource
Assessment Modelling System (AWRA), in: 20th International
Congress on Modelling and Simulation, Adelaide, Australia, 1–6
December 2013, available at: www.mssanz.org.au/modsim2013,
last access: 12 May 2016, 2013.
Viney, N., Vaze, J., Crosbie, R., Wang, B., Dawes, W., and Frost, A.:
AWRA-L v4.5: technical description of model algorithms and
inputs, CSIRO, Australia, 2014.
Vörösmarty, B. C. J., Hoekstra, A. Y., Bunn, S. E., Conway, D.,
and Gupta, J.: What scale for water governance?, Science, 349,
478–479, doi:10.1126/science.349.6247.478-a, 2015.
Wagner, W., Dorigo, W., De Jeu, R., Fernandez, D., Benveniste,
J., Haas, E., and Ertl, M.: Fusion of active and passive mi-
crowave observations to create an essential climate variable data
record on soil moisture, in: ISPRS Annals of the Photogramme-
try, Remote Sensing and Spatial Information Sciences, Volume
I-7, 2012 XXII ISPRS Congress, 25 August–1 September 2012,
Melbourne, Australia, 2012.
Wagner, W., Dech, S. W., and Kuenzer, C.: Remote Sensing Time
Series: Revealing Land Surface Dynamics, Springer, Cham,
Switzerland, 2015.
Westra, T. and De Wulf, R. R.: Modelling yearly flooding extent
of the Waza-Logone floodplain in northern Cameroon based on
MODIS and rainfall data, Int. J. Remote Sens., 30, 5527–5548,
doi:10.1080/01431160802672872, 2009.
Yamazaki, D., Kanae, S., Kim, H., and Oki, T.: A physi-
cally based description of floodplain inundation dynamics in
a global river routing model, Water Resour. Res., 47, 1–21,
doi:10.1029/2010WR009726, 2011.
www.hydrol-earth-syst-sci.net/20/2227/2016/ Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016

2250 V. Heimhuber et al.: Modeling 25 years of spatio-temporal surface water
Yan, K., Di Baldassarre, G., Solomatine, D. P., and Schumann,
G. J. P.: A review of low-cost space-borne data for flood mod-
elling: topography, flood extent and water level, Hydrol. Process.,
29, 3368–3387, doi:10.1002/hyp.10449, 2015.
Ye, N., Walker, J. P., Guerschman, J., Ryu, D., and Gurney, R. J.:
Standing water effect on soil moisture retrieval from L-band pas-
sive microwave observations, Remote Sens. Environ., 169, 232–
242, doi:10.1016/j.rse.2015.08.013, 2015.
Zarfl, C., Lumsdon, A. E., Berlekamp, J., Tydecks, L., and Tockner,
K.: A global boom in hydropower dam construction, Aquat. Sci.,
77, 161–170, doi:10.1007/s00027-014-0377-0, 2015.
Zeileis A.: dynlm: Dynamic Linear Regression, R package version
0.3-3., available at: http://CRAN.R-project.org/package=dynlm,
last access: 12 May 2016, 2014.
Zeileis, A. and Grothendieck, G.: zoo: S3 Infrastructure for Regular
and Irregular Time, J. Stat. Softw., 14, 27 pp., 2005.
Zhao, X., Stein, A., and Chen, X.-L.: Monitoring the dy-
namics of wetland inundation by random sets on multi-
temporal images, Remote Sens. Environ., 115, 2390–2401,
doi:10.1016/j.rse.2011.05.002, 2011.
Hydrol. Earth Syst. Sci., 20, 2227–2250, 2016 www.hydrol-earth-syst-sci.net/20/2227/2016/