remote sensing

Article
Mapping and Monitoring the Multi-Decadal Dynamics of
Australia’s Open Waterbodies Using Landsat
Claire E. Krause * , Vanessa Newey , Matthew J. Alger and Leo Lymburner

Geoscience Australia, Canberra, ACT 2609, Australia; Vanessa.Newey@ga.gov.au (V.N.);

matthew.alger@ga.gov.au (M.J.A.); Leo.Lymburner@ga.gov.au (L.L.)
* Correspondence: claire.krause@ga.gov.au

Abstract: Water detection algorithms are now being routinely applied to continental and global
archives of satellite imagery. However, water resource management decisions typically take place
at the waterbody rather than pixel scale. Here, we present a workflow for generating polygons of
persistent waterbodies from Landsat observations, enabling improved monitoring and management
of water assets across Australia. We use Digital Earth Australia’s (DEA) Water Observations from
Space (WOfS) product, which provides a water classified output for every available Landsat scene,
to determine the spatial locations and extents of waterbodies across Australia. We generated a
polygon set of waterbodies that identified 295,906 waterbodies ranging in size from 3125 m2 to
4820 km2 . Each polygon was used to generate a time series of WOfS, providing a history of the
change in surface area of each waterbody every ~16 days since 1987. We demonstrate the applications
of this new dataset, DEA Waterbodies, to understanding local through to national-scale surface
water spatio-temporal dynamics. DEA Waterbodies provides new insights into Australia’s water



availability and enables the monitoring of important landscape features such as lakes and dams,


improving our ability to use earth observation data to make meaningful decisions.
Citation: Krause, C.E.; Newey, V.;
Alger, M.J.; Lymburner, L. Mapping
Keywords: water; Landsat; WOfS; dams; waterbodies; data cube
and Monitoring the Multi-Decadal
Dynamics of Australia’s Open
Waterbodies Using Landsat. Remote
Sens. 2021, 13, 1437.
1. Introduction
https://doi.org/10.3390/rs13081437
Water availability and security is a key consideration for humans and natural ecosys-
Academic Editor: Yang Hong tems world-wide. Competing demands for limited water resources between human con-
sumptive use and ecosystems is already placing many natural systems under stress [1–5].
Received: 22 February 2021 Water resource conflicts between neighbouring nations are already occurring, and these
Accepted: 3 April 2021 conflicts are likely to worsen as climate change increases the uncertainty in rainfall and
Published: 8 April 2021 evapotranspiration [6–8].
In Australia, the world’s driest inhabited continent, water availability is a key social,
Publisher’s Note: MDPI stays neutral environmental and economic issue affecting the lives of millions of Australians. Australia
with regard to jurisdictional claims in has an extensive surface water monitoring network in place, consisting of stream gauges
published maps and institutional affil- and publicly owned water storage monitoring. While these surface monitoring networks
iations. consist of tens of thousands of instruments, Australia’s large land area and ephemeral
surface water mean that large parts of its surface water network remain unmetered.
Multi-decadal archives of satellite imagery provide insight into water resource avail-
ability in two key ways. Firstly, they have the ability to map and monitor catchment-wide
Copyright: © 2021 by the authors. and continent-wide surface water dynamics, supplementing or replacing surface networks
Licensee MDPI, Basel, Switzerland. in data-sparse regions [9–11]. Secondly, the multi-decadal perspective of these archives
This article is an open access article means that the changes in surface water that are currently taking place can be placed into
distributed under the terms and the context of historical surface water dynamics [12,13]. Satellite data from public-good
conditions of the Creative Commons satellites such as the Landsat series are freely available and global in coverage [14–16],
Attribution (CC BY) license (https:// facilitating global-scale analysis of inland surface water [12,13,17,18]. Australia-specific
creativecommons.org/licenses/by/
surface water studies have also been carried out, which detect the frequency of water
4.0/).

Remote Sens. 2021, 13, 1437. https://doi.org/10.3390/rs13081437 https://www.mdpi.com/journal/remotesensing

Remote Sens. 2021, 13, 1437 2 of 28

across the landscape [19,20], as well as map the locations of seasonal and permanent
waterbodies [21,22].
Whilst information about the presence/absence of water provides valuable insight into
multi-decadal surface water dynamics, further analysis is required to make the information
content more accessible to water managers. This is because water management decisions
are typically based on waterbodies (dams, lakes, river reaches, refugial pools) rather than
pixels. Waterbody mapping with satellite imagery has almost exclusively been done on
a pixel-by-pixel [12,13,18,20,21,23,24] (or sub-pixel [25–28] basis), with only a few studies
delineating waterbodies as vector objects [29–32]. The delineation of waterbodies provides
an object-based analysis, which characterises the dynamics of the whole waterbody, not
its individual pixels. This allows analysis of waterbody characteristics such as the change
in surface area over time [24,33], facilitating study of the events observed to be impacting
individual waterbodies [33]. This in turn allows environmental water managers to eval-
uate the efficacy of environmental flow events in providing aquatic ecosystem provision
objectives [34].
Analysing surface water dynamics with respect to individual waterbodies depends on
the availability of a waterbody polygon set. In previous studies waterbody polygons have
been generated from pre-existing cartographic coverages [35], or from a select number of
cloud-free images [31,36]. The key limitation of these approaches is that the polygon sets
are dependent on line work derived from either a single point in time [35] or reference
imagery from a narrow range of dates [31,36]. This limits the ability of these approaches to
identify ephemeral or recently formed waterbodies.
The combination of ‘analysis ready data’ [37] with large scale Earth observation
analytics platforms such as Google Earth Engine [38] and Open Data Cube [39] now make
it possible to analyse all available water observations to derive a waterbody polygon set.
The aim of this paper is to describe the use of Digital Earth Australia (the Australian
instance of Open Data Cube) [39] to identify inland waterbodies across Australia and to
characterise the change in their surface area from 1987–2018 for the purposes of providing
river regulators, environmental water managers, agricultural water users and catchment
managers with a common, transparent, and shared understanding of Australia’s surface
water resources. The specific objectives are to:
1. Use Water Observations from Space (WOfS) [19] inundation frequency from 1987–
2018 to generate waterbody polygons delineating each waterbody’s maximum surface
extent over this period.
2. Quantify the time series of water surface area for each polygon for all available
Landsat observations (1987–present).
3. Demonstrate how these polygon-specific time series can be aggregated to provide
insight into water availability.
The DEA Waterbodies interface is publicly available via DEA Maps (https://maps.
dea.ga.gov.au/, accessed on 31 March 2021), ensuring that routine, robust, and repeatable
information about Australia’s waterbodies remains current and openly available.

2. Materials and Methods

2.1. Study Area
Australia is the driest inhabited continent on Earth. Australia’s climate spans Koppen
climate zones from tropical rainforest to hot desert and temperate climates (Figure 1).
Australia experiences highly variable rainfall patterns driven by multiple large-scale climate
drivers [40] as well as shorter-lived weather events [41,42]. The amount of surface water
across the Australian continent at any point in time is hard to determine, however some
estimates exist from remotely sensed data [19,43] and surface hydrology networks [44].
Remote Sens. 2021, 13, 1437 3 of 28
Remote Sens. 2021, 13, x FOR PEER REVIEW 3 of 29

Figure 1. Koppen climate classification of Australia adapted from [45]. Copyright Commonwealth of Australia. Repro-
Figure Koppen
duced1.under climate classification of Australia adapted from [45].
CC-BY license.

2.2. Satellite Imagery 2.2. Satellite Imagery

DEA is a data platform created within Geoscience Australia that generates analysis-
DEA is a data platform created
ready satellite data thatwithin
can be used Geoscience Australia
to monitor changes that generates
in Australia’s analysis-
land surface over
ready satellite data that can be used to monitor changes in Australia’s land surface over
time [39,46]. DEA provides free and open earth observation data from two public-good
satellite programs: Landsat [47] and Sentinel 2 [48]. The Landsat program, operated by
time [39,46]. DEA provides free and open earth observation data from two public-good
the United States Geological Survey and National Aeronautics and Space Administration
satellite programs: Landsat [47] and
(NASA), consists Sentinel
of three 2 [48].
multi-spectral The Landsat
satellites/sensors: program,
Landsat operated by
5 TM (1984–2013),
Landsat 7 ETM+
the United States Geological Survey(1999–present),
and Nationaland Landsat 8 OLI (2013–present).
Aeronautics These sensors
and Space have a
Administration
30 m pixel resolution, and image the Earth every 16 days on average. Landsat 5 TM was
(NASA), consists of operational
three multi-spectral
from 1984, however satellites/sensors:
routinely collected data fromLandsat 5 TM
1987 to 2011, before(1984–2013),
being
Landsat 7 ETM+ (1999–present),
decommissioned in and2013Landsat 8 OLI7 (2013–present).
[49]. The Landsat ETM+ and Landsat 8 OLI These sensors
missions are stillhave a
operational, however a failure of the scan-line corrector in Landsat 7 ETM+ in May 2003
30 m pixel resolution, and image the Earth every 16 days on average. Landsat 5 TM was
means that imagery acquired by the satellite after this date has stripes of missing data, as
operational from 1984, however
no gap-fill routinely
algorithm is applied collected datamissing
to this data. These from data
1987havetosome
2011, before
effect on the being
overall quality of the data from this satellite, however
decommissioned in 2013 [49]. The Landsat 7 ETM+ and Landsat 8 OLI missions are still the data present can still provide
valuable insights into surface changes across Australia [50], particularly when combined
operational, however a failure
with data from of the 8scan-line
Landsat OLI. corrector in Landsat 7 ETM+ in May 2003
means that imagery acquired by the satellite after this date has stripes of missing data, as
no gap-fill algorithm is applied to this data. These missing data have some effect on the
overall quality of the data from this satellite, however the data present can still provide
valuable insights into surface changes across Australia [50], particularly when combined
with data from Landsat 8 OLI.
The raw satellite data from the Landsat missions have been corrected, standardised,
and orthorectified to produce an archive of analysis-ready data at 25 m pixel resolution [39,46].
This data can be accessed using the Open Data Cube API [51] via DEA, allowing automatic
extraction and processing of the Earth observation archives.

2.3. Water Observations from Space (WOfS)

Water Observations from Space (WOfS) is a water detection algorithm that classifies
each cloud-masked, observed pixel into wet, dry or invalid (i.e., cloud obscured, high slope)
for all available observations through time [19]. WOfS uses a decision tree based on a
series of Landsat band ratios and ancilliary datasets such as valley bottom flatness and
pixel quality. This additional information provides an extra degree of validation for the
Remote Sens. 2021, 13, 1437 4 of 28

decision tree classifications, over more basic band ratio water classifiers like the normalised
difference water index (NDWI).
The WOfS classifier has a high overall degree of accuracy (97%) in correctly classifying
the presence of water, however this accuracy decreases where there are water and other
targets, particularly vegetation, occurring within the same pixel. WOfS is most effective
where there is a clear, unobstructed water surface that has been sampled by the satellites,
and is known to misclassify areas where water and vegetation co-exist as dry, such as
wetlands and riparian zones.
WOfS is run on the Landsat archive by DEA, providing a 32-year history of water
classified pixels for all of Australia. We make use of this archive to identify where pixels
have been frequently mapped as wet using the WOfS classifier to generate a map of surface
waterbodies across Australia.

3. Workflow
Our methodology takes advantage of WOfS data within DEA to locate and extract
information about persistent waterbodies across Australia. The workflow is written in
Python, and has been almost completely automated to minimise subjective decision making
and ensure reproducibility, allowing the workflow to be run repeatedly with the same
outcome. The code used to produce DEA Waterbodies is open source, and provided for
use under an Apache Licence, Version 2.0 (see Appendix A).
The workflow is built off the WOfS Statistics dataset [52]. The WOfS Statistics
dataset provides a pixel-by-pixel count of clear observations (i.e., not impacted by cloud,
cloud shadow, or other satellite acquisition issues), wet observation counts (total number
of times a pixel was observed as wet over the archive period), and a frequency statistic
that provides the number of wet observations as a percentage of the number of clear
Remote Sens. 2021, 13, x FOR PEER REVIEW
observations over the period 1987–2018. The frequency statistic provides an indicator 5 of of
28
how often water is observed in each pixel over this time period, and it is this product that
forms the basis of this workflow.
9. UseIn order to turn the WOfS
the Polsby–Popper testfrequency statisticand
[53] to identify intomanually
automatically generated
split very polygons
large polygons.
of waterbody locations, a series of steps need to be applied to remove erroneous
The Polsby–Popper statistic was used as an objective means of identifying very data,long
and
to map only the areas of water that meet specified
and/or complicated polygons for manual curation. thresholds. The decision tree used to
produce DEA Waterbodies is shown in Figure 2. A detailed description of
10. Generate a geohash identifier for each polygon. A geohash was used to generate a the methodology
is included
uniquein Appendix
identifier forA.
each mapped waterbody polygon.

2. DEA
Figure 2. DEA Waterbodies
Waterbodiesworkflow
workflowdiagram.
diagram.For
For a detailed
a detailed description
description of the
of the steps
steps involved
involved in producing
in producing the water-
the waterbody
body polygons,
polygons, see Appendix
see Appendix A. A.

The resulting waterbody polygon dataset represents the maximum extent of water-
bodies across the Australian landscape from 1987–2018. The wet frequency thresholds
have been chosen to include locations that are infrequently, but repeatedly wet, while ex-
cluding short-lived and infrequent events like flash flooding (see Appendix A).
Remote Sens. 2021, 13, 1437 5 of 28

The workflow for DEA Waterbodies consists of ten steps:

1. Check each pixel has been observed ≥128 times over the period 1987–2018 (equates
to at least four observations per year of analysis) to exclude pixels that are very
infrequently observed.
2. Check each pixel meets a wet frequency threshold, to define waterbodies only where
water persists in the landscape. We used two thresholds to generate two wet layers: 5%
and 10%. These two thresholds were used in combination to ensure that infrequently
inundated portions of larger waterbodies were included in the final polygons (see
Appendix A).
3. Turn wet pixels into polygons. This facilitates an object-based, rather than a pixel-
based analysis.
4. Merge polygons artificially cut by tile boundaries. Tile boundaries are an artefact of
the way the data are stored and adjoining waterbodies need to be connected.
5. Filter polygons by minimum size. We applied a five Landsat pixel (25 by 25 m)
minimum size limit (3125 m2 ).
6. Remove polygons containing ocean. We applied a custom ocean mask to remove
ocean and ocean-connected waterbodies from the final dataset.
7. Remove polygons within city centres. The WOfS classifer can mistake deep shadows
from high-rise buildings for water, so we remove any waterbodies identified within
high-rise city regions.
8. Find polygons within the 5% wet frequency threshold layer that intersect with the
10% wet frequency threshold layer. The locations of the waterbody polygons are
defined by the 10% wet frequency threshold, but the extent of the polygon is defined
by the 5% wet frequency threshold.
9. Use the Polsby–Popper test [53] to identify and manually split very large polygons.
The Polsby–Popper statistic was used as an objective means of identifying very long
and/or complicated polygons for manual curation.
10. Generate a geohash identifier for each polygon. A geohash was used to generate a
unique identifier for each mapped waterbody polygon.
The resulting waterbody polygon dataset represents the maximum extent of water-
bodies across the Australian landscape from 1987–2018. The wet frequency thresholds have
been chosen to include locations that are infrequently, but repeatedly wet, while excluding
short-lived and infrequent events like flash flooding (see Appendix A).

4. Time Series Extraction for Each Polygon

Each individual waterbody polygon was interrogated to return a time series of wet
surface area through time. The WOfS classifier can be run on individual Landsat scenes
to get a snapshot of water/not water for every satellite acquisition since 1987. We used
the waterbodies dataset to perform polygon drills, whereby the individual waterbody
polygons were used to ‘drill’ through the stack of water classified Landsat imagery through
time, to produce a 32-year history of water surface area (Figure 3). Only scenes where at
least 80% of the total waterbody was clearly observed were included to minimise erroneous
results caused by cloud, Landsat 7 scan-line off missing values or swath boundaries.
Each waterbody time series provides insights into the percentage of each total polygon
area that is observed as wet for each satellite observation. Given that the waterbody
polygons represent a maximum area for each waterbody, it is common for only part of
the whole polygon to be observed as wet. Depending on the shape and nature of each
polygon, some polygons will regularly be observed as 100% wet area or 100% dry area
(e.g., constructed on-farm storages), while some polygons will never be completely dry
(e.g., most large water reservoirs) or completely wet (e.g., large inland lakes like Kati
Thanda–Lake Eyre).
The time series associated with each waterbody is stored in a separate CSV file,
named according to the polygon geohash. Each CSV file is updated as new WOfS data
Remote Sens. 2021, 13, 1437 6 of 28

Remote Sens. 2021, 13, x FOR PEER REVIEW

becomeavailable, providing an operational picture of the changing wet surface 6 of 28 area of

each waterbody over time.

Due to the 2D nature of satellite imagery, this information is only available for surface
area, Due
and to
notthe 2Dwaterbody
for nature of satellite
volume. imagery, this information
For steep-sided is only available
waterbodies for surface
(for example, constructed
area, and
water not forawaterbody
storages), negligible volume.
change in Forsurface
steep-sided waterbodies
area may equate (for
to a example, con- in total
large change
structed water storages), a negligible change in surface area may equate to a large change
waterbody volume. Despite this, the surface area change in water extent provides powerful
in total waterbody volume. Despite this, the surface area change in water extent provides
insights into the timing of filling and emptying events.
powerful insights into the timing of filling and emptying events.

Figure 3. Conceptualisation of the time series extraction for each polygon. (a–c) At each individual
Figure 3. Conceptualisation of the time series extraction for each polygon. (a–c) At each individual
time step, the percentage surface area of water (indicated in blue) within each waterbody polygon
time step,inthe
(outlined percentage
red) surface
was calculated, area ofawater
producing (indicated
time series in blue)
(d) of the changewithin each
in surface waterbody
water extent polygon
(outlined in red) was
for each waterbody calculated,
since 1987. producing a time series (d) of the change in surface water extent for
each waterbody since 1987.
5. Results
5.5.1.
Results
Validation
5.1. Validation
5.1.1. Polygon Validation
5.1.1. Polygon Validation
We compared polygons from DEA Waterbodies to those from the Bureau of Meteor-
ology We compared
Australian polygons Geospatial
Hydrological from DEAFabricWaterbodies
(Geofabric).to The
those from the
Geofabric Bureau of Meteo-
is Australia’s
rology Australian Hydrological
most comprehensive hydrologicalGeospatial
cartographyFabric
product:(Geofabric). The Geofabric
a digital database of surfaceisand
Australia’s
most comprehensive
groundwater hydrological
hydrological cartography
features, including product:
the spatial a digitalbetween
relationships database of [54].
them surface and
It is describedhydrological
groundwater as a ‘digital street directory
features, of Australia’s
including the spatialimportant waterbetween
relationships features’them
([55],[54]. It is
page 2). Waterbody
described polygons
as a ‘digital street contained
directory within the Geofabric
of Australia’s are taken
important from
water the Surface
features’ ([55], p. 2).
Hydrology Database
Waterbody polygons[56], which iswithin
contained composed of datasetsare
the Geofabric from statefrom
taken government agencies
the Surface Hydrology
from satellite imagery, aerial imagery, fieldwork, digital elevation models and
Database [56], which is composed of datasets from state government agencies from satellite other data
sources. The Geofabric represents the best available, nationally recognised and exten-
imagery, aerial imagery, fieldwork, digital elevation models and other data sources. The Geo-
sively quality controlled dataset of Australian waterbodies, and as such, is the most suit-
fabric represents the best available, nationally recognised and extensively quality controlled
able dataset for validation of DEA Waterbodies polygons.
dataset ofoverall
The Australian waterbodies,
distribution and assize
of waterbody such, is the
is very mostbetween
similar suitableDEA
dataset for validation of
Waterbodies
DEA Waterbodies polygons.
and the Geofabric. DEA Waterbodies misses waterbodies smaller than 5 Landsat pixels
(3125Them2)overall
which we distribution of waterbody
explicitly filtered size4).
out (Figure is However,
very similar DEA between DEA maps
Waterbodies Waterbodies
and the Geofabric. DEA Waterbodies misses waterbodies smaller than
more small waterbodies above this threshold. There are likely two reasons for this: first, 5 Landsat pix-
els (3125 m2 ) which we explicitly filtered out (Figure 4). However, DEA Waterbodies
maps more small waterbodies above this threshold. There are likely two reasons for
this: first, complex waterbody boundaries may result in many small discrete waterbodies,
Remote Sens. 2021, 13, x FOR PEER REVIEW 7 of 28
Remote Sens. 2021, 13, 1437 7 of 28

complex waterbody boundaries may result in many small discrete waterbodies, rather
than a single object; second, DEA Waterbodies identifies additional new waterbodies not
rather than a single object; second, DEA Waterbodies identifies additional new waterbodies
mapped in the Geofabric. Other discrepancies may come from the different methods by
not mapped in the Geofabric. Other discrepancies may come from the different methods
which the products were created: waterbodies mapped in DEA Waterbodies must have
by which the products were created: waterbodies mapped in DEA Waterbodies must have
been
been observed
observed to to be
be wet
wet in
in at
at least
least 10%
10% of
of observations
observations since
since 1987,
1987, as
as opposed
opposed to to Geofabric
Geofabric
waterbodies,
Remote Sens. 2021, which
13, x FOR PEER REVIEW are visually interpreted and may be no longer active or
7 of
waterbodies, which are visually interpreted and may be no longer active or may not have
28 may not have

existed
existed at at the
the time
time of
of production.
production.
complex waterbody boundaries may result in many small discrete waterbodies, rather
than a single object; second, DEA Waterbodies identifies additional new waterbodies not
mapped in the Geofabric. Other discrepancies may come from the different methods by
which the products were created: waterbodies mapped in DEA Waterbodies must have
been observed to be wet in at least 10% of observations since 1987, as opposed to Geofabric
waterbodies, which are visually interpreted and may be no longer active or may not have
existed at the time of production.

Figure 4.
Figure Size distribution
4. Size distribution of
of waterbodies
waterbodies in
in DEA
DEA Waterbodies
Waterbodies and
and the
the Geofabric
Geofabric for
for (a)
(a) all
all waterbodies,
waterbodies, (b)
(b) waterbodies
waterbodies
2
less than 0.1 km ..
2
Figure 4. Size distribution of waterbodies in DEA Waterbodies and the Geofabric for (a) all waterbodies, (b) waterbodies
less than 0.1 km2.
We also
We also performed
performed aa closer closer comparison
comparison on on two
two regions,
regions, one one near
near Moree
Moree NSW
NSW with
with
well-defined but We also performed
dynamic farm adams,
closer comparison
and one on two dominated
area regions, one near
by Moree
salt NSW
lakes with
near Cranbrook
well-defined well-defined
but dynamic farm dams, and one area dominated by salt
but dynamic farm dams, and one area dominated by salt lakes near Cran-
lakes near Cran-
WA. The
brook WA. comparison
The
brook between
comparison
WA. The DEA
between
comparison Waterbodies
DEA
between DEA Waterbodiesand
Waterbodies and Geofabric
and Geofabric
Geofabric polygons
polygons Mo-near near
polygons
near MoreeMo-is
shown
ree in Figure
is shown ree is 5 and5the
shown
in Figure in
and comparison
Figure 5 and the with Cranbrook
comparison
the comparison with Cranbrook
with Cranbrooksalt lakes
salt lakes is is shown
shown
salt lakes in in Figure
Figure
is shown A7.
in Figure
A7.
A7.

Figure 5. Difference and intersection between DEA Waterbodies and Geofabric in the Moree test area, dominated by con-
Figure 5. Difference and intersection between DEA Waterbodies and Geofabric in the
structed farm storages. Differences have been smoothed and buffered by 10 m to aid visibility.
Moree test area,
dominated by constructed farm storages. Differences have been smoothed and buffered by 10 m to
The Moree comparison shows that almost all Geofabric waterbodies are also in-
aid visibility. cluded in DEA Waterbodies, with only minor differences along the edge of the polygon.
These minor differences are due to the pixel grid of Landsat compared to the smooth and
The Moree comparison shows that almost all Geofabric waterbodies are also in-
cluded in DEA Waterbodies, with only minor differences along the edge of the polygon.
These minor differences are due to the pixel grid of Landsat compared to the smooth and
manually-curated
Figure 5. Difference and intersection between DEAGeofabric polygons.
Waterbodies The Geofabric-only
and Geofabric polygons
in the Moree test in the northeast
area, dominated by con- are
rarely active
structed farm storages. Differences andsmoothed
have been do not fulfil
and the >10%bywet
buffered 10 mpixel threshold
to aid visibility.to be included in DEA Water-
bodies. Additionally, DEA Waterbodies includes two large reservoirs that are not present
in theThe
Geofabric. The larger reservoir
Moree comparison shows is often
that dry for
almost allentire years,waterbodies
Geofabric and the smallerarereservoir
also in-
was wetinonly
cluded DEAfrom 2016 to 2018,
Waterbodies, explaining
with whydifferences
only minor they were missed in the
along the Geofabric.
edge There is
of the polygon.
also a large creek in the northwest which the Geofabric does not include,
These minor differences are due to the pixel grid of Landsat compared to the smooth andas it manages
streams separately to waterbodies. The Cranbrook comparison shows little difference in
Remote Sens. 2021, 13, 1437 8 of 28
Remote Sens. 2021, 13, x FOR PEER REVIEW 8 of 28

which polygons areGeofabric

manually-curated identified, but someThe
polygons. polygons are smaller
Geofabric-only in DEAinWaterbodies
polygons the northeast than
are
in the Geofabric (Figure A7). This is due to salt lakes being misidentified
rarely active and do not fulfil the >10% wet pixel threshold to be included in DEA Water- as cloud in the
underlying WOfS dataset.
bodies. Additionally, DEA Waterbodies includes two large reservoirs that are not present
To perform a statistical comparison between DEA Waterbodies and the Geofabric,
in the Geofabric. The larger reservoir is often dry for entire years, and the smaller reservoir
we associated the polygons in DEA Waterbodies and Geofabric by searching for their
was wet only from 2016 to 2018, explaining why they were missed in the Geofabric. There
intersections, and assumed that intersecting polygons represented the same waterbody.
is also a large creek in the northwest which the Geofabric does not include, as it manages
If for a given polygon in one dataset there were multiple intersecting polygons in the other
streams separately to waterbodies. The Cranbrook comparison shows little difference in
dataset, we assumed that all intersecting polygons were part of the same physical water-
which polygons are identified, but some polygons are smaller in DEA Waterbodies than
body. We compared the areas covered by the waterbody in each dataset. The distribution
in the Geofabric (Figure A7). This is due to salt lakes being misidentified as cloud in the
of the differences in area for each waterbody is shown in Figure 6. While most waterbodies
underlying WOfS dataset.
are very close in area between the two datasets (with a median deviation of 0.73% for
To perform a statistical comparison between DEA Waterbodies and the Geofabric,
Moree and –8.11% for Cranbrook) DEA Waterbodies and the Geofabric are much more
we associated the polygons in DEA Waterbodies and Geofabric by searching for their in-
consistent on constructed farm storages, compared to salt lakes, which tend to be smaller
tersections,
in and assumed
DEA Waterbodies. Thisthat intersecting
highlights polygons represented
the difficulties in defining athe same waterbody.
boundary for naturalIf
for a given polygon
waterbodies, with DEA in one dataset there
Waterbodies were multiple
defining the polygonintersecting polygons
limit based in the other
on observed wet
dataset, we assumed that all intersecting polygons
history, compared to Geofabric, which is manually curated. were part of the same physical water-
body.WeWe compared
tested the areas covered
the asymmetric by between
difference the waterbody in eachindataset.
all polygons the twoThe testdistribution
areas, and
of the differences in area for each waterbody is shown in Figure 6. While
computed the area of the result. This describes how much of each waterbody is contained most waterbodiesin
are Geofabric
the very closebut
in not
areainbetween the two datasets
DEA Waterbodies and vice(with
versa.a The
median deviationofofthese
distributions 0.73% for
areas
Moree
are andin–8.11%
shown for Cranbrook)
Figure A8. DEA Waterbodies
For the Cranbrook salt lakes, theand the Geofabric
difference in area isare much more
primarily due
consistent on constructed farm storages, compared to salt lakes, which tend
to the difference in the polygon boundary as defined in the Geofabric that is not included in to be smaller
in DEA
DEA Waterbodies.
Waterbodies, largelyThis highlights
due the of
to the edges difficulties
salt lakesin defining
being masked a boundary
as cloud. The for natural
Moree
waterbodies, with DEA Waterbodies defining the polygon limit based
test site is much more symmetric, with similar rates of omission and commission between on observed wet
history, compared
the datasets. to Geofabric, which is manually curated.

Figure
Figure 6. Distribution of
6. Distribution ofdifferences
differencesininarea
areabetween
betweenDEA
DEA Waterbodies
Waterbodies polygons
polygons and
and Geofabric
Geofabric
polygonsfor
polygons forthe
theMoree
MoreeandandCranbrook
Cranbrooktesttestareas.
areas.

5.1.2. Surface Area Time Series Validation

We tested the asymmetric difference between all polygons in the two test areas, and
Validation
computed of the
the area ofDEA Waterbodies
the result. time series
This describes how(amuch
wet surface
of eacharea time series)
waterbody is made
is contained
difficult
in the Geofabric but not in DEA Waterbodies and vice versa. The distributions ofathese
by the lack of a comparable dataset to validate against. We chose to compare sub-
set of DEA Waterbodies time series against gauged reservoir volume time
areas are shown in Figure A8. For the Cranbrook salt lakes, the difference in area is pri- series available
through
marily duethe Bureau of Meteorology’s
to the difference Water Data
in the polygon Online (WDO)
boundary service.
as defined WDO
in the provides
Geofabric vol-
that is
ume and depth measurements for gauges in approximately 360 reservoirs,
not included in DEA Waterbodies, largely due to the edges of salt lakes being masked as dams, and weirs
around
cloud. TheAustralia.
MoreeWe testselected a subset
site is much of 139
more gauges forwith
symmetric, validation,
similar based
rates ofonomission
the lengthandof
the gauge record, the availability
commission between the datasets. and quality of depth and volume measurements, and the
waterbody type (See Appendix B). We compared each volume time series to the surface
area
5.1.2.time series
Surface from
Area DEA
Time Waterbodies
Series Validationusing the Spearman ρ coefficient. The Spearman
coefficient measures the correlation of two datasets which have a positive monotonic rela-
Validation
tionship, regardlessof the DEA Waterbodies
of whether time series
that relationship (a or
is linear wet surface area
otherwise. It cantime series) is
be assumed
made difficult by the lack of a comparable dataset to validate against.
that the relationship between waterbody surface area and volume is positive monotonic, We chose to com-
pare a subset of DEA Waterbodies time series against gauged reservoir volume time series
 erwise. It can be assumed that the relationship between waterbody su
ume is positive monotonic, making this statistic suitable for comparis
measurements (surface area vs. volume), without the need to convert
Remote Sens. 2021, 13, 1437 The locations and ρ values of the 139 gauges are shown in 9 ofFigure
28 7
show good correlation with the gauge data, with ρ coefficients > 0.8. W
mania show an overall poor relationship between DEA Waterbodies
making this statistic suitable for comparison of two different measurements (surface area
The highland
vs. volume), withoutlakes in to
the need Tasmania havethem.
convert between high tannin content, making them
ing The locations
in very and levels
low ρ values of
ofthewater
139 gauges are shown
leaving in Figure 7.When
radiance. Most waterbodies
combined wi
show good correlation with the gauge data, with ρ coefficients > 0.8. Waterbodies in
this
Tasmaniaresults in near-zero
show an overall surface
poor relationship betweenreflectance
DEA Waterbodiesvalues
and WDOacross
gauges. both
near/shortwave infrared
The highland lakes in Tasmania havewavelengths. The
high tannin content, normalised
making difference
them look dark, resulting indic
in very low levels of water leaving radiance. When combined with low solar angles this
classification algorithm are extremely sensitive to noise, resulting in a
results in near-zero surface reflectance values across both the visible and near/shortwave
misclassified as land.
infrared wavelengths. This causes
The normalised the
difference waterbody’s
indices within the WOfSwetclassification
surface area o
algorithm are extremely sensitive to noise, resulting in areas of water being misclassified as
tuate as misclassified land pixels are counted amongst the correctly cla
land. This causes the waterbody’s wet surface area observations to fluctuate as misclassified
Figure
land pixels7d).
are counted amongst the correctly classified water pixels Figure 7d).

Figure 7. Validation of DEA Waterbodies time series. (a) Spearman ρ coefficients from the DEA
Figure 7. Validation of DEA Waterbodies time series. (a) Spearman ρ coefficie
Waterbodies/WDO time series comparison. Waterbodies/gauges used to validate the DEA Water-
Waterbodies/WDO
bodies time series, colouredtime by theseries
Spearmancomparison.
ρ coefficient ofWaterbodies/gauges usedofto valid
the relationship. (b) Distribution
bodies
Spearmantime series,
ρ coefficients forcoloured by individual
all analyses. The the Spearman
gauges andρtheir
coefficient of the
corresponding ρ are relationship.
given in
Spearman ρ coefficients for all analyses. The individual gauges andfrom
Appendix C, in Table A1. (c) Lake Burrendong surface area in DEA Waterbodies and estimated their corr
WDO. (d) Lake St. Clair surface area in DEA Waterbodies and estimated from WDO. The surface
given in Appendix C, in Table A1. (c) Lake Burrendong surface area in DEA W
areas in (c,d) are scaled as a percentage of maximum waterbody extent in DEA Waterbodies.
mated from WDO. (d) Lake St. Clair surface area in DEA Waterbodies and est
The To help visualise
surface areas thein comparison betweenas
(c,d) are scaled WDO and DEA Waterbodies,
a percentage of maximum we estimated
waterbody e
the surface area time series for each WDO volume time series using the volume and
bodies.
depth gauge data. Figure 7c and d show the estimated WDO surface area time series
plotted alongside the DEA Waterbodies surface area time series for Lake Burrendong (New
South To
Wales) andvisualise
help Lake St. Clair
the(Tasmania)
comparisonrespectively.
between The Burrendong
WDO andtime DEAseries
Waterb
(Figure 7c) shows very close agreement between the surface area time series modelled
the surface area time series for each WDO volume time series using the
Remote Sens. 2021, 13, 1437 10 of 28

from the reservoir gauge and the surface area time series retrieved from Landsat. This is
reflected in its high Spearman ρ of 0.99. The time series diverge when the dam is close
to completely full. This may be due to steep slopes at the edges of the reservoir, or trees
obscuring the edges of reservoir blocking the water signal and generating mixed signal
pixels. The Landsat time series does not capture the 2012 filling event as this occurred in
the gap between Landsat 5 and Landsat 8.
The Lake St. Clair time series (Figure 7d) shows poor agreement between the surface
area time series modelled from the reservoir gauge and the surface area time series retrieved
from Landsat. This is reflected in its relatively low Spearman ρ of 0.33, and is a result
of a misclassification of water as land by WOfS. Based on the results of our analysis,
DEA waterbodies surface area time series for Tasmanian highland lakes should be treated
with caution. Such a pattern mainly occurs in the highland lakes of Tasmania, where Lake
St. Clair is located. This spatial pattern is apparent in Figure 7a. In future, the WOfS water
classifier will be revised to address the inaccuracies that can occur over dark waterbodies
at low solar angles.

5.2. Case Studies of DEA Waterbodies Insights

DEA Waterbodies provides an unprecedented level of insights into the spatial and
temporal dynamics of surface water across the Australian continent. The combination of
waterbody polygons with accompanying wet area observational time series provides an
opportunity to explore surface water dynamics at an individual waterbody, through to
national scale. The following section presents a series of case study results that demonstrate
the power of this dataset, though these do not represent an exhaustive analysis of the
insights available from DEA Waterbodies.

5.2.1. Per Waterbody

Individual waterbody time series provide insights into the change in wet surface area
of each waterbody across Australia, providing detailed information about the behaviour of
the waterbody over time and providing indirect insights into the nature of the waterbody
itself. The time series produced for each polygon corresponds well with the observed
percentage of the total surface area of each waterbody observed as wet at each time step
(Figure 8). These time series are most useful for indicating the timing of a change in the
surface wet area, since they do not provide insights into volume. The timing of filling
and emptying events is an important metric for understanding how waterbodies change
over time. These changes are not attributed in DEA Waterbodies, and could result from,
rainfall events, watercourse flooding, human management of a waterbody (e.g., pumping
into/out of a storage), groundwater processes etc.
In constructed waterbodies (e.g., Figure 8) the wet area time series provides an indica-
tive measure of the date the waterbody was constructed. The date that each waterbody
was first observed as having at least 50% of the total waterbody area classified as wet was
used as a proxy for construction date. Waterbodies that are natural and therefore not built,
or waterbodies constructed prior to 1987, can be expected to return a ‘build date’ that
corresponds to the first time they filled since our records began in 1987. A moderate La
Niña event occurred in 1988–1989, which could be expected to fill present waterbodies
across Australia’s east coast. We therefore discard any ‘build dates’ before 1990 to remove
the natural and existing waterbodies from this construction analysis, but recognise that
early dates may reflect water availability, rather than waterbody existence.
Prior to the construction of the storage in Figure 8, the associated time series shows
almost no open water signal. Examination of satellite imagery (not shown) indicates this
waterbody commenced construction in April 2000, and first began to fill with water in
November 2000, registering in the time series as nearly 100% ‘wet’ for the first time in
December 2000. This method can lead to bias towards very wet years, particularly in cases
where waterbodies were constructed, but not filled until a wet season a few years later.
 Remote Sens. 2021, 13, x FOR PEER REVIEW 11 of 28

waterbody commenced construction in April 2000, and first began to fill with water in
Remote Sens. 2021, 13, 1437 11 of 28
November 2000, registering in the time series as nearly 100% ‘wet’ for the first time in
December 2000. This method can lead to bias towards very wet years, particularly in cases
where waterbodies were constructed, but not filled until a wet season a few years later.

Figure 8. Time series of the change in the percentage of total surface area observed as water for waterbody ID r6f07ug9r.
Time
Figure 8. The series of the change in the percentage of total surface area observed as water for waterbody ID r6f07ug9r.
corresponding false colour imagery is shown for three time steps, showing the relationship between the time series
The corresponding
and the rawfalse colour imagery is shown for three time steps, showing the relationship between the time series and
imagery.
the raw imagery.
5.2.2. Per Region
Individual waterbody time series can be combined at the catchment scale to provide
5.2.2.insights
Per Region
into regional-scale catchment behaviours and trends. Water management is done
at multiple scales,
Individual and catchment-scale
waterbody time series water
can bemanagement
combined in at Australia regularly
the catchment corre-
scale to provide
sponds to government policies and decisions that have been implemented at this scale
insights into regional-scale catchment behaviours and trends. Water management is done at
[57].
multiple The
scales, and catchment-scale water management in Australia regularly corresponds
Fitzroy Basin in Queensland is one of the state’s largest basins covering an area
to government policies
of approximately 142,600 andkmdecisions thatinhave
2. The growth been implemented
the number of waterbodies at this scale
mapped within[57].
the Fitzroy Basin reflects the development of agriculture and mining in this region (Figurean area of
The Fitzroy Basin in Queensland is one of the state’s largest basins covering
approximately 142,600
9). The recently km2 . The
constructed growthininthe
waterbodies thenorthern
number ofof
part waterbodies mapped
the Fitzroy Basin within the
are pri-
marily tailings dams from mining operations [58]. Insights into the growth of
Fitzroy Basin reflects the development of agriculture and mining in this region (Figure 9).waterbodies
within a region
The recently providewaterbodies
constructed an independent in assessment
the northernof infrastructure growth,Basin
part of the Fitzroy and canarebeprimarily
used to verify the installation of approved works and potentially flag unapproved works
tailings dams from mining operations [58]. Insights into the growth of waterbodies within
as well. A recent paper by Malerba et al. [59] used DEA Waterbodies to explore the rate
a region provide
of large storagean independent
construction acrossassessment of infrastructure
Australia since 1987. growth, and can be used to
verify the installation of approved works and potentially flag unapproved works as well.
A recent paper by Malerba et al. [59] used DEA Waterbodies to explore the rate of large
storage construction across Australia since 1987.

5.2.3. National/Continental
Analysis of all the waterbody time series together allows us to develop a picture of
the changing nature of open surface water across Australia. The total area of waterbody
polygons is 87,548.7 km2 , which is approximately 1% of the total land area of Australia.
The seasonal average wet area within all mapped waterbodies across Australia suggests
that there is never an occasion where all mapped waterbodies contain at least some water,
and there is a large degree of spatial variability in the spatial and temporal patterns of
waterbody wet events (Figure 10). Two similarly extensive wet events in 1989 and 2010
show large variations in the spatial patterns of wet waterbodies across Australia. The 1989
event shows wet areas dominating in the south of the continent, corresponding with a
weak La Niña event, while the 2010/11 event that coincided with a large La Niña event
shows wet waterbodies mainly in eastern Australia (Figure 10). Future work will explore
 Figure 9. Date each waterbody within the Fitzroy River Catchment, QLD,
having at least 50% of the waterbody area wet. Note that this metric indica
Remote Sens. 2021, 13, 1437 12 of 28
struction date for artificial waterbodies.
x FOR PEER REVIEW
5.2.3.
the National/Continental
spatiotemporal relationships between DEA Waterbodies and large-scale climate drivers
like El Nino Southern Oscillation, Indian Ocean Dipole, and the Southern Annular Mode.
Analysis of all the waterbody time series together allows us t
the changing nature of open surface water across Australia. The to
polygons is 87,548.7 km2, which is approximately 1% of the total
The seasonal average wet area within all mapped waterbodies acr
that there is never an occasion where all mapped waterbodies conta
and there is a large degree of spatial variability in the spatial and
waterbody wet events (Figure 10). Two similarly extensive wet ev
show large variations in the spatial patterns of wet waterbodies acr
event shows wet areas dominating in the south of the continent,
weak La Niña event, while the 2010/11 event that coincided with
shows wet waterbodies mainly in eastern Australia (Figure 10). Fu
the spatiotemporal relationships between DEA Waterbodies and la
ers like El Nino Southern Oscillation, Indian Ocean Dipole, and
Figure 9. Date each waterbody within the Fitzroy River Catchment, QLD, was first observed as having at least 50% of the
Figure 9. Date each Mode.
waterbody area wet. Note that
waterbody within the Fitzroy River Catchment, QLD, was first obse
this metric indicates an approximate construction date for artificial waterbodies.
having at least 50% of the waterbody area wet. Note that this metric indicates an approx
struction date for artificial waterbodies.

5.2.3. National/Continental
Analysis of all the waterbody time series together allows us to develop a
the changing nature of open surface water across Australia. The total area of w
polygons is 87,548.7 km2, which is approximately 1% of the total land area of
The seasonal average wet area within all mapped waterbodies across Australi
that there is never an occasion where all mapped waterbodies contain at least so
and there is a large degree of spatial variability in the spatial and temporal p
waterbody wet events (Figure 10). Two similarly extensive wet events in 1989
show large variations in the spatial patterns of wet waterbodies across Australia
event shows wet areas dominating in the south of the continent, correspondi
weak La Niña event, while the 2010/11 event that coincided with a large La N
shows wet waterbodies mainly in eastern Australia (Figure 10). Future work w
the spatiotemporal 10.relationships
National-scale between
seasonal wetseasonal DEATheWaterbodies
area statistics. and large-scale clim
Figure
Figure 10. National-scale wet bar chart
area at the top indicates
statistics. The bar the total
chartwetat the t
ers like El Nino Southern
surface area (average Oscillation, Indian
of all wet pixels inside waterbodyOcean
polygons Dipole, and
each year), the
wet surface area (average of all wet pixels inside waterbody polygons eac the
total dry Southern
surface
area (average of all dry pixels inside waterbody polygons each year), and the total area of polygon
Mode. surface area
waterbodies that was(average
unobservedof all dry
(average of allpixels inside
unobserved pixelswaterbody
inside waterbodypolygons
polygons eacheach yea
polygon
year waterbodies
e.g., cloud, that
Landsat 7 scan-line off was unobserved
missing (average
values etc.) for all of The
waterbodies. all maps
unobserved
indicate pix
the spatial pattern of the percentage of each waterbody’s surface wet area at selected time steps.
polygons each year e.g., cloud, Landsat 7 scan-line off missing values etc.) f
The data gap in 2011/12 occurs between the Landsat 5 and Landsat 8 missions.
maps indicate the spatial pattern of the percentage of each waterbody’s su
lected time steps. The data gap in 2011/12 occurs between the Landsat 5 an
Remote
Remote Sens. 2021, 13,
Sens. 2021, 13, 1437
x FOR PEER REVIEW 13 of
of 28
28

6.
6. Discussion
Discussion
6.1. Accuracies and Limitations of DEA Waterbodies
limitations of DEA Waterbodies are largely inherited from WOfS,
Accuracies and limitations
with DEA Waterbodies a reanalysis and mapping product built off the WOfS datasets.
WOfS hashas aanumber
numberofofknownknownlimitations
limitationsand and these
these manifest
manifest as incorrectly
as incorrectly mapped
mapped wa-
waterbodies within
terbodies within this
this analysis.
analysis. WOfS
WOfS usesthe
uses thespectral
spectralsignature
signatureofofwater
water to
to classify wet
pixels and is known to be suboptimal in locations where water and vegetation are mixed.
This includes
includeslocations
locationssuch
suchasas rivers with
rivers vegetated
with vegetatedriparian zones
riparian and and
zones vegetated wetlands
vegetated wet-
(Figure 11). The
lands (Figure effect
11). The of this can
effect be seen
of this can beby seen
the discontinuity of narrower
by the discontinuity river features
of narrower river
identifiedidentified
features within this analysis
within this(Figure 11a),
analysis and by11a),
(Figure an under-representation of water within
and by an under-representation of
vegetated wetlands, such as the Macquarie Marshes, NSW (Figure 11b).
water within vegetated wetlands, such as the Macquarie Marshes, NSW (Figure 11b).

DEA Waterbodies
Figure 11. Limitations of DEA Waterbodiesassociated
associatedwith
withmixed
mixedwater
waterand
andvegetation
vegetationtargets.
targets.This
Thisresults
resultsinin(a)
(a)disconti-
discon-
tinuitywithin
nuity withinsome
someriver
river reaches,
reaches, and
and (b)(b)
anan under
under representation
representation ofof waterbodies
waterbodies within
within wetlands.
wetlands.

Other known WOfS limitations have been addressed through the filtering processes
produce the
used to produce the map
mapof ofwaterbodies.
waterbodies.The Thesizesizeofofmapped
mappedwaterbodies
waterbodies is is limited
limited to to
at
at least
least fivefive Landsat
Landsat pixels,
pixels, removing
removing small,
small, isolatedisolated
‘water’ ‘water’
pixels.pixels.
The sizeThe size filtering
filtering reduces
reduces
the noisethe noise associated
associated with the withdatasetthebydataset by only focusing
only focusing on largeron largerofgroups
groups pixels.of pixels.
Misclas-
Misclassification of water in deep shadows in high density cities
sification of water in deep shadows in high density cities has been handled by removing has been handled by
removing any waterbody polygons identified within central
any waterbody polygons identified within central business districts (CBDs). Intermit-business districts (CBDs).
Intermittently
tently misclassified
misclassified features,features,
which returnwhichvalidreturn validonly
results results only a of
a handful handful
times of times
over the
over thestudy
32-year 32-year study
period, areperiod, are also
also filtered outfiltered
by testing outfor
bythetesting
numberfor the number
of valid of valid
observations
observations
returned returned
for each pixel.for each pixel.
Despite this,
Despite this, somesomeerrors
errorsremain
remain in in
thethefinalfinal waterbodies
waterbodies dataset.
dataset. SteepSteep
terrainterrain
shad-
shadows present a known difficulty for the WOfS classifier, due
ows present a known difficulty for the WOfS classifier, due to the shadows produced. to the shadows produced.
While
While WOfS
WOfS hashas attempted
attempted to mitigate this
to mitigate this issue,
issue, some
some misclassification
misclassification remains.
remains. We We have
have
not specifically attempted to address these errors within this workflow, and asassuch,
not specifically attempted to address these errors within this workflow, and such,a
a negligiblenumber
negligible numberofofthetheidentified
identifiedwaterbodies
waterbodiesmay mayin in fact
fact be
be artefacts
artefacts caused
caused by by terrain
terrain
shadow. A
shadow. Asignal-to-noise
signal-to-noiseratio ratioerror
errorover
overdeeper
deeper water
water in in
thethe
WOfSWOfS classifier
classifier hashas
alsoalso
not
been addressed here, and may result in some pixels missing from the centre of deeper
not been addressed here, and may result in some pixels missing from the centre of deeper
waterbodies,
waterbodies, resulting
resulting in in doughnut-shaped
doughnut-shaped mapped mapped polygons. Similarly, different
polygons. Similarly, different water
water
colours may interfere with the decision-tree classifier, resulting in very turbid or
colours may interfere with the decision-tree classifier, resulting in very turbid or coloured
coloured
waterbodies
waterbodies being
being omitted.
omitted. For For aa full
full discussion
discussion of of the
the limitations
limitations andand accuracy
accuracy of of WOfS,
WOfS,
see Mueller et al., 2016.
see Mueller et al., 2016.
Not all waterbodies across Australia have been identified by this product, and there
Not all waterbodies across Australia have been identified by this product, and there
remain a number of known reasons why some waterbodies may have been missed:
remain a number of known reasons why some waterbodies may have been missed:
1. They might be too small: DEA Waterbodies only maps waterbodies larger than
1. They might be too small: DEA Waterbodies only maps waterbodies larger than 3125
3125 m2 (five whole Landsat pixels).
m2 (five whole Landsat pixels).
 sifier does not work well where water is combined with vegetation. If there is vege-
tation obscuring the water (like a tree leaning across a river, or a wetland), the classi-
fier will not see this as water and the waterbody may not be mapped.
4. The waters in the waterbodies are not classified as water: Sediment rich water or
Remote Sens. 2021, 13, 1437 14 of 28
other unusual water spectra [60], like those in algae-rich waterbodies can confuse the
WOfS classifier, resulting in the pixels being incorrectly classified as not water.
5. The waterbodies might be new: Waterbodies that have been constructed or modified
2. They 2016
after might may
notnot be been
have captured within this
wet enough: DEAtoolWaterbodies
as they will only
not have
mapsbeen observed
waterbodies
as wet at least 10% of the time between 1987 and 2018. Future updates of this
that have been observed as wet at least 10% of the time between 1987 and 2018. product
If a
should capture newer waterbodies.
waterbody fills very infrequently, it may not meet this threshold.
3. The automated
waterbodies mightofhave
nature DEAtoo much vegetation
Waterbodies surrounding
identification them: The
and mapping meansWOfS
that
classifier does not work well where water is combined with vegetation.
some artefacts exist in the waterbody outlines, which do not directly match the real-world If there
is vegetation
waterbody shape. obscuring
Waterbodies thehave
water (like
been a tree leaning
mapped by DEAacross a river, using
Waterbodies or a wetland),
the reso-
the
lution ofclassifier
Landsat,will not outlines
so the see this as
of water
everyand the waterbody
waterbody may notand
are pixelated, be mapped.
may not be an
4.
exactThe waters in the
representation of waterbodies
the real-world areshape
not classified as water: Since
of the waterbody. Sediment
DEArich water or
Waterbodies
uses other unusual water
an automatic workflow spectra [60], like
utilising those data
Landsat in algae-rich waterbodies
to objectively can confuse the
map waterbodies, we
WOfS classifier, resulting in the pixels being incorrectly classified
have chosen to leave the final waterbody line work pixelated to transparently identify as not water.
5.
whichThe waterbodies
Landsat pixels might be new:
have been Waterbodies
mapped inside (andthat outside)
have been of constructed or modified
each waterbody (Figure
after 2016 may not be captured within this tool as they will
12). Future iterations of DEA Waterbodies could make use of sub-pixel methods not have been observed as
[25–28]
wet atresolution
or higher least 10%imagery
of the time between
to better 1987the
capture andtrue
2018. Future updates
waterbody shape. of this product
should capturecoarse
The relatively newerspatial
waterbodies.
resolution of Landsat and WOfS has resulted in the loss
The discontinuities
of some automated nature likeof DEA
dam Waterbodies
walls, identification
causing multiple and mapping
waterbodies means that
to be mapped as a
some
singleartefacts
polygonexist in the12a).
(Figure waterbody
This is outlines, which do
most common not directly
where separatingmatch the real-world
features like dam
waterbody shape.
walls or banks Waterbodies
are relatively haveresulting
narrow, been mapped by DEA
in mixed Waterbodies
water/land usingrecorded
pixels being the res-
olution of Landsat,
over these features, so
andthe outlines
losing of every
the clear waterbody
distinction are pixelated,
between and may
two waterbodies not the
within be
an exact representation
Landsat data. Conversely,of the real-world shape
discontinuities in the of the waterbody.
water Sinceby
classifier caused DEA
realWaterbod-
world fea-
ies uses
tures an as
such automatic workflow
small discrete waterutilising
pools, orLandsat data to objectively
spatial variability map waterbodies,
in the frequency of inunda-
we have chosen to leave the final waterbody line work pixelated to transparently
tion within a single waterbody, or limitations in the WOfS classifier as described above, iden-
tify
havewhich Landsat
resulted pixels have
in waterbodies been
being mapped
mapped as inside
a large(and outside)
number of each
of small waterbody
polygons, rather
(Figure 12). Future
than a single iterations
connected waterbodyof DEA Waterbodies
(Figure could
12b), or the make
extent usewaterbody
of the of sub-pixel meth-
being in-
ods [25–28] or higher resolution imagery
completely captured by the polygon (Figure 12c).to better capture the true waterbody shape.

Figure 12. Known limitations with the DEA Waterbodies outlines. (a) multiple waterbodies
Figure 12. Known limitations with the DEA Waterbodies outlines. (a) multiple waterbodies mapped
mapped as a single waterbody; (b) complex waterbody boundaries resulting in many small poly-
as a single waterbody; (b) complex waterbody boundaries resulting in many small polygons rather
gons rather than a single large polygon; (c) incomplete waterbody outlines.
than a single large polygon; (c) incomplete waterbody outlines.
The production of DEA Waterbodies using an objective, coded workflow means that
The relatively coarse spatial resolution of Landsat and WOfS has resulted in the loss
of some data
as new become available,
discontinuities like damthe workflow
walls, causingcan be easily
multiple re-run to update
waterbodies the analysis
to be mapped as a
single polygon (Figure 12a). This is most common where separating features like dam walls
or banks are relatively narrow, resulting in mixed water/land pixels being recorded over
these features, and losing the clear distinction between two waterbodies within the Landsat
data. Conversely, discontinuities in the water classifier caused by real world features such
as small discrete water pools, or spatial variability in the frequency of inundation within a
single waterbody, or limitations in the WOfS classifier as described above, have resulted
in waterbodies being mapped as a large number of small polygons, rather than a single
connected waterbody (Figure 12b), or the extent of the waterbody being incompletely
captured by the polygon (Figure 12c).
The production of DEA Waterbodies using an objective, coded workflow means that
as new data become available, the workflow can be easily re-run to update the analysis
Remote Sens. 2021, 13, 1437 15 of 28

and improve the overall dataset quality. Future work that makes use of Sentinel 2 data
will increase the spatial and temporal resolution of the analysis, and assist in developing a
more accurate DEA Waterbodies map. Discontinuities caused by real world features could
be minimised in future data releases by removing holes from ‘doughnut-shaped’ polygons,
however this would require additional work to explore how much this process would
help improve the quality of the DEA Waterbodies map, rather than introduce new sources
of error.
Identified waterbodies are a combination of natural and man-made waterbodies,
and contain features such as rivers, lakes, water supply dams, on farm storages, wetlands,
oxbow lakes, and ponds. The DEA Waterbodies product does not identify the nature
of the waterbody (natural vs. man-made) nor its purpose. We also do not attempt to
label features that have a known name and instead use the geohash as a unique identifier.
This means that well known features such as Kati Thanda–Lake Eyre are not labelled with
their common name in this product, and are instead referred to just using their geohash
(e.g., r4ctk0hzm). DEA Waterbodies could be spatially joined with the Geofabric [61] to
attach names from previously mapped features, however differences in the extents of
mapped features between these two products may make this task difficult.

6.2. Comparison of DEA Waterbodies against Other Datasets

National-scale water mapping has been done previously over all of Australia, however
DEA Waterbodies represents the first, national-scale, satellite-derived polygon dataset of
waterbodies across the continent. Pekel et al., 2016 and Pickens et al., 2020 both present
global surface water maps, comparable to WOfS. Both of these products look at the fre-
quency of water for each individual pixel, but do not provide vector-based analyses of
surface water, making it more difficult to interrogate whole waterbodies. Pekel et al., 2016
and Pickens et al., 2020 provide the ability to generate a time series of water history for each
individual pixel, adding some time context to the water maps. While this functionality is
similar to DEA Waterbodies, our ability to track the surface area of water over time across a
whole waterbody provides insights into the behaviour of the waterbody as a whole, rather
than an arbitrary pixel within it.
The Blue Dot Water Observatory (https://www.blue-dot-observatory.com/; [62]
accessed on 31 March 2021) provides insights into the change in surface area of water,
much like DEA Waterbodies, however on a much more limited spatial and temporal scale.
The Blue Dot Observatory allows users to interrogate changes in the surface area of water
for 188 identified waterbodies across Australia (40,330 globally) using the archive of Sen-
tinel 2 data. The period of common data coverage between DEA Waterbodies and The Blue
Dot Observatory shows a high degree of similarity between the results derived from the
two products. This acts as an additional validation of our methods, which source different
waterbody extent polygons, use different water detection algorithms and use different
satellites to produce the outputs. The use of Landsat to derive the waterbody extents and
time series histories in DEA Waterbodies extends the length of the analysis back to 1987,
compared to Blue Dot Water Observatory’s five-year record from Sentinel 2.

6.3. DEA Waterbodies Applied Case Study

The development of a satellite-based waterbody map for all of Australia means that
regional and national-scale monitoring of the state of Australia’s surface waterbodies
can be carried out routinely. DEA Waterbodies is being used in the New South Wales
(NSW) Department of Primary Industries Monthly Seasonal Conditions State Seasonal
Update [63]. The State Seasonal Update is a monthly report containing an overview of the
current climate and drought state of NSW provided to government, industry, landholders
and the public. Data from DEA Waterbodies are resampled to a parish scale (to de-identify
the data) and mapped according to the average wet surface area within all storages in
that parish. This update provides an indication of stock water availability, as well as
the potential risk to agricultural water supply in the region. Insights gained from the
Remote Sens. 2021, 13, 1437 16 of 28

State Seasonal Update are being used by government to help inform drought mitigation
strategies and financial support [64].

7. Conclusions
DEA Waterbodies maps and monitors almost 300,000 waterbodies across Australia,
providing updating insights into the change in the wet surface area within each waterbody
over time. DEA Waterbodies also provides insights into the behaviour of individual
waterbodies over time, and can be used to gain insights into drought, regional-scale
development, water management practices, and water spatial variability. Future versions
of this product will make use of the higher temporal and spatial resolution of Sentinel
2 satellites to improve both the quality of the waterbody polygons themselves, and the
frequency with which they are observed. The provision of both the code used to develop
DEA Waterbodies and the outputs datasets under open-source licenses will facilitate future
work in this area, allowing others to build off the results presented here.
While DEA Waterbodies provides new insights into the change in surface water area
within waterbodies over time, water managers need to understand changes in water vol-
ume to input into water balance accounts. Future work will integrate DEA Waterbodies
with LiDAR data to derive surface area/volume curves for individual waterbodies, allow-
ing an uncertainty-bounded estimate of volume changes to be provided alongside changes
in wet surface area. This will make the product much more useful to water managers and
the water industry, who need to monitor and understand the movement of water volumes
within a catchment over time.
National-scale reporting on water availability is a critical metric for the United Nation
Sustainable Development Goal (SDG) six: clean water and sanitation. DEA Waterbodies
provides a mechanism for reporting against SDGs by providing snapshots in time of the
amount and spatial locations of surface water across Australia. Development of the product
to include volume estimates will also increase the relevance of DEA Waterbodies to this
type of national-scale reporting.

Author Contributions: Conceptualization, C.E.K. and L.L.; methodology, C.E.K.; software, C.E.K.,
V.N. and M.J.A.; validation, C.E.K., V.N. and M.J.A.; investigation, C.E.K. and V.N.; writing—original
draft preparation, C.E.K. and L.L.; writing—review and editing, C.E.K., L.L., M.J.A. and V.N. All
authors have read and agreed to the published version of the manuscript.
Funding: This research was supported by a grant from the New South Wales Department of Primary
Industries (now NSW Department of Planning, Industry and Environment).
Data Availability Statement: The data presented in this study are openly available. DEA Waterbod-
ies polygons can be downloaded from http://pid.geoscience.gov.au/dataset/ga/132814 (accessed
on 31 March 2021). Wet surface area time series can be accessed via https://maps.dea.ga.gov.au/
(accessed on 31 March 2021). The code used to produce DEA Waterbodies is open source, provided
under an Apache Licence, Version 2.0, and can be found at https://github.com/GeoscienceAustralia/
dea-waterbodies (accessed on 31 March 2021). Additional documentation on the access and use of
DEA Waterbodies is available at https://www.ga.gov.au/dea/products/dea-waterbodies (accessed
on 31 March 2021).
Acknowledgments: This research was undertaken with the assistance of resources from the National
Computational Infrastructure (NCI Australia), an NCRIS enabled capability supported by the Aus-
tralian Government. This paper is published with the permission of the CEO, Geoscience Australia.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A
Details of the methodology used to produce DEA Waterbodies is provided here to
facilitate reproducibility of the workflow. The code produced for this workflow can be
found here: https://github.com/GeoscienceAustralia/dea-waterbodies (accessed on 31
March 2021), and is open source, provided under an Apache Licence, Version 2.0.
Remote Sens. 2021, 13, 1437 17 of 28

Appendix A.1. Step 1: Check Each Pixel Has Been Observed ≥128 Times
The maximum extent of each waterbody is identified using WOfS classified scenes
from Landsat from 1987 to 2018. The number of valid WOfS observations for each pixel
varies depending on the frequency of clouds and cloud shadow, the proximity to high slope
and terrain shadow, and the seasonal change in solar angle. The ‘count_clear’ parameter,
calculated within the WOfS all time summaries, provides a count of the number of valid
observations each pixel recorded over the analysis period. We can use this parameter to
mask out pixels that were infrequently observed. If this mask is not applied, pixels that
were observed only once could be included, if that observation was wet (i.e., a single wet
observation means the calculation of the frequency statistic would be (1 wet observation)/
(1 total observation) = 100% frequency of wet observations). A threshold for the number
of clear, valid observations needed to be determined to generate a filter to exclude pixels
that were not observed frequently enough over the 32-year analysis period. We chose to
set this threshold at 128 valid observations, equating to roughly four observations per year,
though we do not require that the timing of valid observations meets any temporal regularity.

Appendix A.2. Step 2: Check Each Pixel Meets 5% and/or 10% Wet Frequency Threshold
The WOfS Statistics was filtered to remove pixels that were infrequently wet, as we
wanted to define waterbodies where water persists in the landscape, rather than floodplains.
The threshold used to filter out low frequency wet pixels can be varied based on the required
sensitivity of the output to flood events. Similarly, this threshold could be set higher to
identify the locations of refugial waterbodies.
For this analysis, a sensitivity test was conducted to determine the effect of different
wet frequency thresholds on the number and size of resulting waterbody polygons. The wet
frequency threshold indicates how frequently a pixel needed to be observed as wet in
order to be included. A wet frequency threshold of 0.1 means that at least 10% of the total,
valid observations needed to be classified as ‘wet’. Five wet frequency thresholds were
tested: 0.01, 0.02, 0.05, 0.1 and 0.2. The suitability of each threshold was tested in three
separate locations, representing three different landcover types: desert floodplain, irrigated
cropping, combination of natural water and irrigated cropping.
The influence of the varying wet frequency thresholds on both the number and area of
identified waterbodies was explored (Figure A1). The lower the wet frequency threshold
applied, the larger both the number of total identified polygons, and the area of individual
polygons (Figure A1). The area distribution of identified polygons is consistent between
wet frequency thresholds, with the absolute number of polygons the changing parameter
(Figure A1a). The number and areas of polygons identified by the 0.1 and 0.05 thresholds
are relatively consistent, suggesting that these wet frequency thresholds result in similar
numbers and size distributions of waterbody polygons. The changing sensitivity of the
wet frequency threshold on the total area of waterbodies identified (Figure A1b), shows
an exponential relationship, highlighting the importance of finding the balance between
errors of omission and commission.
While an objective analysis of the suitability of different wet frequency thresholds
is preferred, the final decision on the wet frequency thresholds used was determined
based on looking at the suitability of the individual waterbody polygons identified by
each threshold. The outputs in Figure A2 demonstrate the impact of the wet frequency
thresholds on waterbody polygon locations and shapes, however the thresholding is
applied at the pixel level prior to turning adjoining pixels into polygons. The lowest two
thresholds (0.01 and 0.02), return a lot of noisy pixels that contain some striping from
Landsat 7’s scan line corrector failure, as well as a lot of small polygons that do not clearly
line up with waterbody features (Figure A2a). At the other end of the sensitivity thresholds,
0.2 was shown to be not sensitive enough, missing both the maximum extent of some
identified features, or missing whole waterbodies entirely (Figure A2b).
 wet frequency threshold indicates how frequently a pixel needed to be observed as wet in
order to be included. A wet frequency threshold of 0.1 means that at least 10% of the total,18 of 28
Remote Sens. 2021, 13, x FOR PEER REVIEW
valid observations needed to be classified as ‘wet’. Five wet frequency thresholds were
tested: 0.01, 0.02, 0.05, 0.1 and 0.2. The suitability of each threshold was tested in three
separate
Remote Sens. 2021, 13, 1437 locations,
oldrepresenting three
applied, the larger different
both landcover
the number types: desert
of total identified floodplain,
polygons, irri-
and the area 18 of 28
of indi-
gated cropping, combination of natural water and irrigated cropping.
vidual polygons (Figure A1). The area distribution of identified polygons is consistent
between wet frequency thresholds, with the absolute number of polygons the changing
parameter (Figure A1a). The number and areas of polygons identified by the 0.1 and 0.05
thresholds are relatively consistent, suggesting that these wet frequency thresholds result
in similar numbers and size distributions of waterbody polygons. The changing sensitiv-
ity of the wet frequency threshold on the total area of waterbodies identified (Figure A1b),
shows an exponential relationship, highlighting the importance of finding the balance be-
tween errors of omission and commission.
While an objective analysis of the suitability of different wet frequency thresholds is
preferred, the final decision on the wet frequency thresholds used was determined based
on looking at the suitability of the individual waterbody polygons identified by each
threshold. The outputs in Figure A2 demonstrate the impact of the wet frequency thresh-
olds on waterbody polygon locations and shapes, however the thresholding is applied at
the pixel level prior to turning adjoining pixels into polygons. The lowest two thresholds
(0.01 and 0.02), return a lot of noisy pixels that contain some striping from Landsat 7’s
scan line corrector failure, as well as a lot of small polygons that do not clearly line up
A1. WetA1.
Figure Figure frequency threshold threshold
Wet frequency impacts onimpacts
(a) the total
on number
(a)A2a). of individual
the total number mapped waterbodies, (b)water-
the total area
with waterbody features (Figure At the other of individual
end mapped
of the sensitivity thresholds, 0.2
contained within
bodies, (b)the individual
the totalwas mapped
area towaterbodies.
contained
shown be within Individual
the
not sensitive scatter
individual
enough, points
mapped
missing represent integrations
waterbodies.
both the maximum overofscatter
Individual
extent different time
some identi-
periodspoints
from 1987 to 2018.integrations
represent over different time periods from 1987 to 2018.
fied features, or missing whole waterbodies entirely (Figure A2b).

The influence of the varying wet frequency thresholds on both the number and area
of identified waterbodies was explored (Figure A1). The lower the wet frequency thresh-

Figure
Figure A2.
A2.Waterbody
Waterbodypolygons
polygonsidentified
identifiedusing
usingthethedifferent
different wet
wetfrequency
frequency thresholds.
thresholds. TheThe
color
color of
of the polygon indicates the highest threshold that an area is identified by, e.g., if
the polygon indicates the highest threshold that an area is identified by, e.g., if you see an area inyou see an area
in pink,
pink, this
this hashas been
been detected
detected byby the0.05,
the 0.05,0.02
0.02and
and0.01
0.01thresholds;
thresholds;ififit’s
it’sblue,
blue,then
then itit has
has been
been detected
detected by all thresholds etc. (a) All wet frequency threshold polygons. (b) 0.01 (yellow) and 0.02
by all thresholds etc. (a) All wet frequency threshold polygons. (b) 0.01 (yellow) and 0.02 (green)
(green) thresholds removed. Locations missed by the 0.2 (blue) threshold are shown in pink/or-
thresholds
ange. removed.
(c) Just Locations
0.05 (pink) missed by
and 0.1 (orange) the 0.2 (blue)
thresholds. threshold
Locations missedareby shown
the 0.1in(orange)
pink/orange.
thresh-(c) Just
old are shown in pink. (d) Just 0.05 (pink) and 0.1 (orange) thresholds. Locations missed byare
0.05 (pink) and 0.1 (orange) thresholds. Locations missed by the 0.1 (orange) threshold theshown
0.1 in
(orange) threshold are shown in pink. Legend for all figures as in (a).
pink. (d) Just 0.05 (pink) and 0.1 (orange) thresholds. Locations missed by the 0.1 (orange) threshold
are shown in pink. Legend for all figures as in (a).
Both the 0.05 and 0.1 thresholds showed both errors of omission and commission,
Both the
depending 0.05location
on the and 0.1inthresholds
which theyshowed both errors
were applied. of omission
In a location and commission,
with large areas of
depending on the location in which they were applied. In a location with large areas
of flood irrigation, the 0.05 threshold identified a number of paddocks as waterbodies,
indicating this threshold is too sensitive in these conditions (Figure A2c). In more arid
areas, where waterbodies rarely reach their maximum extent, the 0.1 threshold was shown
to underrepresent the maximum waterbody extent, picking out only partial polygons
within individual waterbodies (Figure A2d). Given the limitations of both wet frequency
thresholds, a hybrid threshold was chosen, whereby the locations of waterbodies were
Remote Sens. 2021, 13, 1437 19 of 28

identified using the 0.1 threshold, but the spatial extent of these waterbodies was charac-
terised using the 0.05 threshold. This minimises commissions caused by flood irrigation,
Remote Sens. 2021, 13, x FOR PEER REVIEW 19 of 28
which are not readily captured by the 0.1 wet frequency threshold, but also captures the
larger spatial extent of individual waterbodies afforded by the 0.05 threshold. While we
have chosen a 0.1/0.05 threshold for our purposes, the workflow has been written to allow
flood
the userirrigation,
to definethe 0.05
their threshold
own single, identified
or hybrid athreshold,
number ofmaking
paddocks as waterbodies,
it customisable indi-
to different
cating this threshold
environmental conditions. is too sensitive in these conditions (Figure A2c). In more arid areas,
where waterbodies rarely reach their maximum extent, the 0.1 threshold was shown to
underrepresent
Appendix the3:maximum
A.3. Step Turn Wet waterbody extent, picking out only partial polygons within
Pixels into Polygons
individual waterbodies (Figure A2d). Given the limitations of both wet frequency thresh-
Inaorder
olds, hybridtothreshold
treat pixelswasfrom common
chosen, whereby waterbodies in the
the locations same way, we
of waterbodies polygonise
were identifiedthe
raster to produce polygons that represent the outlines of wet waterbodies.
using the 0.1 threshold, but the spatial extent of these waterbodies was characterised using Once the wet
frequency threshold had been applied to the data, pixels that did not meet
the 0.05 threshold. This minimises commissions caused by flood irrigation, which are not this threshold
were set captured
readily to ‘NaN’by andthepixels
0.1 wetthat were to threshold,
frequency be includedbutwere set to ‘1’.the
also captures This wasspatial
larger done to
ensure that the boundaries of each waterbody were easily discernable
extent of individual waterbodies afforded by the 0.05 threshold. While we have chosen aprior to converting
them intothreshold
0.1/0.05 polygons,for toourminimise
purposes,errors
thecaused
workflowby the
has polygonisation
been written to process
allow the (Figure
user toA3).
define their own single, or hybrid threshold, making it customisable to different environ-
Appendix A.4. Step 4: Merge Polygons Artificially Cut by Tile Boundaries
mental conditions.
The workflow has been designed to be completed on individual Albers Equal Areas
tiles acrossA.3.
Appendix Step 3: as
Australia Turn Wet Pixels
a means into Polygons
to break up the processing into sensible chunks. A tempo-
rary shapefile
In order was written
to treat pixelsthat included
from common allwaterbodies
of the identified
in thepolygons
same way, forwe
each individual
polygonise
tile.
the This
rastershapefile
to produce was appended
polygons that to following
represent the each tile’s
outlines of processing.
wet waterbodies.This Once
resulted
the in
wet frequency
waterbodies thatthreshold
straddlehad
tilebeen applied being
boundaries to the segmented
data, pixels that did notat
artificially meet
the this thresh-
tile boundary.
Toold were set
correct towe
this, ‘NaN’ and pixels
created that were
a one-pixel to be
buffer included
around the were
AlberssetEqual
to ‘1’. Area
This was done iden-
tile grid, to
ensure
tified anythat the boundaries
polygons of each waterbody
that intersected with thiswere easily
buffer, anddiscernable
merged any prior to converting
selected polygons
them
that into polygons, to minimise errors caused by the polygonisation process (Figure A3).
touched.

Figure
Figure A3.
A3. Polygonisationprocess
Polygonisation processto
totransfer
transfer pixels identified
identifiedas
aswaterbodies
waterbodies(a)
(a)into
intopolygons
polygons(b).
(b).

Appendix A.4. Step 4: Merge Polygons Artificially Cut by Tile Boundaries

Appendix A.5. Step 5: Filter Polygons by Max and Min Size
The workflow has been designed to be completed on individual Albers Equal Areas
A size filter was applied to the mapped waterbody polygons that removed polygons
tiles across Australia as a means to break up the processing into sensible chunks. A tem-
below five Landsat 25 m by 25 m pixels in size (3125 m2 ). This threshold was set to limit
porary shapefile was written that included all of the identified polygons for each individ-
noise associated with very small waterbodies. As a result of this threshold, small farm
ual tile. This shapefile was appended to following each tile’s processing. This resulted in
storages and small
waterbodies sections
that straddle ofboundaries
tile river havebeing
also been excluded
segmented from the
artificially final
at the product.
tile boundary.
The workflow has been designed to allow the user to specify their own
To correct this, we created a one-pixel buffer around the Albers Equal Area tile grid, size thresholds,
iden-
which can be modified depending on the desired outcome.
tified any polygons that intersected with this buffer, and merged any selected polygons
that touched.
Appendix A.6. Step 6: Remove Polygons Containing Ocean
Appendix A.5. Step
A coastline 5: Filter
dataset isPolygons
used to by Max and
remove Min Size containing ocean. The user can
waterbodies
chooseAwhich coastline dataset to use, depending on whether
size filter was applied to the mapped waterbody they
polygons would
that like features
removed polygonslike
estuaries included in the final product. Here, we use a coastline based on the
below five Landsat 25 m by 25 m pixels in size (3125 m ). This threshold was set to
2 Intertidal
limit
Extents Model version
noise associated 2 (ITEM
with very small [65]), which uses
waterbodies. As a Landsat datathreshold,
result of this to generate snapshots
small farm
storages and small sections of river have also been excluded from the final product.
The workflow has been designed to allow the user to specify their own size thresh-
olds, which can be modified depending on the desired outcome.
 Remote Sens. 2021, 13, x FOR PEER REVIEW 20 of 28

Appendix A.6. Step 6: Remove Polygons Containing Ocean

Remote Sens. 2021, 13, 1437 20 of 28
A coastline dataset is used to remove waterbodies containing ocean. The user can
choose which coastline dataset to use, depending on whether they would like features like
estuaries included in the final product. Here, we use a coastline based on the Intertidal
Extents Model version 2 (ITEM [65]), which uses Landsat data to generate snapshots of
of the Australian coastline at different tidal heights. We use the highest tide modelled
the Australian coastline at different tidal heights. We use the highest tide modelled by
by ITEM (pixels exposed at highest 80–100% of the observed tidal range (land) [66]) to
ITEM (pixels exposed at highest 80–100% of the observed tidal range (land) [66]) to repre-
represent the high-tide coastline of Australia. In order to remove estuaries and some
sent the high-tide coastline of Australia. In order to remove estuaries and some inland
inland lakes, we considered any pixel connected to the ocean to be ‘ocean’, and were thus
lakes, we considered any pixel connected to the ocean to be ‘ocean’, and were thus re-
removed by this mask. The mask has a number of limitations, which result in parts of
moved by this mask. The mask has a number of limitations, which result in parts of tidal
tidal river estuaries being mapped as waterbodies. These primarily occur where there
river estuaries being mapped as waterbodies. These primarily occur where there is a dis-
is a discontinuity in the water signal along the river channel, caused by large bridges or
continuity in the water signal along the river channel, caused by large bridges or narrow-
narrowing river channels,
ing river channels, for example
for example (Figure A4).
(Figure A4).

Figure
Figure A4.Ocean
A4. Ocean mask
mask used
usedtotoremove
removeocean andand
ocean ocean-connected waterbodies
ocean-connected from the
waterbodies final
from product.
the The maskThe
final product. incor-
mask
rectly fails
incorrectly to map
fails some
to map somecomplete
complete estuaries where
estuaries the the
where water signal
water from
signal the the
from ocean is not
ocean continuous,
is not likelike
continuous, seenseen
above in in
above
Perth,
Perth, Western
Western Australia,where
Australia, wheretwotwolarge
largebridges
bridges have
have terminated
terminated the
theocean
oceanmask
maskprematurely
prematurely (yellow
(yellowsquares).
squares).

Appendix A.7. Step 7: Remove Polygons within CBDs

Appendix A.7. Step 7: Remove Polygons within CBDs
The WOfS dataset has a known limitation, whereby deep shadows caused by high
The WOfS
rise buildings indataset
CBDs arehas a known
classified as limitation, whereby
water [19]. This meansdeep
that shadows caused
our workflow by high
is defin-
rise buildings in around
ing ‘waterbodies’ CBDs are classified
these as water
misclassified [19].inThis
shadows means
capital cities that ourA5).
(Figure workflow
To ad- is
defining
dress this‘waterbodies’
problem, we around
use the these misclassified
Australian shadows
Bureau of StatisticsinStatistical
capital cities
Area(Figure
3 (SA3)A5).
To address[67]
shapefile thistoproblem, we usefootprint
define a spatial the Australian Bureau of
for Australia’s Statistics
CBD Statistical
areas. The Area 3 (SA3)
SA3 boundaries
shapefile [67] to
are a regional define a spatial
population footprint
grouping for Australia’s
that contain betweenCBD areas.
30,000 The SA3people,
and 130,000 boundaries
typi- are
Remote Sens. 2021, 13, x FOR PEER REVIEW
acally
regional population grouping that contain between
clustered around commercial and transport hubs [68]. 30,000 and 130,000 people, 21 of 28
typically
clustered around commercial and transport hubs [68].
We use the following SA3 statistical areas as our CBD filter: 11703: Sydney Inner City,
20604: Melbourne City, 30501: Brisbane Inner, 30901: Broadbeach–Burleigh, 30910: Surfers
Paradise, 40101: Adelaide City, 50302, Perth City. The CBDs of other cities, such as Can-
berra and Darwin, do not contain enough high-rise buildings to be affected, and so were
not used within this filter.

FigureA5.
Figure A5.Incorrectly
Incorrectlyidentified
identifiedwaterbodies
waterbodieswithin
withinthe
theSydney
SydneyCBD.
CBD.

Appendix A.8. Step 8: Find Polygons within the 5% Wet Frequency Threshold Layer That
Intersect with the 10% Wet Frequency Threshold Layer
The hybrid wet frequency threshold discussed above is applied in this step. Up until
this point, two parallel analyses are being run—the 5% wet frequency threshold and the
Remote Sens. 2021, 13, 1437 21 of 28

We use the following SA3 statistical areas as our CBD filter: 11703: Sydney Inner
City, 20604: Melbourne City, 30501: Brisbane Inner, 30901: Broadbeach–Burleigh, 30910:
Surfers Paradise, 40101: Adelaide City, 50302, Perth City. The CBDs of other cities, such as
Canberra and Darwin, do not contain enough high-rise buildings to be affected, and so
were not used within this filter.

Appendix A.8. Step 8: Find Polygons within the 5% Wet Frequency Threshold Layer That Intersect
with the 10% Wet Frequency Threshold Layer
The hybrid wet frequency threshold discussed above is applied in this step. Up until
this point, two parallel analyses are being run—the 5% wet frequency threshold and the
10% wet frequency threshold. The filtering steps described in the workflow up until this
point are applied independently to these two differently thresholded products, producing
two outputs. At this point, we perform a spatial join on the two polygon datasets to find
locations where the 10% and 5% threshold overlap. Where a polygon in the 10% wet
frequency output is found to overlap with the 5% wet frequency product, the polygon
from the 5% wet frequency product is kept, and the 10% wet frequency polygon is thrown
away. This means that the locations of the waterbody polygons are defined by the 10%
wet frequency threshold, but the extent of the polygon is defined by the 5% wet frequency
threshold. This produces more spatially extensive polygons, but only in locations that have
been identified with the more conservative threshold.

Appendix A.9. Step 9: Use the Polsby-Popper Test to Identify and Split Very Large Polygons
The final waterbodies shapefile was checked for large river segments using the Polsby–
Popper statistical test [53]. The Polsby–Popper test is an assessment of the ‘compactness’ of
a polygon. This method was originally developed to test the shape of congressional and
state legislative districts, to prevent gerrymandering. The Polsby–Popper test examines the
ratio between the area of a polygon, and the area of a circle equal to the perimeter of that
polygon. The result falls between 0 and 1, with values closer to 1 being assessed as more
compact. We selected all polygons with a Polsby–Popper test value ≤ 0.005 for further
interrogation. This resulted in a subset of 186 polygons (Figure A6).
The 186 polygons were buffered with a −50 m (2 pixel) buffer to separate the polygons
where they are connected by two pixels or less. This allows us to split up these very large
polygons by using natural thinning points. The resulting negatively buffered polygons was
run through the multipart to singlepart tool in QGIS, to give the now separated polygons
unique IDs. These polygons were then buffered with a +50 m buffer to return the polygons to
approximately their original size. These final polygons were used to separate the 186 original
polygons identified above.
The process for dividing up the identified very large polygons varied depending
on the polygon in question. Where large waterbodies (like the Menindee Lakes) were
connected, the buffered polygons were used to determine the cut points in the original
polygons. Where additional breaks were required, the Bureau of Meteorology’s Geofabric
v3.0.5 Beta [61] waterbodies dataset was used as an additional source of information for
breaking up connected segments.
The buffering method didn’t work on long segments of river, which became a series
of disconnected pieces when negatively and positively buffered. Instead, we used a
combination of tributaries and man-made features such as bridges and weirs to segment
these river sections. Where no tributaries existed, large river segments were split using the
Albers Equal Area tile boundaries.
Remote Sens. 2021, 13, 1437 22 of 28
Remote Sens. 2021, 13, x FOR PEER REVIEW 22 of 28

Figure
FigureA6.
A6.Initial
InitialDEA
DEAWaterbodies
Waterbodies product,
product, with polygons
polygons with
with aa Polsby-Popper
Polsby-Poppertest value≤≤ 0.005
testvalue
0.005 highlighted
highlighted in red.
in red.

Thisprocess
The method forwas used to
dividing upproduce an objective
the identified flag polygons
very large for waterbodies
varied that may need
depending on
to be split to produce smaller and more meaningful waterbody
the polygon in question. Where large waterbodies (like the Menindee Lakes) were polygons as wellcon-
as a
semi-objective
nected, meanspolygons
the buffered for splitting
werelarge
usedpolygons. As much
to determine as was
the cut pointspossible,
in the decisions were
original poly-
hard coded
gons. Wheretoadditional
make them repeatable
breaks were for future runs
required, of this workflow.
the Bureau of Meteorology’s Geofabric
v3.0.5 Beta [61] waterbodies dataset was used as an additional source of information for
Appendixup
breaking A.10. Step 10: segments.
connected Generate a Geohash ID for Each Polygon
A unique
The identifier
buffering method is required for every
didn’t work polygon
on long to allow
segments it to be
of river, referenced.
which becameThe nam-
a series
ing
of convention for
disconnected generating
pieces unique IDs
when negatively here
and is the Geohash
positively [69].
buffered. A Geohash
Instead, we usedis a geocod-
a com-
ing system
bination used to generate
of tributaries short unique
and man-made featuresidentifiers based on
such as bridges andlatitude/longitude
weirs to segment these coor-
dinates.
river It is aWhere
sections. short combination
no tributariesofexisted,
letters large
and numbers, with the
river segments length
were split of the string
using the Al-a
function
bers EqualofArea
the tile
precision of the location. We use the python package ‘python-geohash’
boundaries.
to generate a Geohash unique
This method was used to produce identifieranfor each polygon.
objective We use precision
flag for waterbodies = 9 geohash
that may need to
characters,
be which represents
split to produce smaller andanmore on the ground accuracy
meaningful waterbody of <20 m. This
polygons asensures
well as athat the
semi-
precisionmeans
objective is highfor enough to differentiate
splitting large polygons. between waterbodies
As much as was located next
possible, to each were
decisions other.
Finally,
hard individual
coded to makepolygon areas were
them repeatable forrecalculated
future runs of and added
this as an attribute to the final
workflow.
waterbody shapefile.
Appendix A.10. Step 10: Generate a Geohash ID for Each Polygon
Appendix B. Geofabric Polygon Comparison
A unique identifier is required for every polygon to allow it to be referenced. The
naming This appendix shows
convention the spatial
for generating differences
unique IDs herebetween
is theDEA Waterbodies
Geohash [69]. A and Geofabric.
Geohash is a
Figure A7 shows the Moree test area, which mostly contains farm
geocoding system used to generate short unique identifiers based on latitude/longitude dams, and Figure A8
shows the Cranbrook test area, which is dominated by salt lakes. The
coordinates. It is a short combination of letters and numbers, with the length of the stringdifference polygons
ahave beenof
function smoothed and buffered
the precision by 10 mWe
of the location. to aid
usevisibility.
the python package ‘python-geohash’
to generate a Geohash unique identifier for each polygon. We use precision = 9 geohash
characters, which represents an on the ground accuracy of <20 m. This ensures that the
precision is high enough to differentiate between waterbodies located next to each other.
Finally, individual polygon areas were recalculated and added as an attribute to the final
waterbody shapefile.
 Appendix B.
Appendix B. Geofabric
Geofabric Polygon
Polygon Comparison
Comparison
This appendix
This appendix shows
shows the
the spatial
spatial differences
differences between
between DEA
DEA Waterbodies
Waterbodies and
and Geofab-
Geofab-
Remote Sens. 2021, 13, 1437 ric. Figure
ric. Figure A7
A7 shows
shows the
the Moree
Moree test
test area,
area, which
which mostly
mostly contains
contains farm
farm dams,
dams, and
and Figure
Figure
23 of 28
A8 shows the Cranbrook test area, which is dominated by salt lakes. The difference
A8 shows the Cranbrook test area, which is dominated by salt lakes. The difference poly- poly-
gons have been smoothed and buffered by 10 m to aid visibility.
gons have been smoothed and buffered by 10 m to aid visibility.

Figure A7.
Figure A7. Difference and intersection
Difference and intersection between
between DEA
DEA Waterbodies
Waterbodies
Waterbodies and
and Geofabric
and Geofabricin
Geofabric inthe
in theCranbrook
the Cranbrooktest
Cranbrook testarea.
test area.
area.

Figure A8.
Figure A8. Distribution
Distribution of
of areas
areas of
of the
the difference
difference polygons
polygons between
between DEA
DEA Waterbodies
Waterbodies polygons
polygons and
and Geofabric
Geofabric polygons
polygons
for the Moree and Cranbrook
for the Moree and Cranbrook test
Cranbrook test areas.
test areas.
areas.

Appendix C. Water
Appendix Water Data
Data Online
Online TimeTime Series
Series Comparison
Comparison
Appendix C. C. Water Data Online Time Series Comparison
This appendix details
This details the water
water level
level and
and volume
volume gauges
gauges used
used to to validate DEADEA Wa-Wa-
This appendix
appendix detailsthe the water level and volume gauges usedvalidate
to validate DEA
terbodies and
terbodies and the
the corresponding
corresponding Spearman
Spearman ρρ values
values of of the comparison.
comparison.
Waterbodies and the corresponding Spearman ρ values theof the comparison.
The Water
The Water Data
Data Online
Online service
service maintained
maintained by by the
the Bureau
Bureau ofof Meteorology
Meteorology provides
provides
The Water Data Online service maintained by the Bureau of Meteorology provides vol-
volume
volume and depth measurements for gauges in approximately 360 reservoirs, dams, and
ume andand depth
depth measurements
measurements for gauges
for gauges in approximately
in approximately 360 reservoirs,
360 reservoirs, dams,dams, and
and weirs
weirs
weirs around
around Australia.
Australia. We selected
We selected a subset
a subset of these gauges for validation based on the
around Australia. We selected a subset of theseofgauges
these gauges for validation
for validation based on based on the
the follow-
following
following criteria:
criteria:
ing criteria:
••• The
The gauge
The gauge needed to
gauge needed
needed to have
to have at
have at least
at least 10
least 10 years
10 years of
years of data.
of data.
data.
••• Depth and
Depth
Depth and volume
and volume measurements
volume measurements needed
measurements needed to
needed to be
to be related,
be related, i.e.,
related, i.e., an
i.e., an increase
an increase in
increase in depth
in depth
depth
needed
needed to
to be
be associated
associated with
with an
an increase
increase
needed to be associated with an increase in volume. in
in volume.
volume.
••• Gauges at
Gauges
Gauges at weirs
at weirs and
weirs and locks
and locks were
locks were removed,
were removed, since
removed, since aaa gauged
since gauged volume
gauged volume at
volume at aaa weir
at weir is
weir is not
is not
not
comparable to a waterbody polygon, and they cannot be meaningfully
comparable to a waterbody polygon, and they cannot be meaningfully compared. compared.
compared.
••• Exceptionally noisy
Exceptionally noisy gauge
gauge time
time series
series were
were removed.
removed. ThisThis was
This was determined
was determined by
determined by man-
by man-
man-
ual inspection
ual inspection
inspectionof of each
ofeach time
eachtime
time series.
series. Gauges
Gauges
series. Gaugeswere
were removed
removed
were where
where
removed therethere
there
where werewere
were erroneous
erroneous
erro-
neous spikes in data, prolonged data gaps and values that didn’t change over time,
where they could be expected to.
• Gauges that did not correspond to a polygon of a waterbody in DEA Waterbodies
were removed.
Remote Sens. 2021, 13, 1437 24 of 28

Thus, 139 volume gauges remained, and were compared against the corresponding
DEA Waterbodies polygon time series using the Spearman coefficient (ρ). These correlations
are shown in Table A1.

Table A1. Storage gauges and the ρ values from comparing their time series to DEA Waterbodies.

Gauge ρ
Arthurs Lake—At Pump Station 0.889
Aroona Creek/Dam 0.845
Brogo Dam 0.506
Bronte Lagoon—At Dam 0.713
Burbury Lake—At Crotty Dam 0.442
Burrendong Dam 0.991
Barkers Creek Storage 0.955
Bjelke-Petersen Dam 0.989
Boondooma Dam 0.986
Burdekin Falls Dam 0.948
Cargelligo Storage 0.950
Cluny Lagoon—At Dam 0.358
Cairn Curran Reservoir 0.993
Callide Dam (Intake) 0.997
Canning Wsl—Logger Data 0.934
Canning Wsl-Ranger 0.934
Cascade Ck Dam No. 2 0.364
Churchman Bk Wsl-Ranger 0.900
Clarendon Head Water 0.973
Coolmunda Dam 0.990
Corin Res. At Dam 0.860
Cotter Res. At Dam 0.723
Devilbend Reservoir 0.787
Dartmouth 0.913
Drakes Bk Wsl 0.729
Echo Lake—At Dam 0.665
Eildon 0.989
Eungella Dam 0.923
Fellmongers C At Res 0.887
Fairbairn Dam 0.992
Fred Haigh Dam 0.974
Great Lake—At Poatina Inlet 0.737
Greaves Creek Dam 0.849
Geehi At Geehi Res 0.326
Glen Mervyn Wsl—Logger 0.907
Googong Res At Dam 0.855
Greens Lake 0.968
Guthega Pondage Rl 0.437
Happy Valley Reservoir (Sa Water) 0.702
Harding Dam Water Levels—Site Readings 0.953
Harris Wsl 0.846
Harvey Dam Water Level 0.974
Harvey Dam Water Level-Logger 0.970
Harvey Dam Water Level-Manual 0.975
Jounama Pondage Rl 0.854
Julius Dam Hw 0.661
Keepit Dam 0.998
Kangaroo Creek Reservoir (Sa Water) 0.959
Kerferd Reservoir 0.720
Kinchant Dam Hw 0.864
Remote Sens. 2021, 13, 1437 25 of 28

Table A1. Cont.

Gauge ρ
Kroombit Dam 0.946
L Pamararoo Copi H 0.943
L Wetherell Tandure 0.962
Lake Cawndilla 0.982
Lal Lal Res. H.G. 0.901
Laughing Jack Lagoon—At Dam 0.879
Lake Awoonga 0.817
Lake Eucumbene Rl 0.955
Lake Mokoan 0.991
Lake Nillahcootie 0.954
Lake Paluma 0.666
Lake Victoria Wl 0.884
Leslie Dam 0.993
Little Para Reservoir (Sa Water) 0.944
Logue Brook Wsl 0.986
Loombah Reservoir 0.369
Mackenzie Lake—At Dam 0.809
Mackintosh Lake—At Dam 0.470
Margaret Lake—At Dam 0.128
Moorabool Wb Res Hg 0.960
Malmsbury Reservoir 0.988
Mccay Storage 0.687
Millbrook Reservoir (Sa Water) 0.880
Moochalabra WSL 0.825
Moochalabra Wsl Logged Dow 0.574
Mt Bold Reservoir (Sa Water) 0.957
Mundaring Wsl-Logger 0.930
Mundaring Wsl-Ranger 0.911
Myponga Reservoir (Sa Water) 0.931
New Victoria Water Level-Ranger 0.930
Nil Gully Reservoir 0.351
North Pine 0.935
Nth Dandalup Wsl—Logger Data 0.968
Pine Tier Lagoon—At Dam 0.692
Paradise Dam 0.925
Peter Faust Dam 0.950
Prospect Reservoir 0.664
Quickup Wsl—Logger 0.768
Rocklands Reservoir 0.986
Ross Dam 0.961
Silvan Reservoir 0.360
Scabby Gully Dam Wsl Logger 0.622
Serpentine Main Dam Wsl—Ranger 0.954
South Para Reservoir (Sa Water) 0.940
St.Clair Lake—at Pump House Point 0.330
Sth Dandalup Wsl-Ranger 0.968
Stirling Wsl—Logger Data 0.856
Thomson Reservoir 0.957
Trevallyn Lake—At Dam 0.389
Talbingo Res—Ro 0.224
Tantangara Dam Rl 0.985
Teemburra Dam 0.712
Upper Coliban Reservoir 0.944
Upper Stony Creek No. 2 0.601
Upper Stony Creek No. 3 0.876
Remote Sens. 2021, 13, 1437 26 of 28

Table A1. Cont.

Gauge ρ
Upper Stony Creek Reservoirs 0.362
Waranga Basin 0.987
Watts Riv-Maroondah 0.868
Wayatinah Lagoon—At Intake 0.621
Windamere Dam 0.962
Woods Lake—At Dam 0.776
Woronora R. @ Dam 0.824
Waroona Dam Water Level Logger 0.842
Warragamba 0.907
Wartook Reservoir 0.752
White Swan 0.928
Wurdee Boluc Reservoir 0.855
Yarra Riv-Uy Res Hg 0.927

References
1. Bakker, K. Water Security: Research Challenges and Opportunities. Science 2012, 337, 914–915. [CrossRef]
2. Grill, G.; Lehner, B.; Thieme, M.; Geenen, B.; Tickner, D.; Antonelli, F.; Babu, S.; Borrelli, P.; Cheng, L.; Crochetiere, H.; et al.
Mapping the World’s Free-Flowing Rivers. Nature 2019, 569, 215–221. [CrossRef]
3. Haddeland, I.; Heinke, J.; Biemans, H.; Eisner, S.; Flörke, M.; Hanasaki, N.; Konzmann, M.; Ludwig, F.; Masaki, Y.; Schewe, J.; et al.
Global Water Resources Affected by Human Interventions and Climate Change. Proc. Natl. Acad. Sci. USA 2014, 111, 3251–3256.
[CrossRef]
4. Vörösmarty, C.J.; McIntyre, P.B.; Gessner, M.O.; Dudgeon, D.; Prusevich, A.; Green, P.; Glidden, S.; Bunn, S.E.; Sullivan, C.A.;
Liermann, C.R.; et al. Global Threats to Human Water Security and River Biodiversity. Nature 2010, 467, 555–561. [CrossRef]
5. Vörösmarty, C.J.; Green, P.; Salisbury, J.; Lammers, R.B. Global Water Resources: Vulnerability from Climate Change and
Population Growth. Science 2000, 289, 284–288. [CrossRef] [PubMed]
6. Gleick, P.H. Water and Terrorism. Water Policy 2006, 8, 481–503. [CrossRef]
7. Rodell, M.; Famiglietti, J.S.; Wiese, D.N.; Reager, J.T.; Beaudoing, H.K.; Landerer, F.W.; Lo, M.-H. Emerging Trends in Global
Freshwater Availability. Nature 2018, 557, 651–659. [CrossRef] [PubMed]
8. Wouters, P. Water Security: Global, Regional and Local Challenges; Institute for Public Policy Research: London, UK, 2010; p. 17.
9. Hou, J.; van Dijk, A.I.J.M.; Beck, H.E. Global Satellite-Based River Gauging and the Influence of River Morphology on Its
Application. Remote Sens. Environ. 2020, 239, 111629. [CrossRef]
10. Sheffield, J.; Wood, E.F.; Pan, M.; Beck, H.; Coccia, G.; Serrat-Capdevila, A.; Verbist, K. Satellite Remote Sensing for Water
Resources Management: Potential for Supporting Sustainable Development in Data-Poor Regions. Water Resour. Res. 2018, 54,
9724–9758. [CrossRef]
11. Uereyen, S.; Kuenzer, C.A. Review of Earth Observation-Based Analyses for Major River Basins. Remote Sens. 2019, 11, 2951.
[CrossRef]
12. Pekel, J.-F.; Cottam, A.; Gorelick, N.; Belward, A.S. High-Resolution Mapping of Global Surface Water and Its Long-Term Changes.
Nature 2016, 540, 418–422. [CrossRef]
13. Pickens, A.H.; Hansen, M.C.; Hancher, M.; Stehman, S.V.; Tyukavina, A.; Potapov, P.; Marroquin, B.; Sherani, Z. Mapping and
Sampling to Characterize Global Inland Water Dynamics from 1999 to 2018 with Full Landsat Time-Series. Remote Sens. Environ.
2020, 243, 111792. [CrossRef]
14. Wulder, M.A.; Masek, J.G.; Cohen, W.B.; Loveland, T.R.; Woodcock, C.E. Opening the Archive: How Free Data Has Enabled the
Science and Monitoring Promise of Landsat. Remote Sens. Environ. 2012, 122, 2–10. [CrossRef]
15. Wulder, M.A.; White, J.C.; Loveland, T.R.; Woodcock, C.E.; Belward, A.S.; Cohen, W.B.; Fosnight, E.A.; Shaw, J.; Masek, J.G.;
Roy, D.P. The Global Landsat Archive: Status, Consolidation, and Direction. Remote Sens. Environ. 2016, 185, 271–283. [CrossRef]
16. Woodcock, C.E.; Allen, R.; Anderson, M.; Belward, A.; Bindschadler, R.; Cohen, W.; Gao, F.; Goward, S.N.; Helder, D.; Helmer, E.;
et al. Free Access to Landsat Imagery. Science 2008, 320, 1011. [CrossRef] [PubMed]
17. Khandelwal, A.; Karpatne, A.; Marlier, M.E.; Kim, J.; Lettenmaier, D.P.; Kumar, V. An Approach for Global Monitoring of Surface
Water Extent Variations in Reservoirs Using MODIS Data. Remote Sens. Environ. 2017, 202, 113–128. [CrossRef]
18. Yamazaki, D.; Trigg, M.A.; Ikeshima, D. Development of a Global ~90m Water Body Map Using Multi-Temporal Landsat Images.
Remote Sens. Environ. 2015, 171, 337–351. [CrossRef]
19. Mueller, N.; Lewis, A.; Roberts, D.; Ring, S.; Melrose, R.; Sixsmith, J.; Lymburner, L.; McIntyre, A.; Tan, P.; Curnow, S.; et al. Water
Observations from Space: Mapping Surface Water from 25 Years of Landsat Imagery across Australia. Remote Sens. Environ. 2016,
174, 341–352. [CrossRef]
Remote Sens. 2021, 13, 1437 27 of 28

20. Tulbure, M.G.; Broich, M.; Stehman, S.V.; Kommareddy, A. Surface Water Extent Dynamics from Three Decades of Seasonally
Continuous Landsat Time Series at Subcontinental Scale in a Semi-Arid Region. Remote Sens. Environ. 2016, 178, 142–157.
[CrossRef]
21. Tulbure, M.G.; Broich, M. Spatiotemporal Dynamic of Surface Water Bodies Using Landsat Time-Series Data from 1999 to 2011.
ISPRS J. Photogramm. Remote Sens. 2013, 79, 44–52. [CrossRef]
22. Xu, N.; Ma, Y.; Zhang, W.; Wang, X.H. Surface-Water-Level Changes during 2003-2019 in Australia Revealed by ICESat/ICESat-2
Altimetry and Landsat Imagery. IEEE Geosci. Remote Sens. Lett. 2020, 1–5. Available online: https://ieeexplore.ieee.org/
document/9104913 (accessed on 24 November 2020). [CrossRef]
23. Feng, M.; Sexton, J.O.; Channan, S.; Townshend, J.R. A Global, High-Resolution (30-m) Inland Water Body Dataset for 2000: First
Results of a Topographic–Spectral Classification Algorithm. Int. J. Digit. Earth 2016, 9, 113–133. [CrossRef]
24. Schwatke, C.; Scherer, D.; Dettmering, D. Automated Extraction of Consistent Time-Variable Water Surfaces of Lakes and
Reservoirs Based on Landsat and Sentinel-2. Remote Sens. 2019, 11, 1010. [CrossRef]
25. Halabisky, M.; Moskal, L.M.; Gillespie, A.; Hannam, M. Reconstructing Semi-Arid Wetland Surface Water Dynamics through
Spectral Mixture Analysis of a Time Series of Landsat Satellite Images (1984–2011). Remote Sens. Environ. 2016, 177, 171–183.
[CrossRef]
26. Bishop-Taylor, R.; Sagar, S.; Lymburner, L.; Alam, I.; Sixsmith, J. Sub-Pixel Waterline Extraction: Characterising Accuracy and
Sensitivity to Indices and Spectra. Remote Sens. 2019, 11, 2984. [CrossRef]
27. Souza, C.M.; Kirchhoff, F.T.; Oliveira, B.C.; Ribeiro, J.G.; Sales, M.H. Long-Term Annual Surface Water Change in the Brazilian
Amazon Biome: Potential Links with Deforestation, Infrastructure Development and Climate Change. Water 2019, 11, 566.
[CrossRef]
28. Wang, X.; Ling, F.; Yao, H.; Liu, Y.; Xu, S. Unsupervised Sub-Pixel Water Body Mapping with Sentinel-3 OLCI Image. Remote Sens.
2019, 11, 327. [CrossRef]
29. Sun, F.; Sun, W.; Chen, J.; Gong, P. Comparison and Improvement of Methods for Identifying Waterbodies in Remotely Sensed
Imagery. Int. J. Remote Sens. 2012, 33, 6854–6875. [CrossRef]
30. Verpoorter, C.; Kutser, T.; Seekell, D.A.; Tranvik, L.J. A Global Inventory of Lakes Based on High-Resolution Satellite Imagery.
Geophys. Res. Lett. 2014, 41, 6396–6402. [CrossRef]
31. Verpoorter, C.; Kutser, T.; Tranvik, L. Automated Mapping of Water Bodies Using Landsat Multispectral Data. Limnol. Oceanogr.
Methods 2012, 10, 1037–1050. [CrossRef]
32. Zhang, W.; Pan, H.; Song, C.; Ke, L.; Wang, J.; Ma, R.; Deng, X.; Liu, K.; Zhu, J.; Wu, Q. Identifying Emerging Reservoirs along
Regulated Rivers Using Multi-Source Remote Sensing Observations. Remote Sens. 2018, 11, 25. [CrossRef]
33. Liu, X.; Shi, Z.; Huang, G.; Bo, Y.; Chen, G. Time Series Remote Sensing Data-Based Identification of the Dominant Factor for
Inland Lake Surface Area Change: Anthropogenic Activities or Natural Events. Remote Sens. 2020, 12, 612. [CrossRef]
34. Thomas, R.F.; Kingsford, R.T.; Lu, Y.; Cox, S.J.; Sims, N.C.; Hunter, S.J. Mapping Inundation in the Heterogeneous Floodplain
Wetlands of the Macquarie Marshes, Using Landsat Thematic Mapper. J. Hydrol. 2015, 524, 194–213. [CrossRef]
35. Lehner, B.; Döll, P. Development and Validation of a Global Database of Lakes, Reservoirs and Wetlands. J. Hydrol. 2004, 296,
1–22. [CrossRef]
36. Sheng, Y.; Song, C.; Wang, J.; Lyons, E.A.; Knox, B.R.; Cox, J.S.; Gao, F. Representative Lake Water Extent Mapping at Continental
Scales Using Multi-Temporal Landsat-8 Imagery. Remote Sens. Environ. 2016, 185, 129–141. [CrossRef]
37. Dwyer, J.L.; Roy, D.P.; Sauer, B.; Jenkerson, C.B.; Zhang, H.K.; Lymburner, L. Analysis Ready Data: Enabling Analysis of the
Landsat Archive. Remote Sens. 2018, 10, 1363. [CrossRef]
38. Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine: Planetary-Scale Geospatial
Analysis for Everyone. Remote Sens. Environ. 2017, 202, 18–27. [CrossRef]
39. Dhu, T.; Dunn, B.; Lewis, B.; Lymburner, L.; Mueller, N.; Telfer, E.; Lewis, A.; McIntyre, A.; Minchin, S.; Phillips, C. Digital Earth
Australia—Unlocking New Value from Earth Observation Data. Big Earth Data 2017, 1, 64–74. [CrossRef]
40. Pui, A.; Sharma, A.; Santoso, A.; Westra, S. Impact of the El Niño–Southern Oscillation, Indian Ocean Dipole, and Southern
Annular Mode on Daily to Subdaily Rainfall Characteristics in East Australia. Mon. Weather Rev. 2012, 140, 1665–1682. [CrossRef]
41. Dare, R.A.; Davidson, N.E.; McBride, J.L. Tropical Cyclone Contribution to Rainfall over Australia. Mon. Weather Rev. 2012, 140,
3606–3619. [CrossRef]
42. Pepler, A.S.; Dowdy, A.J.; van Rensch, P.; Rudeva, I.; Catto, J.L.; Hope, P. The Contributions of Fronts, Lows and Thunderstorms
to Southern Australian Rainfall. Clim. Dyn. 2020, 55, 1489–1505. [CrossRef]
43. Han, S.-C. Elastic Deformation of the Australian Continent Induced by Seasonal Water Cycles and the 2010–2011 La Niña
Determined Using GPS and GRACE. Geophys. Res. Lett. 2017, 44, 2763–2772. [CrossRef]
44. Bureau of Meteorology. Water in Australia 2018–19; Bureau of Meteorology: Melbourne, VIC, Australia, 2020; p. 68.
45. Bureau of Meteorology. Climate Classification Maps. Available online: http://www.bom.gov.au/jsp/ncc/climate_averages/
climate-classifications/index.jsp?maptype=kpngrp#maps (accessed on 13 October 2020).
46. Lewis, A.; Oliver, S.; Lymburner, L.; Evans, B.; Wyborn, L.; Mueller, N.; Raevksi, G.; Hooke, J.; Woodcock, R.; Sixsmith, J.; et al.
The Australian Geoscience Data Cube: Foundations and Lessons Learned. Remote Sens. Environ. 2017, 202, 276–292. [CrossRef]
47. United States Geological Survey Landsat Missions. Available online: https://www.usgs.gov/land-resources/nli/landsat
(accessed on 31 March 2021).
Remote Sens. 2021, 13, 1437 28 of 28

48. European Space Agency. Sentinel-2 User Handbook; European Commission: Paris, France, 2015; Volume 1.2.
49. United States Geological Survey. USGS Completes Decommissioning of Landsat 5. Available online: https://www.usgs.gov/
land-resources/nli/landsat/usgs-completes-decommissioning-landsat-5?qt-science_support_page_related_con=4#qt-science_
support_page_related_con (accessed on 31 March 2021).
50. Andrefouet, S.; Bindschadler, R.; Brown de Colstoun, E.; Choate, M.; Chomentowski, W.; Christopherson, J.; Doorn, B.; Hall, D.K.;
Holifield, C.; Howard, S.; et al. Preliminary Assessment of the Value of Landsat 7 ETM+ Data Following Scan Line Corrector Malfunction;
United States Geological Survey: Sioux Falls, SD, USA, 2003.
51. Leith, A. What Is the Open Data Cube? Available online: https://medium.com/opendatacube/what-is-open-data-cube-805af6
0820d7 (accessed on 13 October 2020).
52. Geoscience Australia. Water Observations from Space Statistics 25m 2.1.5; 2017. Available online: http://pid.geoscience.gov.au/
dataset/ga/121074 (accessed on 31 March 2021).
53. Polsby, D.D.; Popper, R. The Third Criterion: Compactness as a Procedural Safeguard against Partisan Gerrymandering. Yale Law
Policy Rev. 1991, 9, 301–353. [CrossRef]
54. Bureau of Meteorology. Australian Hydrological Geospatial Fabric (Geofabric) Product Guide v3.0; Bureau of Meteorology: Melbourne,
Australia, 2015.
55. Bureau of Meteorology. Australian Hydrological Geospatial Fabric (Geofabric) Information Sheet; Bureau of Meteorology: Melbourne,
Australia, 2015.
56. Crossman, S.; Li, O. Surface Hydrology Database Specifications; Geoscience Australia: Canberra, Australia, 2017.
57. The Parliament of the Commonwealth of Australia. Co-Ordinating Catchment Management: Report of the Inquiry into Catchment
Management; The Parliament of the Commonwealth of Australia: Canberra, Australia, 2000.
58. Queensland Department of Natural Resources and Mines. Queensland Coal-Mines and Advanced Projects; Geological Survey of
Queensland: Brisbane, QLD, Australia, 2017.
59. Malerba, M.E.; Wright, N.; Macreadie, P.I. A Continental-Scale Assessment of Density, Size, Distribution and Historical Trends of
Farm Dams Using Deep Learning Convolutional Neural Networks. Remote Sens. 2021, 13, 319. [CrossRef]
60. Botha, E.J.; Anstee, J.M.; Sagar, S.; Lehmann, E.; Medeiros, T.A.G. Classification of Australian Waterbodies across a Wide Range of
Optical Water Types. Remote Sens. 2020, 12, 3018. [CrossRef]
61. Bureau of Meteorology. Australian Hydrological Geospatial Fabric (Geofabric) V3.0.5—Beta; Bureau of Meteorology: Melbourne,
Australia, 2015.
62. Zupanc, A.; Zupanc, M.; Peressutti, D.; Aleksandrov, M.; Lubej, M.; Milcinski, G.; Batic, M.; Burja, A.; Kirac, K. Bluedot
Water Observatory—Cost Effective near Real Time Monitoring of Global Water Resources. In Proceedings of the Living Planet
Symposium, Milan, Italy, 13–17 May 2019.
63. New South Wales Department of Primary Industries. Seasonal Conditions: State Seasonal Update. Available online: https:
//www.dpi.nsw.gov.au/climate-and-emergencies/seasonal-conditions (accessed on 16 September 2020).
64. Dhu, T.; Giuliani, G.; Juárez, J.; Kavvada, A.; Killough, B.; Merodio, P.; Minchin, S.; Ramage, S. National Open Data Cubes and
Their Contribution to Country-Level Development Policies and Practices. Data 2019, 4, 144. [CrossRef]
65. Sagar, S.; Roberts, D.; Bala, B.; Lymburner, L. Extracting the Intertidal Extent and Topography of the Australian Coastline from a
28 year Time Series of Landsat Observations. Remote Sens. Environ. 2017, 195, 153–169. [CrossRef]
66. Geoscience Australia. Intertidal Extents Model 25m v. 2.0.0; Geoscience Australia: Canberra, Australia, 2017.
67. Australian Bureau of Statistics. Statistical Area Level 3 (SA3) ASGS Ed 2016 Digital Boundaries 2016. Available online: https:
//www.abs.gov.au/AUSSTATS/abs@.nsf/DetailsPage/1270.0.55.001July%202016?OpenDocument#Data (accessed on 16 March
2020).
68. Australian Bureau of Statistics. Australian Statistical Geography Standard (ASGS). Available online: https://www.abs.gov.au/
websitedbs/D3310114.nsf/home/Australian+Statistical+Geography+Standard+(ASGS) (accessed on 16 July 2020).
69. Niemeyer, G. Geohash. Available online: https://en.wikipedia.org/wiki/Geohash (accessed on 18 June 2019).