Lakes and reservoirs are crucial elements of the
hydrological and biochemical cycle and are a valuable resource for
hydropower, domestic and industrial water use, and irrigation. Although their
monitoring is crucial in times of increased pressure on water resources by
both climate change and human interventions, publically available datasets of
lake and reservoir levels and volumes are scarce. Within this study, a time
series of variation in lake and reservoir volume between 1984 and 2015 were
analysed for 137 lakes over all continents by combining the JRC Global
Surface Water (GSW) dataset and the satellite altimetry database DAHITI. The
GSW dataset is a highly accurate surface water dataset at 30 m resolution
compromising the whole L1T Landsat 5, 7 and 8 archive, which allowed for
detailed lake area calculations globally over a very long time period using
Google Earth Engine. Therefore, the estimates in water volume fluctuations
using the GSW dataset are expected to improve compared to current techniques
as they are not constrained by complex and computationally intensive
classification procedures. Lake areas and water levels were combined in a
regression to derive the hypsometry relationship (d
Reservoirs and lakes cover a small part of the Earth's land surface (
The amount of water in a reservoir results from the balance of inflow (i.e. direct precipitation, inflowing river discharge, discharge from riparian communities and industries, and subsurface inflow) and outflow (i.e. direct evaporation, withdrawals, reservoir outflow and groundwater percolation) (Duan and Bastiaanssen, 2013). A long-term imbalance can result in considerable reductions in water storage, as frequently observed around the globe in, for example, Lake Mead, Lake Powell, Lake Poopo and the Aral Sea (Barnett and Pierce, 2008; Micklin, 2016). Reduced water availability in the reservoirs may then result in reductions in hydropower energy production and/or irrigation water availability and lead to economic and societal damage. Many studies have already pointed out that population and economic growth, together with climate change and increasing energy and food requirements, will put increasing pressure on water resources (Haddeland et al., 2014; Liu, 2016). A proper understanding of the historical dynamics of reservoirs as a source of water for irrigation, drinking water and energy production, as well as a buffer for flood protection, is also essential to improve the quality of future projections on global water resources.
While for individual river basin studies information on reservoirs may be available, especially for larger scale water resource studies at national, continental and global scale, almost no historical records on reservoirs are readily available to run, calibrate and validate hydrological models (Hanasaki et al., 2006). Moreover, in situ lake level and volume measurements are sparse – especially in developing countries – and have even decreased around the globe during recent years (Duan and Bastiaanssen, 2013). Even if water levels or volumes are monitored, the information is rarely freely available due to strategic political, commercial or national legislation reasons. Therefore, only a few comprehensive global lake and reservoir datasets exist (e.g. Downing et al., 2006; Lehner and Döll, 2004; Meybeck, 1995; Verpoorter et al., 2014) and if they provide a water storage estimation, these estimates are not dynamic or do not provide data over a longer time series. Therefore, remotely sensed data may be a valuable alternative to monitor water volumes in lakes and reservoirs over the last few decades.
Monitoring lakes and reservoirs using remote sensing has gained much
attention over the last few years (e.g. Avisse et al., 2017; Crétaux et
al., 2016; Duan and Bastiaanssen, 2013; Frappart et al., 2006b; Gao et al.,
2012; Smith and Pavelsky, 2009). Most of these publications focussed on
volume variations by combining altimetry water level with lake area from a
multispectral sensor. Landsat or MODIS imagery is commonly used to estimate
water surface areas, by classifying the satellite images capturing the water
body. The classification procedure is demanding and computationally intensive
if large areas or many images are classified, and misclassifications may
occur because of the diversity of spectral signatures emitted by water
surfaces. Therefore, calculating lake areas is often a constraining factor in
lake volume calculations. They are predominantly used for the lake hypsometry
relationship (d
The paper is organized as follows. Section 2 presents the data used in this research, providing a description of the DAHITI altimetry database and an overview of the GSW dataset. Section 3 contains a description of the methods applied, while Sect. 4 gives a description of the results. Section 5 presents a discussion, and finally the conclusions and recommendations are presented in Sect. 6.
Satellite altimetry was initially designed for observing the ocean's surface. But for more than 10 years now, satellite altimetry has proven to be a suitable tool for measuring water heights of lakes and rivers. Numerous studies have already shown the potential of estimating water level time series over inland waters using different altimeter missions such as TOPEX/Poseidon (Birkett, 1995), Envisat (Frappart et al., 2006a), Saral (Schwatke et al., 2015b), Cryosat-2 (Villadsen et al., 2015) or ICESat (Zhang et al., 2011). Water levels from satellite altimetry have also been used for hydrological applications such as the estimation of river discharge (Kouraev et al., 2004; Tourian et al., 2017; Zakharova et al., 2006) and lake volumes (Duan and Bastiaanssen, 2013; Tong et al., 2016; Zhou et al., 2016).
Satellite altimetry has the potential to provide reliable water level time series of globally distributed inland water bodies over the last 20 years. TOPEX/Poseidon and Jason-1/-2/-3 have an identical orbit configuration with a 9.9156-day repeat cycle and a track separation of about 300 km at the Equator. ERS-1/-2, Envisat and SARAL flew on an orbit with a 35-day repeat cycle and a track separation of about 80 km at the Equator. The combination of different altimeter missions is essential to increase the temporal resolution, spatial resolution and length of the water level time series. In order to combine altimeter data from different missions, a mission-dependent range bias resulting from a multi-mission crossover analysis has to be taken into account to achieve long-term homogenous water level time series (Bosch et al., 2014).
The estimation of water level time series for small lakes, reservoirs or rivers is very challenging. Due to coarse mission-dependent ground tracks with a cross-track spacing of a few hundred kilometres, larger lakes and reservoirs have a much higher probability to be crossed by a satellite track than smaller ones. Moreover, small water bodies tend to have a relatively big altimeter footprint compared to their size, which will affect the resulting shape of the returning waveform. The diameter of the footprint is mainly influenced by the water roughness (i.e. surface waves) and surrounding topography. In reality, the diameter of the footprint can therefore vary between 2 km over the ocean and up to 16 km for small lakes with considerable surrounding terrain topography (Fu and Cazenave, 2001). These land influences and surface waves within the altimeter footprint can affect the altimeter waveforms and require an additional retracking to achieve more accurate ranges. In order to achieve accurate results for small water bodies, the conditions have to be ideal, meaning a low surrounding topography, low surface waves, and perpendicular crossings of the altimeter track and water bodies' shores. In these ideal cases, satellite altimetry has the capability to observe rivers with a width of about 100–200 m or lakes with a diameter of a few hundred metres. The off-nadir effect is another problem which can occur when investigating smaller water bodies. In general, satellite altimetry measures in the nadir direction, but if the investigated water body is not located in the centre of the footprint, then the radar pulses are not reflected in the nadir direction, which leads to longer corrupted ranges that must be taken into account (Boergens et al., 2016).
In this paper we use water level time series from the “Database for
Hydrological Time Series over Inland Waters” (DAHITI) as input data for the
volume estimation. DAHITI is an altimetry database launched in 2013 by the
“Deutsches Geodätisches Forschungsinstitut der Technischen
Universität München” (DGFI-TUM). The data are accessible through a
user-friendly web service (
The quality of the water level time series from satellite altimetry in DAHITI has been validated with in situ data and varies depending on the extent of the inland water body and length of the crossing altimeter track. For large lakes with ocean-like conditions (such as the Great Lakes), accurate measurements can potentially be achieved with a root-mean-square error (RMSE) as low as 4–5 cm, while for smaller lakes and rivers the RMSE could increase to several decimetres (Schwatke et al., 2015a). However, no clear relationship was observed between lake size and altimetry accuracy, as the quality of water level time series is not only dependent on the target size, but also on many other factors (e.g. surrounding topography, surface waves, winter ice coverage, the position of altimeter track crossings).
The JRC GSW dataset (Pekel et al., 2016) maps the temporal and spatial dynamics of global surface water over a 32-year period (from 16 March 1984 to 10 October 2015) at 30 m resolution. This dataset was produced by analysing the whole L1T Landsat 5, 7 and 8 archive. At the time of the study, it represented 3 066 080 images (1823 terabytes of data) and covered 99.95 % of the landmass. The analysis was performed thanks to a dedicated expert system classifier. The inference engine of the classifier is a procedural sequential decision tree, which used both the multispectral and multitemporal attributes of the Landsat archive as well as ancillary data layers. It assigned – in a consistent way in both space and time – each pixel to one of three target classes, either water, land or non-valid observations (snow, ice, cloud or sensor-related issues). Classification performance, measured using over 40 000 reference points, revealed the high accuracy of the classifier: less than 1 % of false water detections, and less than 5 % of omission (Pekel et al., 2016). Thanks to its technical characteristics, the GSW dataset constitutes a very valuable long-term surface water record.
The stack of classified images constitutes the long-term water history documenting the “when and where” of the water presence. This information is recorded in the monthly water historical dataset – a set of 380 global-scale maps documenting the water presence for each month of the 32-year archive. This monthly information constitutes the most comprehensive and detailed dataset of the GSW. Eight additional information layers, documenting different facets of the surface water dynamics, are also available within the GSW dataset: (1) water occurrence, (2) occurrence change intensity, (3) seasonality, (4) recurrence, (5) transitions, (6) maximum water extent, (7) monthly recurrence and (8) yearly history. In the framework of this study, the monthly water history and maximum water extent (MWE) – a map documenting places where water has been detected at least once over the 32 years were used.
The GSW dataset was completely developed using Google Earth Engine and all of the layers are available through the Earth Engine catalog (Gorelick et al., 2017). Moreover, Earth Engine is used in this research to calculate the monthly lake area time series. Earth Engine is a cloud-based global-scale platform optimized for parallel geospatial analyses and data management in Earth sciences, using Google's computational power (Gorelick et al., 2017). Earth Engine allowed the analysis of lakes at global scale in high detail, while maintaining a high resolution of 30 m.
Geographical distribution of the analysed lakes.
Monthly time series of lake areas have been calculated for 137 lakes over all continents (Fig. 1). These contained nearly all lakes available in the DAHITI altimetry database at the time of processing. No additional criteria were set for this study, as the GSW dataset covers all lakes globally. For 380 months over the period 1984–2015 lake areas were calculated using a dedicated Google Earth Engine script. For each lake, a region of interest (ROI) was set by a manually drawn polygon that was approximately equal to the MWE of the lake (Fig. 2). For every month, lake areas were calculated directly from the GSW monthly historical dataset, by counting the number of water pixels inside the polygon and multiplying this by the pixel area. To improve the accuracy of the area calculations, the amount of non-valid observations (no-data pixels) within the MWE, compared to all MWE pixels within the ROI, has been expressed as the no-data fraction. This no-data fraction has been used to filter accurate and less accurate area observations in the regression analysis and volume calculations (see Sect. 3.2). The white striping observed for Lake Mead in Fig. 2 is an example of a lack of data that is caused by Landsat sensor issues.
An example of the area input data for Lake Mead (US) for February 2015, where the maximum water extent is marked as red, water as blue and a lack of data as white pixels.
Volume variation calculation for Lake Eucumbene (Australia). The
observed monthly pairs of
The volume of a lake or reservoir is a function of the water area (
A 95 % prediction interval (PI) and confidence interval (CI) have been
calculated around the linear regression line using Eqs. (S13) and (S14), taking into account the slope and intercept parameter uncertainty
and the standard deviation of the residuals. The PI around the linear
regression has been propagated to a PI around the estimated volume variations
of
Not all lakes showed considerable area fluctuation, as some lakes and reservoirs are artificially bounded or have very steep banks. The coefficient of variation (CV) has been calculated from the area observations to express lake area variation normalized by mean lake size. The signal-to-noise ratio (SNR) for lakes with a very small CV is likely to be low as errors due to a lack of data, misclassifications and the lake border discretization with 30 m pixels will mask the actual area variations. A linear regression between variable lake levels and nearly constant lake areas would thus not be feasible. Therefore, the areas of lakes with a very small CV were interpreted to be constant. Lake volumes were still calculated, but only by multiplying the mean area with water level variations as observed by altimetry.
The validation has been carried out using the Pearson correlation
coefficient
A total of 137 lakes and reservoirs have been analysed over all continents. The linear
OLS regression analysis resulted in highly variable
Area–level regressions for Lake Powell
Based on the
Illustration of the division of the lakes into lakes with a
constant area (Lc) and a variable area with a good performance (LvG) and with
a poor performance (LvP), based on the
For a total of 100 lakes (58 lakes with variable area and 42 lakes with constant area),
the volume variation time series have been calculated, using both water
levels (
Lake volume variations for Lake Powell
Figure 7 summarizes the volumetric results, by showing lake location, lake
type (Lc, LvG or LvP) and the average volumetric change. The volumetric change
shows the magnitude of reduction (red circles) or increase (blue circles) in
average water storage between the periods 1984–2000 and 2000–2015 for LvG,
and from 2000–2008 to 2008–2016 for Lc. Slightly more lakes showed a
positive change (60) than a negative one (40). Considerable reductions of
water storage were observed in the western US, due to major average volume
declines in Lake Mead (11 km
Lake and reservoir types (constant area (Lc), variable area with good (LvG) and poor (LvP) regression performance) with the average volume changes.
Lake Mead was formed after the construction of the Hoover Dam during the
1930s, in the former steep V-shaped slopes created by the Colorado River. It
is located approximately 50 km east of Las Vegas in the Black Canyon,
Arizona–Nevada (Fig. 7). With a maximum depth of 158 m and a maximum
capacity of 33–35 km
Lake Nasser is a crucial resource for Egypt's population, functioning as a
source for irrigational water and electricity and as an important
flood-control mechanism. With an estimated maximum storage capacity of
162 km
Lake Kariba formed after the construction of the Kariba Dam on the Zambezi
River on the border of Zimbabwe and Zambia (Berg et al., 1996). It has an
average surface area of 5364 km
With an area of 653 km
Lake volume variations have been validated against in situ data that are
based on a full bathymetric survey for 18 lakes. Nine of these lakes are
located in the USA (Richland Chambers Reservoir, Hubbard Creek Reservoir,
Lake Mead, Lake Houston, Lake Powell, O. H. Ivie Lake, Toledo Bend Reservoir,
Lake Walker and Lake Berryessa), one in Africa (Roseires Reservoir), six in
Spain (Serena Reservoir, Puente Nuevo Reservoir, Alcantara Reservoir, Lake
Almanor, Yesa Reservoir and Encoro de Salas Reservoir) and two in Australia
(Lake Argyle and Lake Eucumbene). US lakes have been validated using USGS in
situ lake/reservoir volumes obtained from
The validation analysis has been done for volumes both excluding and
including extrapolation. The non-extrapolated volumes showed average Pearson
Overview of the validation results, excluding and including extrapolated volumes.
Figure 8 shows the relationship between satellite and in situ volume
variations for Lake Mead and illustrates the accuracy of the methodology. The
estimates showed strong linearity with in situ data, as shown by the
correlation coefficient of > 0.99 for both
Relationship in volume variation between satellite-estimated and in
situ observations for Lake Mead, with a Pearson
Figures 9 and 10 show the validation volume time series against the satellite-estimated volumes for Lake Mead and Lake Powell respectively. The time series
showed that both the timing and magnitude of the estimated volume
fluctuations were accurate for these lakes. For the non-extrapolated part of
Lake Mead (2002–2016), estimated volumes were almost equal to the validation
data. The extrapolated part (before 2002) was slightly overestimated (see
Discussion). However, the dynamics over this period were still well captured.
Lake Powell also showed accurate results, for both the seasonal and
inter-annual fluctuations in water storage. Noteworthy for both lakes is the
density of area observations due to a low no-data percentage in the GSW
dataset. For Lake Powell, the
Validation time series plotted with estimated reservoir volumes for Lake Mead. The black triangle line represents the validation storage as measured using the full lake bathymetry.
Validation time series plotted with estimated reservoir volumes for Lake Powell. The black triangle line represents the validation storage as measured using the full lake bathymetry.
This study presented a new methodology to estimate lake and reservoir volumes using remote sensing alone. The validation showed that the method can produce water storage change estimates for many lakes, thus highlighting the potential of combining satellite altimetry and the GSW dataset to develop a global lake and reservoir volume variation dataset. The GSW dataset global coverage, 30 m resolution, high accuracy and monthly surface water observations over a 32-year period increases the number of analysed lakes and the accuracy, quantity and temporal range of lake area calculations. Therefore, volume variations can now also be calculated using GSW lake areas as input independent from altimetry data, which allows for volume calculations further back in time to 1984.
The lake and reservoir volume dataset developed here will help to better understand the behaviour and operations of lakes and reservoirs. As the number of reservoirs is still increasing because of growing energy demands, it is crucial to include their effects in (continental- and global-scale) hydrological models. Zajac et al. (2017) found that the exclusion of lakes and reservoirs often leads to inaccurate downstream discharge estimates. Furthermore, lake or reservoir storage change combined with modelled or observed inflow allows for a better estimation of the outflow (e.g. Muala et al., 2014). These outflow estimates can be used to calibrate hydrological models or estimate hydropower production in areas where in situ observations are lacking. However, due to a lack of storage observations and their availability – often because of commercial reasons – the parametrization and the representation of lakes and reservoirs in many hydrological models – if at all present – is still highly simplified. Our global lake and reservoir volume dataset over 32 years will be very beneficial to calibrate and validate their parameterization to mimic their operational behaviour. This will improve our current understanding of lakes and reservoirs, improve their simulations and consequently the simulations in the rest of the river basin. In addition, a better understanding of reservoirs will also likely improve water and energy production projections of the influence of these reservoirs under climate change, or under different management scenarios (e.g. changing downstream water requirements, flow legislation, changing inflow due to other activities upstream). Moreover, the area time series developed in this study can be included in models to improve on (often fixed) current area estimates and can furthermore improve estimates of open-water evaporation.
The methodology used in this study has a couple of limitations. They arise mainly from limitations in the input data (GSW and DAHITI altimetry dataset) and the volume calculation methodology.
The altimeter footprint can be up to 16 m over land and land influences
inside this footprint can alter the water level accuracy by disturbing the
altimeter waveforms (Fu and Cazenave, 2001). Water level estimations of very
small lakes or reservoirs can therefore have a RMSE of several decimetres or
larger (Schwatke et al., 2015a). We analysed many small lakes in Europe that
did show regressions with large residuals (e.g. Thülsfelder Talsperre,
1.7 km
The classifier of the GSW dataset has been shown to be very accurate, with less than 1 % of false water detections, and less than 5 % omission (Pekel et al., 2016). Therefore, the influence of classification errors in the volume estimations was very limited.
In this methodology, no-data classifications in the GSW dataset play a much
more important role. These are caused by snow, ice, cloud or sensor-related
issues (e.g. white striping in Fig. 2) and are likely to give an
underestimation of the actual lake area if they are inside the MWE.
Therefore, their influence has been reduced with a strict no-data thresholds
applied to each monthly calculation. The 1 % no-data threshold for the
regression and 5 % for the volume calculation resulted in the best
trade-off between the number of observations from the GSW dataset and the
accuracy of the estimates. Higher no-data thresholds introduced too much
variability in the volume estimates, as (1) regression lines became much
noisier and (2) the area of a monthly lack of data exceeded the actual lake area
variation. Locations with frequent cloud cover, lake ice coverage or
sensor-related image failures will often return no-data amounts that exceed
the threshold, resulting in a sparse
Time series of satellite-estimated volume variations compared to validation storages for Puente Nuevo Reservoir.
Besides the varying image quality, the whole stack of historical Landsat
imagery has an unequal global spatial distribution. The US and Australia are
well covered, while other regions, like Africa and southwestern Europe, have
much fewer Landsat observations (Pekel et al., 2016; Wulder et al., 2016). The
volumetric plots of US lakes therefore typically show a highly dense
Many layers in the GSW dataset could be used to reduce the amount of no-data pixels. This would considerably increase the capabilities of the methodology, especially in the regions mentioned above. No-data pixels that are located in the middle of the lake and are surrounded by water pixels (e.g. Fig. 2) have a high probability of being water. Furthermore, the large temporal range (1984–2015) of the GSW dataset could be used to further decrease the no-data percentage. For example, no-data pixels that are classified as water over nearly all the 380 months could be assigned to the water class with high confidence. The GSW seasonality layer could also be used to find permanent water pixels, which can be converted to water if they are not observed during a specific month.
Only 4 out of 137 lakes (Tawakoni, Urmia, Tsimlyansk and Eagle) showed a clear non-linear area–level relation. For these lakes, volume variations were not estimated. Their regressions could be explained by a second- or third-order polynomial and a hyperbolic sine function, as shown for Lake Urmia in Fig. 12. This non-linearity is caused by a considerable change in slope, which will mostly be observed for lakes with extremely low water levels (e.g. Lake Urmia) or during floods.
Relationship between
To reduce data size, the monthly GSW dataset does not include exact dates of the Landsat observations. This causes more uncertainty in the regression, as the level and area observations may refer to different dates (maximum difference of 1 month) and therefore to slightly different lake conditions. For lakes with a highly variable level within a month, this uncertainty therefore increases. The outliers in the regression of Lake Nasser (Fig. 4d) are expected to be largely induced by this uncertainty. For both outliers, the altimeter measurements were taken in the beginning of the month (2nd and 6th day), and the water level changed considerably by the next month. The Landsat observation therefore likely measured different lake conditions than the satellite altimeter.
For 37 lakes with lower performance, the regression showed relatively low
The calculated extrapolated volumes are more uncertain than the
non-extrapolated ones, depending on how much the hypsometry relationship
changes outside the regression range. In general, if the altimetry
measurement period is short compared to the area time series, the range of
level and area values captured by the regression line is likely to be small
and the volume extrapolation using extreme area observations is expected to
be less accurate. This was observed for the Roseires Reservoir and Lake Mead.
The validation results for Lake Mead (Figs. 8 and 9) indicate that for large
areas observed outside the regression range (1984–2002) the found linear
hypsometry relationship produces slightly biased volume variation estimates.
The extrapolated
This study successfully combined the JRC Global Surface Water (GSW) dataset and the DAHITI satellite altimetry dataset to estimate lake and reservoir volume fluctuations over all continents. The GSW dataset records surface water over a 32-year period, containing 3 066 080 monthly images that cover 99.95 % of the landmass. The extensive size and high accuracy of this surface water dataset allowed for detailed volume variation estimations over a very long time period (1984–2015), without being constrained by complex and computationally intensive classification procedures.
Lake areas from the GSW dataset and water levels from the DAHITI altimetry
dataset have been combined in a regression to explain the lake hypsometry of
137 lakes globally. Nearly all lakes showed a linear regression. A total of 58 of these
lakes returned relatively low residuals with
For 100 lakes (58 with variable area and 42 with nearly constant area) volume
variations were calculated by integrating the hypsometric relationships,
using both area and water level observations separately. Decreases in water
storage were found in the western US, where Lake Mead, Lake Powell and the Great Salt
Lake lost 11, 6 and 16 km
The low number of adequate Landsat lake area observations for some regions like Africa and southwestern Europe still remains a limitation. Therefore, it would be highly beneficial for the purpose of this research to include surface water data from other satellites in the GSW dataset and to develop techniques to decrease the no-data percentage in the current dataset. Future plans are to include Sentinel-1 and Sentinel-2 in the GSW dataset. The DAHITI database is continuously growing by analysing new water bodies, and newly available altimetry data will be processed to expand the volumetric dataset.
This lake and reservoir volume dataset will help to improve our current understanding of the behaviour of lakes and reservoirs, their representation in hydrological models and consequently the simulations of the river basin. This will, moreover, improve projections of the river basin under climate change or under different management scenarios and improve hydropower and open-water evaporation estimations.
This study constitutes a proof of concept paving the way for increasing the
number of lakes and reservoirs analysed, which could potentially be included
as an a priori water storage dataset for the Surface Water and Ocean
Topography (SWOT) hydrology and oceanography satellite mission. Launched in
2020, this mission will combine water body contours and accurate water level
estimations to estimate storage changes in lakes and reservoirs with an
average accuracy of 20 cm (Biancamaria et al., 2016; Crétaux et al.,
2015). The SWOT satellite will be unique due to its accuracy and capabilities
on smaller water bodies with a size of at least 250 m
Research outcomes (i.e. lake and reservoir volumes) are not
publicly accessible. However, they may become publicly available in a later
stage through the DAHITI web service
(
Overview of lake properties for the three categories (Lc, LvP and LvG).
Continued.
Continued.
The supplement related to this article is available online at:
TB developed the methodology, the scripts and was the main writer of the paper. AdR helped writing the introduction section and gave comprehensive scientific support over the whole period. EG gave outstanding and extensive support in coding the scripts, developing the methodology and the uncertainty analysis during the feedback process. CS provided DAHITI altimetry data, additionally processed water bodies on demand and, moreover, wrote the altimetry section. JFP and AC contributed in the Google Earth Engine usage and wrote the GSW section. All authors actively contributed to the feedback process after the first draft.
The authors declare that they have no conflict of interest.
We would like to thank the Google Earth Engine team, and especially Matthew Hancher and Noel Gorelick, for support in developing the Earth Engine scripts. We would also like to express our gratitude to Guido Lemoine (Joint Research Centre) for his technical support in Earth Engine and Steven de Jong (Utrecht University) for his feedback on the first report. Edited by: Anas Ghadouani Reviewed by: Renata Romanowicz and two anonymous referees