VANDERHOOF

Remote Sensing of Environment 288 (2023) 113498
Contents lists available at ScienceDirect
Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse
High-frequency time series comparison of Sentinel-1 and Sentinel-2

satellites for mapping open and vegetated water across the United
States (2017–2021)
Melanie K. Vanderhoof a, *, Laurie Alexander b, Jay Christensen c, Kylen Solvik d,
Peter Nieuwlandt a, Mallory Sagehorn a
a
U.S. Geological Survey, Geoscience and Environmental Change Science Center, PO Box 25046, MS 980, Denver Federal Center, Denver, CO 80225, USA
b
Office of Research and Development, U.S. Environmental Protection Agency, 1200 Pennsylvania Avenue, Washington, DC 20460, USA
c
Office of Research and Development, U.S. Environmental Protection Agency, 26 W. Martin Luther King Dr., Cincinnati, OH 45268, USA
d
Department of Geography, Guggenheim 110, 260 University of Colorado, Boulder, CO 80309-0260, USA
A R T I C L E I N F O A B S T R A C T
Edited by Dr. Menghua Wang Frequent observations of surface water at fine spatial scales will provide critical data to support the management
of aquatic habitat, flood risk and water quality. Sentinel-1 and Sentinel-2 satellites can provide such observa
Keywords: tions, but algorithms are still needed that perform well across diverse climate and vegetation conditions. We
Climate developed surface inundation algorithms for Sentinel-1 and Sentinel-2, respectively, at 12 sites across the
Google Earth Engine
conterminous United States (CONUS), covering a total of >536,000 km2 and representing diverse hydrologic and
Inundation
vegetation landscapes. Each scene in the 5-year (2017–2021) time series was classified into open water, vege
Lakes
python tated water, and non-water at 20 m resolution using variables from Sentinel-1 and Sentinel-2, as well as variables
Streams derived from topographic and weather datasets. The Sentinel-1 algorithm was developed distinct from the
Surface water Sentinel-2 model to explore if and where the two time series could potentially be integrated into a single high-
Time series frequency time series. Within each model, open water and vegetated water (vegetated palustrine, lacustrine, and
Wetlands riverine wetlands) classes were mapped. The models were validated using imagery from WorldView and Plan
etScope. Classification accuracy for open water was high across the 5-year period, with an omission and com
mission error of only 3.1% and 0.9% for the Sentinel-1 algorithm and 3.1% and 0.5% for the Sentinel-2
algorithm, respectively. Vegetated water accuracy was lower, as expected given that the class represents mixed
pixels. The Sentinel-2 algorithm showed higher accuracy (10.7% omission and 7.9% commission error) relative
to the Sentinel-1 algorithm (28.4% omission and 16.0% commission error). Patterns over time in the proportion
of area mapped as open or vegetated water by the Sentinel-1 and Sentinel-2 algorithms were charted and
correlated for a subset of all 12 sites. Our results showed that the Sentinel-1 and Sentinel-2 algorithm open water
time series can be integrated at all 12 sites to improve the temporal resolution, but sensor-specific differences,
such as sensitivity to vegetation structure versus pixel color, complicate the data integration for mixed-pixel,
vegetated water. The methods developed here provide inundation at 5-day (Sentinel-2 algorithm) and 12-day
(Sentinel-1 algorithm) time steps to improve our understanding of the short- and long-term response of sur
face water to climate and land use drivers in different ecoregions.
1. Introduction following snowmelt or large precipitation events, or show change

concentrated at the edges as waterbodies expand and contract over days,
Surface water is a highly dynamic resource that responds to land use weeks, months, or years (Donchyts et al., 2016; Parra et al., 2021).
changes and climate variability ranging from drought to flood condi Further, the amount of surface water storage varies greatly across
tions (Prigent et al., 2012; Yamazaki and Trigg, 2016). Change in surface different ecoregions (Poff et al., 2006; Lane and D’Amico, 2016) and the
water extent can occur when areas temporarily store surface water response of that surface storage to changes in precipitation and
* Corresponding author.
E-mail address: [email protected] (M.K. Vanderhoof).
https://doi.org/10.1016/j.rse.2023.113498
Received 9 August 2022; Received in revised form 2 February 2023; Accepted 6 February 2023
Available online 15 February 2023
0034-4257/Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
M.K. Vanderhoof et al. Remote Sensing of Environment 288 (2023) 113498
snowmelt can also be highly variable (Vanderhoof et al., 2018). Because of the challenge of distinguishing vegetated water, some efforts
Continual changes in surface water extent necessitate reliable, accurate, have only sought to distinguish water versus non-water (Hird et al.,
and high spatial and temporal resolution surface water products. Cloud- 2017; Yang et al., 2020), while others have identified multiple water
based platforms, such as Google Earth Engine (Gorelick et al., 2017), classes, such as distinguishing vegetated water from open water
have become a common way to process large amounts of Earth Obser (Lefebvre et al., 2019; Slagter et al., 2020).
vation (EO) imagery to estimate variables of interest, including global Additional variables, such as topographic metrics (e.g., topographic
surface water extent from Landsat at 30 m spatial resolution (Donchyts wetness index, relief, slope, and curvature) have often been found to
et al., 2016; Pekel et al., 2016; Pickens et al., 2020). Dense time series of improve inundation models (Huang et al., 2018; Ludwig et al., 2019).
inundation are essential to support diverse hydrological applications Because emergent vegetation color and structure can be easily confused
(Soulard et al., 2022; Tulbure et al., 2022). Multi-sensor integration with upland vegetation types, incorporating topographic conditions can
efforts can potentially support higher-frequency observations at finer help model where water is likely to collect on the landscape (Mahdavi
spatial resolutions which would provide improved data for management et al., 2018) and where it is less likely to occur such as on steep slopes
applications such as informing questions about aquatic habitat avail (Jones, 2019). Hird et al. (2017), for instance, found topographic met
ability, flood risk, and water quality. rics more important than Sentinel-1 or Sentinel-2 variables to indicate
The launch of Sentinel-1A and 1B (C-band synthetic aperture radar wetland occurrence. In tracking patterns of inundation over time,
(SAR)) in 2014 and 2016, respectively, and Sentinel-2A and 2B (multi weather data can potentially help model when, in a time series, pixels
spectral) in 2015 and 2017, respectively, has greatly improved oppor are more likely to be inundated. Although weather variables have not
tunities for high-frequency observations of inundation dynamics, typically been incorporated into inundation algorithms, coarse-
including enhanced detection of floods, and improved mapping of resolution (4 km) weather variables have previously been successfully
smaller wetlands (<1 ha) and streams. Sentinel-2, for example, has been integrated into Sentinel-2 models to map burned area in wetlands across
used to map surface water extent in small, wetland-dominated sites the southeastern U.S. without sacrificing the mapped spatial resolution
(Lefebvre et al., 2019; Ludwig et al., 2019), flooded calibration/vali (Vanderhoof et al., 2021). Further, time series of inundation extent have
dation sites across Europe (Goffi et al., 2020), and across urban, agri been found to be positively correlated with precipitation and negatively
cultural, and mountainous land cover types in France (Yang et al., correlated with temperature and evapotranspiration (Mishra and Cher
2020). Sentinel-1, meanwhile, has been used to map flood extents across kauer, 2011; Song et al., 2018; Tulbure and Broich, 2019; Xia et al.,
the conterminous U.S., where potential images to process were identi 2019), suggesting that weather data could potentially improve the
fied using stream stage observations and precipitation estimates (Yang detection of flood events and seasonal changes in water extent.
et al., 2021), and to map isolated flood events in urban environments Algorithms derived from Sentinel-1 and Sentinel-2 that can effec
(DeVries et al., 2020). Several efforts have documented improved ac tively track surface water extent across different climate and vegetation
curacy when incorporating Sentinel-2 spectral indices with Sentinel-1 conditions are still needed. Yang et al. (2021) processed Sentinel-1 im
backscatter, for example in mapping wetlands in southwestern Spain agery across the conterminous United States (CONUS), but only when a
(Manakos et al., 2020), central Ireland (Bioresita et al., 2019), and South potential flood event was predicted. Schlaffer et al. (2022) and Huang
Africa (Slagter et al., 2020). Other studies have shown that Sentinel-2 et al. (2018) both developed sub-regional approaches to map inundation
outperformed Sentinel-1 when mapping flood extents at sites across from Sentinel-1, while efforts to map inundation from Sentinel-2 have
the globe (Konapala et al., 2021) as well as in mapping wetland occur been limited in scope within CONUS (Page et al., 2019; Sahour et al.,
rence in densely forested northern Alberta (Hird et al., 2017). While 2022). Sentinel-1 and Sentinel-2 algorithms are needed that (1) can
these prior efforts demonstrated the potential usefulness of both sensors track not only open water but also vegetated water, which can be
for water applications, most were too limited in spatial extent to allow difficult to distinguish from non-inundated vegetation, (2) can be
for the scaling of efforts or to understand how the accuracy of a water applied across a range of climate conditions, and (3) are not developed
classification algorithm might change across different ecosystems, cli for single-site applications, but instead can be applied broadly
mates, and hydrological systems. geographically. To address these needs, we developed surface inunda
The ease with which inundation can be mapped will depend on the tion algorithms for Sentinel-1 and Sentinel-2 for 12 sites across CONUS
land cover type, waterbody size, and vegetation structure (Lechner over a 5-year period (2017–2021). The sites represent a wide range of
et al., 2009; Vanderhoof and Lane, 2019). While both Sentinel-1 and vegetation (forest, grassland, agriculture), topography (coasts, plains,
Sentinel-2 have been shown to effectively map deep, open water (Bio mountains), and aridity (humid to arid). As national and global surface
resita et al., 2019; Ludwig et al., 2019), the presence of shallow water as water products are currently mapped from Landsat imagery (Donchyts
well as vegetation, both emergent and forest, decreases the accuracy et al., 2016; Pekel et al., 2016; Jones, 2019), our initial training points
with which inundation can be mapped (Goffi et al., 2020; Slagter et al., were sampled using a Landsat water algorithm before being edited with
2020; Konapala et al., 2021). Shallow and turbid flood water are more Sentinel-1 and Sentinel-2 imagery. Each scene in the time series was
commonly mapped with Sentinel-1, as radar can penetrate cloud cover classified into open water, vegetated water, non-water (e.g., non-
to map peak flood conditions (Shen et al., 2019; Yang et al., 2021), but inundated vegetation and agricultural fields, bare soil, urban areas),
can also be effectively mapped with Sentinel-2 under cloud-free condi and for Sentinel-2, no-data (e.g., clouds, cloud shadows, snow). Model
tions (Konapala et al., 2021). In areas with vegetated water, for example variables included bands and indices from Sentinel-1 and Sentinel-2, but
in riparian corridors, wetlands with emergent vegetation, and forested also topography and weather datasets to help model spatial and tem
wetlands, inundation can still be detected with Sentinel-2 (Lefebvre poral variability in inundation extent. The Sentinel-1 algorithm (12-day
et al., 2019) or Sentinel-1 and Sentinel-2 together (Bioresita et al., 2019; return interval) was applied distinct from the Sentinel-2 algorithm (5-
Slagter et al., 2020). However, rates of omission error with Sentinel-2 day return interval). Where model performance is comparable, the po
have also been shown to be high in densely vegetated wetlands, low- tential exists to integrate time series derived from each sensor. There
lying forests, and peatlands (Ludwig et al., 2019). Furthermore, across fore, we also explored if and where the algorithms could be used
France, Yang et al. (2020) found Sentinel-2 showed confusion between together to maximize time series density, despite differences in both
wet soil in agricultural fields and wetlands. While a C-band SAR satellite data type and data limitations (Mahdavi et al., 2018), and where the
like Sentinel-1 can penetrate some vegetation better than multispectral algorithms are more effectively used in complement to characterize
sensors (Shen et al., 2019), an L-band SAR is typically necessary to inundation duration and seasonality.
reliably detect water under dense leaf cover (Shen et al., 2019).
Consequently, in densely vegetated boreal forest and peatlands,
Sentinel-1 may not necessarily outperform Sentinel-2 (Hird et al., 2017).
2
2. Methods band Spectral Relationship Visible (MBSRV = (Green + Red) – (NIR +

SWIR1) (Zhang et al., 2011), and the (4) Automated Water Extraction
2.1. Study areas Shadow (AWEsh = Blue +1.5*Green - 1.5*(NIR + SWIR1) –
0.25*SWIR2) (Feyisa et al., 2014) (Jones, 2015). The thresholds were
Inundation extent was mapped across 12 sites within the CONUS organized into six tests which were applied to each pixel (Vanderhoof
(Fig. 1) that represented different land use and land cover types, wetland et al., 2020). In this algorithm?, pixels that pass four or more of the tests
and river densities, and climate regimes. State names were used for ease were classified as high-confidence inundation, while pixels that passed
of referencing sites but were not intended to represent political bound at least two of the tests were classified as moderate- confidence water, or
aries. Annual precipitation averaged 866 mm year− 1 across the sites and one of 3 partial water classes: conservative and aggressive as defined by
ranged from 405 mm year− 1 in Arizona to 1388 mm year− 1 in Arkansas- Jones (2019) and forested, as defined by Vanderhoof et al. (2020) to
Louisiana (Table 1). Maximum monthly vapor pressure deficit averaged target forested wetlands. The partial water classes have been shown to
25.2 hPa and ranged from 19.0 hPa in North Dakota to 48.3 hPa in specifically map pixels with a mix of water and vegetation (Jones, 2019;
Arizona (PRISM Climate Group, 2022). Topography was also variable Taylor et al., 2022). Confusion between urban and water in DSWE's
across the sites with the average slope ranging from 2.0% in Minnesota lower confidence classes was previously identified (Vanderhoof et al.,
to 22.4% in Montana (Table 1). The percentage of area classified as non- 2020), therefore pixels mapped as any of the partial water classes that
riverine wetland by the National Wetland Inventory (NWI) dataset were also mapped as developed land (e.g., residential, commercial, in
averaged 14.1% and ranged from 1.1% in Arizona to 40.5% in Minne dustrial, transportation development and urban cores), as defined by the
sota (US Fish and Wildlife Service (USFWS), 2019), while stream density 2019 National Land Cover Database's (NLCD) low, medium, and high-
averaged 1340 m km− 2 and ranged from 462 m km− 2 in Minnesota to density development classes, were reclassified to non-water (Homer
1939 m km− 2 in Arizona (U.S. Geological Survey, 2019; Table 1). The et al., 2020).
sites represent different land cover types and surface water arrange Within each of the 12 study sites same-date, or near-date Sentinel-1,
ments. Maryland-Delaware, South Carolina, and Texas represent coastal Sentinel-2, and Landsat 8 - DSWE images were identified. Eight image
sites with a high-density of tidal emergent and forested wetlands. The trios per site were selected representing different seasons and multiple
Texas and Maryland-Delaware sites are also characterized by large years within the 2017–2021 period. The average date gap between
urban centers. The Arkansas-Louisiana site represents a mix of agricul Landsat-8 and Sentinel-1 and Landsat-8 and Sentinel-2 was 1.8 days and
ture and evergreen forest. It contains the Mississippi River floodplain 2.2 days, respectively (93% of date gaps were 3 days or less with a
and receives the most amount of precipitation out of the 12 sites. The maximum date gap of 36 days). Within the overlapping area across the
Kansas-Missouri and Iowa sites are both dominated by agriculture, with three images, 150 training points were selected per image: 50 points
river systems, notably the Missouri River and the Mississippi River, from the DSWE upland category to represent non-water; 50 points from
respectively, running through the sites. The North and South Dakota the DSWE high-confidence water category to represent open water; and
sites occur within the Prairie Pothole Region which is a wetland-dense to represent vegetated water, 50 points from the DSWE moderate-
area with agriculture as the primary land cover. The Minnesota site is confidence and the three partial water classes. Examples of the map
one of the colder sites and is dominated by lakes and forested wetlands. ped DSWE classes within vegetated wetlands are shown in Fig. A1. The
The northern California and southwestern Montana sites contain ever sampling resulted in 1200 training points per site for a total of 14,400
green forested mountains in the higher elevations with rivers, lakes, and training points across the 12 sites. The image and training point sample
shrub/scrub and grasslands dominating the lower elevations. And lastly, sizes were selected to balance between the labor-intensive task of
the site in southeastern Arizona is dominated by shrub/scrub and rep generating high-quality training data and the necessity for training data
resents the arid southwest (Fig. 1). to adequately represent variability across landscapes and time.
The variability in land cover, climate, topography, and wetland and To account for error in the DSWE water extent, differences in spatial
river density between sites is shown in Table 1 and was intentional to resolution between the Landsat and Sentinel satellites, and inaccuracies
test if the Sentinel-1 and Sentinel-2-based water algorithms can be potentially introduced by the date gap, all points were visually reviewed
effectively developed and applied across different vegetation types and against the natural color Sentinel-2 images and VV and VH Sentinel-1
climate regimes, as well as flat and mountainous landscapes. In addition backscatter images. The Sentinel-1 images were relied on in particular
to spatial variability, we also wanted to test if the algorithm could to edit training points where snow cover was present and in sites with
respond to temporal variability in wetness conditions. To this end, we forested wetlands. Draft training points were edited to remove mis
calculated the minimum and maximum Palmer Drought Severity Index classified points. Training points were further edited to ensure that (1)
over the 5-year period for each site. All 12 sites showed a range of non-water conditions that could be confused with water conditions (e.g.,
conditions from dry to wet over the time series, with the range averaging shadowed vegetation, topographic shadows, impervious surfaces) were
from − 3 (dry) to 4.7 (wet) (Table 1). adequately represented, (2) small (< 2 ha) open water bodies and rivers
were adequately represented, and (3) vegetated water points were
2.2. Training points visibly inundated and limited to identifiable wetland and river features.
High-resolution base imagery delivered by ArcMap and the NWI dataset
To generate initial, draft training points, Landsat 8 images were (US Fish and Wildlife Service (USFWS), 2019) were used where addi
classified using the Dynamic Surface Water Extent (DSWE; Jones, 2019) tional information on vegetation type and water regime were needed.
algorithm, modified for Landsat 8 (Vanderhoof et al., 2020) in Google Imagery used to generate the training data is shown in Table A1.
Earth Engine. The DSWE algorithm was selected because it was designed
specifically to improve the detection of vegetated water (Jones, 2015; 2.3. Sentinel-1 preprocessing
Jones, 2019). An additional advantage of the DSWE algorithm is that it is
essentially an unsupervised classifier, meaning it does not require We processed all Sentinel-1 imagery that intersected the study area
training data and can therefore be rapidly implemented. The DSWE al extent between January 2017 and December 2021 (10,660 images).
gorithm set a series of thresholds for the blue, green, near infrared (NIR), Sentinel-1 imagery was limited to images collected in interferometric
shortwave infrared one (SWIR1), and shortwave infrared two (SWIR2) wide (IW), ascending swath mode, which acquires data with a 250 km
bands, as well as for spectral indices including: (1) the modified swath, resolution of 20 × 22 m, and pixel spacing of 10 × 10 m. Imagery
Normalized Difference Wetness Index (mNDWI = (Green – SWIR1)/ collected in this mode delivers VV and VH backscatter. Image return
(Green + SWIR1) (Xu, 2006), (2) Normalized Difference Vegetation interval averaged 12 days over the five-year period. The data were pre-
Index (NDVI = (NIR – Red)/(NIR + Red) (Tucker, 1979), (3) the Multi- processed using the European Space Agency's Sentinel Application
3
Fig. 1. The distribution of National Wetland Inventory (NWI) wetlands within each site (in black) and the subset area (in red) where a monthly time series was
generated for, (a) California, (b) North Dakota, (c) Minnesota, (d) Maryland-Delaware, (e) Montana, (f) South Dakota, (g) Iowa, (h) Arkansas-Louisiana, (i) Arizona,
(j) Kanas-Missouri, (k) Texas, (l) South Carolina, and (m) The land cover types within each of the sites, as mapped by the 2019 National Land Cover Database (Homer
et al., 2020), and the distribution of validation images. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of
this article.)
4
Table 1
Number of images processed per site as well as site characteristics including wetland area and stream density, annual precipitation, minimum (min) and maximum
(max) Palmer Drought Severity Index (PDSI) over the 5-year period, average slope, and dominant land cover classes. Wetland density was calculated from the National
Wetland Inventory dataset, excluding ocean and estuarine deep-water habitat, and stream density calculated from the high-resolution National Hydrography Dataset
(U.S. Geological Survey, 2019). Site names are not intended to denote political boundaries. CC: cultivated crops, EF: evergreen forest, SS: shrub/scrub, HP: hay/
pasture, OW: open water, DF: deciduous forest, WW: woody wetlands, H: herbaceous.
Site km2 Sentinel-1 image Sentinel-2 image Wetland Stream Annual PDSI PDSI Slope Dominant land
count count area (%) density (m precipitation (min) (max) (%) cover classes
(2017–2021) (2017–2021) km¡2) (mm) (%)
Arkansas- CC (26%), EF
64,813 840 3724 15.2 1711 1388 − 1.7 5.1 2.1
Louisiana (26%)
Arizona 29,177 501 3625 2.6 1939 405 − 4.4 1.8 16.4 SS (75%)
SS (38%), EF
California 39,152 1425 4883 9.7 1259 575 − 5.3 3.5 11.1
(38%)
CC (57%), HP
Iowa 50,937 949 2420 5.3 1415 913 − 1.8 6.4 6.5
(14%)
Kansas- HP (37%), CC
70,673 981 5741 6.1 1642 1064 − 3.4 5.7 4.4
Missouri (26%)
Maryland- OW (23%), CC
63,587 1011 4020 18.6 1621 1132 − 0.9 6 5.2
Delaware (17%)
DF (28%), WW
Minnesota 14,582 661 1701 40.9 462 688 − 4.1 4.8 2.0
(25%)
EF (39%), SS
Montana 59,614 1012 3835 3.2 1544 533 − 3.7 1.9 22.4
(29%)
North
26,055 669 1679 7.9 695 530 − 5.2 4 1.7 CC (74%)
Dakota
South WW (29%), EF
50,168 1142 6030 30.2 1498 1217 − 0.6 4.7 8.4
Carolina (18%)
South CC (58%), H
18,569 806 1225 10.8 909 603 − 3.2 7.2 2.1
Dakota (16%)
HP (19%), EF
Texas 48,859 663 4107 22.8 1389 1347 − 1.5 5.5 5.4
(15%)
Total/
536,185 10,660 42,990 14.4 1340 866 ¡3.0 4.7 7.3
Average
Platform (SNAP) software package. The orbit files were updated with a 2.4. Sentinel-2 pre-processing
restituted orbit file, providing a geometric accuracy of <10 cm (Prats-
Iraola et al., 2015). Ground Range Detected (GRD) border noise (for We processed all Sentinel-2 imagery that intersected the study area
images collected after January 12, 2018) was masked out to remove extent between January 2017 and December 2021 (42,990 images).
artificially low backscatter pixels that occur near the edge of an image Image return interval averaged 5 days over the five-year period. The
swath (Ali et al., 2018). Thermal noise was masked out to reduce dis Sentinel-2 top-of-atmosphere (TOA) image collection (Level 1C) was
continuities between sub-swaths. The images were radiometrically used, as prior efforts have shown that TOA image collections perform
calibrated to backscatter intensity using sensor calibration parameters similarly to surface reflectance in wetland and water environments
from the GRD metadata (Sabel et al., 2012). The images were then (Pahlevan et al., 2019; Yao et al., 2019). To mask clouds, we used the
converted from ground range geometry to the backscatter coefficient Sentinel-2 cloud probability layer (Guolin et al., 2017). Pixels were
sigma naught (σ0), measured in decibels (dB), using a digital elevation classified as cloud if the cloud probability was >30%. Images were
model (DEM) from the Shuttle Radar Topography Mission (SRTM). The excluded if >70% of the image was classified as cloud cover. Cloud
radiometry of the imagery remained an ellipsoid model based on shadows were masked by modifying code provided through the Google
σ0 (Small, 2011). The GRD data is delivered as log-scaled (i.e., 10*log10 Earth Engine community, “Sentinel-2 Cloud Masking with s2cloudless”
(x)). Extreme values, defined as where VV and VH back-scatter were < (https://developers.google.com/earth-engine/tutorials/communit
− 35 dB and < − 40 dB, respectively (DeVries et al., 2020), were removed y/sentinel-2-s2cloudless). Cloud shadows were defined as the intersec
and filled with pixel-wise median values. A cosine correction or inci tion of the predicted cloud shadow area, derived from the cloud prob
dence angle normalization was calculated for both VV and VH following ability layer and the mean solar azimuth angle, and dark areas, defined
the approach presented in Topouzelis et al. (2016) to create backscatter as near-infrared (NIR) values of <0.17. All identified cloud shadows
values independent of the incidence angle. A 60 m focal median filter were buffered by 5 pixels. Single spectral bands were not considered in
was then applied to the backscatter values. Speckle filtering is a common the selection of variables to reduce the influence of spectral variability
pre-processing step that can reduce backscatter noise; however, moving attributable to atmospheric contamination; instead, we relied primarily
window filters can also reduce the effective spatial resolution which can on normalized spectral indices. For both Sentinel-1 and Sentinel-2,
be disadvantageous when attempting to map narrow streams or small same-date images were mosaicked into a single image to eliminate
wetlands and ponds. We therefore retained both the unfiltered and double-counting along overlapping image edges. As free water and
filtered backscatter values. We tested filtering for high wind conditions frozen water show similar scattering coefficients in Sentinel-1 images (Li
using the Real-Time Mesoscale Analysis (RTMA) hourly wind speed et al., 2020), all Sentinel-1 images, regardless of snow status, were
dataset. A wind mask was not applied because we did not observe a retained. Sentinel-2 image classification, in contrast, cannot detect
consistent improvement in the open water or vegetated water after water under snow and ice conditions (Yang et al., 2020). Therefore,
masking out high wind, but wind speed was still considered as a po Sentinel-2 images within site-specific periods of persistent snow and ice
tential variable for the classification model described in section 2.5. The were excluded from the image processing. A snow-free season was
pre-processing steps for Sentinel-1 are shown in Fig. 2. defined at six sites (snow-free day of year range defined as 53–342 in
California, 59–351 in Iowa, 90–328 in Minnesota, 85–320 in Montana,
87–319 in North Dakota, 80–339 in South Dakota). To account for snow
5
Fig. 2. Flowchart showing the processing steps taken to produce the Sentinel-1 (S1) and Sentinel-2 (S2)-based time series of open water and vegetated water. DSWE:
dynamic surface water extent, BU3: 3-band index, GBC: gradient boosted classifier.
and ice cover outside of these periods, Sentinel-2 pixels were masked backscatter (i.e., non-speckle-filtered), as well as speckle-filtered,
where the corresponding Moderate Resolution Imaging Spectroradi cosine-corrected VV and VH, were considered in the Sentinel-1 model,
ometer (MODIS) snow cover (500 m) product predicted >50% snow as were variables derived from the non-speckle-filtered backscatter (‘VV
cover (Hall and Riggs, 2021). Although error in snow masking was likely + VH’, ‘VV-VH’, ‘VH/VV’, ‘VV/(VV + VH)’, ‘VH/(VV + VH)’, angle).
introduced due to the difference in spatial resolution, the Landsat frac Lefebvre et al. (2019) compiled a comprehensive list of 21 multispectral
tional snow-covered area science product is currently only processed for band combinations and indices commonly used to map inundation. For
the mountainous regions within the U.S. Pre-processing steps for the Sentinel-2 model, all spectral combinations and indices from
Sentinel-2 are shown in Fig. 2. Lefebvre et al. (2019) were considered (Table A2).
We also included the 3-Band Index (BU3 = Red + SWIR1 – NIR;
Gautam et al., 2017). This index was originally designed to target urban
2.5. Variables considered areas but was previously found to be helpful in forested wetland envi
ronments (Vanderhoof et al., 2020) and therefore was tested here for
All variables considered for inclusion in the Sentinel-1 and Sentinel-2 classifying areas of vegetated water. Deep water typically shows low
models are listed in Table A2. Sentinel-1 cosine-corrected VV and VH
6
reflectance over the visible, NIR and SWIR regions. Vegetation, in learn Python module (Pedregosa et al., 2011). The potential predictor
contrast, tends to show low reflectance in the visible bands, but leaf variables (Table A2) were sequentially tested. In each step, an additional
tissue and cell structure result in high reflectance in the NIR band (Ma predictor was included in the model that maximized an increase of the
et al., 2019). Consequently, “R – NIR” creates strongly negative values in AUC (area under the receiver operating characteristic curve), while
upland areas, weakly negative values in wetlands, and weakly positive predictors were removed if they had high (≥ 0.90) correlation with
values in open water (i.e., the inverse of NDVI). SWIR1 reflectance, previously selected predictors (Hawbaker et al., 2020). Predictors were
where water is the dominant factor controlling reflectance values selected until the increase in AUC of adding additional variables was
(Mahdavi et al., 2018; Ma et al., 2019), is then added. SWIR1 shows a <0.001. The variable selection model was parameterized as: boosting
gradient from very low reflectance in deep, dark water, intermediate stages = 300, learning rate = 0.3, subsample = 0.7, maximum leaf
reflectance in vegetated wetlands and conifer forests, and high reflec nodes = 30. The selected predictor variables were then used to train the
tance in agriculture and developed areas. The resulting BU3 values show GEE's version of the GBRM, smileGradientTreeBoost classifier. The
a greater potential, relative to more common greenness and wetness Sentinel-1 and Sentinel-2 model used identical model parameterization
indices, to distinguish between vegetated wetlands and conifer forests, (boosting stages = 300, shrinkage = 0.3, sampling rate = 0.7, maximum
and between upland areas prone to flooding and saturation and those leaf nodes = 30, and loss = “Least Squares”). Only single-scene back
that are persistently dry. scatter and multispectral indices, in addition to the topographic and
Topographic and weather variables were also considered to help weather variables, were tested for inclusion during the variable selection
inform models on where and when water may collect, respectively. process. The initial Sentinel-1 model under-performed in distinguishing
Hydrologically adjusted elevations represent the height above the vegetated water from upland vegetation. Multiple studies have found
nearest drainage (HAND) in meters with the vertical increment set at 10 improved performance when integrating variables from both Sentinel-1
cm and were derived from the Multi-Error-Removed Improved Terrain and Sentinel-2 in a single model (e.g., Hird et al., 2017; Slagter et al.,
hydrology dataset (3 arc-second (~90 m at the equator) MERIT Hydro, 2020), however, this approach reduces the temporal resolution of out
Yamazaki et al., 2019). This dataset was resampled to 20 m resolution puts. To improve the detection of vegetated water without reducing the
using cubic convolution. Slope was calculated from the U.S. Geological temporal resolution of the Sentinel-1 outputs, we tested the inclusion of
Survey National Elevation Dataset (NED, 10 m). Same-day evapotrans an annual snow-free, median BU3 variable. An annual time step was
piration (assuming a static vegetative surface of grass) and same-day used to allow for inter-annual variability and was found to outperform a
vapor pressure deficit were included to characterize seasonality, 7-day monthly BU3 variable. After observing improvement in the Sentinel-1
accumulated precipitation helped characterize recent precipitation model, we tested including an annual BU3 variable in the Sentinel-2
events, and 30-day accumulated precipitation helped characterize model and found this model improved as well. The Sentinel-1 and
antecedent wetness conditions. All weather variables were calculated Sentinel-2 models were then applied to the pre-processed image col
from the daily Gridded Surface Meteorological (GRIDMET) data (Abat lections, respectively.
zoglou, 2013; 4 km). Daily snow water equivalent was also considered to Multiple post-processing filters were applied to the vegetated water
help characterize seasonality for sites that experience consistent snow time series (Fig. 2). As a mixed-pixel product, where water represents a
cover during the winter months and was sampled from DAYMET fraction of a pixel, this class can be expected to show a higher rate of
(Thornton et al., 2022; 1 km). In addition, as wind speed can impact confusion with non-water, relative to the open water class. For both
backscatter over open water, hourly wind speed was considered in the Sentinel-1 and Sentinel-2 vegetated water, we applied a focal median
Sentinel-1 models, derived from the RTMA (De Pondeca et al., 2011; 2.5 filter (3 × 3 pixels) to reduce speckling. Confusion between suburban
km). While spectral and backscatter values reflect ground-based phe development and vegetated water was reduced by masking low, me
nomenon (e.g., land cover type and condition), weather variables reflect dium, and high density developed areas, as well as developed open
atmospheric conditions which show relative consistency over 1 to 4 km spaces, as defined by the 2019 NLCD (30 m resolution, Homer et al.,
scales. Therefore, up-sampling the weather data to match the spatial 2020). Further, the NLCD 2019 evergreen forest class was masked in 7 of
resolution of Sentinel-1 and Sentinel-2 was not considered logically the 12 sites, specifically Arkansas-Louisiana, California, Kansas-
consistent. Our primary concern when considering the inclusion of Missouri, Maryland-Delaware, Minnesota, South Carolina, and Texas
weather variables in the classification was that the coarser spatial res (Homer et al., 2020). This step was to reduce high rates of commission
olution would be visible in the results and impact the minimum mapping error in evergreen forests since we saw minimal overlap at the 12 sites
unit. Weather variables were retained in the models only because im between the NWI dataset and the evergreen forest class. This lack of
pacts to the minimum mapping unit were very rare. overlap is because the NLCD dataset maps forested wetlands as a sepa
rate class. We avoided masking conifer forest in Montana, where streams
2.6. Algorithm development within the minimum mapping unit extended into conifer forests. The
NLCD urban class and evergreen forest class have producer accuracies of
We used the Google Earth Engine (GEE) Python application pro 77% and 75%, respectively (Wickham et al., 2021). Lastly, substantial
gramming interface (API) and the geemap package to access, process, precipitation events can cause temporary, widespread soil saturation
and analyse the image collections and ancillary datasets in the Jupyter that can induce confusion with vegetated water (Bazzi et al., 2019). To
environment (Gorelick et al., 2017; Wu, 2020). A single Sentinel-1 reduce this source of confusion, vegetated water observations were
gradient boosted regression model (GBRM) and a single Sentinel-2 masked (i.e., reclassified to NA) where a same-day precipitation event
GBRM were developed from training data across all 12 sites. Devel was classified as more than light rain (>5 mm as measured by daily
oping a single model across sites enables the algorithms to be applied in GRIDMET; Endarwin et al., 2014). In contrast, minimal post-processing
areas outside of where the models were trained. GBRMs use a sequence steps were applied to the Sentinel-1 and Sentinel-2 open water time
of simple classification and regression tree (CART) models (Hastie et al., series. Only developed areas, defined as low, medium and high-density
2009). In contrast to random forest models, which sample a subset of the development classes (NLCD 2019), were masked from the open water
data for each CART model and average the results, a boosted approach time series. No filters were applied to the open water time series to
sets a learning rate and fits trees in a sequence, where the observations maximize the retention of narrow and small water bodies.
used to fit each sequential CART model are weighted by the residuals of
the previous CART model. Consequently, GBRMs often shows a higher 2.7. Validation
predictive accuracy than that of CART or random forest models for
classification problems (Hastie et al., 2009). Within each of the 12 sites, we randomly selected WorldView-2 (2 m,
To select a subset of variables, a GBRM was first fit using the Scikit- images = 13), WorldView-3 (1.4 m, images = 9) and PlanetScope (3–5
7
m, images = 14) images (total images = 36) representing diverse seasons area at each time step between Sentinel-1 and Sentinel-2 outputs. Shifts
and multiple years (2017–2021, 3 images per site, Table A3). Validation in assigned class for highly dynamic select points were also graphed,
points (n = 200) were randomly generated across each image. Visual where every Sentinel-1 and Sentinel-2 observation was shown. The
cues were used to manually classify the pixel corresponding to each corresponding GRIDMET Palmer Drought Severity Index was reported
point as non-water, open water, or vegetated water. Aerial imagery, the as an indicator of relative site wetness (Abatzoglou, 2013).
NWI wetland type and water regime attributes as well as monthly pre
cipitation (Abatzoglou, 2013) were used as ancillary datasets as needed. 3. Results
However, it is likely that relying on optical imagery for reference data
may have biased the distribution of reference vegetated water points. As 3.1. Inundation algorithms
inundation is often a minority cover class, we employed a stratified
disproportionate sampling scheme, where open water and vegetated The final Sentinel-1 algorithm consisted of nine variables (Fig. 3).
water were required to represent at least 15% of the points, respectively, The annual average 3-Band Index (BU3) derived from Sentinel-2 showed
when feasible. This sampling design enabled increased precision in the greatest importance as a predictor, followed by slope, and single
estimating the accuracy of the rare cover type, in this case, open and scene speckle-filtered VH and speckle-filtered VV (Fig. 4). While annual
vegetated water (Strahler et al., 2006). The validation points from each BU3 was originally included to reduce confusion between urban and
high-resolution image were compared with the nearest date Sentinel-1 water cover (Vanderhoof et al., 2020), we found that while many of the
and cloud-free (<50% cloud cover) Sentinel-2 image. The validation common water indices helped distinguish open water, BU3 helped
data was retained at its higher spatial resolution and was not down distinguish vegetated water. A comparison between the four top vari
sampled. Therefore, an open water validation point, for example, could ables can also be seen in Fig. 4, where BU3 at an annual time-step helps
represent a mixed pixel in the classified image (20 m). The mean ab to distinguish potential vegetated water along riparian corridors, while
solute date gap between the Sentinel-1 and high-resolution images was the VH filtered helps identify open water. Observations are from June
2.7 days, and 1.9 days between the Sentinel-2 and the high-resolution 2020, which is a relatively wet month in this region. The Sentinel-1
images. While it is unlikely that large differences in surface water algorithm also selected two topographic variables, hydrologically
extent will occur in these short date gaps, potential differences attrib adjusted elevation and slope, as well as three weather variables, vapor
utable to the date gap is another potential source of uncertainty. pressure deficit and precipitation accumulated for the last 7 and 30 days,
The primary objective was to map surface water extent. Open water respectively. The Sentinel-2 algorithm included ten variables (Fig. 3).
and vegetated water were mapped as separate classes because they show Three single scene spectral indices showed the greatest variable
distinct color, form, and structure (Mahdavi et al., 2018; Shen et al., importance for the Sentinel-2 algorithm (Modified Index of Free Water
2019). To this end, omission and commission errors were reported first (MIFW), BU3 and Soil Adjusted Vegetation Index (SAVI)) followed by
as the confusion between only the non-water class and both the open both topographic variables. The annual BU3 showed a lower variable
water and vegetated water classes. Omission and commission errors for importance (7th out of 10 variables), however, the single scene BU3
open water and vegetated water, as well as overall accuracy were re showed the 2nd highest variable importance. In Fig. 4, the relative
ported for each of the 12 sites and all sites combined. However, as
confusion between the two water classes is also relevant, we additionally
reported the percent of open water and vegetated water points that were
mapped as vegetated water and open water, respectively for each site.
Validation images and the corresponding Sentinel-1 and Sentinel-2 im
ages are listed in Table A3.
2.8. Analysis
The percent open water and vegetated water were calculated as the
classification count divided by the clear-sky observation count. We note
that the percentiles are not directly comparable between Sentinel-1 and
Sentinel-2 because cloud cover, ice and snow reduced clear-sky obser
vations for Sentinel-2 relative to Sentinel-1. To summarize the propor
tion of water that is highly dynamic versus more persistent over time, we
reported the percent of each site (Fig. 1) that was classified over the time
series (2017–2021) as (1) open water >80%, (2) vegetated water >80%,
(3) open or vegetated water >80%, (4) open water 20–80%, (5) vege
tated water 20–80%, and (6) open or vegetated water 20–80%. To
further demonstrate the mapped variability in surface water extent over
time, the Sentinel-1 and Sentinel-2 time series were consolidated to a
monthly time-step, where the class was the majority assigned value
within the clear-sky observations of each month. The total open and
vegetated water was summed at each time step and graphed over the 5-
year period. As the tabulation function “reduceRegion” is a memory
intensive process for Google Earth Engine, we selected a portion of each
site with the highest concentration of water bodies, ranging from 3298
km2 in California to 5123 km2 in Montana, to represent the temporal
dynamics of that area. The percent area mapped as water in these
smaller areas were correlated using Spearman correlation (Hmisc R
package) to test for the consistency of changes in mapped extent be
tween the Sentinel-1 and Sentinel-2 time series. To account for sites
where a correlation does not occur because water extent shows little Fig. 3. Variables selected for inclusion in the Sentinel-1 and Sentinel-2 models,
change over time, we also calculated the median difference in percent respectively, and the corresponding variable importance in each model.
8
Fig. 4. An example comparison of the model variables showing the greatest importance including (a) annual 3-band index for 2020, (b) Modified Index of Free Water
for June 1, 2020, June 3, 2020 (c) filtered VV and (d) filtered VH, (e) percent slope, and (f) PlanetScope (3 m) natural color image.
9
contribution of the MIFW and annual BU3 can be seen where MIFW 3.3. Inundation dynamics
distinguishes open water and the annual BU3 helps distinguish vege
tated water. Two weather variables were selected in the Sentinel-2 al Examples of inundation frequency across the five-year period for the
gorithm, daily evapotranspiration and precipitation which was open water class are shown for the Sentinel-1 and Sentinel-2 algorithms
accumulated for the past 30 days. The selection of multiple weather at river and tidal-dominated sites (Fig. 6) and wetland and lake domi
variables in both models suggests that data on recent weather conditions nated sites (Fig. 7). The example locations and extents are consistent
can improve efforts to track inundation over time. However, because the between the Sentinel-1 and Sentinel-2 algorithms to enable comparison
variables showing the greatest importance in both models were finer in of the open water class between sensors, while the scale of the examples
spatial resolution (10–20 m), the coarse resolution of the weather var was varied to demonstrate how well the algorithms perform in different
iables (4 km) did not visibly influence the minimum mapping unit. hydrological landscapes. Most sites showed a similar identification of
When the algorithms were applied over the Sentinel-1 and Sentinel-2 open water extent and frequency between the two algorithms. However,
image collections we found that although Sentinel-2 collected imagery in Iowa and Montana, both river-dominated sites, the Sentinel-1 algo
at more than double the frequency of Sentinel-1, after accounting for rithm saw a greater delineation of open water along the river network
cloud cover, Sentinel-1 provided more observations over the 5-year compared to the Sentinel-2 algorithm. Consistent with the finding that
period. The Sentinel-1 observation count averaged 172 across the 12 the Sentinel-2 algorithm commonly mapped open water as vegetated
sites with site averages ranging from 141 in Texas to 217 in California water, the Sentinel-1 algorithm saw more open water in Minnesota,
(Fig. 5). Sentinel-2 clear-sky observation count averaged 148 but North Dakota, and South Dakota (Figs. 7). Another example can be
showed greater variability across sites ranging from Minnesota with one found in Fig. 8, where the Sentinel-1 algorithm showed a greater fre
of the longest persistent snow-cover seasons (155 observations) to a dry, quency of open water extent relative to the Sentinel-2 algorithm in South
warm site like Arizona (466 observations) (Fig. 5). Dakota. Fig. 8 displays examples of both the open water and vegetated
water classes in lake-dominated (a-c), tidal wetland dominated (d-f),
3.2. Accuracy assessment river dominated (g-i), and forested wetland dominated (j-l) landscapes.
Similar to Figs. 6 and 7, we see a clearer delineation of river networks
After excluding pixels classified as no data from the validation sta with the Sentinel-1 algorithm. We also see a difference in the inundation
tistics, the validation of the Sentinel-1 and Sentinel-2 algorithms used frequency for forested wetlands. While high leaf-cover in the growing
7200 and 7085 validation points, respectively. Classification accuracy season obscures forested inundation for the Sentinel-2 algorithm, we see
for open water was high, averaging only 3.1% omission error (OE) and greater rates of detection and frequency with the Sentinel-1 algorithm.
0.9% commission error (CE) for the Sentinel-1 algorithm, and 3.1% OE In Fig. 8, tidal wetlands and inland riparian wetlands, narrow river
and 0.5% CE for the Sentinel-2 algorithm (Table 2), indicating that open networks and forested wetlands are all mapped as vegetated water by
water was typically found by the algorithm. Open water error was <1% both algorithms. However, open water percent is mapped over vege
for both error types in 5 sites for both the Sentinel-1 and Sentinel-2 al tated water extent Fig. 8, so that pixels that switch between open and
gorithms. Open water omission error for the Sentinel-1 algorithm was vegetated water are not visible using this visualization approach.
highest in sites dominated by narrow river systems (i.e., Arizona, Cali Surface water dynamics derived from the 5-year time series (TS)
fornia, and Iowa), and for the Sentinel-2 algorithm was highest in Ari were summarized by calculating the proportion of each site classified as
zona and Minnesota (Table 2). Vegetated water accuracy was lower, as open or vegetated water (1) 20–80% of the time over the TS and (2)
expected given that vegetated water represents mixed pixels. Across the >80% of the time over the TS. Table 3 shows that the surface water
12 sites, vegetated water showed higher accuracy using the Sentinel-2 dynamics were disproportionally greater in vegetated water compared
algorithm, with 10.7% OE and 7.9% CE compared to the Sentinel-1 al to open water. Across the 12 sites, the percent area classified as open
gorithm with 28.4% OE and 16.0% CE (Table 2). For the Sentinel-2 al water >80% of the TS, open water 20–80% of the TS, and vegetated
gorithm, vegetated water accuracy tended to be highest in tidal wetland water >80% of the TS all averaged <3% of the site area for both sensors,
environments (i.e., Maryland-Delaware, South Carolina, Texas), and while the precent area classified as vegetated water 20–80% of the TS
showed greater omission error where forested leaf cover was high (i.e., averaged 20.3% and 22.2% of the site area for the Sentinel-1 and
Arkansas-Louisiana, Minnesota) or vegetated water cover was low (i.e., Sentinel-2 algorithms, respectively. For example, the Arkansas-
Arizona). For the Sentinel-1 algorithm, vegetated water showed higher Louisiana site, which includes a large river floodplain, showed that
omission error in inland riparian corridors (i.e., Montana, Kansas- only 2.3% and 1.7% of the area was classified as vegetated water >80%
Missouri, Arizona) and lower omission error at sites with tidal wet of the TS, while 30.7% and 37.0% of the area was classified as vegetated
lands (i.e., South Carolina, Texas, Maryland-Delaware). Commission water 20–80% of the TS, for the Sentinel-1 and Sentinel-2 algorithms,
error for the Sentinel-1-based vegetated water was highest in the respectively (Table 3). However, the inland open water dynamics were
mountainous portions of the California site and the highly vegetated not insubstantial. Averaged across all sites, approximately equivalent
Texas site. Overall accuracy averaged 85.8% for the Sentinel-1 algo amounts of open water were present >80% of the TS (2.5% and 1.4% for
rithm (ranged from 79.8% to 93.8%) and 93.3% for the Sentinel-2 al the Sentinel-1 and Sentinel-2 algorithms, respectively) compared to
gorithm (ranged from 84.8% to 97.9%). 20–80% of the TS (1.7% and 2.1% for the Sentinel-1 and Sentinel-2
Confusion between the two water classes is also important to algorithms, respectively) (Table 3).
consider. Although open water was typically found by either the open The stability of surface water extent was also variable across sites.
water or vegetated water classes, at a subset of sites, including Minne Sites with large river floodplains relative to site area, including
sota, Montana, and North and South Dakota, the Sentinel-1 algorithm Arkansas-Louisiana (i.e., Mississippi River), Kansas-Missouri (i.e., Mis
commonly mapped vegetated water as open water (Table 2). Across the souri and Kansas Rivers) and North Dakota (i.e., Red River) showed the
12 sites, 14.4% of the Sentinel-1-based vegetated water was mapped as lowest proportion of water present >80% of the TS, while sites with tidal
open water, compared to only 2.4% of open water that was mapped as wetlands (e.g, Texas, South Carolina) and sites with large lakes (e.g,
vegetated water. The Sentinel-2 algorithm showed the opposite pattern, California, South Dakota) showed the highest proportion of water pre
with 20.4% of open water mapped as vegetated water, with the highest sent >80% of the TS (Table 3). Comparing between sensors, the
rates of confusion in California, Minnesota, Montana, and North and Sentinel-2 algorithm mapped more open water (present 20–80% of the
South Dakota, but only 6% of vegetated water mapped as open water TS), vegetated water (present >80% of the TS) and vegetated water
(Table 2). The between class confusion for the Sentinel-2 algorithm (present 20–80% of the TS), while the Sentinel-1 algorithm mapped
reduced the overall accuracy by a greater amount (93.3% to 86.1%) more open water (present >80% of the TS) (Table 3).
compared to the Sentinel-1 algorithm (85.8% to 82.1%). If we assume that surface water extent varies at a time-scale
10
Fig. 5. Clear-sky observation count for (a) Sentinel-1 and (b) Sentinel-2 across the 12 sites (2017–2021), after masking same-day large rain events, and for Sentinel-2
only, also masking cloud and cloud shadow and snow cover.
11
Table 2
Omission error (OE), commission error (CE), overall accuracy (OA), and between class confusion, reported as percent, for open water and vegetated (veg) water by site.
Site Sentinel-1
open open veg water veg water open water mapped veg water mapped as OA (water and OA (non-water, open
water OE water CE OE CE as veg water open water non-water) water, veg water)
Arizona 19.6 0 48.6 2.7 1.1 8.2 84.5 83.3

Arkansas-
1.4 0.7 38 16.8 10.4 0 81.3 78.8
Louisiana
California 7.2 6.6 39.8 29.8 4.3 7.5 82 80.0
Iowa 7.4 0 22.9 3.9 5.8 11.5 89.8 85.7
Kansas-
0 0 54.4 19 0.7 6.3 88.8 88.1
Missouri
Maryland-
0.9 0 19.7 14.5 1.4 7.1 84.2 81.0
Delaware
Minnesota 0 0 22.5 17.2 0 47.9 86 72.5
Montana 3.5 0 54.3 7.1 0.6 34.1 79.8 74.5
North Dakota 2.6 1.3 50.8 24.7 2 31.1 85.2 81.5
South Carolina 0.5 0 7.3 9.5 0.6 2.6 92.8 91.7
South Dakota 0.7 0 5.6 13.3 0.7 18.8 93.8 88.3
Texas 1.1 0.5 13.9 30.5 3.2 8.6 81.5 79.5
All 3.1 0.9 28.4 16 2.4 14.4 85.8 82.1
Sentinel-2
Site open open veg water veg water open water mapped veg water mapped as OA (water and OA (non-water, open
water OE water CE OE CE as veg water open water non-water) water, veg water)
Arizona 12.5 0 31.7 5.8 11.2 9.3 89.1 85.7
Arkansas-
1.4 1.4 20.6 2.3 18 7.6 91.2 84.7
Louisiana
California 0.7 0 8.3 0 35.1 0 97.9 88.9
Iowa 5.6 1.5 2.1 3.6 11.1 10.2 96.4 90.6
Kansas-
0 0 14.6 6.4 0.7 12.5 96.5 94.5
Missouri
Maryland-
0.5 0 0.4 8.7 9.4 1.5 95.3 91.2
Delaware
Minnesota 6.9 0.5 9.6 21.5 42.1 8.1 84.8 68.0
Montana 2.9 1.8 16.6 2.4 33.9 6.6 92.4 81.1
North Dakota 5.3 0 20 12.7 25.4 7.3 92.2 85.0
South Carolina 0.5 0 6.6 1.3 1.1 2.2 96.6 95.5
South Dakota 3.4 0 6.9 11.5 31.9 12.3 93.5 82.5
Texas 0.5 0.5 5.3 11.4 21.6 1.5 93.6 86.7
All 3.1 0.5 10.7 7.9 20.4 6 93.3 86.1
detectable with a 5-year monthly dataset, then the correlation between water and open water. In the California example, during the dry 2021
the Sentinel-1 and Sentinel-2-based mapped water can further speak to period, indicated by a low Palmer Drought Severity Index (PDSI value),
the consistency between the two datasets, and therefore, how seamlessly the point transitioned from open water to primarily non-water. Simi
the Sentinel-1 and Sentinel-2-based time series can be integrated for this larly, in the Texas wetland, during the driest period, and consequently
objective. Montana and North Dakota, for example, both showed a the lowest PDSI values (Jan 2020 – July 2021), the point was more
strong seasonal signal in surface water extent (Fig. 9), and in turn frequently mapped as vegetated water instead of open water by both
showed a significant correlation (p < 0.01) for both open water and algorithms. Further, in addition to tracking individual points, we can
open/vegetated water (Table 4). Sites with the highest density of also track seasonal and interannual variability in surface water for areas
forested wetlands, Arkansas-Louisiana and Minnesota (Fig. 1), where where dynamic changes are expected, like floodplains (Fig. 11). The Red
inundation may be obscured by leaf cover for both Sentinel-1 and River showed substantial flooding in spring 2019, which expanded the
Sentinel-2, showed significant correlation with open water, but not open and vegetated water, as mapped by both Sentinel-1 and Sentinel-2,
open/vegetated water. The nine inland sites, however, showed signifi relative to a non-flood period (Fig. 11). We note, however, that cloud
cant positive correlations between either open water or both open water cover and seasonal snow cover limited the number Sentinel-2 observa
and open/vegetated water (Table 4). Alternatively, we found that the tions relative to the number of Sentinel-1 observations (Fig. 11).
three sites with extensive tidal wetlands, Maryland-Delaware, Texas,
and South Carolina (Fig. 1), all showed more stable proportions of water 4. Discussion
over time (Fig. 9) and no significant correlations between mapped
variability in open water or open/vegetated water (Table 4). We 4.1. Variability in surface water extent over time
therefore also provided the median absolute difference in percent area
mapped as water by the Sentinel-1 and Sentinel-2 algorithms. The me An abundance of inland water is distributed in narrow river networks
dian difference in the Sentinel-1 and Sentinel-2-based percent open and small (<1 ha), frequently seasonal wetlands that play a critical role
water was <2% at 9 of the 12 sites. Difference in open and vegetated in the biogeochemistry, biodiversity, and hydrological conditions of
water percent depended on the amount of vegetated water present in the downstream waters (Benstead and Leigh, 2012; Leibowitz et al., 2018)
site with larger differences observed in Texas, Minnesota and Arkansas- but are at greater risk for loss due to human activities (Van Meter and
Louisiana, and smaller differences observed in Arizona, Montana, Cali Basu, 2015). Consequently, despite the production of global surface
fornia, and Kansas-Missouri (Table 4). These results indicate that time water products at 30 m, improved detection of inland water bodies and
series can be used to track changes (non-water versus open or vegetated their dynamics is still an outstanding challenge for the remote sensing
water) in condition over time in response to drivers such as weather community. Processing dense time series of imagery in cloud-based
variability. In Fig. 10, we show an example from California (a-e) and platforms further supports efforts to monitor surface water extent,
Texas (f-j) of two points that transition primarily between vegetated which often shows seasonal and inter-annual variability (Pekel et al.,
12
Fig. 6. For sites dominated by river and tidal wetlands the percentage of time (2017–2021) a pixel was classified as open water in Arkansas-Louisiana with (a)
Sentinel-1 algorithm and (b) Sentinel-2 algorithm, Iowa with (c) Sentinel-1 algorithm and (d) Sentinel-2 algorithm, Kansas-Missouri with (e) Sentinel-1 algorithm and
(f) Sentinel-2 algorithm, Maryland-Delaware with (g) Sentinel-1 algorithm and (h) Sentinel-2 algorithm, South Carolina with (i) Sentinel-1 algorithm and (j) Sentinel-
2 algorithm, and Texas with (k) Sentinel-1 algorithm and (l) Sentinel-2 algorithm.
13
Fig. 7. For sites dominated by lake and inland wetlands the percentage of time (2017–2021) a pixel was classified as open water in Arizona with (a) Sentinel-1
algorithm and (b) Sentinel-2 algorithm, California with (c) Sentinel-1 algorithm and (d) Sentinel-2 algorithm, Minnesota with (e) Sentinel-1 algorithm and (f)
Sentinel-2 algorithm, Montana with (g) Sentinel-1 algorithm and (h) Sentinel-2 algorithm, North Dakota with (i) Sentinel-1 algorithm and (j) Sentinel-2 algorithm,
and South Dakota with (k) Sentinel-1 algorithm and (l) Sentinel-2 algorithm.
14
Fig. 8. Examples of different water body types including (a) inland wetlands in South Dakota PlanetScope), (b) the corresponding Sentinel-1 and (c) Sentinel-2
algorithm classification, (d) tidal wetlands in South Carolina (Sentinel-2_, (e) the corresponding Sentinel-1 and (f) Sentinel-2 algorithm classification, (g) rivers in
Missouri (PlanetScope), (h) the corresponding Sentinel-1 and (i) Sentinel-2 algorithm classification, and (j) forested wetlands in Maryland-Delaware (PlanetScope), as
well as (k) the corresponding Sentinel-1 and (l) Sentinel-2 algorithm classification. Left panels show natural color Sentinel-2 images. Classifications were derived
from all images collected in 2018 and 2019 and reported as a percent.
15
Table 3
Percent of each area classified as open water or vegetated water >80% of the time series (2017–2021) and 20–80% of the time series (2017–2021) as mapped by the
Sentinel-1 (S1) and Sentinel-2 (S2) algorithms. Water refers to open and vegetated water. Site-scale summaries enable comparisons of how dynamic surface water is
between sites, and if the two algorithms are mapping this variability differently or similarly.
Site Percent of area classified Percent of area classified Percent of area classified Percent of area classified as Proportion of water
as open water > 80% of as open water 20–80% of as vegetated water > 80% vegetated water 20–80% of present > 80% compared to
the time series the time series of the time series the time series water present 20–80%
S1 S2 S1 S2 S1 S2 S1 S2 S1 S2
Arkansas-Louisiana 1.5 1.2 1 2.3 2.3 1.7 30.7 37 0.12 0.08

Arizona 0.1 0.04 0.5 0.36 0.1 0.2 1 2.1 0.14 0.08
California 2.2 1.4 3.2 2.8 0.2 1.7 7.6 7.8 0.26 0.21
Iowa 0.9 0.4 1.3 1 0.5 0.6 9 6.2 0.14 0.16
Kansas-Missouri 1.3 1.1 0.6 0.9 0.6 0.5 19.3 14.9 0.10 0.11
Maryland-Delaware 1.5 1.4 0.6 1.6 2.5 4.4 25.9 30 0.15 0.19
Minnesota 11.4 3.6 3.1 8.6 2.3 5.3 31 49.6 0.44 0.18
Montana 0.5 0.3 1.8 1.8 0.3 0.7 5.1 4.8 0.10 0.16
North Dakota 1.5 1 1.9 0.7 0.1 0.9 30.2 20.3 0.05 0.09
South Carolina 1.2 1.1 0.5 1 6.1 6.8 27.5 37.1 0.26 0.21
South Dakota 5.8 3.5 5.1 3.3 0.3 1.4 23.6 16.3 0.23 0.30
Texas 1.8 1.8 0.7 1.1 11.8 6.4 33.1 41.3 0.41 0.20
Average 2.5 1.4 1.7 2.1 2.3 2.6 20.3 22.2 0.22 0.17
Fig. 9. Tracking monthly change in open water (OW) and vegetated water (VW) over time (2017–2021) as mapped by the Sentinel-1 and Sentinel-2 algorithms
across the portion of all 12 sites with the highest concentration of waterbodies (average 4475 km2, total: 53,699 km2). Subset areas are shown in more detail in Fig. 1.
Accuracy statistics are provided to support interpretation. Gaps in the record represent months with inadequate clear-sky observation coverage.
2016; Park et al., 2022). We found that surface water dynamics were such as the Prairie Pothole Region (Vanderhoof et al., 2018; Yeo et al.,
much larger within the vegetated water class compared to the open 2019) or regions with an abundance of floodplain habitat (Heimhuber
water class. This finding suggests that remote sensing approaches which et al., 2017; Fuentes et al., 2019). In this effort, we saw seasonal vari
primarily track open water will underestimate much of the surface water ability in the portion of Montana and North Dakota where the time series
dynamics occurring over time. At regional to national scales, surface was mapped at a monthly time-step. In South Dakota, the time series
water dynamics typically do not respond uniformly to increases or de reflected interannual variability, where 2019 was a very wet year for
creases in precipitation or snowmelt (Vanderhoof et al., 2018). Instead, this region (Fig. 9). Similarly, both seasonal and interannual variability
the magnitude of the change to surface water extent can depend on the was evident in tracking floodplain dynamics along the Red River
amount of surface storage (i.e., wetlands and lakes) as well as the ge (Fig. 11). At other sites, vegetated water often showed much greater
ology and topography of a site. More dramatic changes in surface water variability relative to open water and was identified inconsistently be
storage are typically observed in regions with more wetlands and lakes, tween sensors. In these cases, the per-site validation statistics may help
16
Table 4 determine if one algorithm should be used over the other, or if the al
Spearman correlation values between the Sentinel-1 (S1)-based mapped pro gorithms should be used in complement instead of integrated together.
portion and Sentinel-2 (S2)-based mapped proportion, as derived from the
monthly time series datasets. Additionally, median absolute (abs) difference
4.2. Model variable importance
(diff) in the percent of area classified as open water (OW) or open and vegetated
water (OW + VW) by the Sentinel-1 and Sentinel-2 algorithms. *p < 0.05, **p <
0.01. In addition to the Sentinel-1 and Sentinel-2 derived variables,
topographic and weather variables were consistently selected as
Site Open Open and Median Diff in Median Diff in
important contributors to both the Sentinel-1 and Sentinel-2 algorithms.
Water Vegetated S1-S2 Percent S1-S2 Percent
Water (OW, abs, %) (OW þ VW, abs, Topographic variables, such as slope and hydrologically adjusted
%) elevation, which were selected by both the Sentinel-1 and Sentinel-2
Arizona 0.51** 0.14 0.2 1.0
algorithms, help models determine where water collects on the land
Arkansas- scape, while weather variables help models determine if water is present
Louisiana 0.88** − 0.46 0.5 13.8 at a given point in time. While other efforts have found topographic
California 0.76** 0.52** 2.9 3.1 variables helpful (e.g., Ludwig et al., 2019), weather variables are not
Iowa 0.54** − 0.05 0.7 4.8
typically included in inundation algorithms, despite wide acknowl
Kansas-
Missouri 0.50** 0.30* 0.2 3.2 edgement that surface water extent is often correlated with weather
Maryland- inputs (Mishra and Cherkauer, 2011; Song et al., 2018; Tulbure and
Delaware 0.03 − 0.32 1.1 7.8 Broich, 2019; Xia et al., 2019). Our biggest concern in including coarse
Minnesota 0.47* − 0.31 2.9 9.2
resolution weather variables was that these variables would impact the
Montana 0.70** 0.63** 0.5 1.6
North
effective spatial resolution of the outputs. Although weather variables
Dakota 0.36* 0.56** 1.3 4.1 were consistently selected by both the Sentinel-1 and Sentinel-2 algo
South rithms, these variables showed far less importance relative to Sentinel-1
Carolina 0.03 − 0.25 1.5 6.2 and Sentinel-2 variables. Therefore, given limited evidence that the in
South
clusion of the coarse resolution weather variables impacted the mini
Dakota 0.36* 0.25 2.1 6.0
Texas 0.15 − 0.23 0.3 25.2 mum mapping unit, we hypothesize that the weather variables
accounted for some time series variability in spectral or backscatter
thresholding.
In the Sentinel-2 algorithm, the MIFW, BU3, and SAVI showed the
greatest variable importance. SAVI uses the red edge effect to detect
green vegetation, while MIFW uses the green and shortwave infrared
bands, which are sensitive to soil moisture and vegetation moisture (Xu,
Fig. 10. The time series data can be used to track the mapped class of individual pixels over time. The intersection of a lake and river in northern California is shown
(a) comparing the Sentinel-1 and Sentinel-2 algorithm classes with the Palmer Drought Severity Index (PDSI, positive and negative values reflect wet and dry
conditions, respectively), (b) Sentinel-2-based open water, (c) Sentinel-2-based vegetated water, (d) Sentinel-1-based open water, and (e) Sentinel-1-based vegetated
water. A second example of a lake edge in southern Texas is shown (f) comparing the Sentinel-1 and Sentinel-2 algorithm classes with the PDSI, (g) Sentinel-2-based
open water, (h) Sentinel-2-based vegetated water, (i) Sentinel-1-based open water, and (j) Sentinel-1-based vegetated water. PlanetScope images of the two sites from
a dry and wet date (1–4) were included for context.
17
Fig. 11. The time series can be used to track seasonal and interannual variability in open and vegetated water, such as floodplain dynamics in the Red River, which
runs north along the North Dakota-Minnesota border. (a) Sentinel-1 algorithm and (b) Sentinel-2 algorithm observations over time, and examples of (c) PlanetScope
imagery during a flood event, and corresponding (d) Sentinel-1 algorithm classification, and (e) Sentinel-2 algorithm classification, and (f) PlanetScope imagery in a
drier period (white strip indicates a gap in collection coverage), and corresponding (g) Sentinel-1 algorithm classification, and (h) Sentinel-2 algorithm classification.
Stars show the timing of the classification images relative to the time series. (For interpretation of the references to color in this figure legend, the reader is referred to
the web version of this article.)
2006) and have been effective in mapping water in coastal salt marshes Further, using the annual time-step instead of a 5-year average value
(Davranche et al., 2013). BU3, a relatively under-used variable in helped the models account for interannual variability in vegetated
remote sensing of surface waters, combines the red edge contrast with wetland extent. Its high variable importance in the Sentinel-1 model
the shortwave infrared band. Its contribution to improved model per suggests that annual BU3 improved the model's ability to identify lo
formance indicates that this index of all three bands was helpful for cations where vegetated water is likely to occur, however, the annual
detecting vegetated water in this study. time-step may have reduced the model's ability to track seasonal or
In the Sentinel-1 algorithm, in addition to VV and VH backscatter, episodic variability in vegetated water (Fig. 9). This could account for
the annual BU3 index derived from Sentinel-2 also showed high variable lower Sentinel-1 vegetated water accuracy (Table 2) and between-sensor
importance. Inclusion of BU3 helped the Sentinel-1 algorithm better correlation in this class (Table 4) (although not in detection of seasonal
detect and classify vegetated water and riparian corridors (Fig. 4). While variability in Sentinel-1-based vegetated water in floodplains; Fig. 11).
BU3 was originally developed for urban applications (Gautam et al.,
2017) and has been previously used to distinguish vegetated water from
urban development (Vanderhoof et al., 2020), the high variable 4.3. Comparisons between the Sentinel-1 and Sentinel-2 models
importance of BU3 in both the Sentinel-1 and Sentinel-2 algorithms
support a greater use of this index in future efforts to track vegetated Quantifying variability in the performance of the Sentinel-1 and
water (e.g., riparian vegetation, emergent and forested wetlands). Prior Sentinel-2 algorithms is an essential pre-cursor to scaling up inundation
studies have found it helpful to integrate Sentinel-1 and Sentinel-2 products from these satellites to national or global scales. In our effort,
variables to derive a single, inundation product, albeit at a lower tem the coordination of training and validation datasets across both the
poral resolution (e.g., Hird et al., 2017; Slagter et al., 2020). Our Sentinel-1 and Sentinel-2 algorithm efforts enabled direct comparisons
approach of using BU3 at an annual time-step enabled us to integrate of the two models' performance. We note, however, that as our training
data from Sentinel-1 and Sentinel-2 while maintaining the 12-day and validation datasets relied on optical algorithms and visual inter
temporal resolution of the Sentinel-1 algorithm. However, requiring pretation of optical imagery, these datasets likely contain some propa
an annual variable will limit the near-real time capabilities of the al gated misclassification errors that could bias interpretation of algorithm
gorithm. While use of a monthly average, instead of an annual average accuracy towards the Sentinel-2 algorithm. For example, Fig. 8, shows
will be limited in months with snow cover, future efforts could consider that the Sentinel-1 algorithm more consistently mapped inundation in
using a previous year average or monthly averages where and when Maryland forested wetlands than the Sentinel-2 algorithm, however, in
snow cover is limited to reduce the time lag needed before the algorithm validating inundation at this site, we prioritized leaf-off high-resolution
can process additional imagery. The variables VV and VH backscatter images where forested wetlands could be more accurately mapped. This
both easily identify open water (Fig. 4), but can lead to model in likely improved the accuracy statistics of the Sentinel-2-based vegetated
efficiencies, e.g., in requiring disproportionate effort to learn to identify water class relative to the Sentinel-1-based vegetated water class as we
vegetated water. As with the topographic variables, inclusion of the were not able to validate cases where the Sentinel-1 algorithm mapped
annual BU3 index improved model predictions with information about water that was omitted by the Sentinel-2 algorithm under forested, leaf-
where vegetated wetlands are likely to occur on the landscape and ex on conditions.
plains its high variable importance within the Sentinel-1 algorithm. Despite this uncertainty, open water showed high accuracy across
both models, while vegetated water showed lower accuracy. A diversity
18
of vegetation morphologies across wetlands and riparian corridors cre decreased the effective accuracy of the Sentinel-1 training points in
ates a challenge to discriminate vegetated water from terrestrial eco training images where vegetated water is limited to narrow, riparian
systems (Lefebvre et al., 2019). In our effort we found the Sentinel-1 and corridors and wetland edges.
Sentinel-2 algorithms most often confused vegetated water (e.g., inun Site-specific rates of confusion between the open and vegetated
dated riparian vegetation, emergent, and forested wetlands) with other water classes can provide valuable insights into what spectral back
mixed pixels, such as heavily shadowed conifer forests, linear conifer scatter attributes the Sentinel-1 and Sentinel-2 algorithms keyed in on.
wind breaks with shadows, irrigated agriculture with a high moisture For example, it was evident in the Dakotas and Minnesota sites that
content, or dense suburban areas with a mix of houses and trees. In open-water lakes and wetlands with high chlorophyll levels would often
addition, the vegetated water class also mapped areas that were satu be mapped as vegetated water by the Sentinel-2 algorithm (Fig. 8b, c,
rated, but may not be considered a wetland, such as water-logged, fallow Table 2), suggesting that this algorithm focused on color, while the
agricultural fields or temporarily inundated floodplains (e.g., Yang Sentinel-1 algorithm mapped these same water bodies as open water,
et al., 2020). Commission error in conifer and urban environments was suggesting that it responded to the presence or absence of plant cover
reduced by applying land cover masks to vegetated water classifications and plant structure (Yuan et al., 2015). Similarly, in generating training
but persisted in fallow agricultural fields, especially for the Sentinel-1- data, it was often challenging to optically distinguish sub-surface hy
based vegetated water. This condition, while not necessarily an error, drophytes (e.g., free-floating algae) from low-stature surface hydro
was difficult to validate as distinguishing soil moisture from optical phytes (e.g., duckweed, water lilies), which may have both been mapped
image interpretation was challenging. Omission error, in contrast, often as open water by the Sentinel-1 algorithm. As evidence, the Sentinel-1
occurred when vegetated water was concentrated in narrow bands along algorithms between-class confusion was low at sites dominated by
the edge of open water features as was often the case in sites such as forested wetlands, tidal wetlands, and rivers, while vegetated water was
North Dakota, South Dakota, and Kansas-Missouri (Fig. 8g-i; Yang et al., commonly mapped as open water at sites dominated by open water
2020). wetlands and lakes (Table 2).
In summary, we were able to validate open and vegetated water Given these sensor-specific differences, if open water is the primary
across different ecosystems, where we could be confident of the refer class of interest, the Sentinel-1 algorithm is the preferred sensor as it
ence point accuracy. We anticipate that as commercial, high-resolution showed similar accuracy (relative to non-water) as the Sentinel-2 algo
SAR imagery becomes more widely available, this approach will provide rithm, but less confusion with vegetated water and, is not limited by
a preferable mechanism for validating Sentinel-1 based algorithms that darkness or cloud cover. Further, as the nine inland sites in this study
track inundation dynamics in forested as well as non-forested locations. showed a significant, positive correlation in open water between the
The exploration of if and where the two algorithms can be integrated Sentinel-1 and Sentinel-2 algorithms (Table 4), time series for the open
will depend on the site-specific accuracy. Model accuracy varied pre water class can generally be integrated between the two sensors. For the
dictably with climate and landscape conditions, with some notable ex vegetated water class, despite a lower reported accuracy for the
ceptions. For example, in sites like Minnesota, where much of the year Sentinel-1-based vegetated water class (Table 2), most of the sites
has snow cover or deciduous leaf cover, the Sentinel-1 algorithm may showed either low differences in the proportion of mapped vegetated
provide a better ability to track inundation compared to a multispectral water or a positive, significant correlation when a monthly total vege
sensor like Sentinel-2 (Fig. 9). In forested areas like Arkansas, the tated water was tracked across the time series (Table 4, Fig. 9). Tidal
Sentinel-2 algorithm showed very low estimates of vegetated water wetlands showed a similar between sensor performance (Fig. 8), but
extent in summer months when leaf-cover was at its maximum, a pattern inland vegetated wetlands and riparian corridors showed a more mixed
not documented by the Sentinel-1 algorithm. However, despite the po performance. For instance, Fig. 12 shows a riparian corridor in south
tential advantages of the Sentinel-1 algorithm, overall the Sentinel-2 west Montana with adjacent irrigated agriculture where the soil mois
algorithm, although limited by cloud cover and dense leaf cover, ture status may be difficult to determine using visual cues alone. While
showed a higher accuracy in mapping vegetated water (Table 2). There both algorithms showed a decrease in surface water extent between
are several potential reasons for this finding. First, as a C-band SAR, wetter and drier conditions, differences between sensors, particularly in
penetration of backscatter signals through woody vegetation is not as how vegetated water was mapped, were larger than the response to
effective as it is for longer wavelengths such as P- and L-band SAR weather variability. In Minnesota and Arkansas-Louisiana we observed a
(Mahdavi et al., 2018; Shen et al., 2019). In addition, Sentinel-1 does not decline in the Sentinel-2-based vegetated water during peak growing
currently collect HH backscatter over terrestrial mid-latitudes, yet HH- season months when high leaf cover obscured water (Fig. 9). Addi
polarized backscatter typically outperforms VV in discriminating tionally, Arkansas-Louisiana and Texas both showed no correlations and
wetland type and mapped flooded forests (Hird et al., 2017; Mahdavi substantial mapped differences in vegetated water meaning that caution
et al., 2018). The lack of HH in mid-latitudes, compared to SAR satellites should be used if integrating the Sentinel-1 and Sentinel-2 vegetated
such as RadarSat-2, which delivers quad-polarized data, has been used water time series at these sites.
to help explain why Sentinel-2 may outperform Sentinel-1 in mapping
wetlands (Konapala et al., 2021). 4.4. Comparison to other Sentinel-1 and Sentinel-2 efforts
Further, while accessing Sentinel-1 imagery through Google Earth
Engine provides an efficient mechanism to apply a SAR-based algorithm, Comparison of accuracy across studies is an essential but compli
Google Earth Engine does not currently support Sentinel-1 Single Look cated activity as studies use different sensors, algorithms, apply algo
Complex (SLC) data, where each image pixel contains both amplitude rithms over different study areas, and map different numbers of water
and phase information and is represented by a complex (I and Q) classes (e.g., binary versus multiple water classes). Efforts to track
magnitude value. Instead, Google Earth Engine serves Sentinel-1 GRD inundation or flooding with Sentinel-1 saw 8–43% OE, 1–14% CE in the
products that contain the detected amplitude and are multi-looked to Prairie Pothole Region using a random forest model (Huang et al.,
reduce the impact of speckle. Using only GRD data limits the imple 2018), 14% OE, 1% CE in a study area in China with a thresholding
mentation of SAR-specific data tools such as decompositions or the approach (Li et al., 2020), and 8% OE, 12% CE for flood events across
normalized Kennaugh scattering matrix, which have been found to be CONUS (Yang et al., 2021), but these efforts all validated binary water
helpful in mapping surface water (Touzi et al., 2007; Henderson and classes. In comparison, our Sentinel-1 algorithm accuracy for open water
Lewis, 2008; Schmitt and Brisco, 2013). Lastly, speckle is a well know (3% OE, 1% CE) improved upon these values, while our accuracy of
issue of SAR data that degrades the radiometric quality of the image and vegetated water (28% OE, 16% CE) was lower in comparison. Yang et al.
can necessitate training samples that are coarser in resolution (Mahdavi (2020) used an unsupervised rule-based approach to track inundation
et al., 2018; Shen et al., 2019), relative to Sentinel-2, which may have from Sentinel-2 and reported 11% OE, 4% CE across France (Yang et al.,
19
Fig. 12. An example of open and vegetated water extent along a river riparian corridor and adjacent irrigated agriculture, derived from (a) the Sentinel-1 algorithm
in May 2018 (PDSI = 1.1) compared to (b) the Sentinel-1 algorithm in August 2020 (PDSI = − 0.2), (c) the Sentinel-2 algorithm in May 2018, (d) the Sentinel-2
algorithm in August 2020, (e) natural color Sentinel-2 image from May 2018 and (f) natural color Sentinel-2 image from August 2020.
2020). Goffi et al. (2020), in turn, tested single index thresholds from accuracy using Sentinel-2 relative to Sentinel-1 with a deep convolu
Sentinel-2 across multiple countries and found 19–34% OE, 22–35% CE tional neural network. While unsupervised algorithms (e.g., Jones,
within the best-performing single-index models, but both efforts also 2019; Yang et al., 2020) are more easily scalable, machine learning
validated binary water classes. More similar to our effort, Ludwig et al. approaches are also effective and common (e.g., Huang et al., 2018;
(2019) saw higher accuracy mapping open water with Sentinel-2 (1–4% Mahdianpari et al., 2020; Konapala et al., 2021) and enable algorithms
OE, 0–3% CE), relative to mapping mixed-pixel water (10–26% OE, to be trained to detect specific wetland and water types. The gradient
5–35% CE), reporting accuracy values similar to our Sentinel-2 algo boosted classifier, employed in this effort, relied on training data and
rithms accuracy for open water (3% OE, 0.5% CE) and vegetated water used a learning-based algorithm. Given the number of wetland types
classes (11% OE, 8% CE), but used a linear regression, tile-based that our algorithm was trained on (e.g., tidal, forested, emergent, and
approach. Mahdianpari et al. (2020) with a random forest, object- riparian wetlands, open water) the approach developed here can likely
based algorithm also observed lower accuracy when mapping multiple be extended to wetlands outside of the U.S.
wetland classes across Canada, reporting overall accuracy ranging from
74 to 84% by Canadian province. Further, similar to our finding that the
Sentinel-2 algorithm outperformed the Sentinel-1 algorithm when 4.5. Challenges to remotely detecting surface water
mapping vegetated water, Konapala et al. (2021) reported higher
Accurate identification of surface water from dense time series of
20
satellite imagery faces multiple challenges, one of which is that the map and monitor variability in surface water dynamics, particularly
variable duration of water features necessitates a per-image classifica under leaf-cover conditions such as forested wetlands, vegetated emer
tion approach. However, it is widely acknowledged across the remote gent wetlands, and rivers with dense riparian corridors (Papa et al.,
sensing community that image classification is typically more accurate 2022).
when mapped as the change between multiple images (Chuvieco et al.,
2019). While classification of flood extent can be identified as an 5. Conclusion
anomalous condition relative to images collected pre-flood (Cian et al.,
2018; DeVries et al., 2020), mapping total water extent, where some In this effort we developed two algorithms based on Sentinel-1 and
water bodies are permanent and others are variably inundated, requires Sentinel-2, respectively, at 20 m resolution to map open water and
a per-image classification. Consequently, most large-scale efforts to vegetated water areas for 2017–2021 across 12 sites representing
remotely detect surface water or wetlands spend an extended time different hydrological landscapes, totaling >536,000 km2. Globally, the
applying post-processing filters to reduce commission error in areas that United Nations Sustainable Development Goals include targets to ensure
are easily confused with water (e.g., topographic or vegetation shadows, sustainable supplies of freshwater, implement transboundary water
urban areas, irrigated agriculture) (Xie et al., 2016; Mahdavi et al., resource management, and protect and restore aquatic and water-
2018). In addition to our pre-processing steps in this effort, for example, dependent ecosystems, all of which require accurate data on surface
we reduced commission error in the vegetated water class by masking water dynamics. Sentinel-1 and Sentinel-2 satellites provide opportu
select land cover classes with the highest rates of confusion, applying a nities to improve upon existing surface water products. We found that
moving window filter, and excluding images with same day substantial we were able to track open and vegetated water across different vege
rain-events that can cause widespread, temporary soil saturation. tation and hydrological landscapes and that model performance was
However, initiatives to improve our understanding of where water col improved by including topographic and weather variables in the algo
lects, including lidar-derived and routinely updated stream networks, rithms. While open water showed a higher accuracy for both algorithms
wetland datasets, and floodplain maps (Kloiber et al., 2015; Moore and relative to vegetated water, the inclusion of both categories allows not
Dewald, 2016) could help per-image classifications more accurately just lakes and ponds to be monitored, but also wetland systems to be
track the presence of water at a given time. Nevertheless, our approach tracked over time. This work and other efforts to expand and improve
demonstrates that adequate accuracy can be achieved without the surface water products will enable more of the world's freshwater re
additional effort required to integrate those datasets or when such data sources, much of which is in narrow rivers and small wetlands, to be
are not available, more accurately tracked and better managed.
Efforts to track inundation extent from satellite imagery are
increasingly produced in cloud-based platforms. While cloud-based ap Author responsibilities
proaches enable more imagery to be processed in a time-efficient
manner, users can still be limited by pre-determined data formats or MV, LA, and JC conceived of the paper, MV and KS developed the
pre-processing steps performed on image collections delivered in the methods with assistance from LA and JC, PN, MS, KS, and MV processed
cloud. For example, in addition, to Google Earth Engine's Sentinel-1 the data, MV led the writing of the paper with support from LA, JC, PN
image collection lacking phase information, memory and computation and MS.
limits are still limiting factors. Moving from a 20-m Sentinel-2 product to
a 10-m Sentinel-2 product, for instance, quadruples the number of Declaration of Competing Interest
pixels, greatly slowing the processing and exporting of classified
imagery. The authors declare no conflicts of interest.
4.6. Alternative sensors to move remote mapping of water forward Data availability
Sentinel-1 and Sentinel-2, relative to Landsat, greatly improve the The training points, validation points, and 5-year percentile of open
temporal frequency of observations. While Sentinel-1 observation fre water and vegetated water for the Sentinel-1 and Sentinel-2 algorithms
quency was more spatially consistent, Sentinel-2 observations were across the 12 sites will be available for download at Vanderhoof (2022;
reduced by snow and cloud cover (Fig. 5), particularly in northern sites https://doi.org/10.5066/P9RSXQ2U) concurrent with the manuscript
that experience persistent seasonal snow cover and sites that experience publication.
substantial precipitation (Fig. 1). More frequent observations will enable
improved detection of short-lived events, such as peak flood extents, but Acknowledgements
also provides an opportunity to use a multi-image approach to reduce
classification uncertainty. Continued progress towards more complete This research was funded by the U.S. Geological Survey's National
remote monitoring of river networks and small wetlands may also be Land Imaging and Land Change Science Programs and the U.S. Envi
supported by high spatial (<5 m) and temporal resolution (on-demand ronmental Protection Agency's, Office of Research and Development
to near-daily) commercial, multi-spectral imagery that is increasingly through an interagency agreement (DW-014-92558501-0, “Multisource
becoming available, e.g., from sensors hosted on small satellites remote sensing to enhance national mapping of aquatic resources”).
(CubeSats) (Cooley et al., 2019). However, many of these CubeSats Maxar imagery was made available through the NextView License.
contain a limited number of multispectral bands which has been shown PlanetScope imagery was made available through the NASA CSDA-
to elevate spectral confusion with other land cover classes (Feng et al., Planet partnership. We appreciate comments on earlier versions from
2019). In addition, without a similarly high-resolution DEM, it is diffi Jessica Walker, David Williams, and the anonymous reviewers. We also
cult to obtain a high enough georeferencing accuracy to automate time appreciate the graphic development support from Jeremy Havens. Any
series analysis, particularly along narrow features, such as streams and use of trade, firm, or product names is for descriptive purposes only and
rivers. In the near-future, the NASA-ISRO SAR (NISAR) satellite, a dual does not imply endorsement by the U.S. Government. This publication
L-band and S-band satellite and the Surface Water Ocean Topography represents the views of the authors and does not necessarily reflect the
(SWOT) satellite, a Ka-band SAR instrument, will help further efforts to views or policies of the U.S. EPA.
21
Appendix
Table A1
Dates of image trios for each site from which training points were generated.
Landsat-8 Sentinel-1 Sentinel-2 Landsat-8 Sentinel-1 Sentinel-2 Landsat-8 Sentinel-1 Sentinel-2
Arizona Kansas-Missouri North Dakota

5/5/2017 5/6/2017 5/5/2017 7/21/2018 7/21/2018 7/21/2018 11/7/2016 11/15/2016 11/7/2016
10/21/2017 10/16/2017 10/22/2017 11/10/2018 11/6/2018 11/13/2018 8/6/2017 7/1/2017 7/5/2017
5/17/2018 5/20/2018 5/20/2018 3/18/2019 3/18/2019 3/18/2019 5/21/2018 5/21/2018 5/21/2018
10/24/2018 10/23/2018 10/27/2018 4/21/2020 4/17/2020 4/21/2020 8/5/2018 8/13/2018 8/9/2018
6/21/2019 6/20/2019 6/19/2019 8/27/2020 8/27/2020 8/24/2020 4/22/2019 4/22/2019 4/26/2019
7/14/2019 7/19/2019 7/14/2019 10/14/2020 10/14/2020 10/13/2020 10/12/2019 10/7/2019 10/8/2019
1/6/2020 1/15/2020 1/5/2020 2/19/2021 2/16/2021 2/20/2021 3/16/2020 3/11/2020 3/16/2020
1/31/2020 2/3/2020 1/30/2020 3/7/2021 3/7/2021 3/2/2021 6/4/2020 6/3/2020 6/4/2020
Arkansas-Louisiana Maryland-Delaware South Carolina
1/5/2018 1/5/2018 1/1/2018 7/2/2016 7/2/2016 6/30/2016 1/20/2018 1/22/2018 1/19/2018
4/11/2018 4/11/2018 4/11/2018 11/23/2016 11/23/2016 11/27/2016 3/9/2018 3/4/2018 3/10/2018
7/23/2018 7/21/2018 7/23/2018 4/16/2017 4/16/2017 4/16/2017 5/12/2018 5/10/2018 5/9/2018
1/24/2019 1/24/2019 1/21/2019 3/18/2018 3/18/2018 3/22/2018 7/18/2019 7/21/2019 7/18/2019
8/11/2019 8/9/2019 8/12/2019 8/9/2018 8/9/2018 8/9/2018 9/19/2019 9/19/2019 9/16/2019
11/17/2020 11/19/2020 11/19/2020 3/14/2019 3/13/2019 3/12/2019 2/27/2020 2/22/2020 2/28/2020
11/26/2020 11/26/2020 11/26/2020 10/15/2019 10/15/2019 10/18/2019 6/1/2020 5/28/2020 6/2/2020
4/19/2021 4/19/2021 4/20/2021 10/17/2020 10/21/2020 10/17/2020 4/2/2021 4/5/2021 4/3/2021
California Minnesota South Dakota
6/30/2017 7/1/2017 7/4/2017 8/1/2017 7/27/2017 8/1/2017 6/5/2017 6/7/2017 6/2/2017
8/17/2017 8/18/2017 8/18/2017 5/16/2018 5/15/2018 5/15/2018 4/28/2018 4/27/2018 4/28/2018
9/21/2018 9/18/2018 9/17/2018 12/10/2018 12/13/2018 12/6/2018 7/17/2018 7/20/2018 7/17/2018
10/7/2018 10/6/2018 10/12/2018 3/16/2019 3/19/2019 3/16/2019 10/21/2018 10/24/2018 10/20/2018
1/27/2019 1/28/2019 1/30/2019 3/2/2020 3/1/2020 2/27/2020 7/13/2019 7/15/2019 7/12/2019
3/16/2019 3/17/2019 3/11/2019 4/19/2020 4/18/2020 4/19/2020 9/6/2019 9/1/2019 9/5/2019
6/4/2019 6/3/2019 6/4/2019 9/10/2020 9/9/2020 9/11/2020 8/23/2020 8/26/2020 8/25/2020
5/5/2020 5/4/2020 5/4/2020 10/12/2020 10/15/2020 10/9/2020 10/10/2020 10/13/2020 10/9/2020
Iowa Montana Texas
5/21/2016 5/21/2016 5/19/2016 8/11/2016 8/4/2016 8/4/2016 4/25/2018 4/21/2018 4/24/2018
7/8/2016 7/20/2016 7/11/2016 7/4/2017 7/5/2017 7/3/2017 7/14/2018 7/19/2018 7/18/2018
4/22/2017 4/22, 4/16/17 4/17/2017 8/21/2017 8/22/2017 8/27/2017 9/16/2018 9/12/2018 9/16/2018
6/9/2017 6/9/2017 6/3/2017 5/20/2018 5/20/2018 5/24/2018 3/27/2019 4/4/2019 3/20/2019
9/13/2017 9/13/2017 9/11/2017 7/23/2018 7/24/2018 7/23/2018 11/6/2019 11/6/2019 11/10/2019
10/21/2019 10/21, 10/15/19 10/14/2019 3/20/2019 3/21/2019 3/20/2019 12/26/2020 12/30/2020 12/24/2020
3/6/2020 3/1/2020 3/4/2020 10/30/2019 10/30/2019 11/5/2019 4/1/2021 4/5/2021 3/29/2021
9/21/2020 9/21/2020 9/20/2020 12/3/2020 11/29/2020 12/4/2020 8/23/2021 8/27/2021 8/21/2021
22
Table A2
Variables considered for potential inclusion in the Sentinel-1 and Sentinel-2 models. Variables considered for inclusion in both models are highlighted in gray.
23
Table A3
Validation images and corresponding Sentinel-1 and Sentinel-2 images.
Site Sensor Reference Date Sentinel-1 Date Sentinel-2 Date Sentinel-1 Date Gap Sentinel-2 Date Gap
Arizona PlanetScope 4/7/2020 4/8/2020 4/9/2020 1 2

Arizona PlanetScope 4/24/2020 4/20/2020 4/24/2020 − 4 0
Arizona PlanetScope 11/27/2021 11/29/2021 11/25/2021 2 − 2
Arkansas-Louisiana WorldView− 2 11/20/2017 11/18/2017 11/22/2017 − 2 2
Arkansas-Louisiana WorldView-3 4/15/2019 4/30/2019 4/16/2019 15 1
Arkansas-Louisiana WorldView− 2 8/11/2020 8/15/2020 8/11/2020 4 0
California WorldView-3 10/7/2017 10/5/2017 10/5/2017 -2 -2
California WorldView-3 4/3/2020 4/4/2020 4/2/2020 1 − 1
California WorldView-2 10/14/2020 10/14/2020 10/16/2020 0 2
Iowa WorldView-3 10/21/2018 10/20/2018 10/21/2018 − 1 0
Iowa WorldView-3 5/4/2019 5/6/2019 5/4/2019 2 0
Iowa WorldView-2 6/17/2020 6/17/2020 6/17/2020 0 0
Kansas-Missouri WorldView-2 4/9/2019 4/4/2019 4/15/2019 − 5 6
Kansas-Missouri WorldView-2 7/27/2019 7/26/2019 7/29/2019 − 1 2
Kansas-Missouri WorldView-2 11/7/2019 11/6/2019 11/6/2019 − 1 − 1
Maryland-Delaware WorldView-2 2/20/2017 2/22/2017 2/25/2017 2 5
Maryland-Delaware PlanetScope 2/26/2019 3/1/2019 2/25/2019 3 − 1
Maryland-Delaware PlanetScope 4/5/2021 4/2/2021 4/5/2021 − 3 0
Minnesota PlanetScope 5/5/2019 5/6/2019 5/15/2019 1 10
Minnesota PlanetScope 4/30/2020 4/30/2020 4/29/2020 0 − 1
Minnesota PlanetScope 10/17/2021 10/22/2021 10/16/2021 5 − 1
Montana WorldView-3 6/20/2017 6/18/2017 6/20/2017 − 2 0
Montana PlanetScope 5/29/2019 5/27/2019 5/29/2019 − 2 0
Montana PlanetScope 6/2/20 & 6/4/20 6/2/2020 6/2/2020 0 0
North Dakota WorldView-2 8/7/2017 8/6/2017 8/7/2017 − 1 0
North Dakota WorldView-2 10/27/2017 10/29/2017 10/28/2017 2 1
North Dakota WorldView-2 10/8/2019 10/7/2019 9/28/2019 − 1 − 10
South Carolina WorldView-2 3/13/2018 3/16/2018 3/15/2018 3 2
South Carolina WorldView-2 4/3/2019 3/23/2019 4/4/2019 − 11 1
South Carolina WorldView-3 4/30/2019 4/28/2019 4/29/2019 − 2 − 1
South Dakota WorldView− 3 8/15/2019 8/20/2019 8/16/2019 5 1
South Dakota PlanetScope 6/1/2020 6/3/2020 6/1/2020 2 0
South Dakota PlanetScope 7/12/2020 7/9/2020 7/11/2020 -3 − 1
Texas PlanetScope 4/9/2019 4/4/2019 4/9/2019 − 5 0
Texas WorldView-3 8/18/2020 8/20/2020 8/19/2020 2 1
Texas PlanetScope 10/16/2021 10/14/2021 10/5/2021 − 2 − 11
24
Fig. A1. The Dynamic Surface Water Extent (DSWE) mapped classes were used to stratify initial draft training points for the open and vegetated water classes.
Examples of water class 1 (high confidence), class 2 (moderate confidence), class 3 (partial surface water – conservative), and class 4 (partial surface water –
aggressive), as mapped by Jones (2019) and class 5 (partial surface water – forest, as added by Vanderhoof et al. (2020)) for April 2021 are shown for (a) tidal
emergent wetlands, (b) Landsat 8 image collected April 20, 2021, (c) forested wetlands, and (d) Landsat image collected April 4, 2021.
Appendix References
Abatzoglou J. T., 2013. Development of gridded surface meteorological data for ecological applications and modeling. Int. J. Climatol. 33 (1),
121–131.
Adell, C., Puech, C., 2003. L'analyse spatiale des plans d'eau extraits par télédétection satellitale permet-elle de retrouver la marque cynégétique en
Camargue? Bull. Soc. Fr. Photogramm. Teledetect. 172, 76–86.
Broge, N.H., Leblanc, E., 2001. Comparing prediction power and stability of broadband and hyperspectral vegetation indices for estimation of
green leaf area index and canopy chlorophyll density. Remote Sens. Environ. 76, 156–172.
Caillaud, L., Guillaumont, B., Manaud, F., 1987. Essai de discrimination des modes d'utilisation des marais maritimes par analyse multitemporelle
d'images spot. Application aux marais maritimes du Centre Ouest. https://archimer.ifremer.fr/doc/00446/55728/. Accessed September 20, 2019.
Crippen, R.E., 1990. Calculating the vegetation index faster. Remote Sens. Environ. 34, 71–73.
Davranche, A., Poulin, B., Lefebvre, G., 2013. Mapping flooding regimes in Camargue wetlands using seasonal multispectral data. Remote Sens.
25
Environ. 138, 165–171.

De Pondeca, M.S., Manikin, G.S., DiMego, G., Benjamin, S.G., Parrish, D.F., Purser, R.J., Wu, W.S., Horel, J.D., Myrick, D.T., Lin, Y., Aune, R.M.,
Keyser, D., Colman, B., Mann, G., Vavra, J., 2011. The real-time mesoscale analysis at NOAA's national centers for environmental prediction: Current
status and development. Weather Forecast. 26, 593–612.
Feyisa, G.L., Melby, H., Fehsholt, R., Proud, S.R., 2014. Automated Water Extraction Index: A new technique for surface water mapping using
Landsat imagery. Rem. Sens. Environ. 140, 23–25.
Gao, B., 1996. NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens. Environ. 58,
257–266.
Gautam, V.K., Murugan, P., Annadurai, M., 2017. A new three band index for identifying urban areas using satellite images. ICGCSC Proc. Vol.
Inter. Conf. 27–31.
Gond, V., Bartholomé, E., Ouattara, F., Nonguierma, A., Bado, L., 2004. Surveillance et cartographie des plans d'eau et des zones humides et
inondables en régions arides avec l'instrument VEGETATION embarqué sur SPOT-4. Int. J. Remote. Sens. 25, 987–1004.
Huete, A.R., 1989. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25, 295–309.
Hunt Jr., E., Rock. B., 1989. Detection of changes in leaf water content using Near- and Middle-Infrared reflectances. Remote Sens. Environ. 30,
43–54.
Kameyama, S., Yamagata, Y., Nakamura, F., Kaneko, M., 2001. Development of WTI and turbidity estimation model using SMA — application to
Kushiro Mire, eastern Hokkaido. Jpn. Remote Sens. Environ. 77, 1–9.
Lillesand, T.M., Kifer, R.W., 1987. Remote Sensing and Image Interpretation, second ed. New York, New York.
McFeeters, S.K., 1996. The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features. Int. J. Remote Sens. 17,
1425–1432.
Richardson, A.J., Everitt, J.H., 1992. Using spectral vegetation indices to estimate rangeland productivity. Geocarto Int. 7, 63–69.
Rikimaru, A., Roy, P.S., Miyatake, S., 2002. Tropical forest cover density mapping. Trop. Ecol. 43, 39–47.
Rondeaux, G., Steven, M., Baret, F., 1996. Optimization of soil-adjusted vegetation indices. Remote Sens. Environ. 55, 95–107.
Shen, L., Li, C., 2010. Water body extraction from Landsat ETM+ imagery using adaboost algorithm. Proc. 2010 18th Int. Conf. Geoinf. Beijing,
China 1–4.
Thornton, M.M., Shrestha, R., Wei, Y., Thornton, P.E., Kao, S., Wilson, B.E., 2022. Daymet: daily surface weather data on a 1-km grid for North
America, version 4. ORNL DAAC, Oak Ridge, Tennessee, USA. https://doi.org/10.3334/ORNLDAAC/1840.
Tucker C.J., 1979. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 8, 127–150.
Xu, H., 2006. Modification of normalized difference water index (NDWI) to enhance open water features in remotely sensed imagery. Int. J.
Remote Sens. 27, 3025–3033.
Yamazaki, D., Ikeshima, D., Sosa, J., Bates, P.D., Allen, G.H., Pavelsky, T.M., 2019. MERIT Hydro: A high-resolution global hydrography map based
on latest topography datasets. Water Resour. Res. 55, 5053–5073.
References Feng, D., Gleason, C.J., Yang, X., Pavelsky, T., 2019. Comparing discharge estimates
made via the BAM algorithm in high-order Arctic rivers derived solely from optical
CubeSat, Landsat, and Sentinel-2 data. Water Resour. Res. 55, 7753–7771.
Abatzoglou, J.T., 2013. Development of gridded surface meteorological data for
Feyisa, G.L., Melby, H., Fehsholt, R., Proud, S.R., 2014. Automated water extraction
ecological applications and modelling. Int. J. Climatol. 33 (1), 121–131.
index: a new technique for surface water mapping using Landsat imagery. Rem. Sens.
Ali, I., Cao, S., Naeimi, V., Paulik, C., Wagner, W., 2018. Methods to remove the border
Environ. 140, 23–25.
noise from Sentinel-1 synthetic aperture radar data: implications and importance for
Fuentes, I., Padarian, J., Van Ogtrop, F., Vervoort, R., 2019. Spatiotemporal evaluation of
time-series analysis. IEEE JSel. Top. Appl. Earth Obs. Remote Sens. 11, 777–786.
inundated areas using MODIS imagery at a catchment scale. J. Hydrol. 573,
Bazzi, H., Baghdadi, N., El Hajj, M., Zribi, M., 2019. Potential of Sentinel-1 surface soil
952–963.
moisture product for detecting heavy rainfall in the south of France. Sensors 19 (4),
Gautam, V.K., Murugan, P., Annadurai, M., 2017. A new three band index for identifying
802.
urban areas using satellite images. ICGCSC Proc. Vol. Inter. Conf. 27–31.
Benstead, J.P., Leigh, D.S., 2012. An expanded role for river networks. Nat. Geosci. 5,
Goffi, A., Stroppiana, D., Brivio, P., Bordogna, G., Boschetti, M., 2020. Towards an
678–679.
automated approach to map flooded areas from Sentinel-2 MSI data and soft
Bioresita, F., Puissant, A., Stumpf, A., Malet, J.P., 2019. Fusion of Sentinel-1 and
integration of water spectral features. Int. J. Appl. Earth Obs. Geoinf. 84, 101951.
Sentinel-2 image time series for permanent and temporary surface water mapping.
Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., Moore, R., 2017.
Int. J. Remote Sens. 40 (23), 9026–9049.
Google Earth Engine: planetary-scale geospatial analysis for everyone. Remote Sens.
Chuvieco, E., Mouillot, F., van der Werf, G.R., Miguel, J., Tanasse, M., Koutsias, N.,
Environ. 202, 18–27.
Garcia, M., Yebra, M., Padilla, M., Gitas, I., Heil, A., Hawbaker, T.J., Giglio, L., 2019.
Guolin, K., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., Liu, T.Y., 2017.
Historical background and current developments for mapping burned area from
LightGBM: a highly efficient gradient boosting decision tree. Adv. Neural Inf.
satellite earth observation. Remote Sens. Environ. 225, 45–64.
Process. Syst. 30, 3149–3157.
Cian, F., Marconcini, M., Ceccato, P., 2018. Normalized difference flood index for rapid
Hall, D.K., Riggs, G.A., 2021. MODIS/Terra Snow Cover Daily L3 Global 500m SIN Grid,
flood mapping: taking advantage of EO big data. Remote Sens. Environ. 209,
Version 6.1. NASA National Snow and Ice Data Center Distributed Active Archive
712–730.
Center, Boulder, Colorado. https://doi.org/10.5067/MODIS/MOD10A1.061.
Cooley, S.W., Smith, L.C., Ryan, J.C., Pitcher, L.H., Pavelsy, T.M., 2019. Arctic-boreal
Hastie, T., Tibshirani, R., Friedman, J., 2009. The Elements of Statistical Learning; Data
lake dynamics revealed using CubeSat imagery. Geophys. Res. Lett. 46 (4),
Mining, Inference, and Prediction, 2nd ed. Springer, New York, NY, USA.
2111–2120.
Hawbaker, T.J., Vanderhoof, M.K., Schmidt, G.L., Beal, Y.J., Picotte, J.J., Takacs, J.D.,
Davranche, A., Poulin, B., Lefebvre, G., 2013. Mapping flooding regimes in Camargue
Falgout, J.T., Dwyer, J.L., 2020. The Landsat burned area algorithm and products for
wetlands using seasonal multispectral data. Remote Sens. Environ. 138, 165–171.
the conterminous United States. Remote Sens. Environ. 244, 111801.
De Pondeca, M.S., Manikin, G.S., DiMego, G., Benjamin, S.G., Parrish, D.F., Purser, R.J.,
Heimhuber, V., Tulbure, M., Broich, M., 2017. Modeling multidecadal surface water
Wu, W.S., Horel, J.D., Myrick, D.T., Lin, Y., Aune, R.M., Keyser, D., Colman, B.,
inundation dynamics and key drivers on large river basin scale using multiple time
Mann, G., Vavra, J., 2011. The real-time mesoscale analysis at NOAA's national
series of earth-observation and river flow data. Water Resour. Res. 53, 1251–1269.
centers for environmental prediction: current status and development. Weather
Henderson, F., Lewis, A., 2008. Radar detection of wetland ecosystems: a review. Int. J.
Forecast. 26, 593–612.
Remote Sens. 29, 5809–5835.
DeVries, B., Huang, C., Armston, J., Huang, W., Jones, J., Lang, M., 2020. Rapid and
Hird, J.N., DeLancey, E.R., McDermid, G.J., Kariyeva, J., 2017. Google Earth Engine,
robust monitoring of flood events using Sentinel-1 and Landsat data on the Google
Open-Access satellite data, and machine learning in support of large-area
Earth Engine. Remote Sens. Environ. 240, 111664.
probabilistic wetland mapping. Remote Sens. 9, 1315.
Donchyts, G., Baart, F., Winsemius, H., Gorelick, N., Kwadijk, J., van de Giesen, N., 2016.
Homer, C., Dewitz, J., Jin, S., Xian, G., Costello, C., Danielson, P., Gass, L., Funk, M.,
Earth’s surface water change over the past 30 years. Nat. Clim. Chang. 6, 810–813.
Wickham, J., Stehman, S., Auch, R., Riiters, K., 2020. Conterminous United States
Endarwin, E., Hadi, S., Bayong, H.K., Gunawan, D., Siswanto, S., 2014. Modified
land cover change patterns 2001–2016 from the 2016 National Land Cover
convective stratiform technique (CSTm) performance on rainfall estimation in
Database. ISPRS J. Photogramm. Remote Sens. 162, 184–199.
Indonesia. J. Math. Fundam. Sci. 46, 251–268.
26
Huang, W., DeVries, B., Huang, C., Lang, M.W., Jones, J.W., Creed, I.F., Carroll, M.L., Prigent, C., Papa, F., Aires, F., Jimenez, C., Rossow, W.B., Matthews, E., 2012. Changes in
2018. Automated extraction of surface water extent from Sentinel-1 data. Remote land surface water dynamics since the 1990s and relation to population pressure.
Sens. 10 (5), 797. Geophy. Res. Let. 39 (8), L08403.
Jones, J.W., 2015. Efficient wetland surface water detection and monitoring via Landsat: PRISM Climate Group, 2022. Oregon State University (accessed 7 April 2022). http://pr
comparison with in situ data from the Everglades Depth Estimation Network. ism.oregonstate.edu.
Remote Sens. 7, 12503–12538. Sabel, D., Bartalis, Z., Wagner, W., Doubkova, M., Klein, J.P., 2012. Development of a
Jones, J.W., 2019. Improved automated detection of subpixel-scale inundation—revised global backscatter model in support to the Sentinel-1 mission design. Remote Sens.
Dynamic Surface Water Extent (DSWE) partial surface water tests. Remote Sens. 11, Environ. 120, 102–112. https://doi.org/10.1016/j.rse.2011.09.028.
374. Sahour, H., Kemink, K.M., O’Connell, J., 2022. Integrating SAR and optical remote
Kloiber, S.M., Macleod, R.D., Smith, A.J., Knight, J.F., Huberty, B.J., 2015. A semi- sensing for conservation-targeted wetlands mapping. Remote Sens. 14, 159.
automated, multi-source data fusion update of a wetland inventory for east-Central Schlaffer, S., Chini, M., Dorigo, W., Plank, S., 2022. Monitoring surface water dynamics
Minnesota,USA. Wetlands 35, 335–348. in the prairie pothole region of North Dakota using dual-polarised Sentinel-1
Konapala, G., Kumar, S., Ahmad, S., 2021. Exploring Sentinel-1 and Sentinel-2 diversity synthetic aperture radar (SAR) time series. Hydrol. Earth Syst. Sci. 26, 841–860.
for flood inundation mapping using deep learning. ISPRS J. Photogramm. Remote Schmitt, A., Brisco, B., 2013. Wetland monitoring using the curvelet-based change
Sens. 180, 163–173. detection method on polarimetric SAR imagery. Water. 5, 1036–1051.
Lane, C.R., D’Amico, E., 2016. Identification of putative geographically isolated wetlands Shen, X., Wang, D., Mao, K., Anagnostou, E., Hong, Y., 2019. Inundation extent mapping
of the conterminous United States. J. Am. Water Resour. Assoc. 52 (3), 705–722. by Synthetic Aperture Radar: a review. Remote Sens. 11 (7), 879.
Lechner, A.M., Stein, A., Jones, S.D., Ferwerda, J.G., 2009. Remote sensing of small and Slagter, B., Tsendbazar, N., Vollrath, A., Reiche, J., 2020. Mapping wetland
linear features: quantifying the effects of patch size and length, grid position and characteristics using temporally dense Sentinel-1 and Sentinel-2 data: a case study in
detectability on land cover mapping. Remote Sens. Environ. 113 (10), 2194–2204. the St. Lucia wetlands, South Africa. Int. J. Appl. Earth Obs. Geoinf. 86, 1–11.
Lefebvre, G., Davranche, A., Willm, L., Campagna, J., Redmond, L., Merle, C., Small, D., 2011. Flattening gamma: radiometric terrain correction for SAR imagery. IEEE
Guelmami, A., Poulin, B., 2019. Introducing WIW for detecting the presence of water Trans. Geosci. Remote Sens. 49, 3081–3093.
in wetlands with Landsat and Sentinel satellites. Remote Sens. 11, 2210. Soulard, C.E., Waller, E.K., Walker, J.J., Petrakis, R.E., Smith, B.W., 2022. DSWEmod —
Leibowitz, S.G., Wigington Jr., P.J., Schofield, K.A., Alexander, L.C., Vanderhoof, M.K., the production of high-frequency surface water map composites from daily MODIS
Golden, H.E., 2018. Connectivity of streams and wetlands to downstream waters: an images. J. Am. Water Resour. Assoc. 58, 248–268.
integrated systems framework. J. Am. Water Resour. Assoc. 54 (2), 298–322. Song, C., Ke, L., Pan, H., Zhan, S., Liu, K., Ma, R., 2018. Long-term surface water changes
Li, Y., Niu, Z., Xu, Z., Yan, X., 2020. Construction of high spatial-temporal water body and driving cause in Xiong-an, China: from dense Landsat time series images and
dataset in China based on Sentinel-1 archives and GEE. Remote Sens. 12, 2413. synthetic analysis. Sci.Bull. 63 (11), 708–716.
Ludwig, C., Walli, A., Schleicher, C., Weichselbaum, J., Riffler, M., 2019. A highly Strahler, A.H., Boschetti, L., Foody, G.M., Friedl, M.A., Hansen, M.C., Herold, M.,
automated algorithm for wetland detection using multi-temporal optical satellite Mayaux, P., Morisette, J.T., Stehman, S.V., Woodcock, C.E., 2006. In: Global Land
data. Remote Sens. Environ. 224, 333–351. Cover Validation: Recommendations for Evaluation and Accuracy Assessment of
Ma, S., Zhou, Y., Gowda, P.H., Dong, J., Zhang, G., Kakani, V.G., Wagle, P., Chen, L., Global Land Cover Maps. European Commission Joint Research Centre EUR, Ispra,
Flynn, K.C., Jiang, W., 2019. Application of the water-related spectral reflectance Italy, p. 22156.
indices: a review. Ecol. Indic. 98, 68–79. Taylor, J.B., Sullivan, J.D., Teitelbaum, C.S., Reese, J.G., Prosser, D.J., 2022. Comparing
Mahdavi, S., Salehi, B., Granger, J., Amani, M., Brisco, B., Huang, W., 2018. Remote Landsat Dynamic Surface Water Extent to alternative methods of measuring
sensing for wetland classification: a comprehensive review. GISci. Remote Sens. 55 inundation in developing waterbird habitats. Remote Sens. App.: Soc. Environ. 28,
(5), 623–658. 100845.
Manakos, I., Kordelas, G.A., Marini, K., 2020. Fusion of Sentinel-1 data with Sentinel-2 Thornton, M.M., Shrestha, R., Wei, Y., Thornton, P.E., Kao, S., Wilson, B.E., 2022.
products to overcome non-favourable atmospheric conditions for the delineation of Daymet: daily surface weather data on a 1-km grid for North America, version 4.
inundation maps. Eur. J. Remote Sens. 53, 53–66. ORNL DAAC, Oak Ridge, Tennessee, USA. https://doi.org/10.3334/ORNLDAAC/
Mahdianpari, M., Salehi, B., Mohammadimanesh, F., Brisco, B., Homayouni, S., Gill, E., 1840.
DeLancey, E., Bourgeau-Chavez, L., 2020. Big data for a big country: the first Topouzelis, K., Singha, S., Kitsiou, D., 2016. Incidence angle normalization of wide swath
generation of Canadian wetland inventory map at a spatial resolution of 10-m using SAR data for oceanographic applications. Open Geosci. 8 (1), 450–464.
Sentinel-1 and Sentinel-2 data on the Google Earth Engine cloud computing Touzi, R., Deschamps, A., Rother, G., 2007. Wetland characterization using polarimetric
platform. Can. J. Remote. Sens. 46, 1–19. RADARSAT-2 capability. Can. J. Remote. Sens. 33 (1), S56–S67.
Mishra, V., Cherkauer, K., 2011. Influence of cold season climate variability on lakes and Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring
wetlands in the Great Lakes region. J. Geophys. Res. 116, 1–21. vegetation. Remote Sens. Environ. 8, 127–150.
Moore, R.B., Dewald, T.G., 2016. The road to NHDPlus – advancements in digital stream Tulbure, M., Broich, M., 2019. Spatiotemporal patterns and effects of climate and land
networks and associated catchments. J. Am. Water Resour. Assoc. 52 (4), 890–900. use on surface water extent dynamics in a dryland region with three decades of
Page, B.P., Olmanson, L.G., Mishra, D.R., 2019. A harmonized image processing Landsat satellite data. Sci. Total Environ. 658, 1574–1585.
workflow using Sentinel-2/MSI and Landsat-8/OLI for mapping water clarity in Tulbure, M.G., Broich, M., Perin, V., Gaines, M., Ju, J., Stehman, S.V., Pavelsky, T.,
optically variable lake systems. Remote Sens. Environ. 231, 111284. Masek, J.G., Yin, S., Mai, J., Betbeder-Matibet, L., 2022. Can we detect more
Pahlevan, N., Chittimalli, S.K., Balasubramanian, S.V., Vellucci, V., 2019. Sentinel-2/ ephemeral floods with higher density harmonized Landsat Sentinel 2 data compared
Landsat-8 product consistency and implications for monitoring aquatic systems. to Landsat 8 alone? ISPRS J. Photogramm. Remote Sens. 185, 232–246.
Remote Sens. Environ. 220, 19–29. US Fish and Wildlife Service (USFWS), 2019. National Wetlands Inventory – Version 2 –
Papa, F., Crétaux, J.F., Grippa, M., Robert, E., Trigg, M., Tshimanga, R.M., Kitambo, B., Surface Waters and Wetlands inventory. U.S. Department of the Interior, Fish and
Paris, A., Carr, A., Fleischmann, A.S., deFleury, M., Gbetkom, P.G., Calmettes, B., Wildlife Service, Washington, D.C. (last accessed October 5, 2021). http://www.fws.
Calmant, S., 2022. Water resources in Africa under global change: monitoring gov/wetlands
surface waters from space. Surv. Geophys. https://doi.org/10.1007/s10712-022- U.S. Geological Survey, 2019. National Hydrography Dataset. U.S. Department of the
09700-9. Interior, U.S. Geological Survey, Reston, VA (last accessed October 5, 2021).
Park, J., Kumar, M., Lane, C.R., Basu, N.B., 2022. Seasonality of inundation in https://www.usgs.gov/national-hydrography/national-hydrography-dataset.
geographically isolated wetlands across the United States. Environ. Res. Let. 17 (5), Van Meter, K.J., Basu, N.B., 2015. Signatures of human impact: size distributions and
054005. spatial organization of wetlands in the prairie pothole landscape. Ecol. Appl. 25 (2),
Parra, G., Guerrero, F., Armengol, J., Brendonck, L., Brucet, S., Finlayson, C.M., Gomes- 451–465.
Barbosa, L., Grillas, P., Jeppesen, E., Ortega, F., Vega, R., Zohary, T., 2021. The Vanderhoof, M.K., Lane, C., McManus, M., Alexander, L., Christensen, J., 2018. Wetlands
future of temporary wetlands in drylands under global change. Inland Waters. 11 inform how climate extremes influence surface water expansion and contraction.
(4), 445–456. Hydrol. Earth Syst. Sci. 22, 1851–1873.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Vanderhoof, M.K., Lane, C., 2019. The potential role of very high-resolution imagery to
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., 2011. Scikit-learn: machine characterize lake, wetland, and stream systems across the prairie pothole region,
learning in Python. J. Mach. Learn. Res. 12, 2825–2830. United States. Int. J. Remote Sens. 40, 5768–5798.
Pekel, J., Cottam, A., Gorelick, N., Belward, A., 2016. High-resolution mapping of global Vanderhoof, M.K., Christensen, J., Beal, Y.G., DeVries, B., Lang, M.W., Hwang, N.,
surface water and its long-term changes. Nature 540, 418–436. Mazzarella, C., Jones, J.W., 2020. Isolating anthropogenic wetland loss by
Pickens, A., Hansen, M., Hancher, M., Stehman, S., Tyukavina, A., Potapov, P., concurrently tracking inundation and land cover disturbance across the mid-Atlantic
Marroquin, B., Sherani, Z., 2020. Mapping and sampling to characterize global region, U.S. Remote Sens. 12, 1464.
inland water dynamics from 1999 to 2018 with full Landsat time-series. Remote Vanderhoof, M.K., Alexander, L., Christensen, J., Solvik, K., Nieuwlandt, P.,
Sens. Environ. 243, 111792. Sagehorn, M., 2022. Sentinel-1 and Sentinel-2 based frequency of open and
Poff, N.L., Bledsoe, B.P., Cuhaciyan, C.O., 2006. Hydrologic variation with land use vegetated water across the United States (2017-2021). U.S. Geological Survey data
across the contiguous United States: geomorphic and ecological consequences for release, Sciencebase, 10.5066/P9RSXQ2U.
stream ecosystems. Geomorphology 79 (3–4), 264–285. Vanderhoof, M.K., Hawbaker, T.J., Ku, A., Teske, C., Noble, J., Picotte, J., 2021. Mapping
Prats-Iraola, P., Nannini, M., Scheiber, R., De Zan, F., Wollstadt, S., Minati, F., wetland burned area from Sentinel-2 across the southeastern United States and its
Vecchioli, F., Costantini, M., Borgstrom, S., De Martino, P., Siniscalchi, V., contributions relative to landsat 8 (2016–2019). Fire 4 (3), 52.
Walter, T., Foumelis, M., Desnos, Y.L., 2015. In: Sentinel-1 assessment of the Wickham, J., Stehman, S.V., Sorenson, D.G., Gass, L., Dewitz, J.A., 2021. Thematic
interferometric wide-swath mode. In: Int. Geosci. Remote Sens. Symp. 2015-Novem, accuracy assessment of the NLCD 2016 land cover for the conterminous United
pp. 5248–5251. https://doi.org/10.1109/IGARSS.2015.7327018. States. Remote Sens. Environ. 257, 112357.
Wu, Q., 2020. geemap: A Python package for interactive mapping with Google Earth
Engine. J. Open Source Softw. 5, 2305.
27
Xia, H., Zhao, J., Qin, Y., Yang, J., Cui, Y., Song, H., Ma, L., Jin, N., Meng, Q., 2019. Yang, Q., Shen, X., Anagnostou, E.N., Mo, C., Eggleston, J.R., Kettner, A.J., 2021. A high-
Changes in water surface area during 1989–2017 in the Huai River Basin using resolution flood inundation archive (2016-the present) from Sentinel-1 SAR imagery
Landsat data and Google Earth Engine. Remote Sens. 11 (15), 1824. over CONUS. Bull. Am. Meteor. Soc. 102, E1064–E1079. https://doi.org/10.1175/
Xie, H., Luo, X., Xu, X., Pan, H., Tong, X., 2016. Automated subpixel surface water BAMS-D-19-0319.1.
mapping form heterogeneous urban environments using Landsat 8 OLI imagery. Yao, F., Wang, J., Wang, C., Crétaux, J.F., 2019. Constructing long-term high-frequency
Remote Sens. 8 (7), 584. time series of global lake and reservoir areas using Landsat imagery. Remote Sens.
Xu, H., 2006. Modification of normalised difference water index (NDWI) to enhance open Environ. 232, 111210.
water features in remotely sensed imagery. Int. J. Remote Sens. 27, 3025–3033. Yeo, I., Lang, M., Lee, S., McCarty, G., Sadeghi, A., Yetemen, O., Huang, C., 2019.
Yamazaki, D., Trigg, M.A., 2016. The dynamics of Earth’s surface water. Nature 540, Mapping landscape-level hydrological connectivity of headwater wetlands to
348–349. downstream waters: a geospatial modeling approach - part 1. Sci. Total Environ.
Yamazaki, D., Ikeshima, D., Sosa, J., Bates, P.D., Allen, G.H., Pavelsky, T.M., 2019. 653, 1546–1556.
MERIT hydro: a high-resolution global hydrography map based on latest topography Yuan, T., Lee, H., Jung, H.C., 2015. Toward estimating wetland water level changes
datasets. Water Resour. Res. 55, 5053–5073. based on hydrological sensitivity analysis of PALSAR backscattering coefficients over
Yang, X., Qin, Q., Yesou, H., Ledauphin, T., Koehl, M., Grussenmeyer, P., Zhu, Z., 2020. different vegetation fields. Remote Sens. 7, 3153–3318.
Monthly estimation of the surface water extent in France at a 10-m resolution using Zhang, F., Zhang, B., Li, J., Shen, Q., Wu, Y., Song, Y., 2011. Comparative analysis of
Sentinel-2 data. Remote Sens. Environ. 244, 111803. automatic water identification method based on multispectral remote sensing.
Procedia Environ. Sci. 11, 1482–1487.
28

VANDERHOOF

Uploaded by

Copyright:

Available Formats

VANDERHOOF

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

VANDERHOOF

Uploaded by

Copyright:

Available Formats

Remote Sensing of Environment 288 (2023) 113498

Contents lists available at ScienceDirect

Remote Sensing of Environment

High-frequency time series comparison of Sentinel-1 and Sentinel-2

1. Introduction following snowmelt or large precipitation events, or show change

2. Methods band Spectral Relationship Visible (MBSRV = (Green + Red) – (NIR +

Arizona 19.6 0 48.6 2.7 1.1 8.2 84.5 83.3

Arkansas-Louisiana 1.5 1.2 1 2.3 2.3 1.7 30.7 37 0.12 0.08

Landsat-8 Sentinel-1 Sentinel-2 Landsat-8 Sentinel-1 Sentinel-2 Landsat-8 Sentinel-1 Sentinel-2

Arizona Kansas-Missouri North Dakota

Arizona PlanetScope 4/7/2020 4/8/2020 4/9/2020 1 2

Environ. 138, 165–171.

You might also like