Remote Sensing of Environment 287 (2023) 113466

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Time series soil moisture retrieval from SAR data: Multi-temporal

constraints and a global validation
Liujun Zhu a, b, Shanshui Yuan a, *, Yi Liu c, Cheng Chen d, Jeffrey P. Walker b
a
Yangtze Institute for Conservation and Development, Hohai University, Nanjing 210098, China
b
Department of Civil Engineering, Monash University, Clayton, Vic 3800, Australia
c
College of Hydrology and Water Resources, Hohai University, Nanjing 210098, China
d
State Key Laboratory of Hydrology-Water Resources and Hydraulic Engineering, Nanjing Hydraulic Research Institute, Nanjing 210029, China

A R T I C L E I N F O A B S T R A C T

Edited by Jing M. Chen Time series algorithms for soil moisture retrieval from synthetic aperture radar (SAR) data have steadily
increased in popularity over the past decade due to the feasibility of decoupling the effect of other surface
Keywords: variables, and the increasing availability of dense time series SAR data. While soil moisture inversion from time
Soil moisture series data can utilize more independent observations, the value of further constraints on the inversion process
Synthetic aperture radar
are widely acknowledged. However, how to constrain a time series retrieval for global soil moisture mapping is
Multi-temporal
still unresolved. In this study, three kinds of time series constraints were further developed and evaluated,
Global validation
including the use of 1) temporal behavior of soil moisture and soil moisture bounds; 2) temporal behavior of
vegetation or time-invariant vegetation; and 3) time series ensemble skill. The effect of these constraints was
investigated using 4 years (2016–2019) C-band Sentinel-1 data collected over 547 worldwide stations from 17
networks available on the international soil moisture network (ISMN) and intensive ground samples collected
during the Fifth Soil Moisture Active and Passive Experiment (SMAPEx-5). While the effect of these temporal
retrieval skills varies in time and space, the global validation yielded four general suggestions: 1) the assumption
of time-invariant vegetation contributed negatively even for a short retrieval period of ≤12 days; 2) reliable soil
moisture bounds of each retrieval period can substantially improve the retrieval statistics at the cost of an
underestimated soil moisture range; 3) the temporal constraints of soil moisture and vegetation need to be used
together with the soil moisture bounds for reliable estimation; 4) the use of an ensemble retrieval could partly
remove the retrieval uncertainties at the expense of underestimating soil moisture variation. The use of these
constraints resulted in a competitive correlation coefficient (R: 0.64), root mean square error (RMSE: 0.072 m3/
m3) and unbiased RMSE (ubRMSE: 0.052 m3/m3) at a spatial grid of 100 m, with similar performance achieved
across a retrieval window up to 132 days.

1. Introduction (NISAR, Kellogg et al., 2020), BIOMASS (Quegan et al., 2019), Sentinel-
1 (Torres et al., 2012) and Chinese Terrestrial Water Resources Satellite
Surface soil moisture (top 5 cm) is important for its impact on land- (TWRS, Zhao et al., 2020), being an operational source of frequent
atmosphere interactions and its partitioning of rainfall into runoff global soil moisture mapping.
through regulation of the infiltration capacity of the soil (Demargne In the past four decades, great efforts have been made to retrieve soil
et al., 2014). Economic, social and environmental planning for a water- moisture from SAR data (Kornelsen and Coulibaly, 2013). Most studies
limited future requires a capacity to monitor soil moisture content at a have focused on developing an “optimal” scattering model (Dubois
level of spatial (0.1–1 km) and temporal (< 3 days) detail that does not et al., 1995; Oh, 2004; Shi et al., 1997) and/or inversion process for a
currently exist (Peng et al., 2020). Despite the modest retrieval accu­ specific retrieval scenario (see Kornelsen and Coulibaly, 2013 for a re­
racy, the recent investments in space-borne Synthetic Aperture Radars view). Various parameterization solutions of rough soil surfaces were
(SAR) enables access to massive free SAR data with an enhanced data proposed to account for the effect of roughness (e.g., Lievens et al., 2011;
revisit of a few days, e.g., the NASA-ISRO Synthetic Aperture Radar Zribi et al., 2014), which commonly includes 1–3 parameters (Fung,

* Corresponding author.
E-mail address: [email protected] (S. Yuan).

https://doi.org/10.1016/j.rse.2023.113466
Received 23 July 2022; Received in revised form 5 January 2023; Accepted 17 January 2023
Available online 25 January 2023
0034-4257/© 2023 Elsevier Inc. All rights reserved.
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

1994), e.g., the root mean square height, correlation length and their using common auxiliary data and evaluated independently and collec­
equivalents or invariants. Similarly, the complex vegetation layers were tively for an improved time series retrieval, including the use of 1)
simplified and described as a few parameters, with the most famous temporal behavior of soil moisture and the soil moisture bounds; 2)
parameterization being the water cloud (Attema and Ulaby, 1978). temporal behavior of vegetation or time-invariant vegetation; and 3)
Based on these simplifications and/or parameterizations, the complex time series ensemble skills. These constraints were treated as extensions
earth surface scattering can be modeled using a limited number of pa­ of the time series retrieval algorithm of (Zhu et al., 2020), being also
rameters (Chen et al., 2003; Fung, 1994; Gu et al., 2021; Huang and compatible with other model based time series methods. The in-situ soil
Tsang, 2012). moisture collected from 547 worldwide stations available on the inter­
Despite the advances, soil moisture retrieval as an inversion process national soil moisture network (ISMN, Dorigo et al., 2021) and the
of the above models remains a challenge because of the large number of ground samplings made in the Fifth Soil Moisture Active and Passive
unknowns to be determined. Consequently, multi-angular (Baghdadi Experiment (SMAPEx-5, Ye et al., 2020) were used as the ground truth.
et al., 2006; Merzouki and McNairn, 2015; Rahman et al., 2008; Sahebi The effects of these constraints on soil moisture retrieval were investi­
and Angles, 2010; Shi et al., 2021) and/or multi-frequency (Bindlish and gated using time series C-band Sentinel-1 data, being expected to pro­
Barros, 2000; Pierdicca et al., 2008; Zhu et al., 2019a) SAR data were vide some straightforward principles on how to constrain scattering-
used to introduce more independent observations. Moreover, a prior model based time series inversion.
information has been used to constrain the inversion process, such as an
initial guess of soil moisture and/or roughness values (Joseph et al., 2. Methodology
2008; Mattia et al., 2009); the range, possibility and probability distri­
butions of soil moisture and roughness (Pierdicca et al., 2008; Verhoest 2.1. Time series retrieval scheme
et al., 2007; Vernieuwe et al., 2010). Various ancillary data have also
been used to constrain the retrieval or as inputs of scattering models. The time series retrieval scheme used in this study was built on the
Soil moisture from coarse passive microwave data and hydrological stochastic ensemble inversion method (Zhu et al., 2020) and the concept
models were confirmed to be an effective guess of high-resolution soil of using sliding window processing (Balenzano et al., 2021) for opera­
moisture (Kim and Van Zyl, 2009; Mattia et al., 2006; Mattia et al., 2009; tional soil moisture mapping. Given a Sentinel-1 acquisition k, it was
Zhan et al., 2006), while optical data (e.g., Landsat series and Sentinel- processed together with the previous Nt − 1 acquisitions (e.g., Nt = 4 in
2) and passive data have been widely used as a source of vegetation Fig. 1). The Ne sub-time-series were then randomly generated from the
information required in vegetation scattering models (Bousbih et al., original VV and HV Sentinel-1 data with random polarization combi­
2018; El Hajj et al., 2016; Wang et al., 2021b). nations, resulting in Ne sub-retrievals for each time instance (Ne = 2 in
The use of time series SAR data provides a promising alternative to Fig. 1). The sliding window was then moved forward to retrieve the soil
address the problem of ill-posed inversion. This so-called change moisture of the following time instances. Consequently, there were Nt ×
detection technique assumes that the vegetation canopy and soil Ne retrievals for each time instance except the first and last Nt - 1 time
roughness change little in the retrieval period, and thus directly relates instances, which were ensemble averaged as the output. The multiple
the variation of backscatter to that of soil moisture (Balenzano et al., sub-retrievals with different inputs are expected to improve the retrieval
2011; Ouellette et al., 2017; Wagner et al., 1999a; Wagner et al., 1999b; accuracy by reducing the effect of uncertainties (Lee et al., 2021; Zhu
Zribi et al., 2020; Zribi et al., 2007). The snapshot methods used to in­ et al., 2020). The parameter of Ne was set to 10 in this study according to
verse a scattering model have been similarly extended for time series the sensitivity analysis of Zhu et al. (2020), with the special case of Ne =
data with the roughness and/or vegetation being assumed time- 1 being the single retrieval using all the input data (i.e., time series VV
invariant (Fan et al., 2021; Kim et al., 2014; Kim et al., 2012; Mattia and VH for Sentinel-1).
et al., 2009; Zhu et al., 2019b). These multi-temporal algorithms have In each sub-retrieval, Nt soil moisture values were achieved by
been increasingly popular in the past decade because of: 1) the minimizing:
simplicity of decoupling the effect of soil moisture and other surface √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√
variables; 2) the increasing availability of dense time series data; and 3) ∑Nt √
√1
∑ Ni ( )
fσ = σ 0ij − σ 0model,ij (Si , Vi , mvi ) (1)
the convenience to retrieve soil moisture over a large spatial area i=1
Ni j=1
(Balenzano et al., 2021; Bauer-Marschallinger et al., 2018).
The challenge is that more observations also mean more unknowns where subscript i refers to the ith acquisition within the sliding window
and so the uncertainties are not fully addressed (Ulaby et al., 2014; Zhao and Ni is the number of selected polarizations from the ith acquisition,
et al., 2021). Moreover, a globally well calibrated forward scattering being 1 or 2 for Sentinel-1. σ0model, ij and σ 0ij are the modeled and observed
model is not available due to the absence of extensive ground roughness backscattering coefficients in dB respectively, with the Si, Vi, and mvi
and vegetation samples, with only the SMAP radar baseline algorithm being the soil surface roughness parameters, vegetation parameters, and
used to model L-band backscattering globally (Kim et al., 2014). This soil moisture.
means large scale soil moisture retrieval from SAR data can only use In this study, look up tables (LUTs) built by the Oh model (Oh, 2004)
imperfect scattering models with large uncertainties. Consequently, and the distorted Born approximation (DBA, Lang and Sighu, 1983)
constraining the inversion process with a prior information and ancillary were used to present the scattering of bare soil and vegetated area (Zhu
data still plays an important role in time series algorithms. An accurate et al., 2019a). Only one independent roughness parameter was consid­
guess of maximum and minimum soil moisture is key to the performance ered, being the root mean square (s) height ranging from 0.5 to 4 cm.
of short-term change detection methods (Ouellette et al., 2017; Palm­ The vegetation was approximated as a layer of randomly distributed
isano et al., 2020) and empirical relationships (Kim and Van Zyl, 2009), cylinders and the vegetation parameters required in the DBA were all
while an assumption of dry down soil moisture for a period after a collected in SMAPEx-5. These vegetation parameters were all related to
rainfall can substantially improve the retrieval accuracy (Zhu et al., vegetation water content (VWC) using allometric relationships. The
2019a, 2019b). The knowledge of temporal vegetation and roughness VWC and mv of LUTs ranged from 0 to 4 kg/m2 and 0.03 to 0.47 m3/m3
variation is critical for determining a proper retrieval time window (Zhu respectively. Evaluation based on C-band RADARSAT-2 data showed an
et al., 2019a). RMSE of <2.2 dB on three general landcover types (bare soil, wheat and
In contrast to the snapshot methods with numerous studies on grass), with consistent performance for an incidence angle of 22.4–39.5◦
inversion constraints and skills, how best to constrain the inversion of (Zhu et al., 2019a). Since soil surface roughness in agricultural areas has
time series retrieval remains unresolved. Consequently, the aforemen­ shown negligible variation from tens days to a whole crop season after
tioned soil moisture and vegetation constraints were further developed

2
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

Fig. 1. The retrieval scheme for time se­

ries soil moisture maps: a sliding window
(e.g., Nt = 4) was used to determine the
input time series data; an ensemble (e.g.,
Ne = 2) of sub-series were arbitrarily
selected with random polarization combi­
nations; soil moisture was retrieved inde­
pendently using each sub-series. The
sliding window was moved forward,
resulting in Nt × Ne soil moisture maps at
each time instance k except the first and
last Nt − 1 instances. The soil moisture
maps were averaged for each time instance
as the output.

soil tillage (Callens et al., 2006; Njoku et al., 2002; Ye et al., 2020), Eq. proposed in this study, inspired by the modelling of vegetation dy­
(1) becomes: namics. Given any two time-instances i and j with NDVIi ≤ NDVIj, the
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ VWCi and VWCj in Eq. (2) was forced to meet:
∑Nt √
√1 ∑ Ni ( )
fσ = √ σ0ij − σ 0model,ij (s, VWCi , mvi ) (2) VWCmin ≤ VWCi ≤ VWCj ≤ VWCmax , i, j ∈ (1, 2, …, Nt ) (4)
i=1
Ni j=1
where VWCmin and VWCmax were taken as 0 and 4 kg/m2 respectively,
with 2Nt + 1 unknowns and Nt to 2Nt independent observations for each being consistent with the bounds of LUTs. Eq. (4) is a safer but weaker
sub-retrieval, being still ill-posed. Therefore, more constraints about constraint compared to Eq. (3), because the time-invariant vegetation is
vegetation and soil moisture are still required for a stable solution. replaced by observed vegetation evolutions but Nt unknows of VWC
need to inverted from Eq. (2).
2.2. Vegetation constraints
2.3. Soil moisture constraints
In general, vegetation has seasonal variation and undergoes smooth
evolution in time. The periodic variation can be modeled to relate the The most widely used soil moisture constraint is the minimum and
vegetation dynamic and σ 0 changes, being either used to correct the time maximum soil moisture of the retrieval scenario. While the range of soil
series σ0 (Pierdicca et al., 2010) or integrated into the long-term change moisture can vary substantially across a watershed, a fixed minimum
detection algorithms (Wagner et al., 1999a). Vegetation indices, e.g., the value (e.g., 0.03 m3/m3) and/or a maximum value (e.g., 0.5 or 0.6 m3/
Normalized Difference Vegetation Index (NDVI), were commonly used m3) have been commonly applied (Balenzano et al., 2021; Zhu et al.,
to provide the vegetation dynamic. Despite the success of these re­ 2022). Such bounds have been either used to constrain the numerical
lationships, the vegetation dynamic cannot be directly integrated into inversion of scattering models or used as the bounds of LUTs. Since the
the proposed multi-temporal inversion scheme because existing vege­ soil moisture range of the LUTs was 0.03–0.47 m3/m3, Eq. (2) includes
tation scattering models and Eq. (2) are not compatible with such in­ an inherent constraint of:
formation. Alternatively, Kim et al. (2017) estimated VWC climatology
0.03 ≤ mvi ≤ 0.47 (5)
using time series NDVI and took the retrieved VWC values as the inputs
of a cost function like Eq. (2). Reliable estimation of VWC climatology Obviously, this can substantially overestimate the range of soil moisture
however is as challenging as the soil moisture retrieval. for dry and/or wet seasons, with many studies reporting the importance
In contrast, the existing short-term multi-temporal methods assume of an accurate soil moisture bound for short-term change detection
time-invariant vegetation for a short retrieval window (Balenzano et al., methods (Al-Khaldi et al., 2019; Balenzano et al., 2021; He et al., 2017;
2011; He et al., 2017; Ouellette et al., 2017; Zhu et al., 2019b), resulting Ouellette et al., 2017; Zhu et al., 2022). Consequently, its effect on
in: scattering-model-based methods (Eq. (2)) was investigated in this study.
VWC1 = VWC2 = … = VWCNt (3) The sophisticated coarse resolution passive microwave data and hy­
drological models can provide an effective guess of soil moisture in
Eq. (2) thus becomes well-constrained at the expense of introducing regional soil moisture retrieval (Kim and Van Zyl, 2009; Mattia et al.,
extra uncertainties from the potential vegetation changes. Although 2006; Mattia et al., 2009; Zhan et al., 2006), with the SMAP L3 passive
most of these studies were confined to a short retrieval period of 1–5 product (36 km) being used in this study considering its relatively better
weeks (2–8 Sentinel-1 acquisitions), substantial VWC changes can occur accuracy than other coarse resolution soil moisture products (Cui et al.,
with their impact on retrieval being unclear. Apart from investigating 2017; Wang et al., 2021a). Given a retrieval time window of k to k + Nt,
the effect of VWC temporal variation, a “soft” variant of Eq. (3) was all of the SMAP soil moisture products collected were used to determine

3
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

the: is 1 when the temporal trend of VWC completely matches that of NDVI,
( ) reaching the maximum value of 0.5N2t + 1 when the two trends are
mvmin = min mvSMAP
i , …, mvSMAP
j , mvSMAP
ave , k ≤ i ≤ j ≤ k + Nt (6) reversed. A genetic algorithm was used to find the optimal solution of
Eq. 10 for time series soil moisture retrieval. The number of chromo­
( )
mvmax = max mvSMAP , …, mvSMAP , mvSMAP , k ≤ i ≤ j ≤ k + Nt (7) somes and the maximum generation (iteration) required in the genetic
algorithm was 30 and 100 respectively. The outputs of the 100th iter­
i j ave

ation were treated as the retrieved results regardless of the value of the
where mvSMAP ave is the mean of all SMAP measurements of the corre­
cost function. The retrieval performance was evaluated using 4 widely
sponding retrieval area. Eq. (6) and Eq. (7) were demonstrated to esti­
used indicators including bias, correlation coefficient (R), root mean
mate reliable soil moisture bounds in the Yanco area, Australia (Zhu
square error (RMSE) and unbiased RMSE (ubRMSE). The Student’s t-test
et al., 2022).
was used to check the significance of performance difference.
Similar to the temporal constraint of vegetation (Eq. (4)), the soil
moisture can be further constrained by the temporal evolution of the
3. Data and preprocessing
SMAP passive products. Given any two time-instances i and j with mvS­ i
MAP
≤ mvSMAP
j , the mvi and mvj in Eq. (2) was forced to meet:
3.1. Soil moisture
mvmin ≤ mvi ≤ mvj ≤ mvmax (8)
In this study, the soil moisture measurements available on the ISMN
Since irrigation or small-scale rainfall events lead to different temporal (Dorigo et al., 2021) were considered for validations. Since the forward
variation of soil moisture within a SAMP passive grid, Eq. (8) was LUTs were built on a single layer of random cylinders with a maximum
removed when σ i − σ j ≥ 1 dB to avoid the exclusion of soil moisture VWC of 4 kg/m2, only the stations over bare soil, grassland, cropland or
“anomalies” (Zhu et al., 2022). shrubland were used. The Copernicus Global Land Cover Layers was
used to provide the landcover type of each station (Buchhorn et al.,
2020). All the 2016–2019 recordings with a sensing depth ≤ 5 cm were
2.4. Inversion and evaluation methods downloaded using the ISMN batch service (Dorigo et al., 2021). The
Shuttle Radar Topography Mission Digital Elevation Data Version 3
The main contribution of this study is a global validation of the po­ (SRTM DEM V3) was used to calculate the slope of each station, with
tential constraints on multi-temporal soil moisture retrieval methods. A these having a slope of >5◦ being removed, considering the challenging
total of 4 constraints were presented above, with the time invariant VWC of terrain correction and flattening. Moreover, only high-quality re­
(Eq. 3) conflicting with the VWC trend (Eq. 4). Accordingly, 9 multi- cordings were used, being these with a flag of “G”.
temporal retrieval algorithms were considered as outlined in Table 1, As the measuring depths of shallow surface soil moisture (≤ 5 cm)
and compared with their ensemble alternative (Fig. 1). Each algorithm varied across networks (e.g., at 0–5 cm, 2 cm and/or 5 cm), they were
invariant was expressed as four letters of T (True) and/or F (False) for treated as independent estimations of the ≤5 cm soil moisture.
simplification (Table 1), with the applied constraints labeled as T. For Accordingly, all the recordings of a station collected within 5 cm was
example, the algorithm without any constraint was labeled as FFFF. averaged as the daily averaged soil moisture. These beyond the LUTs
Apart from the constraint of soil moisture bounds, the use of multiple (0.03–0.47 m3/m3) were then discarded. In order to get interpretable
other constraints results in a conventional multi-objective optimization station-specific accuracy statistics, only the stations with ≥30 valid daily
problem, with some sophisticated algorithms available (Deb, 2014). soil moisture measurements were used. After applying these filters, a
However, a multi-objective optimization can be extremely time total of 547 stations from 17 networks were used in this study (Table 2).
consuming, and thus was simplified as a single objective optimization by Notably, the 34 OzNet stations were all located in the Yanco area, NSW,
integrating the cost-function of vegetation (fv) and/or soil moisture (fmv) Australia (Fig. 2), being different from these in the ISMN.
into Eq. 2: Extensive soil moisture samples collected from a 5 × 20 km area of
F = fσ fV fmv (9) SMAPEx-5 (the red rectangle in Fig. 2) was also used. The campaign was
conducted in the Yanco area, Australia, as part of the SMAP global Cal/
∑
k+Nt ⃒ ⃒ Val scheme (Ye et al., 2020). The Yanco area is characterized as semiarid
fV = ⃒rVWC,i − rNDVI,i ⃒ + 1 (10) with an average annual precipitation of 300 mm. The soil texture is
i=k
mainly clay to sandy soil, with the dominated land use being grazing and
cropping. The vegetation type for grazing areas is natural grass, while
∑
k+Nt ⃒ ⃒
fmv = ⃒rmv,i − rSMAP,i ⃒ + 1 (11) the cropping area had seasonal transitions of crops and/or fallow (bare
i=k soil). The averaged soil moisture of all OzNet sites is depicted in Fig. 3,
with clear inter-annual dry-wet cycle and droughts in 2018 and 2019.
where rVWC, i, rNDVI, i, rmv, i, rSMAP, i are the order of ith VWC, NDVI, mv Nearly concurrent soil moisture samples were collected during the
and SMAP soil moisture in the corresponding time series, respectively. fv acquisition dates of two Sentinel-1 images (Sep. 15 and 27), with the
time difference being <20 h. These soil moisture samples were made on
Table 1 an east-west oriented grid using the Hydraprobe sensor (Merlin et al.,
Variants of the multi-temporal retrieval algorithm, with the applied constraint 2007), with three replicate soil moisture readings of 0–5 cm at each plot
being labeled as True. (Fig. 2). Similarly, measurements <0.03 or > 0.47 m3/m3 were dis­
Algorithms Time invariant VWC trend mv bound (Eq. mv trend carded, with a total of 566 valid samples. Other surface parameters, e.g.,
VWC (Eq. 3) (Eq. 4) 6 and 7) (Eq. 8) VWC, vegetation structure parameters and soil surface roughness, were
TFFF True False False False also measured in this area (Ye et al., 2020) and were used in the
FFFF False False False False parametrization and calibration of surface and vegetation scattering
FTFF False True False False models (Zhu et al., 2019a).
FFTF False False True False
FFFT False False False True
FFTT False False True True 3.2. Remote sensing data and preprocessing
FTTF False True True False
FTFT False True False True The SMAP L3 passive soil moisture product (Version 6, Neill et al.,
FTTT False True True True
2019), Moderate Resolution Imaging Spectroradiometer (MODIS) NDVI

4
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

Table 2
the 17 soil moisture networks used for evaluation. The A/D in the orbit pass refers to the ascending/descending.
Network # stations # mv records # orbits Orbit pass Average revisit (days)* Incidence angle Reference

AMMA-CATCH 5 474 1–2 A 5.2–12.4 31.0–44.7 (Galle et al., 2018)

ARM 15 2490 1–2 A 5.7–15.1 33.0–44.9 (Cook, 2016)
HOAL 18 4007 2 A/D 3.4–3.4 38.3–39.2 (Blöschl et al., 2016)
HOBE 21 5577 4 A/D 1.8–2.3 30.0–45.2 (Jensen and Refsgaard, 2018)
MAQU 11 376 2–3 A/D 6–12 32.8–46.1 (Su et al., 2011)
NAQU 9 795 3 A/D 6.1 31.5–45.2 (Su et al., 2011)
NGARI 15 1392 2–3 A/D 5.3–7.6 34.2–45.5 (Su et al., 2011)
OzNet 34 4080 1 D 12 35.3–40.0 (Smith et al., 2012)
PBO_H2O 94 4752 1–4 A/D 3.7–17 30.3–45.7 (Larson et al., 2008)
REMEDHUS 16 5556 2–3 A/D 3.2–6.1 30.5–42.1 (González-Zamora et al., 2019)
RISMA 23 2354 2–3 A/D 6–12 31.1–43.7 (Ojo et al., 2015)
RSMN 12 5678 2–4 A/D 1.7–3.3 30.2–45.1
SCAN 127 16,814 1–4 A/D 4.5–16.5 30.4–45.6 (Schaefer et al., 2007)
SMN-SDR 27 1770 3–4 A/D 3.8–5.2 30.6–44.1 (Zhao et al., 2020)
SMOSMANIA 11 4931 2–4 A/D 2.2–3.9 30.5–45.4 (Calvet et al., 2016)
SOILSCAPE 35 449 2–3 A/D 5.7–6.1 30.8–43.4 (Moghaddam et al., 2016)
USCRN 74 10,055 1–4 A/D 3.5–16.5 30.4–45.6 (Bell et al., 2013)
Total 547 68,468
*
Average values for 2016–2019. The absence of a few acquisitions resulted in a revisit > the 12-day nominal revisit.

Fig. 2. The Yanco area and OzNet soil moisture stations as well as the SMAPEx-5 ground soil moisture samples collected nearly concurrently with the two Sentinel-1
acquisitions; Sep. 15 and Sep. 272,015.

products and Sentinel-1 Interferometric Wide (IW) ground range maximum time interval of two successive acquisitions was generally
detected (GRD) products were used in this study. All the Sentinel-1 IW ≤12 days expect some short special periods without data. In this study,
GRD acquisitions available on Google Earth Engine (GEE) were used in the GEE GRD data with a pixel size of 10 m was resampled to 100 m
this study. The GEE Sentinel-1 GRD collection was produced by the using the “reduceResolution” method provided by GEE (https
Sentinel-1 Toolbox with the main steps including GRD border and ://developers.google.com/earth-engine/guides/resample). In brief, the
thermal noise removal, radiometric calibration, and terrain correction 10 m pixels were aggregated to larger pixels of 100 m on the Universal
(orthorectification). Radiometric terrain correction and flattening is not Transverse Mercator (UTM) local projection used in the Sentinel-1 data,
made because of the large uncertainty in the available DEM. The GRD e.g., the UTM 55S for the Yanco area. This is expected to substantially
data has two polarizations (VV and VH) and a spacing of 10 m. Since the reduce the effect of speckle noise and result in a satisfactory radiometric
IW data over a station can be collected from varying orbit passes accuracy of ~0.4 dB (Torres et al., 2012).
(ascending/descending) and/or multiple relative orbits, the average The Terra and Aqua MODIS vegetation indices products (MOD13Q1
revisit and the time interval between two successive acquisitions varied and MYD13Q1, V6) were used to provide the time series NDVI required
timely and spatially (Table 2). This means that the same Nt can have in Eq. 4. Since the two NDVI data sets were from two identical sensors,
different lengths of retrieval widow in time across the 547 stations. The they were merged to an 8-day composite with an initial resolution of

5
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

Fig. 3. Time series of OzNet, SMAP L3 soil moisture, NDVI and Sentinel-1 observations on the acquisition dates of Sentinel-1 (a) and their correlation heatmap (b).
The solid line and envelope represent the average and standard deviation of the 34 OzNet sites.

250 m. This composite was then resampled and re-projected to the grid
of the pre-processed Sentinel-1 data using the nearest neighbor method.
The time series NDVI for all Sentinel-1 acquisition dates then were
interpolated using a using a spline function. The time series SMAP L3
passive soil moisture over each station was extracted first based on the
36 km SMAP grids. Similarly, the soil moisture of each station on all
Sentinel-1 acquisition dates were then interpolated with a spline
function.
Fig. 3 shows the mean and standard deviation of OzNet, SMAP,
NDVI, Sentinel-1 VV and VH across the 34 OzNet sites as well as the
correlation heatmap. As expected, the OzNet soil moisture time series
showed high correlation with Sentinel VV backscatter and SMAP soil
moisture. The OzNet soil moisture also had a high correlation with the
NDVI, being the main stress for the vegetation in this area. An abrupt
increase of soil moisture can result in an abrupt increase of NDVI in dry
seasons e.g., April to October 2018.

4. Results

4.1. The effect of vegetation constraint and assumption

The retrieval algorithm with time invariant vegetation (TFFF), time

variant vegetation (FFFF) or temporal vegetation constraint (FTFF) was
evaluated and compared across the 547 stations, first to show the effect
Fig. 4. Comparison of the retrieval algorithms with time invariant vegetation
of vegetation constraints (Fig. 4). The length of retrieval time window
(TFFF), time variant vegetation (FFFF) and vegetation temporal constraint
(Nt) ranged from 2 (≤ 12 days) to 12 (≤ 132 days) Sentinel-1 acquisi­ (FTFF); a-f refer to the results using a Nt varying from 2 to 12 with an interval of
tions with an interval of 2 (≤ 24 days) acquisitions. The TFFF variant 2. The results for grass, crop and shrub and the results for all networks can be
achieved the worst results in all cases of Nt with the ubRMSE being found in the supplementary document (Fig. S1 and S3), with similar patterns.
>0.11 m3/m3. The retrieval performance was improved after removing
the assumption of time invariant vegetation (i.e., the FFFF variant) even averaging filter with a size of Nt. The retrieved VWC maintained the four
for a short retrieval period of ≤12 days (Nt = 2). As Nt was increased, the main vegetation peaks using a Nt of 4, while only the largest peak of Jul.
ubRMSE of the FFFF variant slightly decreased at the expense of – Oct. 2016 was captured with a Nt of 12. As a result, all the short-term
underestimating the standard deviation of the retrieved soil moisture. vegetation variations were removed by the TFFF variant and the effect of
The algorithm using a temporal vegetation constraint (FTFF) achieved such variations on backscattering was mis-interpreted as the change of
similar but slightly better results compared to the FFFF variant, with a roughness and/or soil moisture, being the main reason for the poor re­
larger decrease in retrieval standard deviation for a larger Nt. sults and the relatively larger soil moisture variations (Fig. 4). In
The retrieved time series soil moisture, VWC and RMS heights of the contrast, the time series VWC retrieved by the FFFF method can be
34 OzNet stations are depicted in Fig. 5, further illustrating the joint decomposed into seasonal variations and abrupt changes. The later was
effect of Nt and the vegetation constraints. The VWC retrieved by the observed to be accompanied by abrupt soil moisture changes (Fig. 3),
TFFF variant showed a smoother temporal evolution than that of the which can be partly explained by the abrupt water supply from the soil
other two methods. The time invariant vegetation together with the moisture. Moreover, VWC and soil moisture can be overestimated
sliding window processing (Fig. 1) was similar to a 1-dimension sliding concurrently to meet the real radar observations in the ill-posed

6
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

Fig. 5. Average soil moisture (mv), vegetation water content (VWC) and RMS height (s) of OzNet sites retrieved by the algorithm with time invariant vegetation
(TFFF), time variant vegetation (FFFF,) and vegetation temporal constraint (FTFF) as well as the average NDVI and OzNet measurements. The left and right column
are the results of using Nt = 4 and 12 respectively.

inversion of the FFFF variant, being another reason of the abrupt effective roughness values in the optimization process.
changes.
After applying the temporal vegetation constraint (i.e., the FTFF
4.2. The effect of soil moisture constraint and assumption
variant), the short-term temporal variation of retrieved VWC was further
increased. In the inversion process of the FTFF variant, the cost function
The retrieval algorithms with soil moisture bounds (FFTF), soil
term representing VWC trends converged faster (Eq. 10) than the fσ, and
moisture temporal constraint (FFFT) and both soil moisture constraints
the VWC values were forced to meet the “correct” trend first. The range
(FFTT) were evaluated across the 547 stations, with the FFFF being
of VWC was observed to increase gradually before convergence, leading
included as the benchmark (Fig. 6). The comparison of the FFFF and
to a large inner-window variation. A larger Nt led to a larger variation in
FFTF methods showed that the use of reliable soil moisture bounds can
VWC which resulted in a larger inter-annual variation in soil moisture.
substantially improve the retrieval performance. The average ubRMSE
The three vegetation temporal constraints thus act like “zoom in” to
decreased from 0.103 to 0.087 m3/m3, while the average R increased
“zoom out” operations of soil moisture variation, with their effects being
from 0.40 to 0.59. In contrast, single use of the temporal soil moisture
enhanced using a larger Nt.
constraint led to deterioration in accuracy, with an average decrease of
The retrieved roughness showed limited temporal variation (Fig. 5)
0.06 in R. The joint use of the two soil moisture constraints (FFTT)
except the peak in the middle of 2016, generally ranging from 0.5 to 0.8
achieved the best results, with an average ubRMSE and R of 0.086 m3/
cm. Notably, the retrieved roughness values were effective values and
m3 and 0.61 respectively. While the difference in ubRMSE and R was
can therefore be substantially different from the ground measurements,
marginal for different Nt, the standard deviation of retrieved soil mois­
as a result of the uncertainty of the scattering models (Baghdadi and
ture gradually decreased as Nt increased for all algorithms, being
Zribi, 2011; Lievens et al., 2011; Zhu et al., 2016). The retrieved RMS
consistent with the results of Fig. 4. The FFTT method showed the
heights followed the trend of the VWC in the results of the TFFF and
smallest sensitivity to Nt, suggesting a relatively stable performance for a
FFFF method, playing a similar role of VWC in the optimization process.
longer retrieval window.
A more comprehensive analysis on the 457 stations is provided in
Fig. 7 shows the average soil moisture, VWC and RMS heights of the
Fig. S4. This confirmed that assuming a time invariant roughness for a
34 OzNet stations retrieved by three algorithms using different soil
long retrieval period is also questionable, not only for the potential
moisture constraints. The FFFT failed to capture the general trend of soil
changes in ground roughness, but also for the requirement of varying
moisture and substantially overestimated the short-term (2–4 weeks)

7
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

which is 0.286, 0.029 m3/m3 and 0.023 m3/m3 for the ensemble re­
trievals. This further suggests that the soil moisture bounds are the most
useful constraint.
Among the four algorithms using the soil moisture bounds, the FFTT
method achieved the best results in median RMSE and R (0.072 m3/m3
and 0.643), followed by the FTTT (RMSE: 0.072 m3/m3, R: 0.641), FTTF
(RMSE: 0.073 m3/m3, R: 0.609) and FFTF (RMSE: 0.074 m3/m3,
R:0.603). This suggests that extra use of the soil moisture and/or
vegetation temporal constraint can lead to better but insignificant re­
sults. However, the joint use of two temporal constraints without the soil
moisture bounds (i.e., the FTFT method) led to the worst results among
the algorithms that used multiple constraints. This can be explained by
that either the soil moisture or the vegetation temporal constraint can
increase the short-term soil moisture variations (Fig. 5 and Fig. 7) which
was further enhanced by the joint use of two temporal constraints. The
ensemble retrievals outperformed the single retrievals on all algorithm
variants in R, RMSE and ubRMSE, with the average improvement in
median values being ~0.04, ~ 0.005 m3/m3 and ~ 0.005 m3/m3,
respectively. As expected, the difference between single and ensemble
retrievals in bias was negligible (< 0.001) as the ensemble average of
multiple sub-retrievals retain the system bias. The difference in RMSE
was only significant for 4 algorithm (TFFF, FFFF, FTFF and FTFT) and all
these methods showed relatively higher RMSEs than the others. Simi­
Fig. 6. Same as Fig. 4 but for soil moisture bounds (FFTF), soil moisture larly, the difference in ubRMSE was insignificant for the two most
temporal constraint (FFFT) or both soil moisture bound and temporal con­ powerful variants (FFTT and FTTT), suggesting that weaker algorithms
straints (FFTT). The results for grass, crop and shrub can be found in the sup­ can probably benefit more from ensemble retrievals. Same to the
plementary document (Fig. S3), with similar patterns. ensemble retrievals, the single retrieval using the FFTT method achieved
the best results followed by the FTTT, FTTF and FFTF method, sug­
temporal variations. Specifically, the retrieved soil moisture within each gesting that the use of ensemble retrieval cannot change the relative
retrieval window achieved a similar range of 0.1–0.3 m3/m3 for the case performance of the 9 variants.
of Nt = 4, which extended to 0.1–0.5 m3/m3 when Nt = 12. This suggests
that the effect of the soil moisture temporal constraint on retrieval is
similar to that of the vegetation temporal constraint, substantially 4.4. Evaluation over the SMAPEx-5 focus areas
increasing the short-term temporal variation of soil moisture and thus
VWC. In contrast, the FFTF method well captured the general trend and A further investigation was made on the SMAPEx-5 data set (Fig. 2).
accurately reflected the soil moisture ranges of various periods. For Similar to the results on the 547 stations, the ensemble retrievals ach­
example, the FFTF had smaller underestimation of the high soil moisture ieved better results than the corresponding single retrievals on all
observed in the Australian winter of 2016 than retrievals without the methods except the FFTF method (Fig. 9). The average improvement of
soil moisture bounds. However, the retrieved soil moisture of FFTF using the ensemble concept was ~0.007 m3/m3 in ubRMSE and ~ 0.05
showed smaller short-term variations compared to that of the FFFF in R. These improvements were mainly from the reduced random un­
method (Fig. 5) and the OzNet observations (Fig. 3). Therefore, the time certainty contained in data, scattering models and assumptions, with the
series soil moisture retrieved by the FFTT seemed to be a compromise of pattern of ground measured versus retrieved being less scattered
the FFTF and FFFT methods. In view of a signal, the time series retrieved (Fig. 10). This however is not always positive, probably overestimating
by the FFFT and FFTF methods were like the high and low frequency the low values and underestimating the high values. Consequently, the
components of the FFTT time series respectively. Moreover, the ampli­ improvement in R can be larger than that in RMSE as the bias can be
tude of the high frequency components could be amplified using a longer increased, being consistent with Fig. 8. Different from the results on the
retrieval window, while the low frequency components could also be global dataset, the use of soil moisture bounds and the two vegetation
enhanced by depressing the high frequency components with a longer constraints contributed little, and even negatively to the SMAPEx-5
retrieval window. scenario. For example, the TFFF method showed a competitive perfor­
mance to the FFTT method. The potential reasons include: 1) the range
4.3. The joint effect of constraints and ensemble of soil moisture in SMAPEx-5 was close to the global soil moisture range
of 0.03–0.47 m3/m3 and thus the use of SMAP based soil moisture
The bias, R, RMSE and ubRMSE of all sites were calculated for the bounds could not add any useful constraint and 2) the VWC during the
single retrievals and ensemble retrievals (Fig. 8), with the single re­ SMAPEx-5 was nearly time-invariant (Ye et al., 2020) and thus the
trievals using the full time series of Sentinel-1 (i.e., Ne = 1 in Fig. 1). For assumption of time-invariant VWC can be more reliable.
simplicity, only the results of Nt = 4 are presented as it is the default Fig. 10 shows the soil moisture maps for the SMAPEx-5 focus area on
value for various time series methods (Al-Khaldi et al., 2019; Balenzano Sep. 14, 2015. The retrieved soil moisture maps of the 9 algorithms had
et al., 2021; He et al., 2017; Ouellette et al., 2017; Zhu et al., 2022). The similar spatial patterns, which generally matched the pattern of ground
4 algorithms with soil moisture bounds (i.e., FFTF, FFTT, FTTF and measurements. The bare soil had smaller soil moisture values, followed
FTTT) achieved a near-zero median bias (<0.007 m3/m3) on both single by grass land and the areas with crops (mainly wheat). Consistent with
and ensemble retrieval modes, being significantly lower than that of the the evaluation results on the global data set (Fig. 4 and 6), the soil
other methods without soil moisture bounds (0.026–0.042 m3/m3). moisture maps retrieved by different algorithms showed a similar
Apart from a smaller bias, the use of soil moisture bounds resulted in average value but a large difference in standard deviation. The ensemble
significant higher R, lower RMSE and ubRMSE. The average improve­ retrievals showed smaller spatial variations due to the ensemble average
ment in R, RMSE and ubRMSE from using soil moisture bounds is 0.277, of multiple retrievals, in line with the evaluation on the SMAPEx-5 data
0.033 m3/m3 and 0.027 m3/m3 respectively for the single retrievals, set (Fig. 9).

8
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

Fig. 7. Same as for Fig. 5 but for the algorithm with soil moisture bounds (FFTF), soil moisture temporal constraint (FFFT) and both soil moisture bound and
temporal constraints (FFTT).

5. Discussion The proposed two temporal constraints are qualitative descriptions

of the temporal evolution of vegetation and soil moisture, being rela­
The effect of four temporal constraints and the ensemble skill on time tively weaker constraints compared to the assumptions of time-invariant
series soil moisture retrieval was investigated, providing some vegetation. Since these relationships are easier to be achieved and more
straightforward suggestions on how to constrain scattering-model based reliable than empirical modelling of temporal vegetation or soil mois­
time series retrievals. The widely used assumption of time-invariant ture dynamics (Kim et al., 2017; Wagner et al., 1999a), they were ex­
vegetation contributed negatively to the retrieval accuracy in the pected to provide safer prior knowledge for improved soil moisture
global validation, even for a short retrieval period of ≤12 days (Nt = 2 in retrievals. However, single use of either constraint was found to
Fig. 4 and Fig. S3). However, it still improved the retrieval accuracy for contribute negatively to the retrieval accuracy in the global validation
specific areas and periods, such as the SMAPEx-5 scenario (Fig. 9). The (Fig. 4 and 6), introducing substantial short-term variations in soil
negative effect of the time-invariant vegetation can be partly ascribed to moisture. A conflicting conclusion was made in two previous studies
the vegetation changes, with a longer retrieval window meaning poorer where soil moisture after a rainfall event was forced to decrease
results (Palmisano et al., 2020). Moreover, the assumption of time- monotonously (Zhu et al., 2019a, 2019b). The temporal constraints
invariant roughness and vegetation forced a smooth evolution of VWC contributed positively if they were used together with soil moisture
and roughness in the retrieval process (Fig. 6), with the short-term bounds (Fig. 8 and Fig. 10). The two previous studies inherently used the
variation of radar observations and scattering model uncertainties all soil moisture bounds, because the real soil moisture bounds were close
being ascribed to the variation of soil moisture. The retrieved time series to those of the LUTs (Zhu et al., 2019a, 2019b), and thus the temporal
VWC may accurately reflect the variation of real vegetation but this is constraint of soil moisture contributed positively. Moreover, time series
not necessary to result in better soil moisture retrieval, because effective multi-angular and multi-frequency data was used in the two previous
values can be more favorable for inversing a moderate scattering model. studies with larger uncertainties in data and scattering models, being
Similarly, many studies found that the direct use of ground measured more challenging than the time series retrieval from a single data source.
roughness in soil moisture retrieval is questionable and effective values Consequently, a further investigation using a long time series of multi-
based on scattering models can be more useful (Baghdadi and Zribi, angular and multi-frequency data (e.g., the joint time series of
2011; Lievens et al., 2011; Zhu et al., 2016). This suggests that an over- Sentinel-1 and SAOCOM-1) can be valuable for the use of these temporal
constrained inversion does not necessarily result in good results when an constraints.
imperfect scattering model was used. The soil moisture range was confirmed to be critical for a successful

9
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

Fig. 8. Performance of the 9 algorithms using the single (Ne = 1 in Fig. 1 with full VV/VH time series) and ensemble retrieval mode (Ne = 10) on the 547 stations,
with the top to bottom panel being bias, R, RMSE and ubRMSE respectively. The default Nt = 4 was used.

retrieval, being consistent with the short-term change detection algo­ soil moisture values. This range generally covers the natural soil con­
rithms (He et al., 2017; Ouellette et al., 2017). In this study, the global ditions observed from very dry to wet, with only a limited number of
soil moisture range used is that of the LUTs (0.03 to 0.47 m3/m3), observations (150 out of 68,468; ~0.22%) from the 547 stations
considering the typical calibration error of ground measurements (Smith exceeding 0.47 m3/m3 for the study period. The range of time series
et al., 2012) and the fact that the scattering model uncertainty (e.g., 2 coarse soil moisture products of SMAP (36 km) was used for a more
dB) can be much larger than the sensitivity of the radar signal to high reliable estimation of a specific retrieval window. Other model-based

10
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

Fig. 9. Performance of ensemble (first and third column) and single (second and fourth column) retrievals on the SMAPEx-5 data set using a Nt = 4 retrieved
according to the 9 retrieval algorithms (see Table 1).

11
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

Fig. 10. The soil moisture maps of Sep. 14, 2015 retrieved by the 9 algorithms (see Table 1) with ensemble (top panel) and without ensemble (bottom panel). The
landcover was from Ye et al. (2020).

soil moisture products (e.g., the fifth generation of ECMWF atmospheric smaller spatial variation. The joint use of these constraints and the
reanalyzes, ERA-5) can be promising alternatives, with the advantage of ensemble can further reduce the variation, e.g. the results of the FFTF
providing long-term continuity; refer to the Fig. S7-S9 for a comparison method in Fig. 9.
of using ERA5-Land and SMAP L3 in soil moisture retrieval. Neverthe­ The length of retrieval window (Nt) is a key parameter of the pre­
less, use of such coarse soil moisture products could not fully retain the sented multi-temporal inversion scheme as well as existing time series
large spatial and temporal variations of soil moisture at a grid size of methods. Different from the simple conclusion that a larger Nt means
100 m. Consequently, the retrieved soil moisture showed a smaller larger retrieval error (Palmisano et al., 2020), the effect of Nt on soil
standard deviation both spatially (Fig. 10) and temporally (Fig. 6). moisture retrieval varied when different constraints were applied, as
Fortunately, extra use of the soil moisture temporal constraint could summarized in Table 3. The retrievals without constraint achieved
partly recover the temporal variation (Fig. 4 and 6). better accuracy statistics from an increased Nt, which however led to a
The effectiveness of the ensemble retrieval concept was demon­
strated in Zhu et al. (2020) using a synthetic data set with various
incidence angles, polarizations, frequencies and uncertainty sources. Its Table 3
The effect of an increased Nt on the performance of algorithms using various
effectiveness for a real scenario of a long time series data was first
constraints.
confirmed in this study, with the improvement in RMSE being up to
0.013 m3/m3. While the improvement on 3 of the 9 algorithms was Constraint Accuracy Variation of the retrieved soil
moisture
insignificant and the absolute improvement is limited, the ensemble skill
has no interaction with the 4 constraints in the inversion process and No constraint Increase Decrease
Soil moisture bounds Decrease Decrease
thus can be compatible with any combination of the 4 constraints.
Time-invariant vegetation Unclear Unclear trend
However, the ensemble retrievals were prone to underestimate the trend
range and spatial variation of soil moisture by removing the largest and/ Vegetation temporal Increase Decrease
or smallest estimations in the ensemble average (Fig. 9 and Fig. 10). The constraint
use of soil moisture bounds and the assumption of time-invariant Soil moisture temporal Decrease Increase
constraint
vegetation also tended to result in a smoother temporal evolution and

12
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

smaller and underestimated soil moisture variation. A similar behavior the work reported in this paper.
was observed for the retrievals with vegetation temporal constraints.
Accordingly, the algorithms with a single constraint cannot benefit from Data availability
an increased Nt. However, the joint use of two soil moisture constraints
resulted in negligible difference in accuracy statistics and retained the Data will be made available on request.
variation of retrieved soil moisture for a Nt up to 12, suggesting that the
FFTT and FTTT method can be used in a much longer retrieval window Acknowledgments
beyond the current experience of 1–4 weeks.
Apart from the perspectives on how to constrain the inversion pro­ This work was supported by the National Natural Science Foundation
cess of time series retrieval, this study provided a near-operational time of China (42101374 and 52121006), the Fundamental Research Funds
series scheme for soil moisture retrieval from Sentinel-1 IW GRD data. for the Central Universities (B220201009) and the Basic Research
Some of the 9 algorithm variants achieved a competitive median RMSE Project of Jiangsu Province (BK20210377 and BK20210368). The
and ubRMSE of ~0.072 and ~ 0.051 m3/m3 at a grid size of 100 m, with SMAPEx-5 field campaign was supported by an Australian Research
the performance of the best variant (FFTT) on the 547 stations being Council Discovery Project (DP140100572). The authors express sincere
provided in the supplementary document (Fig. S7). The accuracy sta­ thanks to the data providers and the International Soil Moisture Network
tistics can be further improved when retrieval or evaluated at a coarser for the network data listed in Table 1. Finally, the authors are grateful to
grid of e.g., 1 km. Although the parametrization and calibration of the the Reviewers for their valuable comments which helped to improve the
scattering models still required ground measurements, the LUTs initially quality of this paper.
built for the SMAPEx-5 were successfully applied for the entire period of
2016–2019 and 547 worldwide stations without further calibration or Appendix A. Supplementary data
modification. However, the large performance discrepancy among the
547 stations (Fig. 8 and S7) and 17 networks (Fig. S6), suggested that the Supplementary data to this article can be found online at https://doi.
LUTs based on the SMAPEx-5 cannot fully represent the surface condi­ org/10.1016/j.rse.2023.113466.
tions of all networks. The current LUTs only covered grass and crop
types that were dominated by a single layer of vertical structure, with References
further extensions for other types being possible (Kim et al., 2014). The
VWC was limited to <4 kg/m2, considering the limited soil contributions Al-Khaldi, M.M., Johnson, J.T., O’Brien, A.J., Balenzano, A., Mattia, F., 2019. Time-
series retrieval of soil moisture using CYGNSS. IEEE Trans. Geosci. Remote Sens. 57,
at larger values (Zhu et al., 2019a). However, this boundary and the 4322–4331.
validity of C-band for dense vegetated areas are still questionable, Attema, E., Ulaby, F.T., 1978. Vegetation modeled as a water cloud. Radio Sci. 13,
requiring a location mask based on landcover maps and thresholds for 357–364.
Baghdadi, N., Holah, N., Zribi, M., 2006. Soil moisture estimation using multi-incidence
volume scattering dominated areas (Balenzano et al., 2021). and multi-polarization ASAR data. Int. J. Remote Sens. 27, 1907–1920.
Baghdadi, N., Zribi, M., 2011. Evaluation of radar backscatter models IEM, OH and
6. Conclusion Dubois using experimental observations. Int. J. Remote Sens. 27, 3831–3852.
Balenzano, A., Mattia, F., Satalino, G., Davidson, M.W., 2011. Dense temporal series of C-
and L-band SAR data for soil moisture retrieval over agricultural crops. IEEE J. Sel.
The effect of four temporal constraints and the ensemble skill on time Top. Appl. Earth Obs. Remote Sens. 4, 439–450.
series soil moisture retrieval was investigated using the soil moisture Balenzano, A., Mattia, F., Satalino, G., Lovergine, F.P., Palmisano, D., Peng, J.,
Marzahn, P., Wegmüller, U., Cartus, O., Dąbrowska-Zielińska, K., 2021. Sentinel-1
measurements of 547 stations from 17 networks along with intensive
soil moisture at 1 km resolution: a validation study. Remote Sens. Environ. 263,
ground measurements from the SMAPEx-5. While the effect of these 112554.
temporal retrieval skills varied across stations and networks, and was Bauer-Marschallinger, B., Freeman, V., Cao, S., Paulik, C., Schaufler, S., Stachl, T.,
also related to the robustness of the scattering models, four empirical Modanesi, S., Massari, C., Ciabatta, L., Brocca, L., 2018. Toward global soil moisture
monitoring with Sentinel-1: harnessing assets and overcoming obstacles. IEEE Trans.
suggestions for global applications are made: 1) assuming time-invariant Geosci. Remote Sens. 57, 520–539.
vegetation contributed negatively in the global validation as a result of Bell, J.E., Palecki, M.A., Baker, C.B., Collins, W.G., Lawrimore, J.H., Leeper, R.D.,
potential vegetation changes and over-constrained inversion of moder­ Hall, M.E., Kochendorfer, J., Meyers, T.P., Wilson, T., 2013. US climate reference
network soil moisture and temperature observations. J. Hydrometeorol. 14,
ate scattering models; 2) reliable information of soil moisture bounds is 977–988.
critical for successful retrieval with the risk of underestimating the soil Bindlish, R., Barros, A.P., 2000. Multifrequency soil moisture inversion from SAR
moisture variation; 3) the two temporal constraints of vegetation and measurements with the use of IEM. Remote Sens. Environ. 71, 67–88.
Blöschl, G., Blaschke, A., Broer, M., Bucher, C., Carr, G., Chen, X., Eder, A., Exner-
soil moisture must be used together with the soil moisture bounds to Kittridge, M., Farnleitner, A., Flores-Orozco, A., 2016. The hydrological open air
maintain the variation of soil moisture; and 4) the use of an ensemble laboratory (HOAL) in petzenkirchen: a hypothesis-driven observatory. Hydrol. Earth
can benefit the soil moisture retrieval by partly removing uncertainties Syst. Sci. 20, 227–255.
Bousbih, S., Zribi, M., El Hajj, M., Baghdadi, N., Lili-Chabaane, Z., Gao, Q., Fanise, P.,
at the expense of underestimating the soil moisture variation. The pro­ 2018. Soil moisture and irrigation mapping in a semi-arid region, based on the
posed time series inversion scheme with these inversion skills achieved a synergetic use of Sentinel-1 and Sentinel-2 data. Remote Sens. 10, 1953.
competitive retrieval accuracy (median RMSE: ~0.072 m3/m3, median Buchhorn, M., Lesiv, M., Tsendbazar, N.-E., Herold, M., Bertels, L., Smets, B., 2020.
Copernicus global land cover layers—collection 2. Remote Sens. 12, 1044.
R: 0.64) at a high resolution of 100 m, with its generalization capability
Callens, M., Verhoest, N.E., Davidson, M.W., 2006. Parameterization of tillage-induced
confirmed for a long period of 4 years. single-scale soil roughness from 4-m profiles. IEEE Trans. Geosci. Remote Sens. 44,
878–888.
Calvet, J.-C., Fritz, N., Berne, C., Piguet, B., Maurel, W., Meurey, C., 2016. Deriving
CRediT authorship contribution statement
pedotransfer functions for soil quartz fraction in southern France from reverse
modeling. Soil 2, 615–629.
Liujun Zhu: Conceptualization, Methodology, Writing – original Chen, K.-S., Wu, T.-D., Tsang, L., Li, Q., Shi, J., Fung, A.K., 2003. Emission of rough
draft. Shanshui Yuan: Writing – original draft, Writing – review & surfaces calculated by the integral equation method with comparison to three-
dimensional moment method simulations. IEEE Trans. Geosci. Remote Sens. 41,
editing. Yi Liu: Writing – review & editing. Cheng Chen: Writing – 90–101.
review & editing. Jeffrey P. Walker: Writing – review & editing. Cook, D.R., 2016. Soil temperature and moisture profile (STAMP) system handbook.. In:
DOE Office of Science Atmospheric Radiation Measurement (ARM) Program.
Cui, C., Xu, J., Zeng, J., Chen, K.-S., Bai, X., Lu, H., Chen, Q., Zhao, T., 2017. Soil
Declaration of Competing Interest moisture mapping from satellites: an intercomparison of SMAP, SMOS, FY3B,
AMSR2, and ESA CCI over two dense network regions at different spatial scales.
The authors declare that they have no known competing financial Remote Sens. 10, 33.
interests or personal relationships that could have appeared to influence

13
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

Deb, K., 2014. Multi-objective optimization. In: Search methodologies. Springer, and Temperature Data from SoilSCAPE Sites, USA, ORNL DAAC, Oak Ridge,
pp. 403–449. Tennessee, USA.
Demargne, J., Wu, L., Regonda, S.K., Brown, J.D., Lee, H., He, M., Seo, D.-J., Neill, O.E., Chan, S., Njoku, E.G., Jackson, T., Bindlish, R., Chaubell, J., 2019. SMAP L3
Hartman, R., Herr, H.D., Fresch, M., 2014. The science of NOAA’s operational Radiometer Global Daily 36 km EASE-Grid Soil Moisture. N.N.S.a.I.D.C.D.A.A.
hydrologic ensemble forecast service. Bull. Am. Meteorol. Soc. 95, 79–98. Center.
Dorigo, W., Himmelbauer, I., Aberer, D., Schremmer, L., Petrakovic, I., Zappa, L., Njoku, E.G., Wilson, W.J., Yueh, S.H., Dinardo, S.J., Li, F.K., Jackson, T.J., Lakshmi, V.,
Preimesberger, W., Xaver, A., Annor, F., Ardö, J., 2021. The international soil Bolten, J., 2002. Observations of soil moisture using a passive and active low-
moisture network: serving earth system science for over a decade. Hydrol. Earth frequency microwave airborne sensor during SGP99. IEEE Trans. Geosci. Remote
Syst. Sci. 25, 5749–5804. Sens. 40, 2659–2673.
Dubois, P.C., Van Zyl, J., Engman, T., 1995. Measuring soil moisture with imaging Oh, Y., 2004. Quantitative retrieval of soil moisture content and surface roughness from
radars. IEEE Trans. Geosci. Remote Sens. 33, 915–926. multipolarized radar observations of bare soil surfaces. IEEE Trans. Geosci. Remote
El Hajj, M., Baghdadi, N., Zribi, M., Belaud, G., Cheviron, B., Courault, D., Charron, F., Sens. 42, 596–601.
2016. Soil moisture retrieval over irrigated grassland using X-band SAR data. Ojo, E.R., Bullock, P.R., L’Heureux, J., Powers, J., McNairn, H., Pacheco, A., 2015.
Remote Sens. Environ. 176, 202–218. Calibration and evaluation of a frequency domain reflectometry sensor for real-time
Fan, D., Zhao, T., Jiang, X., Xue, H., Moukomla, S., Kuntiyawichai, K., Shi, J., 2021. Soil soil moisture monitoring. Vadose Zone J. 14.
moisture retrieval from Sentinel-1 time-series data over croplands of northeastern Ouellette, J.D., Johnson, J.T., Balenzano, A., Mattia, F., Satalino, G., Kim, S.-B.,
Thailand. IEEE Geosci. Remote Sens. Lett. 19, 4011105. Dunbar, R.S., Colliander, A., Cosh, M.H., Caldwell, T.G., 2017. A time-series
Fung, A.K., 1994. Microwave scattering and emission models and their applications. approach to estimating soil moisture from vegetated surfaces using L-band radar
Artech House. backscatter. IEEE Trans. Geosci. Remote Sens. 55, 3186–3193.
Galle, S., Grippa, M., Peugeot, C., Moussa, I.B., Cappelaere, B., Demarty, J., Mougin, E., Palmisano, D., Mattia, F., Balenzano, A., Satalino, G., Pierdicca, N., Guarnieri, A.V.M.,
Panthou, G., Adjomayi, P., Agbossou, E., 2018. AMMA-CATCH, a critical zone 2020. Sentinel-1 sensitivity to soil moisture at high incidence angle and the impact
observatory in West Africa monitoring a region in transition. Vadose Zone J. 17, on retrieval over seasonal crops. IEEE Trans. Geosci. Remote Sens. 59, 7308–7321.
1–24. Peng, J., Albergel, C., Balenzano, A., Brocca, L., Cartus, O., Cosh, M.H., Crow, W.T.,
González-Zamora, Á., Sánchez, N., Pablos, M., Martínez-Fernández, J., 2019. CCI soil Dabrowska-Zielinska, K., Dadson, S., Davidson, M.W., 2020. A roadmap for high-
moisture assessment with SMOS soil moisture and in situ data under different resolution satellite soil moisture applications–confronting product characteristics
environmental conditions and spatial scales in Spain. Remote Sens. Environ. 225, with user requirements. Remote Sens. Environ. 112162.
469–482. Pierdicca, N., Castracane, P., Pulvirenti, L., 2008. Inversion of electromagnetic models
Gu, W., Tsang, L., Colliander, A., Yueh, S.H., 2021. Wave propagation in vegetation field for bare soil parameter estimation from multifrequency polarimetric SAR data.
using a hybrid method. IEEE Trans. Antennas Propag. 69, 6752–6761. Sensors 8, 8181–8200.
He, L., Qin, Q., Panciera, R., Tanase, M., Walker, J.P., Hong, Y., 2017. An extension of the Pierdicca, N., Pulvirenti, L., Bignami, C., 2010. Soil moisture estimation over vegetated
alpha approximation method for soil moisture estimation using time-series SAR data terrains using multitemporal remote sensing data. Remote Sens. Environ. 114,
over bare soil surfaces. IEEE Geosci. Remote Sens. Lett. 14, 1328–1332. 440–448.
Huang, S., Tsang, L., 2012. Electromagnetic scattering of randomly rough soil surfaces Quegan, S., Le Toan, T., Chave, J., Dall, J., Exbrayat, J.-F., Minh, D.H.T., Lomas, M.,
based on numerical solutions of Maxwell equations in three-dimensional simulations D’alessandro, M.M., Paillou, P., Papathanassiou, K., 2019. The European Space
using a hybrid UV/PBTG/SMCG method. IEEE Trans. Geosci. Remote Sens. 50, Agency BIOMASS mission: measuring forest above-ground biomass from space.
4025–4035. Remote Sens. Environ. 227, 44–60.
Jensen, K.H., Refsgaard, J.C., 2018. HOBE: the danish hydrological observatory. Vadose Rahman, M., Moran, M., Thoma, D., Bryant, R., Collins, C.H., Jackson, T., Orr, B.,
Zone J. 17, 1–24. Tischler, M., 2008. Mapping surface roughness and soil moisture using multi-angle
Joseph, A.T., van der Velde, R., O’Neill, P.E., Lang, R.H., Gish, T., 2008. Soil moisture radar imagery without ancillary data. Remote Sens. Environ. 112, 391–402.
retrieval during a corn growth cycle using L-band (1.6 GHz) radar observations. IEEE Sahebi, M., Angles, J., 2010. An inversion method based on multi-angular approaches for
Trans. Geosci. Remote Sens. 46, 2365–2374. estimating bare soil surface parameters from RADARSAT-1. Hydrol. Earth Syst. Sci.
Kellogg, K., Hoffman, P., Standley, S., Shaffer, S., Rosen, P., Edelstein, W., Dunn, C., 14, 2355–2366.
Baker, C., Barela, P., Shen, Y., 2020. NASA-ISRO synthetic aperture radar (NISAR) Schaefer, G.L., Cosh, M.H., Jackson, T.J., 2007. The USDA natural resources conservation
mission. In: 2020 IEEE Aerospace Conference. IEEE, pp. 1–21. service soil climate analysis network (SCAN). J. Atmos. Ocean. Technol. 24,
Kim, S.-B., Moghaddam, M., Tsang, L., Burgin, M., Xu, X., Njoku, E.G., 2014. Models of L- 2073–2077.
band radar backscattering coefficients over global terrain for soil moisture retrieval. Shi, H., Zhao, L., Yang, J., Lopez-Sanchez, J.M., Zhao, J., Sun, W., Shi, L., Li, P., 2021.
IEEE Trans. Geosci. Remote Sens. 52, 1381–1396. Soil moisture retrieval over agricultural fields from L-band multi-incidence and
Kim, S.-B., Tsang, L., Johnson, J.T., Huang, S., Van Zyl, J.J., Njoku, E.G., 2012. Soil multitemporal PolSAR observations using polarimetric decomposition techniques.
moisture retrieval using time-series radar observations over bare surfaces. IEEE Remote Sens. Environ. 261, 112485.
Trans. Geosci. Remote Sens. 50, 1853–1863. Shi, J., Wang, J., Hsu, A.Y., O’Neill, P.E., Engman, E.T., 1997. Estimation of bare surface
Kim, S.-B., van Zyl, J.J., Johnson, J.T., Moghaddam, M., Tsang, L., Colliander, A., soil moisture and surface roughness parameter using L-band SAR image data. IEEE
Dunbar, R.S., Jackson, T.J., Jaruwatanadilok, S., West, R., 2017. Surface soil Trans. Geosci. Remote Sens. 35, 1254–1266.
moisture retrieval using the L-band synthetic aperture radar onboard the soil Smith, A., Walker, J., Western, A., Young, R., Ellett, K., Pipunic, R., Grayson, R.,
moisture active-passive satellite and evaluation at Core validation sites. IEEE Trans. Siriwardena, L., Chiew, F., Richter, H., 2012. The Murrumbidgee soil moisture
Geosci. Remote Sens. 55, 1897–1914. monitoring network data set. Water Resour. Res. 48.
Kim, Y., Van Zyl, J.J., 2009. A time-series approach to estimate soil moisture using Su, Z., Wen, J., Dente, L., Van Der Velde, R., Wang, L., Ma, Y., Yang, K., Hu, Z., 2011. The
polarimetric radar data. IEEE Trans. Geosci. Remote Sens. 47, 2519–2527. tibetan plateau observatory of plateau scale soil moisture and soil temperature
Kornelsen, K.C., Coulibaly, P., 2013. Advances in soil moisture retrieval from synthetic (Tibet-Obs) for quantifying uncertainties in coarse resolution satellite and model
aperture radar and hydrological applications. J. Hydrol. 476, 460–489. products. Hydrol. Earth Syst. Sci. 15, 2303–2316.
Lang, R.H., Sighu, J.S., 1983. Electromagnetic backscattering from a layer of vegetation: Torres, R., Snoeij, P., Geudtner, D., Bibby, D., Davidson, M., Attema, E., Potin, P.,
a discrete approach. IEEE Trans. Geosci. Remote Sens. 62–71. Rommen, B., Floury, N., Brown, M., 2012. GMES Sentinel-1 mission. Remote Sens.
Larson, K.M., Small, E.E., Gutmann, E.D., Bilich, A.L., Braun, J.J., Zavorotny, V.U., 2008. Environ. 120, 9–24.
Use of GPS receivers as a soil moisture network for water cycle studies. Geophys. Ulaby, F.T., Long, D.G., Blackwell, W.J., Elachi, C., Fung, A.K., Ruf, C., Sarabandi, K.,
Res. Lett. 35. Zebker, H.A., Van Zyl, J., 2014. Microwave radar and radiometric remote sensing.
Lee, J.H., Budhathoki, S., Lindenschmidt, K.-E., 2021. Stochastic bias correction for The University of Michigan Press, Ann Arbor.
RADARSAT-2 soil moisture retrieved over vegetated areas. Geocarto International Verhoest, N., De Baets, B., Mattia, F., Satalino, G., Lucau, C., Defourny, P., 2007.
1–14. A possibilistic approach to soil moisture retrieval from ERS synthetic aperture radar
Lievens, H., Verhoest, N., De Keyser, E., Vernieuwe, H., Matgen, P., Álvarez-Mozos, J., De backscattering under soil roughness uncertainty. Water Resour. Res. 43.
Baets, B., 2011. Effective roughness modelling as a tool for soil moisture retrieval Vernieuwe, H., Verhoest, N.E., Lievens, H., De Baets, B., 2010. Possibilistic soil roughness
from C-and L-band SAR. Hydrol. Earth Syst. Sci. Discuss. 7, 4995–5031. identification for uncertainty reduction on SAR-retrieved soil moisture. IEEE Trans.
Mattia, F., Satalino, G., Dente, L., Pasquariello, G., 2006. Using a priori information to Geosci. Remote Sens. 49, 628–638.
improve soil moisture retrieval from ENVISAT ASAR AP data in semiarid regions. Wagner, W., Lemoine, G., Borgeaud, M., Rott, H., 1999a. A study of vegetation cover
IEEE Trans. Geosci. Remote Sens. 44, 900–912. effects on ERS scatterometer data. IEEE Trans. Geosci. Remote Sens. 37, 938–948.
Mattia, F., Satalino, G., Pauwels, V., Loew, A., 2009. Soil moisture retrieval through a Wagner, W., Lemoine, G., Rott, H., 1999b. A method for estimating soil moisture from
merging of multi-temporal L-band SAR data and hydrologic modelling. Hydrol. Earth ERS scatterometer and soil data. Remote Sens. Environ. 70, 191–207.
Syst. Sci. 13, 343–356. Wang, Z., Che, T., Zhao, T., Dai, L., Li, X., Wigneron, J.-P., 2021a. Evaluation of SMAP,
Merlin, O., Walker, J., Panciera, R., Young, R., Kalma, J., Kim, E., 2007. Calibration of a SMOS, and AMSR2 soil moisture products based on distributed ground observation
soil moisture sensor in heterogeneous terrain. In: Oxley, L., Kulasiri, D. (Eds.), 2007 network in cold and arid regions of China. IEEE J. Sel. Top. Appl. Earth Obs. Remote
International Congress on Modelling and Simulation (MODSIM). Modelling and Sens. 14, 8955–8970.
Simulation Society of Australia and New Zealand, pp. 2604–2610. Wang, Z., Zhao, T., Qiu, J., Zhao, X., Li, R., Wang, S., 2021b. Microwave-based
Merzouki, A., McNairn, H., 2015. A hybrid (Multi-angle and multi-Polarization) vegetation descriptors in the parameterization of water cloud model at L-band for
approach to soil moisture retrieval using the integral equation model: preparing for soil moisture retrieval over croplands. GISci. Remote Sens. 58, 48–67.
the RADARSAT constellation Mission. Can. J. Remote. Sens. 00. Ye, N., Walker, J.P., Wu, X., Jeu, R.D., Gao, Y., Jackson, T.J., Jonard, F., Kim, E.,
Moghaddam, M., Silva, A., Clewley, D., Akbar, R., Hussaini, S., Whitcomb, J., Merlin, O., Pauwels, V., Renzullo, L.J., Rüdiger, C., Sabaghy, S.C., Hebel, V., Yueh, S.
Devarakonda, R., Shrestha, R., Cook, R., Prakash, G., 2016. Soil Moisture Profiles H., Zhu, L., 2020. The soil moisture active passive experiments: validation of the
SMAP products in Australia. IEEE Trans. Geosci. Remote Sens. 59, 2922–2939.

14
L. Zhu et al. Remote Sensing of Environment 287 (2023) 113466

Zhan, X., Houser, P.R., Walker, J.P., Crow, W.T., 2006. A method for retrieving high- Zhu, L., Walker, J.P., Tsang, L., Huang, H., Ye, N., Rüdiger, C., 2019b. Soil moisture
resolution surface soil moisture from hydros L-band radiometer and radar retrieval from time series multi-angular radar data using a dry down constraint.
observations. Geosci. Remote Sens. IEEE Trans. 44, 1534–1544. Remote Sens. Environ. 231, 111237.
Zhao, T., Shi, J., Entekhabi, D., Jackson, T.J., Hu, L., Peng, Z., Yao, P., Li, S., Kang, C.S., Zhu, L., Walker, J.P., Ye, N., Rudiger, C., 2016. The effect of radar configuration on
2021. Retrievals of soil moisture and vegetation optical depth using a multi-channel effective correlation length. In: Electromagnetics in Advanced Applications (ICEAA),
collaborative algorithm. Remote Sens. Environ. 257, 112321. 2016 International Conference on. IEEE, pp. 820–823.
Zhao, T., Shi, J., Lv, L., Xu, H., Chen, D., Cui, Q., Jackson, T.J., Yan, G., Jia, L., Chen, L., Zribi, M., Foucras, M., Baghdadi, N., Demarty, J., Muddu, S., 2020. A new reflectivity
2020. Soil moisture experiment in the Luan River supporting new satellite mission index for the retrieval of surface soil moisture from radar data. IEEE J. Sel. Top.
opportunities. Remote Sens. Environ. 240, 111680. Appl. Earth Obs. Remote Sens. 14, 818–826.
Zhu, L., Si, R., Shen, X., Walker, J., 2022. An advanced change detection method for time Zribi, M., Gorrab, A., Baghdadi, N., 2014. A new soil roughness parameter for the
series soil moisture retrieval from Sentinel-1. Remote Sens. Environ. 279, 113137. modelling of radar backscattering over bare soil. Remote Sens. Environ. 152, 62–73.
Zhu, L., Walker, J.P., Shen, X., 2020. Stochastic ensemble methods for multi-SAR-mission Zribi, M., Saux-Picart, S., André, C., Descroix, L., Ottlé, C., Kallel, A., 2007. Soil moisture
soil moisture retrieval. Remote Sens. Environ. 251, 112099. mapping based on ASAR/ENVISAT radar data over a sahelian region. Int. J. Remote
Zhu, L., Walker, J.P., Tsang, L., Huang, H., Ye, N., Rüdiger, C., 2019a. A multi-frequency Sens. 28, 3547–3565.
framework for soil moisture retrieval from time series radar data. Remote Sens.
Environ. 235, 111433.

15