You are on page 1of 13

6780 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO.

12, DECEMBER 2017

Triple Collocation-Based Merging of


Satellite Soil Moisture Retrievals
Alexander Gruber, Wouter Arnoud Dorigo, Wade Crow, and Wolfgang Wagner, Senior Member, IEEE

Abstract— We propose a method for merging soil moisture mounted on satellites, which are available since 1978. Since
retrievals from spaceborne active and passive microwave instru- then, the number of available instruments has been growing
ments based on weighted averaging taking into account the steadily, including missions that are specifically dedicated
error characteristics of the individual data sets. The merging
scheme is parameterized using error variance estimates obtained to soil moisture monitoring such as the Soil Moisture and
from using triple collocation analysis (TCA). In regions where Ocean Salinity (SMOS) [2] and the Soil Moisture Active Pas-
TCA is deemed unreliable, we use correlation significance levels sive (SMAP) mission [3]. However, most available microwave
( p-values) as indicator for retrieval quality to decide whether to missions were not designed to observe soil moisture, and hence
use active data only, passive data only, or an unweighted average. their instruments have very different characteristics including
We apply the proposed merging scheme to active retrievals from
advanced scatterometer and passive retrievals from the Advanced different frequency bands, polarizations, and wide incidence
Microwave Scanning Radiometer—Earth Observing System angle ranges. In the view of soil moisture monitoring, all these
using Global Land Data Assimilation System-Noah to comple- antenna characteristics provide individual advantages and
ment the triplet required for TCA. The merged time series is eval- disadvantages, which make them more or less sensitive to
uated against soil moisture estimates from ERA-Interim/Land soil moisture and other surface conditions such as vegetation
and in situ measurements from the International Soil Moisture
Network using the European Space Agency’s (ESA’s) current density or radio frequency interference (RFI) contamination.
Climate Change Initiative—Soil Moisture (ESA CCI SM) product The long-term availability of various satellite-based sen-
version v02.3 as benchmark merging scheme. Results show that sors with different spatiotemporal coverage and measurement
the p-value classification provides a robust basis for decisions characteristics provides an opportunity for generating a value-
regarding using either active or passive data alone, or an
added product by merging soil moisture retrievals from differ-
unweighted average in cases where relative weights cannot be
estimated reliably, and that the weights estimated from TCA in ent instruments into a single data set following three main
almost all cases outperform the ternary decision upon which the purposes: 1) extending the temporal coverage beyond the
ESA CCI SM v02.3 is based. The proposed method forms the life time of single satellites by sequentially concatenating
basis for the new ESA CCI SM product version v03.x and higher. data sets from different sensors; 2) increasing the temporal
Index Terms— Accuracy, least squares approximation, radar, measurement density in periods where multiple instruments
radiometers, remote sensing, soil moisture, uncertainty. are available; and 3) reducing random errors with respect to
I. I NTRODUCTION the single input data sets by averaging observations acquired

S OIL moisture is one of the key internal state variables


of the global hydrological cycle. Consistent global soil
moisture records are essential for studying hydrology driven
from different instruments with close observation times.
In 2012, the European Space Agency (ESA) initiated the
Climate Change Initiative (CCI) project for soil moisture,
phenomena of the earth system such as climate, vegetation in which global long-term multisatellite soil moisture data
growth, and many others [1]. The importance and feasibility records have been generated by merging various available
of global soil moisture monitoring through earth observation active and passive microwave-based soil moisture data sets [4].
were acknowledged by the Global Climate Observing This product, henceforth referred to as ESA CCI SM, not
System (GCOS) via its endorsement as an essential climate only aims for providing a consistent soil moisture record from
variable. 1978 to the present, it also attempts to optimally combine
Some of the most important sources for global soil moisture various products by applying a merging scheme when multi-
data records are microwave radar and radiometer instruments ple instruments are available. The merging scheme that was
Manuscript received May 17, 2017; revised June 26, 2017 and July 19, employed in all ESA CCI SM data set versions up to the
2017; accepted July 19, 2017. Date of publication September 19, 2017; latest (v02.3) is based on the following steps: 1) all active data
date of current version November 22, 2017. This work was supported in
part by the eartH2Observe project through the European Union’s Seventh sets are combined into a merged active microwave product
Framework Programme under Grant 603608 and in part by the European by harmonizing their climatology; 2) all passive data sets
Space Agency Climate Change Initiative for Soil Moisture under Contract are combined into a merged passive microwave product by
4000104814/11/I-NB. (Corresponding author: Alexander Gruber.)
A. Gruber, W. A. Dorigo, and W. Wagner are with the Department harmonizing their climatology; and 3) a merged active–passive
of Geodesy and Geoinformation, TU Wien, 1040 Vienna, Austria (e-mail: data set is created based on the correlation between the merged
alexander.gruber@geo.tuwien.ac.at). active and the merged passive time series as follows. If the
W. Crow is with the Hydrology and Remote Sensing Laboratory, U.S.
Department of Agriculture, Beltsville, MD 20705-2350 USA. correlation at a certain location is above 0.65, then both active
This paper has supplementary downloadable material available at and passive data are combined using an unweighted averaging.
http://ieeexplore.ieee.org., provided by the author. If the correlation is lower, then the merged active data are
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. used if the multiyear average vegetation optical depth (VOD)
Digital Object Identifier 10.1109/TGRS.2017.2734070 at this location exceeds a certain threshold while the merged
0196-2892 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
GRUBER et al.: TRIPLE COLLOCATION-BASED MERGING OF SATELLITE SOIL MOISTURE RETRIEVALS 6781

passive data are used when the mean VOD is below this where x is the averaged soil moisture estimate; x =
threshold. To harmonize the individual time series, they are [x 1 , . . . , x N ] is a vector containing the temporally collocated
rescaled against the supposedly best active or passive input soil moisture estimates from the N different data sets that are
data set in the merged active and merged passive product, merged; and w = [w1 , . . . , w N ] is a vector containing the
respectively, and against long-term soil moisture estimates weights  that are given to these estimates where wi ∈ [0, 1]
from a land surface model in the merged active–passive data and i wi = 1. If the weights are chosen correctly, then
set using a cumulative distribution function (CDF)-matching the random errors of the merged time series will be lower
approach (for more details, see [4]–[7]). than those of the individual input data sets. The magnitude of
However, the current ESA CCI SM merging scheme up the weights that lead to a maximum random error reduction
to version v02.3 is suboptimal in that it does not take into is determined by the individual random error characteristics
account the (random) error characteristics of the individual of the input data sets and can be derived from their error
input data sets. For example, even though active soil moisture covariance matrix
retrievals are typically more accurate over more densely vege-
tated regions and passive soil moisture retrievals are typically w = (1C−1 1)−1 1C−1
more accurate over more sparsely vegetated areas [5], [8], σε2x = (1C−1 1)−1 (2)
this does not mean that the respective other product does
not contain useful complementary soil moisture information where C is the N × N error covariance matrix that holds
as well. Likewise, a high correlation between active and the random error variances σε2x of the input data sets in
passive data does not necessarily mean that both data sets i
the diagonals and their error covariances σεxi εxi in the off-
are equally accurate. In this paper, we propose an improved diagonals; 1 = [1, . . . , 1] is a ones-vector of length N;
weighted-average based method for merging active and passive and σε2x is the resulting random error variance of the merged
satellite soil moisture retrievals. It takes into account the estimates.
individual error characteristics of the merged data sets, which
are estimated using triple collocation analysis (TCA) [9]. This
TCA-based merging scheme forms the basis of the next
B. Weight Estimation
ESA CCI SM product versions v03.x and later.
The proposed merging scheme and its theoretical back- In principle, the covariance matrix C and hence also
ground are presented in Section II. Data sets used for test- the weights w are time dependent and represent the error
ing the proposed novel merging approach are described in (co)variance properties of each individual measurement. How-
Section III. An evaluation of the proposed method is shown ever, in practice, time-variant error (co)variance estimates are
in Section IV. Conclusions and an outlook to future work are seldom available, but it is often reasonable to assume them
provided in Section V. to be stationary [11]. Following this assumption allows us to
estimate a single set of weights that is representative of the
II. T HEORY entire merging period based on the individual (temporal) mean
Merging multiple satellite-based soil moisture data sets into error (co)variances of the data sets.
a single time series follows two main purposes: 1) increasing Recall that the purpose of this paper is to merge satellite soil
the temporal sampling density and/or coverage and 2) reducing moisture retrievals from active and passive microwave instru-
random retrieval errors. The former can be achieved by simply ments. Since errors in such data sets are commonly assumed
concatenating measurements from the different satellites while to be independent [12], we will neglect error covariances in
averaging temporally close observations is required for the this paper. A more detailed discussion on this assumption
latter. The requirements and consequences of both of these is provided in Section V. As a consequence, C becomes a
approaches will be discussed in detail in the following sec- diagonal matrix and only random error variance estimates are
tions. Note that even though we focus on merging satellite required for estimating w. According to (2), the weights can
soil moisture estimates in this paper, many of the concepts then be written as
are of general applicability.
A. Optimal Averaging σε2p σε2a
wa = wp = (3)
Optimality may be defined in different ways. Here we use σε2a + σε2p σε2a + σε2p
the term “optimal” in a least squares sense, i.e., that the
variance of residual random errors shall be minimized, which where the subscripts denote the active (a) and the pas-
is achieved through averaging observations from different sive ( p) data set, respectively. Note that the weight estimator
sources. If the magnitudes of the random errors inherent to in (2)—and hence also (3)—implicitly assumes that there
the data sets that are to be merged are not equal—which are no systematic differences between the data sets. This
is usually the case—such averaging requires a careful rela- assumption can be circumvented by deriving the weights from
tive weighting [10]. Formally, such weighted average can be signal-to-noise ratio (SNR) properties instead from random
written as error variances [11], [13]
N
 SNRa SNR p
x =w x= wi x i (1) wa = wp = (4)
i=1 SNRa + SNR p SNRa + SNR p

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
6782 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 12, DECEMBER 2017

with masking procedures. However, such concatenation results in


βa2 σ2 temporally varying random error properties of the merged time
SNRa = series. That is, in cases where observations in the merged data
σε2a
set are (weighted) averages of the input data sets, random
β 2p σ2 errors are reduced according to (2). In contrast, in cases
SNR p =
σε2p where single-sensor measurements are used, random errors
correspond to those of that sensor.
where βa2 σ2 and β 2p σ2 are the variances of the jointly This may become problematic if one data set is of sig-
observed true signal and the βi parameter represents a nificantly lower quality but with superior temporal sampling
systematic error in the active and the passive data set, such that it is regularly used to fill gaps in the other higher
respectively. However, before these weights can be actually quality data set. In this case, the (overall) quality of the merged
applied to merge the data sets, at least relative systematic time series may reduce significantly. To avoid this, one can
differences have to be corrected for. This is usually done set thresholds, for instance, a minimum SNR or a minimum
by rescaling the data sets into a common data space in relative weight that one data set has to have in order to
which case deriving the weights from the rescaled data be used for such gap filling. However, some tasks such as
sets using (3) converges—if done properly—to the opti- runoff modeling or drought monitoring may favor an improved
mal weights as estimated by (4). In this paper, we derive temporal coverage over highest possible SNRs because they
both rescaling parameters and random error variances using depend more strongly on the observation update frequency
TCA [11], [12], [14], [15]. (see [18], [19]). Therefore, it is a highly application-dependent
task to find a best tradeoff between the lowest possible random
C. Triple Collocation Analysis errors and the highest possible measurement density.
TCA is the most popular method for estimating random To support the process of finding such tradeoff, we present
error variances or SNRs of remotely sensed soil moisture data here an analytical investigation of the relationship between
sets over larger scales in the absence of reliable ground refer- error properties of the input and merged data sets and the
ence data [16]. Furthermore, it allows for estimating relative ratio of averaged and single-sensor measurements, which may
rescaling parameters between data sets that are optimal in a provide guidance for selecting thresholds in specific use cases.
sense that they take into account the individual random error Therefore, consider the overall mean error variance of a
properties and match the variability of the jointly observed merged data set that contains both averaged measurements of
signal [17]. TCA requires three collocated data sets with two data sets and single-sensor measurements of one of these
orthogonal and mutually independent errors. Here we apply data sets
TCA to the active and the passive data sets that are to be
n x σε2x + n s σε2s
merged, together with soil moisture estimates from a land σ 2ε = (6)
surface model that serves as the third data set in the triplet. The n
rescaling coefficients and random error variances are estimated where σ 2ε denotes the overall mean error variance; n x is the
from these data sets as number of days where both data sets are (weighted) averaged;
σa,m
β ap = n s is the number of days where only single-sensor measure-
σ p,m
σa, p σa,m ments are used; and n is the total number of measurements
σε2a = σa2 − with n = n x + n s . We use only one sensor, denoted with the
σ p,m
σ p,a σ p,m superscript s, for gap filling. Since extending this analysis with
2 2
σε p = σ p − (5) single-sensor measurements of the second data set is trivial,
σa,m
we—for simplicity of presentation—omit this scenario.
where β ap is used to linearly scale the passive against the active From (2), (3), and (6), it follows that changes in σ 2ε can
data set [ pa = β ap ( p − p)+a]; and σi2 and σi, j are the data set be directly derived from changes in the relative weighting—
variances and covariances with the superscripts denoting the regardless of the absolute error variance or SNR of the data
active, the passive, and the modeled (m) data set, respectively, sets—and the ratio between n x and n s . Let us therefore
as well as errors therein (εi ). For a detailed derivation of the express such relative change as a factor k in the random error
TCA estimators, we refer the reader to [11]. Note that it does variance, which since k · σ 2ε ≡ k −1 · S N R ε corresponds to an
not matter which data set serves as scaling reference as this additive reduction of the SNR by 10 · log(k) in decibel units.
does not impact the SNR properties of the rescaled or of the We choose the SNR as the main performance criterion because
merged time series. it provides a more meaningful—and better intercomparable—
representation of data set quality than random error variances
D. Merging Noncollocated Observations alone, and decibel units to normalize its nonlinear behavior
As mentioned earlier, merging data sets can also serve the for a more intuitive interpretation. For more details on the
purpose of increasing temporal measurement density or overall SNR metric, we refer the reader to [13] and [11].
data coverage through concatenating the measurements of the Fig. 1 shows the logarithmic SNR degradation (z-axis) as a
individual data sets instead of averaging them, for exam- function of the relative number of single-sensor measurements
ple, if they are not close together due to different overpass in the merged time series, i.e., gaps in the higher quality data
times, or if one product has data gaps due to RFI or other set (x-axis), and the weight that is given to the sensor that

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
GRUBER et al.: TRIPLE COLLOCATION-BASED MERGING OF SATELLITE SOIL MOISTURE RETRIEVALS 6783

uncertainties when applied to small samples. Following other


studies (see [8], [12], [13]), locations with less than 100 col-
located measurement triplets are removed from the analyses.
The merged time series is evaluated against soil moisture esti-
mates from the European Centre for Medium-Range Weather
Forecasts (ECMWF) Reanalysis (ERA-) Interim/Land model
and in situ measurements drawn from the International Soil
Moisture Network (ISMN).

A. ASCAT
Soil moisture estimates from ASCAT are retrieved
from backscatter measurements of MetOp-A using the
TU Wien Water Retrieval Package (WARP) algorithm
version 5.6.5 R1 [21], [22]. ASCAT operates at
C-band (5.26 GHz); retrieved soil moisture estimates
Fig. 1. Relative SNR degradation when using single-sensor measurements are provided as degree of saturation on a 0.25◦ regular grid
to fill gaps in a merged product. The x-axis shows the relative number of and are temporally resampled to 00:00 UTC. The WARP
single-sensor measurements in the merged time series, i.e., gaps in the higher
quality data set, the y-axis shows the weight that is given to the sensor that surface state flag [23] is used to remove measurements taken
is used for filling these gaps, and the z-axis shows the SNR degradation in under frozen or freezing/thawing conditions.
decibel units.

is used for filling these gaps according to the relative SNR B. AMSR-E
properties of both data sets (y-axis). If the weights are equally Soil moisture estimates from AMSR-E are provided by the
distributed between the sensors (i.e., 0.5 for both sensors), then Vrije Universiteit Amsterdam and retrieved using the Land
the SNR reduction upon gap filling is bounded to 3 dB, which Parameter Retrieval Model version 5 [24]. Data are provided
is reached if the number of gaps converges to 100% (i.e., if one in volumetric units on a regular grid with 0.25◦ grid spacing.
of the data sets is used alone). If one data set has a weight Only night time observations are used, which are temporally
of 0.25, the SNR has already decreased by 3 dB if 33% of resampled to 00:00 UTC. VOD estimates are used to mask out
observations in the other (better) data set are missing and filled retrievals with a high uncertainty due to dense vegetation [25].
with that data set, and it can further decrease by up to another RFI estimates are used to switch from C-band (6.9 GHz) to
3 dB when even more observations are missing and gap filled. X-band (10.7 GHz) retrievals in RFI-contaminated areas [24].
If the weight further decreases to 0.1, having 11% of data
missing and gap filled is already enough to cause a decline C. GLDAS-Noah
of the SNR by 3 dB, and having more filled gaps leads to The GLDAS-Noah model version 2.1 provides continuous
an SNR degradation of up to 10 dB, i.e., the error variance soil moisture data for four different depth layers at a spatial
increases by more than a factor of eight. resolution of 0.25◦ in a 3-h sampling rate [26]. Only soil
As mentioned before, finding an optimal tradeoff in terms of moisture estimates from the top layer (0–10 cm) at 00:00 UTC
SNR properties and measurement density is highly application are used in this paper. In addition, GLDAS-Noah provides esti-
dependent. The above-described analysis shall help finding mates for soil temperature and snow water equivalent (SWE),
such tradeoff in specific use cases. For the remainder of this which are used to mask out satellite observations if the
paper, we focus on evaluating the random error reduction upon soil temperature is below 0 ◦ C and the SWE is above 0%
weighted averaging of the proposed merging scheme in the as no reliable soil moisture retrieval is possible under such
real data scenario presented below. For this purpose, we do conditions.
not consider gap filling with single-sensor observations and
use temporally collocated observations only. D. ERA-Interim/Land
III. DATA S ETS ERA-Interim/Land is a global reanalysis of land-surface
The proposed TCA-based weighted average merging parameters that provides continuous soil moisture data for
scheme is tested globally using soil moisture retrievals from four different depth layers in a 6-h sampling rate on the
the active Advanced Scatterometer (ASCAT) and the passive approximately 0.7◦ ECMWF T255 grid [27], [28]. The data are
Advanced Microwave Scanning Radiometer—Earth Observ- resampled to 0.25◦ using a nearest neighbor approach. Only
ing System (AMSR-E). All collocated measurements avail- soil moisture estimates from the top layer (0–7 cm) at
able within their common observation period, i.e., Janu- 00:00 UTC are used in this paper.
ary 2007 to October 2011, are used. Relative weights are
estimated from (3). Soil moisture measurements from the E. ISMN
Global Land Data Assimilation System (GLDAS)-Noah model The ISMN is a data hosting facility that collects and
serve as third data set to estimate the required error variances harmonizes data from networks and field campaigns world-
as well as relative rescaling parameters by means of TCA wide and makes them available on a centralized Web por-
using (5). As shown in [20], TCA is prone to large estimation tal (http://ismn.geo.tuwien.ac.at) [29], [30]. In this paper,

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
6784 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 12, DECEMBER 2017

Fig. 2. Spatial locations of the ISMN stations that are used for evaluation.

we consider all stations that lie within the temporally over- TABLE I
lapping period of ASCAT and AMSR-E, i.e., January 2007 to M ERGING S CHEME BASED ON THE O NE -TAILED P EARSON
C ORRELATION S IGNIFICANCE B ETWEEN A CTIVE (a),
October 2011, and during this period have at least 50 mea- PASSIVE ( p), AND M ODELED (m) S OIL M OISTURE
surements coinciding with both ASCAT and AMSR-E within W ITH A 0.05 S IGNIFICANCE L EVEL
a ±12 h window. Measurements from sensors that are placed
deeper than 10 cm below the surface are excluded. The ISMN
also flags suspicious measurements such as spikes or signal
saturation as well as measurements taken under frozen condi-
tions or exceeding physically meaningful value ranges, based
on automated quality control procedures [31]. Measurements
flagged as suspicious are excluded in this paper. In total,
789 stations from 31 networks meet the above-described
requirements. References to the individual networks can be
found in Table S1. Stations locations are shown in Fig. 2.
Note the strong spatial bias toward the United States. Correlations are summarized for eight classes accord-
ing to the binary coded p-value combination and shown
IV. R ESULTS AND D ISCUSSION
in Fig. 3. The resulting decisions as to which class to use
A. Weight Estimation single sensors, unweighted averages, or weighted averages
Relative weights for ASCAT and AMSR-E soil moisture are shown in Table I; p-values are tested for a 0.05 sig-
retrievals are derived from their respective error variances nificance level. Since a one-tailed test is used, all negative
which are estimated using TCA. However, TCA is known correlations are deemed not significant. The resulting decisions
to have limited reliability in certain regions, such as high follow intuition. If all three data sets correlate significantly
latitudes, due to a lower temporal coverage or excessively with each other, then the error variances, and hence the
poor retrieval quality for one or more data sets in the averaging weights, can be estimated reliably from TCA, and
triplet [11]. Therefore, TCA results are often deemed performing a weighted averaging provides the best correlations
unreliable if one or more members of the triplet are not with respect to the evaluation data set. On the other hand,
significantly correlated according to a (one-tailed) student’s if neither of the data sets correlates significantly with any of
t-test (see [15], [32]). the others, then correlations with the evaluation data set are
Here we follow this approach and use correlation signifi- close to zero both for the single-sensor data and weighted
cance levels between all three data sets, hereinafter referred to or unweighted averages. Since in such cases we do not
as p-values, to select areas where to use TCA to calculate rela- know which of the data sets—if any—has a potential skill,
tive weights and, furthermore, to decide how to proceed in the we disregard these pixels.
remaining geographic regions, i.e., to use only active retrievals, If ASCAT correlates significantly with GLDAS-Noah
to use only passive retrievals, or to apply an unweighted but AMSR-E correlates neither with ASCAT nor with
averaging. Therefore, we empirically investigate for which GLDAS-Noah significantly, then ASCAT also shows signif-
p-value combination each of these cases performs best in icantly better correlations with the evaluation data set. In such
terms of Pearson correlation with respect to our global evalu- areas, we use ASCAT soil moisture retrievals only. If ASCAT
ation data set, i.e., ERA-Interim/Land. Note that the quality of correlates significantly with both GLDAS-Noah and AMSR-E
the chosen evaluation data set does not influence the relative but AMSR-E does not correlate significantly with the model,
ranking of the estimated correlation coefficients as long as its then using ASCAT data alone still seems to be the best choice.
errors are not correlated with those of the satellite retrievals, The reason for this may be that the relatively low SNR of
which is—for ERA-Interim/Land—commonly assumed to be AMSR-E requires a low weight in a weighted average and
true [12], [33]. provides only a marginal SNR improvement compared with

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
GRUBER et al.: TRIPLE COLLOCATION-BASED MERGING OF SATELLITE SOIL MOISTURE RETRIEVALS 6785

Fig. 3. Boxplot showing the correlations (y-axis) of ASCAT, AMSR-E, an unweighted average of both, and a weighted average of both with respect to
ERA-Interim/Land. Values on the x-axis represent the integer value associated with the binary combination of the decision “ p-value below 0.05” for the
correlations between AMSR-E and GLDAS, ASCAT and GLDAS, and ASCAT and AMSR-E, respectively (see Table I) and the numbers in brackets indicate
the number of pixels that were classified as such. Boxes represent the median and the interquartile range, and whiskers represent the 5% and 95% percentiles.
The text in red indicates the merging decision that was made according to Table I.

Fig. 4. Boxplot showing the correlations (y-axis) of ASCAT, AMSR-E, an unweighted average of both, and a weighted average of both with respect to
the ISMN stations. Values on the x-axis represent the integer value associated with the binary combination of the decision “ p-value below 0.05” for the
correlations between AMSR-E and GLDAS, ASCAT and GLDAS, and ASCAT and AMSR-E, respectively (see Table I) and the numbers in brackets indicate
the number of pixels that were classified as such. Boxes represent the median and the interquartile range, and whiskers represent the 5% and 95% percentiles.
The text in red indicates the merging decision that was made according to Table I.

using ASCAT alone (see Section IV-B), which makes the average still seems to be the best choice yet the overall
performance of the weighted average more sensitive to weight correlation to the evaluation data set is much lower. This
estimation uncertainties. may be an indication that ERA-Interim/Land and GLDAS-
If both ASCAT and AMSR-E correlate significantly with Noah are not completely independent (see Section V) but
GLDAS-Noah but not with each other, an unweighted average also that both ERA-Interim/Land and GLDAS-Noah may be
performs much better than the other options. If ASCAT cor- unreliable in these regions. However, such cases occur mainly
relates significantly with AMSR-E but neither of the satellites at very high latitudes, where land surface models are known
correlates significantly with GLDAS-Noah, an unweighted to have lower quality [7]. For this reason, we do not disregard

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
6786 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 12, DECEMBER 2017

such pixels but consider the significant correlation between B. Analytical Evaluation of the Merging Scheme
ASCAT and AMSR-E as an indication that both may still Remember that the main purpose of averaging soil moisture
contain valuable soil moisture information even though the estimates is the reduction of random errors with an optimum
correlation with respect to the evaluation data set is generally reduction (in a least squares sense) being achieved if the
low. Therefore, we perform an unweighted average in these weights are derived from the random error variances of the
regions. input data sets according to (2). The resulting random error
To independently evaluate the above described decisions, variance of the merged data set can be predicted using standard
we further calculate for each class the correlations between error propagation and—for one active and one passive data set
all ISMN stations and the active-only, the passive-only, with mutually uncorrelated errors—written as
the unweighted averaged, and the weighted averaged time
σε2x = wa2 σε2a + w2p σε2p . (7)
series, respectively, which are shown in Fig. 4. Due to their
very inhomogeneous spatial distribution (see Fig. 2), most of From (7), it follows that the random error reduction with
the stations are located within pixels where all three data sets respect to the single-sensor input products is—in multiplicative
are significantly correlated with each other (class 7). For the terms—simply the inverse of the weight that is assigned to the
other classes, only between 2 and 35 pixels are available, respective data set, which corresponds to an additive increase
which reduces the reliability of the quantile statistics due to of the logarithmic SNR by 10 · log(wi−1 ).
estimation uncertainties. Nonetheless, Fig. 4 confirms that the Fig. 6 shows such expected SNR improvement (in decibel
decisions from Table I (use active-only, use passive-only, use units) of the merged data set with respect to the better
an unweighted average, or use a weighted average) do indeed performing input data set, i.e., either ASCAT or AMSR-E
represent—on average—the best choice for all classes except depending on the location. No significant improvements are
for pixels where only the active and the passive data correlate achieved in areas where ASCAT or AMSR-E has a weight
significantly with each other (class 1). However, the mar- close to one, i.e., where one data set is much better than
ginal superiority of the median correlation between weighted the respective other, whereas the largest improvements are
averages and ground measurements over that between obtained in regions where the weights are equally distributed
unweighted averages and ground measurements might well be among the sensors (i.e., 0.5 for both). As it was already shown
a result of estimation uncertainties due to the small sample in Fig. 1, the expected magnitude of improvement in such
size (i.e., only 13 pixels are available for that class). regions is 3 dB, meaning that the merging halves the random
We can now convert the classification in Table I into: 1) rel- error variance relative to that of the input data sets.
ative weights of 0 and 1 for areas where ASCAT or AMSR-E Note that this analytical analysis of SNR improvement
are used alone; 2) weights of 0.5 and 0.5 where an unweighted upon merging assumes that the error variance estimates used
average is used; and 3) TCA-based weights derived from (3) to calculate the weights are error free. However, in reality,
where a weighted average is used. This results in the global the estimated error variances and hence weights will also be
map of relative weighting between ASCAT and AMSR-E soil subject to estimation uncertainties due to the limited number
moisture retrievals, which is shown in Fig. 5 together with of observations, possible violations of underlying assumptions
the weighting map that is currently employed in the ESA CCI such as nonstationary errors, or representativeness errors in
SM v02.3 data set. both the spatial and the temporal domain resulting from
In general, weight patterns agree very well. In areas where different spatial resolutions of the sensors and temporal decor-
the ESA CCI SM v02.3 scheme uses ASCAT alone, the relation of the observed signal due to different overpass times.
TCA-based merging scheme assigns relatively high weights As a consequence, actual SNR improvements may fluctuate
to ASCAT. Likewise, in areas where the ESA CCI SM v02.3 more or less around the analytical predictions in Fig. 6.
scheme uses AMSR-E alone, the TCA-based merging scheme Unfortunately, a global estimation of actual SNR improve-
also assigns relatively high weights to AMSR-E. The most ments is not straightforward. As mentioned earlier, TCA is
significant difference is that the ESA CCI SM v02.3 data currently the only method capable of providing absolute SNR
set uses a ternary (i.e., three-class based) decision of using estimates for soil moisture data sets over larger scales without
ASCAT only, AMSR-E only, or an arithmetic mean of both, requiring highly accurate reference data. However, it requires
whereas the TCA-based merging scheme allows for contin- three data sets with mutually uncorrelated errors, which are
uous transitions between those classes. Remember that the usually only available as one active, one passive, and one
ESA CCI SM v02.3 merging scheme bases the sensor selection modeled soil moisture product [11]. Therefore, the errors of
mainly on VOD levels following the findings of recent studies a merged active–passive soil moisture data set will almost
that active soil moisture retrievals tend to perform better over certainly be correlated to those of other active or passive
more densely vegetated areas whereas passive soil moisture products. Moreover, the random errors of different land surface
retrievals tend to perform better over more sparsely vegetated models have been found to show a certain degree of cross
regions [6], [8]. Therefore, it is not surprising that the weights correlation [10], [34]. That is, it is impossible to form a
estimated from TCA also closely follow global vegetation data set triplet with mutually uncorrelated random errors
patterns, which in turn provides evidence for their reliability. that allows us to estimate absolute SNRs of the merged
In the following sections, we investigate the proposed blending active–passive data set. We therefore further evaluate the
scheme in terms of general performance both analytically and merging scheme indirectly by calculating relative correlation
against evaluation data. improvements with respect to two independent evaluation

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
GRUBER et al.: TRIPLE COLLOCATION-BASED MERGING OF SATELLITE SOIL MOISTURE RETRIEVALS 6787

Fig. 5. (Top) Relative weights between ASCAT and AMSR-E soil moisture retrievals as used in the ESA CCI SM v02.3 merging scheme and the proposed
TCA-based merging scheme. (Bottom) The weight difference between the two merging schemes. White areas masked out because less than 100 collocated
measurements are available from ASCAT and AMSR-E.

data sources, i.e., ERA-Interim/Land and in situ soil moisture the ESA CCI SM v02.3 data set. For this benchmark scheme,
stations. instead of rescaling the data using a CDF-matching approach
as it is done in official product, we match the mean and
C. Evaluation of the Merging Scheme Against the standard deviation only in order to avoid disturbances
Reference Data in the correlation scores by matching higher order statistical
As benchmark for correlation improvements with respect to moments as this would disturb the estimated impact of the
the evaluation data, we use the merging scheme employed in actual merging.

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
6788 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 12, DECEMBER 2017

Fig. 6. SNR improvement in decibel units of the merged data set with respect to the superior single input data set, i.e., either ASCAT or AMSR-E, depending
on the geographic region. White areas masked out because less than 100 collocated measurements are available from ASCAT and AMSR-E.

Fig. 7. Difference in Pearson correlation between the proposed TCA-based merging scheme and ERA-Interim/Land, and the ESA CCI v02.3 merging scheme
and ERA-Interim/Land (without CDF matching). White areas masked out because less than 100 collocated measurements are available from ASCAT and
AMSR-E.

Fig. 8. (a) and (b) Difference in Pearson correlation between the proposed TCA-based merging scheme and ISMN stations, and the ESA CCI v02.3 merging
scheme and ISMN stations. (c) Difference in Pearson correlation between the proposed TCA-based merging scheme and ISMN stations, and the ESA CCI v02.3
merging scheme and ERA-Interim/Land.

Fig. 7 shows global correlation improvements in the pro- CDF matching) with respect to ERA-Interim/Land. Obviously,
posed TCA-based merging scheme over the ternary merg- correlation differences are zero if the weights given to the
ing scheme of the ESA CCI SM v02.3 data set (without sensors are equal in both merging schemes (see Fig. 5),

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
GRUBER et al.: TRIPLE COLLOCATION-BASED MERGING OF SATELLITE SOIL MOISTURE RETRIEVALS 6789

and that the employed p-value classification is—with a few


exceptions—capable of making plausible decisions in regions
where the reliability of TCA estimates is questioned. To further
strengthen these conclusions, we repeat the analysis with
in situ soil moisture measurements as second independent
evaluation data source.
Fig. 8 shows the correlation improvements of the pro-
posed TCA-based merging scheme over the ESA CCI SM
v02.3 merging scheme with respect to all in situ soil mois-
ture stations of the ISMN that provide at least 50 measure-
ments coinciding with the merged data set, and compares
these improvements with those relative to ERA-Interim/Land
(at spatially coinciding locations). Note that the spatial avail-
ability of in situ stations is very scarce and they are mainly
located in the contiguous United States, which may bias the
overall picture (see Fig. 2). However, results are consistent
with those from the model based evaluation. Improvements
are almost always positive and spatial patterns correlate with
R = 0.70, which further strengthens the reliability of the
estimated weights and provides further evidence for the supe-
riority of the TCA-based blending scheme.

V. C ONCLUSION
We propose a method for merging soil moisture retrievals
from spaceborne active and passive microwave instruments
based on weighted averaging taking into account the error
Fig. 9. Flowchart of the proposed merging scheme. characteristics of the individual data sets. This method forms
the basis for the new ESA CCI SM product version v03.x
i.e., in regions where both merging schemes use the same and higher. A flowchart of the proposed algorithm is shown
single-sensor data set or an unweighted average. in Fig. 9.
In general, correlations improve in most regions of the The merging scheme is parameterized using error variance
world. The magnitude of improvements follows the weight estimates obtained from using TCA. In regions where TCA is
differences, which provides evidence for the accuracy of the deemed unreliable, we use correlation significance levels
estimated weights. However, negative correlation differences ( p-values) as indicator for retrieval quality to decide
can be found in some areas, indicating a performance decline, whether to use active data only, passive data only, or an
but these are mainly areas where no weighting is employed unweighted average. We apply the proposed merging scheme
as advised by the p-value classification. Such areas are to active retrievals from ASCAT and passive retrievals from
found, for example, in northern latitudes where the p-value AMSR-E in the period from 2007 to 2011 using GLDAS-Noah
classification suggests an unweighted average as opposed to complement the triplet required for TCA, and evaluate the
to the ESA CCI SM v02.3 scheme that uses ASCAT data obtained merged time series against soil moisture estimates
alone, or western China where the ESA CCI SM v02.3 scheme from ERA-Interim/Land and in situ measurements from the
uses passive retrievals only while the p-value classification ISMN using the ESA CCI SM v02.3 data set as benchmark
again suggests the use of unweighted averages. While these are merging scheme. Results show that the p-value classification
examples that indicate flaws in the p-value classification, only provides—with few exceptions especially in northern latitudes
a few regions in which a weighted average is performed show a and desert areas—a robust basis for decisions regarding using
correlation decline compared to the ESA CCI SM v02.3 merg- either active or passive data alone, or an unweighted average
ing scheme. Such cases can be found, for example, in the Great in cases where relative weights cannot be estimated reliably,
Sand Sea Desert in Libya, where the weighting suggests a and that the weights estimated from TCA in almost all cases
strong usage of ASCAT retrievals as opposed to the AMSR-E outperform the ternary decision upon which the ESA CCI
only usage in the ESA CCI SM v02.3 scheme. However, SM v02.3 is based. However, a few limitations and open issues
spaceborne soil moisture retrievals, especially from active should be mentioned.
microwave instruments, are generally difficult for deserts due In this paper, we demonstrate the performance of the pro-
to very low SNRs [35]. However, further investigation why the posed TCA-based merging scheme using one active and one
proposed merging scheme seems to fail in few small regions passive soil moisture data set, respectively, which are assumed
is beyond the scope of this paper. to have uncorrelated errors. However, including multiple active
In general, we can conclude from Fig. 7 that relative and/or passive satellite data sets, as done in the ESA CCI SM
weights are estimated by TCA with reasonable accuracy product, theoretically requires the parameterization of error
and lead to improvements in most regions of the world, cross covariances in addition to error variances in order

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
6790 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 12, DECEMBER 2017

to estimate the optimal magnitude of the relative weights. into the official ESA CCI SM product, and for this pur-
Recently, [34] proposed an extension to TCA where the pose, we are currently investigating solutions to: 1) cope
inclusion of more than three data sets in the analysis allows with nonzero error cross-correlations in the weight parame-
for—at least partly—resolving nonzero error cross-correlation terization; 2) estimate seasonally varying weights using a
structures, yet a demonstration of the robustness of the method nonstationary TCA approach; 3) estimate random errors on
on a global scale is still pending. Therefore, one may for different time scales to further improve relative weightings;
practical reasons neglect error cross correlations between and 4) modify the TCA implementation by excluding model
different active or passive data sets at the cost of nonoptimal data from the analysis in order to obtain a fully model-
SNR improvements, or make a conservative educated guess independent merged multisatellite soil moisture product.
for error cross-correlation levels for data sets where they are Finally, we want to point out that TCA has already been
expected. applied successfully in other fields of hydrometeorology,
Gruber et al. [34] further showed that different land surface oceanography, and ecology (see [40]–[44]). Therefore, even
models—such as GLDAS-Noah and ERA-Interim/Land—may though demonstrated for soil moisture, the proposed method
also be prone to error cross correlation, which puts the is potentially transferable to other biogeophysical variables.
reliability of ERA-Interim/Land as independent evaluation
data set into question. However, in theory, GLDAS-Noah ACKNOWLEDGMENT
serves merely as an instrument in the analysis to resolve The authors would like to thank the in situ data providers
the relationship between the two satellite instruments [36]. for sharing their data with the International Soil Moisture
Therefore, its own inherent error properties should not affect Network.
the magnitude of the estimated merging weights. That said,
TCA estimates random errors with respect to a signal that
R EFERENCES
is jointly observed by all members of the triplet. Therefore,
processes unresolved by the model may indeed change the [1] D. R. Legates et al., “Soil moisture: A central and unifying theme
in physical geography,” Progr. Phys. Geograph., vol. 35, no. 1,
error variance, and hence weight estimates, attributed to the pp. 65–86, 2011. [Online]. Available: http://ppg.sagepub.com/content/
satellite soil moisture retrievals. For this reason, we rely 35/1/65.abstract
on the ground reference measurements as an indicator of [2] Y. H. Kerr et al., “The SMOS mission: New tool for monitoring
key elements of the global water cycle,” Proc. IEEE, vol. 98, no. 5,
whether or not to trust the model-based evaluation keeping pp. 666–687, May 2010.
in mind that certain regional effects may not have been fully [3] D. Entekhabi et al., “The Soil Moisture Active Passive (SMAP) mis-
captured due to limited spatial availability of ground stations, sion,” Proc. IEEE, vol. 98, no. 5, pp. 704–716, May 2010.
[4] W. Wagner et al., “Fusion of active and passive microwave observa-
for example, in high latitudes. tions to create an essential climate variable data record on soil mois-
Another important assumption that we—and most TCA- ture,” in Proc. 22nd ISPRS Congr., Melbourne, VIC, Australia, 2012,
based studies—make is the stationarity of random errors. How- pp. 315–321.
[5] Y. Y. Liu et al., “Developing an improved soil moisture dataset by blend-
ever, while for many applications, e.g., for validation purposes, ing passive and active microwave satellite-based retrievals,” Hydrol.
it might be sufficient to estimate an average random error Earth Syst. Sci., vol. 15, no. 2, pp. 425–436, 2011. [Online]. Available:
variance for a certain investigation period, other applications http://www.hydrol-earth-syst-sci.net/15/425/2011/
[6] Y. Y. Liu et al., “Trend-preserving blending of passive and active
such as data assimilation usually prefer error estimates on microwave soil moisture retrievals,” Remote Sens. Environ., vol. 123,
a per-measurement basis. Attempts have already been made pp. 280–297, Aug. 2012. [Online]. Available: http://www.sciencedirect.
to use moving-window-based TCA approaches to estimate com/science/article/pii/S0034425712001332
[7] W. Dorigo et al., “ESA CCI soil moisture for improved earth system
seasonally varying error variances [37], yet the significantly understanding: State-of-the art and future directions,” Remote Sens.
lower sample size of such methods hampers their reliability Environ., to be published. [Online]. Available: http://dx.doi.org/10.
so that lumped average random error estimates might still be 1016/j.rse.2017.07.001
[8] A. Al-Yaari et al., “Global-scale comparison of passive (SMOS) and
the more reliable choice. active (ASCAT) satellite based microwave soil moisture retrievals with
Moreover, in addition to seasonal variations in the random soil moisture simulations (MERRA-LAND),” Remote Sens. Environ.,
error variance of satellite soil moisture observations, [38] has vol. 152, pp. 614–626, Sep. 2014.
[9] M. T. Yilmaz, W. T. Crow, M. C. Anderson, and C. Hain, “An objec-
shown that random errors generally manifest differently on tive methodology for merging satellite- and model-based soil moisture
different time scales. That is, lower frequency components in products,” Water Resour. Res., vol. 48, no. 11, p. W11502, 2012,
an observation time series, i.e., its seasonal cycle, may have doi: 10.1029/2011WR011682.
[10] W. T. Crow, C.-H. Su, D. Ryu, and M. T. Yilmaz, “Optimal averaging
different random error characteristics than its higher frequency of soil moisture predictions from ensemble land surface model simula-
components, i.e., anomalies. Draper and Reichle [39] have tions,” Water Resour. Res., vol. 51, no. 11, pp. 9273–9289, 2015.
already shown that assimilating soil moisture observations [11] A. Gruber, C.-H. Su, S. Zwieback, W. Crow, W. Dorigo, and
W. Wagner, “Recent advances in (soil moisture) triple collocation
on different time scales may be superior to the assimilation analysis,” Int. J. Appl. Earth Observ. Geoinf., vol. 45, pp. 200–211,
of the lumped time series. Similarly, also the merging of Mar. 2016.
satellite soil moisture data sets could potentially benefit from [12] W. A. Dorigo et al., “Error characterisation of global active and passive
microwave soil moisture datasets,” Hydrol. Earth Syst. Sci., vol. 14,
a multitimescale-based error characterization, i.e., where sep- no. 12, pp. 2605–2616, 2010. [Online]. Available: http://www.hydrol-
arate weights are estimated for different time scales. earth-syst-sci.net/14/2605/2010/
Future work will focus on applying the proposed [13] C. Draper, R. Reichle, R. de Jeu, V. Naeimi, R. Parinussa, and
W. Wagner, “Estimating root mean square errors in remotely sensed
TCA-based merging scheme to the whole fleet of microwave soil moisture over continental scale domains,” Remote Sens. Environ.,
instrument-carrying satellites that are currently being merged vol. 137, pp. 288–298, Oct. 2013.

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
GRUBER et al.: TRIPLE COLLOCATION-BASED MERGING OF SATELLITE SOIL MOISTURE RETRIEVALS 6791

[14] A. Stoffelen, “Toward the true near-surface wind speed: Error mod- [35] M. Vreugdenhil, W. A. Dorigo, W. Wagner, R. A. de Jeu, S. Hahn,
eling and calibration using triple collocation,” J. Geophys. Res., and M. J. van Marle, “Analyzing the vegetation parameterization in the
vol. 103, no. C4, pp. 7755–7766, 1998. [Online]. Available: http://dx.doi. TU-Wien ASCAT soil moisture retrieval,” IEEE Trans. Geosci. Remote
org/10.1029/97JC03180 Sens., vol. 54, no. 6, pp. 3513–3531, Jun. 2016.
[15] K. Scipal, T. Holmes, R. de Jeu, V. Naeimi, and W. Wagner, [36] C.-H. Su, D. Ryu, W. T. Crow, and A. W. Western, “Beyond triple
“A possible solution for the problem of estimating the error structure collocation: Applications to soil moisture monitoring,” J. Geophys. Res.,
of global soil moisture data sets,” Geophys. Res. Lett., vol. 35, no. 24, Atmos., vol. 119, no. 11, pp. 6419–6439, 2014.
p. L24403, Dec. 2008. [Online]. Available: http://dx.doi.org/10.1029/ [37] A. Loew and F. Schlenz, “A dynamic approach for evaluating coarse
2008GL035599 scale satellite soil moisture products,” Hydrol. Earth Syst. Sci., vol. 15,
[16] F. Chen et al., “Application of triple collocation in ground-based no. 1, pp. 75–90, 2011. [Online]. Available: http://www.hydrol-earth-
validation of soil moisture active/passive (SMAP) level 2 data products,” syst-sci.net/15/75/2011/
IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 10, no. 2, [38] C.-H. Su and D. Ryu, “Multi-scale analysis of bias correction of
pp. 489–502, Feb. 2017. soil moisture,” Hydrol. Earth Syst. Sci., vol. 19, no. 1, pp. 17–31,
[17] M. T. Yilmaz and W. T. Crow, “The optimality of potential rescaling 2015.
approaches in land data assimilation,” J. Hydrometeorol., vol. 14, no. 2, [39] C. Draper and R. Reichle, “The impact of near-surface soil mois-
pp. 650–660, 2013. ture assimilation at subseasonal, seasonal, and inter-annual timescales,”
[18] Y. Tramblay, E. Amoussou, W. Dorigo, and G. Mahé, “Flood risk under Hydrol. Earth Syst. Sci., vol. 19, no. 12, p. 4831, 2015.
future climate in data sparse regions: Linking extreme value models [40] J. Vogelzang, A. Stoffelen, A. Verhoef, and J. Figa-Saldaña, “On the
and flood generating processes,” J. Hydrol., vol. 519, pp. 549–558, quality of high-resolution scatterometer winds,” J. Geophys. Res.,
Nov. 2014. vol. 116, no. C10, p. C10033, 2011, doi: 10.1029/2010JC006640.
[19] Y. Liu et al., “Agriculture intensifies soil moisture decline in Northern [41] S. Caires and A. Sterl, “Validation of ocean wind and wave data using
China,” Sci. Rep., vol. 5, Jul. 2015, Art. no. 11261. triple collocation,” J. Geophys. Res., vol. 108, no. C3, p. 3098, 2003,
[20] S. Zwieback, K. Scipal, W. Dorigo, and W. Wagner, “Structural and sta- doi: 10.1029/2002JC001491.
tistical properties of the collocation technique for error characterization,” [42] R. Roebeling, E. Wolters, J. Meirink, and H. Leijnse, “Triple collocation
Nonlinear Process. Geophys., vol. 19, no. 1, pp. 69–80, 2012. [Online]. of summer precipitation retrievals from SEVIRI over Europe with
Available: http://www.nonlin-processes-geophys.net/19/69/2012/ gridded rain gauge and weather radar data,” J. Hydrometeorol., vol. 13,
[21] W. Wagner, G. Lemoine, and H. Rott, “A method for estimating soil no. 5, pp. 1552–1566, 2012.
moisture from ERS scatterometer and soil data,” Remote Sens. Environ., [43] H. Fang, S. Wei, C. Jiang, and K. Scipal, “Theoretical uncertainty
vol. 70, no. 2, pp. 191–207, 1999. [Online]. Available: http://www. analysis of global MODIS, CYCLOPES, and GLOBCARBON LAI
sciencedirect.com/science/article/pii/S003442579900036X products using a triple collocation method,” Remote Sens. Environ.,
[22] V. Naeimi, K. Scipal, Z. Bartalis, S. Hasenauer, and W. Wagner, vol. 124, pp. 610–621, Sep. 2012.
“An improved soil moisture retrieval algorithm for ERS and METOP
[44] P. D’Odorico et al., “Intercomparison of fraction of absorbed photo-
scatterometer observations,” IEEE Trans. Geosci. Remote Sens., vol. 47,
synthetically active radiation products derived from satellite data over
no. 7, pp. 1999–2013, Jul. 2009.
Europe,” Remote Sens. Environ., vol. 142, pp. 141–154, Feb. 2014.
[23] V. Naeimi et al., “ASCAT Surface State Flag (SSF): Extracting informa-
[45] T. Pellarin, J. P. Laurent, B. Cappelaere, B. Decharme, L. Descroix, and
tion on surface freeze/thaw conditions from backscatter data using an
D. Ramier, “Hydrological modelling and associated microwave emission
empirical threshold-analysis algorithm,” IEEE Trans. Geosci. Remote
of a semi-arid region in South-western Niger,” J. Hydrol., vol. 375,
Sens., vol. 50, no. 7, pp. 2566–2582, Jul. 2012.
nos. 1–2, pp. 262–272, 2009.
[24] M. Owe, R. de Jeu, and T. Holmes, “Multisensor historical cli-
[46] J. Ardö, “A 10-year dataset of basic meteorology and soil prop-
matology of satellite-derived global land surface moisture,” J. Geo-
phys. Res., Earth Surf., vol. 113, no. F1, p. F01002, 2008, erties in central sudan,” Dataset Papers Geosci., vol. 2013, 2013,
Art. no. 297973, doi: 10.7167/2013/297973.
doi: 10.1029/2007JF000769.
[25] R. M. Parinussa, A. G. C. A. Meesters, Y. Y. Liu, W. Dorigo, [47] M. Zreda et al., “COSMOS: The cosmic-ray soil moisture observing
W. Wagner, and R. A. de Jeu, “Error estimates for near-real-time system,” Hydrol. Earth Syst. Sci., vol. 16, no. 11, pp. 4079–4099,
satellite soil moisture as derived from the land parameter retrieval 2012.
model,” IEEE Geosci. Remote Sens. Lett., vol. 8, no. 4, pp. 779–783, [48] K. Yang et al., “A multiscale soil moisture and freeze–thaw monitoring
Jul. 2011. network on the third pole,” Bull. Amer. Meteorol. Soc., vol. 94, no. 12,
[26] M. Rodell et al., “The global land data assimilation system,” pp. 1907–1916, 2013.
Bull. Amer. Meteorol. Soc., vol. 85, no. 3, pp. 381–394, [49] T. Tagesson et al., “Ecosystem properties of semiarid savanna grassland
2004. in west africa and its relationship with environmental variability,” Global
[27] C. Albergel et al., “Skill and global trend analysis of soil moisture from Change Biol., vol. 21, no. 1, pp. 250–264, 2015.
reanalyses and microwave remote sensing,” J. Hydrometeorol., vol. 14, [50] S. Bircher, N. Skou, K. H. Jensen, J. Walker, and L. Rasmussen, “A soil
no. 4, pp. 1259–1277, 2013. moisture and temperature network for SMOS validation in western
[28] G. Balsamo et al., “ERA-Interim/Land: A global land surface reanaly- denmark,” Hydrol. Earth Syst. Sci., vol. 16, no. 5, pp. 1445–1463, 2012.
sis data set,” Hydrol. Earth Syst. Sci., vol. 19, no. 1, pp. 389–407, [51] S. E. Hollinger and S. A. Isard, “A soil moisture climatology of illinois,”
2015. J. Climate, vol. 7, no. 5, pp. 822–833, 1994.
[29] W. Dorigo et al., “A new international network for in situ soil moisture [52] Z. Su et al., “The Tibetan Plateau observatory of plateau scale soil
data,” Eos Trans. Amer. Geophys. Union, vol. 92, no. 17, pp. 141–142, moisture and soil temperature (Tibet-Obs) for quantifying uncertain-
2011. ties in coarse resolution satellite and model products,” Hydrol. Earth
[30] W. A. Dorigo et al., “The international soil moisture network: A data Syst. Sci., vol. 15, no. 7, pp. 2303–2316, 2011. [Online]. Available:
hosting facility for global in situ soil moisture measurements,” Hydrol. http://www.hydrol-earth-syst-sci.net/15/2303/2011/
Earth Syst. Sci., vol. 15, no. 5, pp. 1675–1698, 2011. [53] A. B. Smith et al., “The Murrumbidgee soil moisture monitoring
[31] W. Dorigo et al., “Global automated quality control of in situ soil network data set,” Water Resour. Res., vol. 48, no. 7, p. W07701, 2012,
moisture data from the international soil moisture network,” Vadose Zone doi: 10.1029/2012WR011976.
J., vol. 12, no. 3, 2013. [54] K. M. Larson, E. E. Small, E. D. Gutmann, A. L. Bilich, J. J. Braun,
[32] A. Gruber, W. Dorigo, S. Zwieback, A. Xaver, and W. Wagner, and V. U. Zavorotny, “Use of GPS receivers as a soil moisture network
“Characterizing coarse-scale representativeness of in situ soil moisture for water cycle studies,” Geophys. Res. Lett., vol. 35, no. 24, p. L24405,
measurements from the international soil moisture network,” Vadose 2008, doi: 10.1029/2008GL036013.
Zone J., vol. 12, no. 2, 2013. [55] C. Rüdiger et al., “Goulburn river experimental catchment data
[33] A. Gruber, W. Crow, W. Dorigo, and W. Wagner, “The potential of set,” Water Resour. Res., vol. 43, no. 10, p. W10403, 2007,
2D Kalman filtering for soil moisture data assimilation,” Remote Sens. doi: 10.1029/2006WR005837.
Environ., vol. 171, pp. 137–148, Dec. 2015. [56] C. Albergel et al., “From near-surface to root-zone soil moisture
[34] A. Gruber, C.-H. Su, W. T. Crow, S. Zwieback, W. A. Dorigo, and using an exponential filter: An assessment of the method based
W. Wagner, “Estimating error cross-correlations in soil moisture data sets on in-situ observations and model simulations,” Hydrol. Earth Syst.
using extended collocation analysis,” J. Geophys. Res., Atmos., vol. 121, Sci., vol. 12, no. 6, pp. 1323–1337, 2008. [Online]. Available:
no. 3, pp. 1208–1219, 2016. http://europepmc.org/abstract/AGR/IND44347237

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.
6792 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 12, DECEMBER 2017

[57] G. Leavesley et al., “A modeling framework for improved agricultural Wade Crow received the Ph.D. degree from Prince-
water supply forecasting,” in Proc. AGU Fall Meet. Abstracts, vol. 1. ton University, Princeton, NJ, USA, in 2001.
2008, p. 0497. He is currently a Research Physical Scientist and
[58] W. Marczewski et al., “Strategies for validating and directions a Project Lead Scientist with the Hydrology and
for employing SMOS data, in the Cal-Val project SWEX (3275) Remote Sensing Laboratory, Agricultural Research
for wetlands,” Hydrol. Earth Syst. Sci. Discuss, vol. 7, no. 5, Service, U.S. Department of Agriculture, Beltsville,
pp. 7007–7057, 2010. [Online]. Available: http://www.hydrol-earth-syst- MD, USA. He has been with the science teams for
sci-discuss.net/7/7007/2010/ the NASA GPM, Hydros, SMAP, and AirMOSS
[59] S. Zacharias et al., “A network of terrestrial environmental observatories missions. His research interests include the devel-
in Germany,” Vadose Zone J., vol. 10, no. 3, pp. 955–973, 2011. opment of hydrologic and agricultural applications
[60] F. Schlenz, J. T. dall’Amico, A. Loew, and W. Mauser, “Uncertainty for remote sensing data and the development of
assessment of the SMOS validation in the upper danube catchment,” appropriate data assimilation approaches to facilitate this goal—with a special
IEEE Trans. Geosci. Remote Sens., vol. 50, no. 5, pp. 1517–1529, emphasis on techniques that fuse information from various disparate remote
May 2012. sensing sources.
[61] L. Brocca et al., “Soil moisture estimation through ASCAT and Dr. Crow was an Editor of the American Meteorological Society’s Journal
AMSR-E sensors: An intercomparison and validation study across of Hydrometeorology.
Europe,” Remote Sens. Environ., vol. 115, no. 12, pp. 3390–3408, 2011.
[Online]. Available: http://www.sciencedirect.com/science/article/pii/
S0034425711002756
[62] J. E. Bell et al., “U.S. climate reference network soil moisture
and temperature observations,” J. Hydrometeorol., vol. 14, no. 3,
pp. 977–988, 2013.
[63] T. J. Jackson et al., “Validation of advanced microwave scanning
radiometer soil moisture products,” IEEE Trans. Geosci. Remote Sens.,
vol. 48, no. 12, pp. 4256–4272, Dec. 2010.

Alexander Gruber was born in Vienna, Austria,


in 1988. He received the Dipl.-Ing. (M.Sc.) degree
in geodesy and geophysics and the Dr.techn. (Ph.D.)
degree in surveying and geoinformation, both with
excellence, from the Vienna University of Technol-
ogy, Vienna, in 2013 and 2016, respectively.
Since 2011, he has been a Researcher and a
Lecturer with the Research Group Remote Sensing,
Department of Geodesy and Geoinformation, Vienna
University of Technology, and has been involved
in a large number of research projects. He has
authored or co-authored a large number of peer-reviewed journal papers,
conference papers, and book chapters. His research interests include geophys-
ical parameter retrieval with a focus on soil moisture retrieval from synthetic
aperture radar and scatterometer instruments, error characterization of earth
observation datasets, up and downscaling techniques, and data assimilation.
Dr. Gruber is a reviewer for many SCI journals in the field of remote sensing
and earth observation.
Wolfgang Wagner (M’98–SM’07) was born in
Wouter Arnoud Dorigo received the M.Sc. degree Wels, Austria, in 1969. He received the Dipl.Ing.
in physical geography with an emphasis on remote (M.Sc.) degree in physics and the Dr.techn. (Ph.D.)
sensing from Utrecht University, Utrecht, The degree in remote sensing, both with excellence, from
Netherlands, in 2000, and the Ph.D. degree in remote the Vienna University of Technology (TU Wien),
sensing from the Technical University of Munich, Vienna, Austria, in 1995 and 1999, respectively.
Munich, Germany, in 2008. From 1999 to 2001, he was with the German
From 2002 to 2006, he was with the Imag- Aerospace Agency. In 2001, he was a Professor of
ing Spectroscopy Group, German Remote Sens- remote sensing with TU Wien. Since 2012, he has
ing Data Center, German Aerospace Center, been the Head of the Department of Geodesy and
Oberpfaffenhofen, Germany. Since 2007, he has Geoinformation, TU Wien. He is the Co-Founder
been a Researcher, a Lecturer, and a Scientific and the Head of Science of the Earth Observation Data Center for Water
Project Coordinator with the Research Group Remote Sensing, Department Resources Monitoring, Vienna, Austria. His research interests include geo-
of Geodesy and Geoinformation, Vienna University of Technology (TU physical parameter retrieval techniques from remote sensing data such as
Wien), Vienna, Austria, where he is currently a Professor in climate and scatterometry, SAR, full-waveform airborne laser scanning, and application
environmental remote sensing. He has authored or co-authored more than development.
60 publications in refereed journal papers and conference proceedings. Using Dr. Wagner is a member of the Science Advisory Groups for
earth observation data, he tries to understand and quantify the dynamics and Sentinel-1 (ESA), METOP ASCAT, METOP-SG SCA (EUMETSAT and
interactions of vegetation and the water cycle in a changing climate. His ESA), and GCOS/GTOS/WCRP Terrestrial Observation Panel for Climate.
research interests include remote sensing of soil moisture and vegetation, with From 2008 to 2012, he was an ISPRS Commission VII President. From
a focus on geophysical parameter retrieval through physical and semiempirical 2009 to 2011, he was the Editor-in-Chief of the Open Access Journal Remote
methods, calibration and validation, process automation, up and downscaling, Sensing. During his master’s and Ph.D. studies, he received fellowships to
image classification, data merging, and time series analysis. carry out research at the University of Bern, Atmospheric Environment Service
Dr. Dorigo was a recipient of the 2015 Science Award of TU Wien for the Canada, the NASA Goddard Space Flight Center, the European Space Agency,
water cycle in a changing climate. and the Joint Research Center of the European Commission.

Authorized licensed use limited to: Harvard Library. Downloaded on July 18,2022 at 22:06:47 UTC from IEEE Xplore. Restrictions apply.

You might also like