IEEE Geoscience and Remote Sensing Magazine - June 2022 - 20

0.06 cm3/cm3) and New South Wales (RMSE 0.06 cm3/
cm3). Florida was found to be a quite problematic area for
SM retrieval, even for the SMAP mission, because of the
large number of lakes and swamps (bodies of water).
The retrievals from CyGNSS were also compared with
L2 SMOS data for Florida, Texas, and New South Wales,
also gridded daily over the EASE-Grid, at a resolution of
36 km × 36 km, to facilitate the comparison. The results
(not shown here, more details can be found in [79]) were
slightly worse than those for the tests using SMAP, but this
was expected because the regression was trained on SMAP
data. Thus, overall, the SMOS comparison showed good
consistency with the previous one.
The second approach for retrieving the SM from CyGNSS
consisted in estimating the regression fit between the
CyGNSS reflectivity and the SMAP SM for each pixel of
the considered grid. In this case, a simple linear regression
using only the CyGNSS reflectivity was performed, as the
vegetation opacity and roughness coefficient gathered from
the SMAP products do not exhibit a temporal dynamic in
a single pixel but are almost " static " parameters. This is
quite understandable for roughness, but the same situation
was also observed for the VOD as, in the considered SMAP
product, VOD was auxiliary data, not retrieved from the radiometer
observations. This does not happen on a global
scale as the spatial variability of those parameters is meaningful.
In the pixel-by-pixel approach, the difference in the
surface conditions, other than SM, are implicitly accounted
for by the different regression coefficients. Here, the pixel
is chosen to be 1° × 1° of latitude/longitude since this was
a good compromise between the accuracy of the retrievals
and the need to have a large enough number of samples
in each pixel to form the two separate subsets of training
data and test data, as for the RVR algorithm. The coefficient
of the linear regression for each pixel was derived from a
training set, e.g., a random half of the data falling into that
Density Plot
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
CyGNSS SM (cm3/cm3)
FIGURE 8. The density scatterplot in log scale of true (SMAP) and
retrieved (CyGNSS) SM using the RVR algorithm on a pixel base.
(Source: [78].)
20
4.5
4
3.5
3
2.5
2
1.5
1
0.5
pixel, and the algorithm performance was estimated using
the test data, e.g., the other half. In principle, this algorithm
can be applied considering even smaller pixels, if more data
are available. This methodology is mathematically equivalent
to the approach presented in [35]. A difference between
the two approaches is the use in our case of the reflectivity
as an observed quantity instead of the SNR.
The test of the retrieval on a global scale assuming the
SMAP SM as ground truth is presented in Figure 8. The
density plot highlights a good agreement between CyGNSS
and SMAP, and the agreement is slightly improved
compared to the RVR results, exhibiting on a global scale
an RMSE of 0.06 cm3/cm3, which is close to the typical
RMS differences obtained by other satellites. On a local
scale, the performance in Florida was the worst (RMSE
0.09 cm3/cm3); it was slightly better in Texas (RMSE
0.05 cm3/cm3). The comparison with SMOS data confirmed
the previous considerations.
Finally, the results from CyGNSS were compared with
the SM measured by in situ SCAN stations over Florida
and Texas and in the Merguellil catchment in central
Tunisia. The estimates provided from SMAP were also
considered for comparison. Both retrievals were mean
daily quantities that were extracted in the pixels closest
to the in situ probe location. Unfortunately, this
analysis did not provide reliable conclusions. In some
cases, the number of matchups was quite small, or the
dynamic range of the observed SM was very small; compared
to SMAP, CyGNSS exhibited retrievals that were
sometimes better and sometimes worse. In general, the
performances were not good, and SMAP outperformed
CyGNSS for higher SM values.
SENSITIVITY TO VEGETATION BIOMASS
The sensitivity of TDS-1 to the forest parameters related
to biomass was evaluated on five test areas chosen as representative
of the most important forest types on Earth:
Brazil (Manaus), Uruguay (Algorta), Alaska (Fairbanks),
Finland, and Argentina (Asuncion). Advanced Land Observing
Satellite (ALOS) radar images have also been considered
for comparison, and the sites were chosen with an extension
of approximately 350 km × 350 km, corresponding to
the size of ALOS ScanSAR images, except that the Uruguay
one had an extension of 70 km × 70 km. In addition to
ALOS backscatter images, additional data providing forest
characteristics were selected; specifically, the VOD available
from the SMAP L3 products, the improved pantropical
biomass map providing the AGB as illustrated in [62],
and the tree height derived from the ICE-GLAS lidar acquisitions
at 1-km resolution on a global scale as described in
[61]. It is pointed out here that AGB and tree height were
" static " data, related to a specific epoch, whereas the VOD
has some temporal dynamics.
The TDS-1 and CyGNSS sensitivities to AGB as well as
ALOS backscatter were first evaluated by comparing the
reflectivity to the improved pantropical AGB map. The
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2022
SMAP SM (cm3/cm3)
IEEE Geoscience and Remote Sensing Magazine - June 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2022
