IEEE Geoscience and Remote Sensing Magazine - June 2022 - 17

results show a good sensitivity to the SM and VOD (the latter
being negative, as expected). The sensitivity to the SM is
also quite independent of the incidence angle as the multivariate
regression probably better compensates empirically
for the VOD effect with respect to the physical model in
(2). The sensitivity to the VOD is higher at larger angles,
confirming previous outcomes.
SM RETRIEVAL APPROACHES
SINGLE-PASS RETRIEVAL
The feasibility to perform single-pass SM retrievals using
GNSS-R observations and a physically based approach was
investigated in [64] and [73]. The aim was to exploit data
from a single receiver overpass, in contrast with the next
sections, which discuss the retrieval of space and temporal
means carried out based on empirical approaches. A
single-pass approach does not need to resample and colocate
observations over time; thus, it enables better temporal
and spatial resolution. Another important advantage
is the possibility of considering accurate forward models,
describing the nonlinear relationship between reflectivity
and SM, which exhibits a saturation effect for wet soils (see
the previous discussion about the dependence of sensitivity
on the SM dynamic range). The drawbacks are the need of
an accurate absolute calibration and the collection of ancillary
data to correct roughness and vegetation effects on
each data sample.
The proposed physical SM retrieval algorithm exploits
the model in (2) and inverts it to retrieve the Fresnel reflectivity
using independent ancillary information on the VOD
and roughness. The Fresnel reflection formula is used to derive
the soil permittivity, and finally a permittivity model is
exploited to retrieve the SM (among other variables, such as
the clay fraction and physical temperature) through a minimization
process.
The approach was first attempted on data acquired during
an airborne experiment that took place on 22 October
2018 in an agricultural region around Balaguer, Catalonia,
Spain. The plane was a Cessna Caravan flying at 1,200 m
of height. The AISA instrument (a pushbroom hyperspectral
sensor collecting 126 bands in the visible and nearinfrared
spectrum) flew together with the ARIEL L-band
radiometer and the CORTO GNSS reflectometer at L1/L2
GPS bands [18].
The VOD used to compensate for the vegetation effects
was derived following the SMAP approach [74] applied to
the high-resolution NDVI available from AISA. A field campaign
was carried out to collect data on surface roughness
by a laser profiler and in situ SM measurements performed
the same day as the airborne campaign. Ground samples
were collected from five different areas corresponding to
different agricultural fields. Surface roughness measurements
were revealed to be quite critical, as usual, and the
sampling was not dense enough. The ARIEL L-band radiometer
[75] was used to provide a complete SM map with a
JUNE 2022 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
final resolution of 0.8 m. The retrieved SM maps were then
validated using the in situ SM samples.
Because of the limited estimation of the soil surface
roughness from in situ data, the results show a disparity
between the SM data retrieved by the ARIEL radiometer
and the CORTO GNSS-R instrument. For a single track of
reflections over a relatively flat area, the retrieved SM values
are highly variable with the incidence angle and are quite
different with respect to the SM derived from the L-band
radiometer. These results testify that the surface roughness
correction was not appropriate in either the value of the
HR parameter and/or the power of the cosine function appearing
in (2). Then, the surface roughness correction term
was tuned to match as closely as possible the CORTO SM
retrieval and the SM from the ARIEL data. The best value
for the power of the cosine function was found to be n 0=
(n 2= is the theoretical value).
The ad hoc roughness parameter revealed itself to be
quite spread, notably more for the highest values. There is no
evident correlation with the different fields or with the SM
values to explain the variability of the ad hoc roughness. Of
course, the anomalous values of the ad hoc roughness can
be related to a wrong compensation of the vegetation effect
based on the NDVI and/or to the estimation of the noise in
the DDM needed to compute the reflectivity. This led to the
conclusion that soil surface roughness is a key parameter
to correctly estimate the SM, but in situ measurements of
surface roughness are not suitable to predict the value to
be used; the use of a constant value for a large region is not
suitable either. The approach of forcing a tunable roughness
correction per pixel is not feasible in an operational system.
An alternative approach can consist in deriving empirically
the best roughness parameter as a function of the incidence
angle from a suitable training set. The " ad hoc " roughness
derived in this way provided reasonable results and could
be further exploited in the future.
The approach tested on the Balaguer data has also
been applied to TDS-1 reflectivity data collected over the
REMEDHUS experimental site, where in situ SM measurements
and other data are routinely collected. Similar conclusions
can be outlined through the Balaguer experiment.
An ad hoc roughness parameter is needed to force a correct
SM retrieval. It can be adjusted locally or regionally, but this
TABLE 1. REGRESSION COEFFICIENTS AS A FUNCTION OF
INCIDENCE ANGLE RANGES COMPUTED USING THE DATA
SET WHEN THE RR FLAG FOR SMAP DATA IS APPLIED.
REGRESSION
COEFFICIENTS
FOR CYGNSS
All i
0° <= i < 30°
30° <= i < 60°
i >= 60°
SM
26.00
27.09
25.51
25.66
VEGETATION
OPACITY
-16.02
-13.30
-16.50
-21.77
ROUGHNESS
-15.22
-16.41
-13.50
-13.45
17
IEEE Geoscience and Remote Sensing Magazine - June 2022

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