Abstract:
Soil moisture (SM) is one of the key parameters in the engineering, agronomic,
geological, ecological, biological and hydrological functions of soil. Its needed to
support decision making in agriculture for irrigation and trafficability, and is a key
input to hydrological and meteorological models. Remote sensing has significant
potential but there remains a challenge to improve the accuracy and usefulness
of the data for end users. The objective of this thesis is to develop a field-scale
soil moisture estimate from Sentinel-1 C-band synthetic aperture radar (SAR)
data, that is sufficiently accurate to be useful to farmers and agronomists and the
research community.
SAR is attractive as an all-weather remote sensing solution with potential to
estimate soil moisture over large scales and at high spatial resolution. But because
SAR backscatter is strongly a ected by overlying vegetation and crops, and to
some extent by soil surface roughness and soil texture, the process of SM retrieval
from SAR is very complex. Among a number of potential solutions, one approach
is to quantify the non-SM contributions by processing additional data alongside
SAR data using models that are trained against in-situ SM observations. This is
very resource intensive, with the results being limited in scope and accuracy by
the ancillary and training data. The alternative change detection (CD) algorithm
avoids the use of additional data or training sites by assuming, instead, that the
e ects of soil surface roughness and vegetation are relatively static on the timescale
of soil moisture variation. This can be highly e ective for estimating relative SM
in many areas but not, crucially, in areas of farmland under arable cultivation
where rapid changes in vegetation and soil surface roughness are common. This is
mitigated in current published SM products by restricting their spatial resolution
to around 1 km, averaging out the e ects of anthropogenic change over a diversity
of land uses. A 1 km pixel area is an order of magnitude larger than the average
field size in the UK, for example. Such products do not satisfy the needs of farmers
in arable areas and do not exploit the high spatial resolution available from SAR
sensors such as Sentinel-1 C-band (20 m). A key objective of this thesis is to achieve
a better understanding of the capabilities of the change detection algorithm and
enhance it to provide more accurate results for farming areas at field scale.
Whilst achieving field-scale resolution and better accuracy from the CD approach
would be significant, the results would still not very useful to farmers. The
CD algorithm is not trained against ground measurements, so its output is a Soil
Moisture Index (SMI) which is a relative estimate of SM within the boundaries
of previous SAR measurements at each pixel. To obtain absolute values such as
volumetric water content or soil moisture deficit (e.g. for irrigation management),
the SMI is calibrated against the expected range of SM, often taken to be the di erence
between the soil’s field capacity (FC) and permanent wilting point (PWP)
predicted by the soil texture at each location. This study shows that a 10 to 20 vol.%
mean absolute error may be introduced by uncertainties in selecting appropriate
soil texture so the importance of using a reliable soil map is underlined. It is
further shown that that the highly dynamic nature of SM in the shallow surface
layer penetrated by C-band SAR (1 to 2 cm) means that the Van Genuchten model
parameters,θS and θR should be used as wet and dry references to define the
expected range of SM. The common practice of using FC and PWP is shown to
contribute an additional 2 to 10 vol.% error. Taken together, these errors are large
compared to user requirements of 4 to 5 vol.% and a typical seasonal range of 20
to 40 vol.% depending on soil texture.
Validation is important for users to have confidence in the remote sensing of
soil moisture from SAR. A further objective of this study was to address the issue
of depth mismatch between the penetration of the SAR and the ground observations
used for validation. The study concludes that C-band SAR SM estimates
should not be validated directly against ground sensors at 5 to 10 cm depth. A
novel validation method is proposed for validation of SAR SM estimates against
simulated soil moisture profiles at 2 cm depth using a soil hydraulic model fitted to
ground observations. In this thesis, the latter were obtained from the COSMOS-UK
network of soil monitoring stations using Cosmic Ray Neutron Sensors with an
approximately 200 m radius measurement zone and average measurement depth
calculated to be around 10 cm. As a case study of using this improved method,
the performance of the published Copernicus SSM soil moisture product across
13 COSMOS-UK test sites is shown to be in the range 8 to 20 vol.% mean error. It
confirms, as expected, that the worst performance is in areas of arable agriculture,
justifying the focus of the thesis in such areas.
As a stage to achieve the key objective, an SM estimate, spatially aggregated
to field boundaries, is demonstrated by the first known implementation of the
CD algorithm in Google Earth Engine. It was found that the increased speckle
noise at this scale is typically balanced by reductions in noise from excluded
clutter sources and increased sensitivity to SM. Periodic noise due to satellite
orbit geometries (ascending versus descending) remains evident but temporal
smoothing was shown to be e ective against it. The performance, in terms of
mean error, at field-scale varies from 8 vol.% in grass pasture to 20 vol.% in some
arable fields during periods of rapid crop growth. To improve the accuracy in
arable fields, a process has been developed to use multispectral remote sensing
data to assign levels of confidence in the performance of the CD algorithm, to each
field on every measurement date. An alternative method is proposed to achieve a
more reliable soil moisture estimate by using two-dimensional interpolation using
inverse distance and confidence weighting (IDCW) across a range of neighbouring
fields within a local zone. By this method it is shown that, during the peak growing
season, the mean absolute error in the soil moisture estimate for wheat fields is
reduced from 20 vol.% to less than 5 vol.%, retaining field-scale resolution. This is
the first time such levels of accuracy have been reported at field scale from Sentinel-
1 SAR without the training of a model. The method is applicable anywhere where
the remote sensing data is available alongside a suitable soil texture map. The
output meets the operational requirements of farmers for 5 vol.% SM accuracy
and is very close to research requirements of 4 vol.%. It may be used to create a
mapping product for use by farmers and agronomists using to inform decision
making in near real-time. For scientists, engineers, infrastructure engineers and
environmentalists the data will be valuable for research into flooding, soil erosion,
ground movement and landslip.