In recent years, the amount of water used for agricultural purposes has been rising due to an increase in food demand. However, anthropogenic water usage, such as for irrigation, is still not or poorly parameterized in regional- and larger-scale land surface models (LSMs). By contrast, satellite observations are directly affected by, and hence potentially able to detect, irrigation as they sense the entire integrated soil–vegetation system. By integrating satellite observations and fine-scale modelling it could thus be possible to improve estimation of irrigation amounts at the desired spatial–temporal scale.
In this study we tested the potential information offered by Sentinel-1 backscatter observations to improve irrigation estimates, in the framework of a data assimilation (DA) system composed of the Noah-MP LSM, equipped with a sprinkler irrigation scheme, and a backscatter operator represented by a water cloud model (WCM), as part of the NASA Land Information System (LIS). The calibrated WCM was used as an observation operator in the DA system to map model surface soil moisture and leaf area index (LAI) into backscatter predictions and, conversely, map observation-minus-forecast backscatter residuals back to updates in soil moisture and LAI through an ensemble Kalman filter (EnKF).
The benefits of Sentinel-1 backscatter observations in two different polarizations (VV and VH) were tested in two separate DA experiments, performed over two irrigated sites, the first one located in the Po Valley (Italy) and the second one located in northern Germany. The results confirm that VV backscatter has a stronger link with soil moisture than VH backscatter, whereas VH backscatter observations introduce larger updates in the vegetation state variables. The backscatter DA introduced both improvements and degradations in soil moisture, evapotranspiration and irrigation estimates. The spatial and temporal scale had a large impact on the analysis, with more contradicting results obtained for the evaluation at the fine agriculture scale (i.e. field scale). Above all, this study sheds light on the limitations resulting from a poorly parameterized sprinkler irrigation scheme, which prevents improvements in the irrigation simulation due to DA and points to future developments needed to improve the system.
Irrigation has been applied by humans for as long as they have been cultivating plants. However, in the last century irrigation has become one of the most impactful human activities on the terrestrial water cycle, accounting for nearly 85 % of global water consumption (Gleick, 2003; Kustu et al., 2010) and for 40 % of the world's food production (Foley et al., 2011; Massari et al., 2021; Siebert and Döll, 2010).
In the next decades, the foreseen increase in population (FAO, 2009) and climate change is expected to ask for larger amounts of water resources for food production and consequently increases in irrigation applications
(Matthews and Germain, 2007; Ozdogan et al., 2010b).
Busschaert et al. (2022) found that under a high CO
Therefore, many studies have focused on detecting and quantifying irrigation through modelling and remote-sensing (RS) observations. From the modelling perspective, Girotto et al. (2017) highlighted the key role in representing anthropogenic processes into land surface models (LSMs). The authors realized a study in India, where irrigation provides a large contribution to winter crop production, and they found that data assimilation (DA) of total water storage (TWS) RS observations into the catchment land surface model (CLSM) introduces a negative trend in groundwater due to pumping for irrigation, along with an associated erroneous negative trend in modelled evapotranspiration when irrigation is unmodelled. Other studies have attempted to incorporate irrigation schemes into global LSMs, including the Interaction between Soil, Biosphere, and Atmosphere (ISBA) LSM (Druel et al., 2021), the Community Land Model (CLM; Pokhrel et al., 2012), and the Organizing Carbon and Hydrology in Dynamic Ecosystems (ORCHIDEE) model (De Rosnay et al., 2003), demonstrating the regional impact of irrigation on different water storages and on energy partitioning between sensible and latent heat fluxes (Lawston et al., 2015). Different types of irrigation systems were also implemented into the NASA Land Information System (LIS; Kumar et al., 2006; Peters-Lidard et al., 2007) framework and coupled with several LSMs, such as the Noah LSM (Chen et al., 1996; Chen and Dudhia, 2001; Mahrt and Pan, 1984) and the Noah-MP LSM (Niu et al., 2011). Drip and flood irrigation systems (Evans and Zaitchik, 2008) as well as a sprinkler scheme (Ozdogan et al., 2010a) are currently part of the LIS framework. The sprinkler irrigation scheme, in particular, adds irrigation water as fictitious rainfall based on a root-zone soil moisture threshold, and it was recently improved accounting for groundwater extraction in Nie et al. (2019).
Beyond these efforts, the human influence on the water distribution is still poorly described in LSMs due to simplifying assumptions, such as limitations related to soil texture maps, static crop maps or irrigation intensity inputs (Modanesi et al., 2021; Monfreda et al., 2008; Ozdogan et al., 2010a; Salmon et al., 2015). Another key issue is related to the farmer's irrigation application decision, which is not necessarily related to crop irrigation requirements or based on the root-zone soil moisture availability but mainly on water resource availability (Massari et al., 2021).
On the other hand, RS observations can indirectly disclose the presence of irrigation activities when they sense the entire integrated soil–vegetation system. For instance, visible and near-infrared measurements were mainly used in previous studies for developing irrigation mapping techniques (Ambika et al., 2016; Ozdogan and Gutman, 2008; Peña-Arancibia et al., 2014; Salmon et al., 2015), and, in recent years, optical data have also been combined with microwave (MW) observations (Ferrant et al., 2019) or with thermal sensor data (i.e. land surface temperature data) via energy and water balance models (van Eekelen et al., 2015; Olivera-Guerra et al., 2020; Brombacher et al., 2022), to investigate the advantages of multi-sensor approaches. On the other hand, MW satellite data were exploited in the last decade for both detecting (Dari et al., 2021; Gao et al., 2018; Kumar et al., 2015) and quantifying irrigation (Brocca et al., 2018; Dari et al., 2020; Jalilvand et al., 2019; Zaussinger et al., 2019). All these studies highlighted the importance of high spatial resolution of RS observations for a better estimation of irrigation quantities. In this context, the Copernicus Sentinel-1 mission represents the new era of satellite observations, providing Synthetic Aperture Radar (SAR) backscatter data at a fine spatial resolution (up to 10–20 m) since 2014, which are freely accessible. A recent study by Jalilvand et al. (2021) has highlighted the potential of high-resolution Sentinel-1-based soil moisture data, such as the 1 km SMAP/Sentinel-1 product (Das et al., 2019) to detect the irrigation signal over agricultural areas. However, despite the advances in RS, satellite observations are still characterized by limitations, such as the low revisit time typically associated with higher spatial resolution data, noise in the measurements, and the uncertainties of satellite retrieval algorithms.
The optimal integration of fine-scale modelling and satellite observations using DA in LSMs could be a promising solution to account for anthropogenic activities alongside improving the estimation of irrigation amounts and model predictions. Lawston et al. (2017) and Jalilvand et al. (2021) suggested the use of MW-based surface soil moisture retrievals from SMAP or SMAP/Sentinel-1 respectively, to incorporate the irrigation signal into models via DA. In light of this, Abolafia-Rosenzweig et al. (2019) designed an innovative system to assimilate RS-based soil moisture into the VIC (variable infiltration capacity) model through a particle batch smoother in order to improve irrigation estimates. Further studies investigated the use of surface soil moisture retrievals and vegetation indices such as leaf area index (LAI) or vegetation optical depth to improve model predictions (Albergel et al., 2018; De Lannoy and Reichle, 2016b; Kumar et al., 2020). However, MW-based retrievals could also add unreliable information into LSM systems due to the RS observation preprocessing and the retrieval algorithm. More specifically, passive MW retrievals are produced with ancillary data that might be inconsistent with those used in the LSM (De Lannoy and Reichle, 2016a), and active MW retrieval products based on change detection methods often rely on a climatological approximation of the vegetation signal (Wagner et al., 1999). To avoid this limitation, previous studies investigated the direct assimilation of MW observations, such as brightness temperature (Tb) derived from Soil Moisture and Ocean Salinity (SMOS) or Soil Moisture Active Passive (SMAP) missions (De Lannoy and Reichle, 2016a, b; Reichle et al., 2019), radar backscatter from the Advanced Scatterometer (ASCAT; Lievens et al., 2017a), or the joint assimilation of Sentinel-1 backscatter and SMAP Tb (Lievens et al., 2017b), through the use of calibrated observation operators. The assimilation of Level-1 observations has the potential to limit inconsistencies in the DA system and to address the climatological bias correction through the observation operator calibration, as compared to classical soil moisture bias correction techniques (i.e. cumulative distribution function (CDF) matching).
The earlier studies on the direct assimilation of MW observations did not
include any irrigation or dynamic vegetation modelling. Consequently, they
did not investigate the benefits of jointly updating soil moisture and
vegetation or any consequences on irrigation estimation. This work aims at
filling those gaps using the Noah-MP v.3.6 (Niu et al.,
2011) equipped with a dynamic vegetation model, a sprinkler irrigation
scheme (Ozdogan et al., 2010a) and a calibrated backscatter
forward operator, the Water Cloud Model (WCM; Attema
and Ulaby, 1978), within the NASA LIS framework. The main target of this
study is then to assimilate 1 km Sentinel-1 radar backscatter observations
for a joint update of modelled soil moisture and vegetation in order to
correct them for actual irrigation applications and thereby improve the
initial land surface conditions for subsequent irrigation forecasts. The
backscatter from Sentinel-1 contains information on both soil moisture
(Bauer-Marschallinger et al., 2018; Liu and Shi, 2016;
Zribi et al., 2011) and vegetation (Vreugdenhil et al., 2020,
2018), and we hypothesize that assimilating these data could correct for
misrepresented or missed irrigation events. Furthermore, we assume that when
coupling the Noah-MP DA system with a poorly parameterized sprinkler
irrigation module, irrigation water amounts can be optimally forecasted when
optimal soil moisture and vegetation estimates are available. In this
framework, the WCM calibration is a necessary step to obtain unbiased
predictions over intensively irrigated areas, and this topic was deeply
investigated in Modanesi et al. (2021). The ensemble Kalman filter (EnKF; Evensen, 1994) algorithm is
selected to perform the DA analysis. The EnKF was used in previous studies
for non-linear dynamics, and it is popular in hydrological and land surface
modelling studies (Reichle et al., 2002; De Lannoy and Reichle, 2016a; Kumar
et al., 2019, 2020; De Santis et al., 2021, to cite a few). It uses an
ensemble of model trajectories to represent the background error covariance
at each time of an update. With the support of a near-optimal DA system,
this study aims at improving irrigation estimates, and the following points
are discussed:
the improvement (or deterioration) of LSMs simulations in terms of irrigation, soil moisture, vegetation and evapotranspiration due to the
sequential assimilation of Sentinel-1 backscatter in vertical transmit and
receive (VV) or in vertical transmit and horizontal receive (VH)
polarizations and the differences in results obtained using co- and
cross-polarization observations to update both soil moisture and vegetation; the limitations due to the spatial scale and shortcomings of the system in
terms of model parameterization and DA consistency.
The analysis was carried out over two pilot sites: (i) the Po Valley, one of
the most intensively cultivated and irrigated areas in Italy (Po River
Watershed Authority, 2016), and (ii) an irrigated area located in northern
Germany, in more humid conditions.
The paper is organized as follows. Section 2 describes the study areas, the datasets used (i.e. in situ benchmark data and RS observations), methods (including the LSM description and the DA system) and the experimental setup. Section 3 presents the results obtained from the DA experiments: first the quality of the DA system design is evaluated in terms of DA diagnostics, and then the estimates of irrigation and model state variables and fluxes are evaluated. In Sect. 4 we provide a discussion, and conclusions are reported in Sect. 5.
Two European pilot sites characterized by different climatic conditions were
selected (Fig. 1). The first one is located in the Lower Saxony region in
northern Germany and has an extent of 160 km
Study area composed of the German and Italian pilot sites
(represented by the red boxes). The pilot sites include three test sites:
The second pilot site is located within the Po Valley, one of the most
important agricultural areas in Italy, intensively equipped for irrigation
(Salmon et al., 2015). According to the Köppen–Geiger
climate classification, this area can be ascribed to the Cfa class
(temperate climate, without dry season and with hot summers). The simulation
area (red box on Fig. 1b–c) is mainly characterized by croplands, except
the south, south-western area where forests and more complex topography are
dominant. This pilot site has an extension of 1800 km
The European Space Agency (ESA) and Copernicus Sentinel-1 mission collects
active microwave backscatter data at C-band (5.4 GHz) at a high spatial
(
Additional RS observations were used for the evaluation of Noah-MP LSM
simulations.
The Metop ASCAT surface soil moisture (SSM) Climate Data Records H115 and
its extension H116 are provided by the European Organization for the
Exploitation of Meteorological Satellites (EUMETSAT) Support to Operational
Hydrology and Water Management (H SAF; The PROBA-V leaf area index (LAI) is provided by the Copernicus Global Land Service programme (CGLS, The Moderate Resolution Imaging Spectroradiometer (MODIS) is a multispectral sensor on board TERRA and AQUA satellites acquiring image data of the Earth's surface simultaneously in visible and infrared bands. For this work, the
MOD16A2 Version 6 Evapotranspiration product was used for evaluation. This
is an 8 d composite product characterized by 500 m spatial resolution.
ASCAT SSM, PROBA-V LAI and MODIS ET were extracted over the test sites and
re-gridded over the LIS grid spatial domain (0.01
The LSM selected for this study is the Noah-MP.v.3.6 (Niu et al., 2011). This model is able to dynamically simulate vegetation and soil moisture in four layers, i.e. 0–10, 10–40, 40–100 and 100–200 cm depth. For this study, the Noah-MP model was coupled to a sprinkler irrigation module (Ozdogan et al., 2010a) embedded within the NASA's LIS version 7.3 (Kumar et al., 2008). For a more detailed description of the Noah-MP parameterization used in this study, the reader can refer to Modanesi et al. (2021).
The irrigation module adds water as pseudo-precipitation to mimic sprinkler
systems (Ozdogan et al., 2010a) and does not further change
processes related to, for example, vapour fluxes (which would be needed for highly
efficient drip irrigation systems; Evans and Zaitchik, 2008). In
order to identify the irrigation season, timing and location of irrigation,
four conditions need to be met: (i) irrigable land cover (i.e. croplands),
(ii) irrigated land fraction, (iii) growing season and (iv) dry enough
root-zone soil moisture. The first two conditions are tested against static
land cover (LC) and irrigation intensity (areal fraction) maps. The growing
season was defined based on a user-defined threshold of simulated LAI
(LAI
Parameters needed to activate the irrigation scheme in Noah-MP, based on the two pilot sites.
The modelled irrigation estimates are thus primarily controlled by five
datasets: static LC, irrigation fraction, soil texture, crop type and
dynamic meteorological forcing. In this study, the static 1 km LC is derived
from the CGLS 100 m global LC map for the year 2015 (Buchhorn et
al., 2020; available at
Soil texture and the associated parameters were extracted from the 1 km Harmonized Soil World Database (HWSD v1.21) and mapped to discrete soil classes with their associated soil hydraulic parameters as in Modanesi et al. (2021). Given the lack of European or large-scale dynamic crop map datasets, the crop type was set to a generic type, with a maximum rooting depth of 1 m in Italy (Modanesi et al., 2021) and 0.8 m Germany. In particular, in the Lower Saxony test site, an averaged rooting depth was calculated based on the main crop types cultivated on the irrigated fields.
The dynamic meteorological forcing data were extracted from Modern-Era
Retrospective analysis for Research and Applications, version 2 (MERRA-2;
Gelaro et al., 2017), which is not corrected for surface or
screen-level observations (and are thus unlikely to contain feedback from
– unmodelled and unobserved – irrigation). The meteorological data, at
original spatial resolution of 0.5
The observation operator used to ingest Sentinel-1
The DA system was developed in order to directly assimilate Sentinel-1
The EnKF assumes unbiased observations and forecasts. This is achieved by
running the Noah-MP with a poor guess of irrigation activated and using
calibrated WCM parameters to produce unbiased
Perturbation parameters for forcing (i.e. rainfall, incident and shortwave radiation) and state variables (i.e. SSM and LAI).
The main equation of the EnKF can then be written as follows:
The above method will update forecasted irrigation estimates by correcting random errors in land surface state forecasts. An alternative method would be to run the Noah-MP without irrigation activated and to derive irrigation estimates from the amount of water or vegetation added to the system via DA (i.e. the increments), if the true precipitation is used as input and if the observation operator is also not compensating for irrigation. More specifically, positive and autocorrelated increments in the growing season would be indicative of irrigation. However, this would also be indicative of a suboptimal assimilation system. In this study, we use imperfect reanalysis precipitation input, and we simulate irrigation through an irrigation scheme. Increments of any sign could thus be related to over- or underestimation of irrigation.
In this study we considered two different experimental lines, (i) the
assimilation of Sentinel-1
Workflow of the DA experiments realized within the LIS framework.
The scheme describes how the Sentinel-1
For each pilot site, the Noah-MP model was previously spun up from January
1982 to May 2014. Then, an ensemble spin-up was realized in open loop (OL)
mode using 24 ensemble members from May 2014 to January 2015, in order to
obtain optimal initial conditions. The OL run was continued thereafter from
January 2015 onwards through December 2020. Similarly, the Sentinel-1
The evaluation aimed at (i) verifying the goodness of the DA system in terms
of DA diagnostics; (ii) highlighting benefits of the Sentinel-1 DA for
irrigation, soil moisture and LAI estimation, as well as testing the
differences between the Sentinel-1
To achieve those targets, two types of evaluation were carried out.
The optimality of the DA system design was evaluated regionally for each pilot site, for the period January 2015–December 2020. Following
Reichle et al. (2017), three different filter diagnostics
were analysed. First, the difference between the temporal mean ensemble
standard deviation (or ensemble spread) of the DA and OL runs was computed
to test whether the DA system successfully reduces the uncertainties (i.e.
ensemble spread) as compared to the OL run. Second, the time series standard
deviation of the normalized The OL and DA estimates of irrigation, SSM, LAI and ET were evaluated using independent reference data for the years 2015–2017 and 2016–2017 at the Budrio and Faenza (Italian) sites and for the year 2018 at the German sites. The evaluation of ensemble mean irrigation simulations was assessed in terms
of correlation and percentage bias (Pearson
The two DA experiments (Sentinel-1
Maps of the difference in time–mean ensemble spread between the DA
and OL runs for
Maps of the standard deviation (SD) of the normalized
Figure 4 shows the standard deviation of the normalized
Same as Fig. 3 but for the standard deviation (SD) of the
increments in
Finally, we analysed the increments of SSM and LAI over the study area. In a
well-calibrated DA system, the long-term mean of the increments is expected
to be close to zero at each pixel. As explained in Reichle et
al. (2019), values close to zero indicate that no long-term net addition or
subtraction of water (or vegetation) is generated by the analysis. As
expected, the temporal mean values of SSM and LAI analysis increments vanish
in the regional average (results not shown). However, the standard deviation
of the analysis increments provides valuable information, which has been
summarized in Fig. 5. Maps of SSM (Fig. 5a–d) and LAI (Fig. 5e–h) show
generally small increment standard deviations for both DA experiments and
both study areas. Note that zero standard deviations are found where no data
were assimilated (see Fig. S3). For both areas, complementary patterns are
observed for SSM and LAI increments. Larger SSM variances can be observed
over the cropland area characterized by sandy soil (e.g. north-west in
Italy) and where a higher number of observations are assimilated (especially
in Italy), whereas larger LAI updates are related to the silty-loam soil
type (e.g. ellipsoid-shaped area in Italy), corresponding to the patterns
in
Although the DA diagnostics show satisfactory results in terms of
consistency of the system, input parameters, such as soil texture and
vegetation (i.e. lack of dynamic crop maps in the model), seem to have a
strong impact on the performance of the DA system. In this context, the red
squares in Figs. 4a–b and 5a–b show areas where the test sites are
located. In Italy (Figs. 4a or 5a) the Budrio test site (north-west
square) is placed over an area where results are close to optimal in terms
of standard deviation of normalized
In this section we show the results of the DA and OL runs in terms of irrigation, SSM, LAI and ET. We first discuss the Italian pilot site, where the longest record (2–3 year) of benchmark data is available for three test sites with different spatial extents. This allows us to focus on the impact of the spatial scale on the performance of the DA system. Next, we discuss the pilot site in Germany, where the availability of 1 year of data for 49 small irrigated fields (24 LIS pixels, of which only 8 were retained considering a percentage of irrigated area larger than 25 %) allows for a statistical interpretation of the results.
Figure 6 shows an example time series of SSM, LAI, ET and irrigation at the
Budrio farm for the OL and Sentinel-1
Evaluation of the OL (blue lines) and DA (orange lines) results at
the Budrio test site for the Sentinel-1
The irrigation simulations (Fig. 6e) show a general deterioration of the
performance with DA, with a decrease in Pearson
Figure 6b shows the impact of DA in the SSM dynamics during both the summer
and winter seasons. The Sentinel-1
When comparing the ET results against MODIS ET, a slight improvement is
observed when DA is performed (Fig. 6d; Pearson
Additional time series analyses for the other Italian sites and for
Evaluation results at all the Italian test sites and for both the DA experiments.
Table 3 summarizes the results obtained for the three test sites within the
Italian pilot site. For the Faenza sites, the results are first spatially
aggregated over 3 or 8 pixels before computing time series metrics. The
uncertainty in the reference data and the relatively short data records
prevent a statistically significant evaluation, but overall, the DA runs
provide a slight improvement of SSM and LAI for both the DA experiments
compared to the OL run. As expected, Sentinel-1
Evaluation results in terms of irrigation simulations for the
three Italian test sites based on the extent of the fields in terms of LIS
pixels. Blue bars refer to the OL run, while orange bars refer to the DA run.
The Pearson
Figure 7 summarizes the Pearson
Even though DA deteriorates the irrigation results in terms of
In the Lower Saxony test site, the analysis was conducted over 8 LIS pixels
based on the preprocessing described at Sect. 2.7. The evaluation in
terms of SSM and LAI using ASCAT SSM and PROBA-V LAI (not shown) does not
display substantial differences in terms of Pearson
Evaluation in terms of
Unlike the Italian site, the Lower Saxony site suffers from an irrigation
underestimation by the model, in line with a previous study by
Zappa et al. (2021), which used Sentinel-1 SSM retrievals to detect
and quantify irrigation at the Lower Saxony test site. Figure 8b shows that
the median bias is reduced from a value of
In this study, we built a DA system for the assimilation of Sentinel-1
In the test site analysis, we focused on three different aspects: (i) the
added value of the DA experiments compared to the OL runs and the role of
the Sentinel-1
In Italy, we found an improvement due to DA in terms of SSM, LAI and ET,
compared against RS retrievals (i.e. ASCAT SSM, PROBA-V LAI and MODIS ET).
Sentinel-1
Following the rationale that weather forecasts would be improved if land surface conditions are better constrained, the hope was to also improve irrigation forecasts with better constrained land surface conditions. However, the latter is only true if the assumption holds that the irrigation model produces the best irrigation estimates for the best estimates of land surface state variables. The latter assumption strongly depends on a good characterization of soil, vegetation and irrigation parameters, which was found to be a limitation for the DA system. For instance, the OL run provided a large overestimation or underestimation of the irrigation quantities (depending on the study area) that can be attributed to limited parameterization of the irrigation scheme like, for instance, detailed information on irrigation fraction input, dynamic crop rotation and rooting depths, as well as a poor description of the crop phenology in Noah-MP. The DA experiments helped in reducing the irrigation overestimation (or underestimation) at some, but not all, sites. This means that if the sprinkler irrigation scheme is not well parameterized, the DA system is not able to strongly correct the OL runs. In the case of biased state variables or flux simulations, it is generally more interesting to study the effect of DA on anomalies from (multi-year) climatological conditions, but such an analysis could not be performed with the limited amount of available benchmark data.
The limited spatial coverage and scale of the benchmark data is another
reason of concern in the evaluation of the DA results. The Sentinel-1 DA
appears to degrade the temporal dynamics (Pearson
The evaluation highlighted many aspects that can be improved for a more
reliable irrigation estimation in a DA system which involves the Noah-MP
LSM, with an irrigation scheme and innovative Sentinel-1
In terms of irrigation simulations, we found that absolute irrigation amounts and timing of irrigation estimated by the irrigation scheme strongly depend on soil (e.g. texture), vegetation (e.g. crop type) and irrigation (e.g. “intensity” or area fraction) parameters. These parameters are now based on global datasets that might not be ideal for regional to local applications, not being dynamic and updated. Likewise, irrigation estimates depend on a correct representation of natural forcing input, here reanalysis data, which is unlikely to be accurate at the local scale (where the weather is itself influenced by irrigation).
Furthermore, the irrigation estimates obtained with inclusion of DA do not
always outperform the model-only estimates. The main reason is that the
irrigation model does not necessarily produce the best irrigation estimates
for the best estimates of land surface state variables at the test sites.
In addition, if the soil moisture is updated to wetter conditions to
include irrigation before the model would forecast it, then the irrigation
simulation will be skipped or delayed in the DA results. Thus, in line with
the suggestions by Lawston et al. (2017), besides optimizing the DA itself,
future research should also focus on improving the irrigation model to
optimally use the observational information contained in the Sentinel-1
Another important aspect is related to the temporal and spatial variability
of irrigation. We found that irrigation results become increasingly
uncertain (and depend more on the irrigation parameterization) at shorter,
e.g. daily, timescales. Soil moisture and vegetation increments can indeed
affect the irrigation dynamics at short-term periods (i.e. daily timescale), and the benchmark data are also not representative of the effective
irrigation needs in the short term because of water management policies. It
can be expected that the interannual variability in irrigation can be better
estimated, as also suggested by Lawston et al. (2017). Furthermore, we found
that the agreement between benchmark data and simulations of irrigation
quantities increases in an analysis at pixel level or at small-district
spatial scale (Italy), showing the limitation of the system in providing
information at plot scale, when simulations and RS observations are provided
at coarser spatial resolution (i.e. 0.01
Three last aspects need to be highlighted: (i) a robust evaluation analysis
of land surface variables is not straightforward considering that RS
observations at coarse spatial resolution (i.e. ASCAT) or constrained by
reanalysis data (i.e. MODIS ET) do not necessarily provide accurate
information on irrigation (Zaussinger et al., 2019); (ii) additionally, evaluating irrigation estimates is also more challenging due
to the scarce availability of information on the irrigation management
(Massari et al., 2021); and (iii) finally, the disagreement between
in situ reference data and irrigation estimates obtained from the model only
can be partly explained by the actual in situ irrigation system management,
which depends on water availability and policies unknown by the modelling
system. In this context, although irrigation data are compared at biweekly
timescale, and DA is overall expected to improve the simulation of
irrigation temporal dynamics, the temporal resolution of Sentinel-1
(
To test the goodness of the EnKF assumptions over the study areas, future research could benefit from an experiment using precipitation plus known irrigation as modified input forcing. However, high-quality gridded irrigation products are not yet available, and the difference between the spatial resolution of MERRA-2 forcings and irrigation input will complicate such an experiment. As a final note, future research should also focus on investigating different DA techniques. In particular, the DA analysis could benefit from the use of particle filtering, which has proven useful from a mass-balance perspective, also for irrigation applications (Abolafia-Rosenzweig et al., 2019).
Information on the actual irrigation quantities used for agricultural purposes is still missing, and a correct quantification of irrigation is a challenging topic. The joint use of models and RS observations (which contain irrigation information) can help to fill this gap while also providing irrigation estimates at high temporal scale and medium–high spatial resolution.
In this study we assimilated, in two different experiments, 1 km Sentinel-1
The main conclusions drawn from our evaluation highlight shortcomings of the
system and can be summarized as follows.
The developed DA system is consistent and close to optimality, but it could
benefit from enhanced model inputs, such as more reliable soil texture maps
or the introduction of dynamic and high-resolution crop maps, which could
improve both soil moisture and vegetation simulations (input for the WCM
calibration). Additional effort will also be needed in future research to
account for different Sentinel-1 orbits, both in the WCM calibration and in
the DA system, which will provide a gain of the signal-to-noise ratio, with a general benefit in time series analysis. The Sentinel-1 The Noah-MP LSM input and irrigation parameterization affect the OL and DA
estimates, providing strong over- or underestimations of irrigation,
depending on the study area. These limitations are mainly related to soil
texture uncertainties, lack of crop-type inputs and outdated irrigated
fractional area information, which affect the results of the reference OL
run. In this context, the DA can only correct the estimates of irrigation
amounts, if the irrigation simulation is not excessively biased, meaning
that future research should focus on improving the irrigation model.
Alternatively, even with biased irrigation simulations, the DA should be
able to correct for the interannual variability in irrigation estimates, but the record of available benchmark data on irrigation is insufficient at this time to confirm this hypothesis. When comparing irrigation simulations and benchmark data, the spatial and
temporal resolution play an important role. Irrigation estimates were here
evaluated at a biweekly scale, to limit the influence of short-term
analysis increments on the activation of the irrigation scheme and to
reduce mismatches with benchmark data due to human choices in the timing of
irrigation application. In any case, results in terms of temporal dynamics
and bias could also be affected by the temporal resolution of Sentinel-1
observations (
This study points out that future efforts will be needed to improve
irrigation estimates through the joint use of LSMs and Sentinel-1
observations, to allow for a more realistic description of the hydrological
cycle and more reliable irrigation simulations over irrigated areas.
This Appendix has the objective to describe the WCM
equation, stated in Sect. 2.4, in more detail. The
Equation (A4) accounts for the soil-related term, which is described in a
simple linear approach, as a function of SSM, following the work by Lievens
et al. (2017a). The
The ASCAT surface soil moisture products H115 and H116 can be downloaded from
The supplement related to this article is available online at:
SM designed and coordinated the study and did the analyses. CM, GJMDL and MB designed and coordinated the study and helped in the data analysis and interpretation. HL, AT, LB and LZ helped in the interpretation of the results and the data processing and collection. All authors contributed to the editing of the manuscript.
The contact author has declared that none of the authors has any competing interests.
Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
The authors would like to thank the European Space Agency (ESA) for the
funding support as part of the IRRIGATION+ project (contract no. 4000129870/20/I-NB). For details, please visit
This research has been supported by the European Space Agency (grant no. 4000129870/20/I-NB) and the Belgian Science Policy Office (BELSPO, project SR/00/376 EO-DAHR).
This paper was edited by Narendra Das and reviewed by two anonymous referees.