ArticlePDF Available

Estimation of Corn Yield by Assimilating SAR and Optical Time Series Into a Simplified Agro-Meteorological Model: From Diagnostic to Forecast

Authors:
  • Association climatologique de la Moyenne-Garonne
  • ACMG Association climatologique de la Moyenne-Garonne et du Sud-Ouest - Agen France

Abstract and Figures

The estimation of crop yield plays a major role in decision making and management of food supply. This paper aims to estimate corn dry masses and grain yield at field scale using an agro-meteorological model. The SAFY-WB model (simple algorithm for yield model combined with a water balance) is controlled by green area index (GAI) derived from optical satellite images (GAI opt_\text{opt} ), and the GAI derived from synthetic aperture radar (SAR) satellite images (GAI sar_\text{sar} ) acquired over two crop seasons (2015 and 2016) in the south-west of France. Landsat-8 mission provides the optical data. SAR information ( σVV\sigma _{{\rm{VV}}}^\circ , σVH\sigma _{{\rm{VH}}}^\circ , and σVH/VV\sigma _{{\rm{VH/VV}}}^\circ ) is provided by Sentinel-1A mission through two angular normalized orbits (30 and 132) allowing a repetitiveness from 12 to 6 days. σVH/VV\sigma _{{\rm{VH}}/{\rm{VV}}}^\circ is successfully used to derive GAI sar_\text{sar} (R2 = 0.72, relative root mean square error (rRMSE) = 10.4%) over the leaf development stages of the crop cycle from a nonlinear function. Others SAR signal ( σVV\sigma _{{\rm{VV}}}^\circ and σVH\sigma _{{\rm{VH}}}^\circ ) are too much related to soil moisture changes. At the opposite of GAI opt_\text{opt} , GAI sar_\text{sar} cannot be used alone in the model to accurately estimate vegetation parameters. Finally, the robustness of the results comes from the combination of GAI derived from SAR and optical data. In this condition, the model is able, thanks to the inclusion of a new “production module,” to simulate dry masses and yield (R2 > 0.75 and rRMSE < 12.75%) with good performances in the diagnostic approach. In the context of forecast, results offer lower performances but stay acceptable, with relative errors inferior to 13.95% (R2 > 0.69).
Content may be subject to copyright.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
1
AbstractThe estimation of crop yield plays a major role in
decision making and management of food supply. This work aims
to estimate corn dry masses and grain yield at field scale using an
agro-meteorological model. The SAFY-WB model (Simple
Algorithm For Yield model combined with a Water Balance) is
controlled by Green Area Index (GAI) derived from optical
(GAIopt) and Synthetic Aperture Radar (SAR) (GAIsar) satellite
images acquired over two crop seasons (2015 and 2016) in the
South-West of France. Landsat-8 mission provides the optical
data. SAR information (σ°VV, σ°VH, σ°VH/VV) is provided by the
Sentinel-1A mission through two angular normalized orbits (30
and 132) allowing a repetitiveness from 12 to 6 days. The σ°VH/VV
is successfully used to derive GAIsar (R² = 0.72, rRMSE = 10.4%)
over the leaves development stages of the crop cycle from a non-
linear function. Others SAR signal (σ°VV, σ°VH) are too much
related to soil moisture changes. At the opposite of GAIopt, the
GAIsar cannot be used alone in the model to accurately estimate
vegetation parameters. Finally, the robustness of the results comes
from the combination of GAI derived from SAR and optical data.
In this condition, the model is able, thanks to the inclusion of a new
“Production module”, to simulate dry masses and yield (R² > 0.75
and rRMSE < 12.75 %) with good performances in the diagnostic
approach. In the context of forecast, results offer lower
performances but stay acceptable, with relative errors inferior to
13.95 % (R² > 0.69).
Index Terms SAFY-WB, GAI, Sentinel-1, Landsat, biomass,
maize, corn, remote sensing, yield forecast.
I. INTRODUCTION
ULTIVATED throughout the world, corn (Zea mays L.) is the
most produced cereal (almost 1.3 billion of tons in 2016,
[1]). In France, it is the second crop production and is
mainly cultivated in areas that are located in the south west
(40% of the national production). Fine scale information about
corn production over large area is needed for farmers and
decision-makers to forecast food security or economic impact
from local to global scale [2]. To this end, numerous crop
models have been developed over the last few decades (e.g.,
Manuscript submitted January 5, 2018. This work is part of the PRECIEL
project, supported by ACMG, Agralis, Nouvelle-Aquitaine Region, European
Union, and CESBIO and certified by Agri Sud-Ouest Innovation.
M. Ameline, R. Fieuzal, F. Baup are with CESBIO, Université de Toulouse,
CNES, CNRS, IRD, UPS, Toulouse, France
CERES [3], CROPSYST [4], GRAMI [5], STICS [6],
WOFOST [7] (for more details see review on crop modeling
production: [8][11]). Many of them simulate plenty of
variables (i.e., nitrogen consumption, phenological stages,
energy fluxes, water balance, carbon allocation, Green Area
Index (GAI), grain yield...) and require an accurate description
(complete set of parameters used as input) of the surface. Such
a list of crop parameters cannot be realistically measured at a
regional or national scale (too much time consuming). This
constraint is reinforced with the increase of fields and spatial
heterogeneities intrinsic to large areas.
To overcome this limitation, several studies shown the
interest of combining simplified crop models (inducing less
input parameters), well-selected field data (lower time
requirements) and remote-sensing images (large areas covered)
to estimate crop production parameters (Aquacrop [12], SAFY
[13], GRAMI [5]). These models are all based on the efficiency
theories of the plant using the Monteith equations [14]. Based
on the growing degrees day base, they simulate the GAI from
which the production of dry mass is directly derived. The
oldest, Aquacrop is developed by the Food and Agriculture
Organization (FAO) and offers the best described water balance
module (runoff, infiltration, deep percolation). GRAMI and
SAFY-WB (Simple Algorithm For Yield model combined with
a Water Balance) are respectively developed by the United
States Department of Agriculture (USDA) and by the CESBIO
laboratory. For both, the vegetation part is driven by the same
equations of Monteith. SAFY-WB integrates a simple water
balance model since no hydrological model is implemented in
GRAMI. The SAFY-WB model is retained since it has been
validated for different crop types (wheat [13], [15], corn [16],
[17], sunflower [18], soybean [19]), for contrasting climatic
conditions (semiarid [20] and temperate), and controlled by
optical and/or SAR remotely sensed. Currently, with the high
availability of Sentinel and Landsat missions, satellites are fully
adapted for field monitoring [21] by providing timely and
continuous information (few days of repeat cycle) at high
spatial resolution (pixel size from 10 to 30m for optical and
J-F. Berthoumieu and M. Ameline are with Association Climatologique de
Moyenne-Garonne et du Sud-ouest (ACMG)
J. Betbeder is with Cirad, UPR BSEF, F-34398 Montpellier, France
Estimation of corn yield by assimilating SAR
and optical time series into a simplified agro-
meteorological model: from diagnostic to
forecast
Maël Ameline, Rémy Fieuzal, Julie Betbeder, Jean-François Berthoumieu, Frédéric Baup
C
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
2
microwave), over large areas (several hundred kilometers
wide).
Often combined with simplified model, the contribution of
optical images have been widely proved in literature in order to
monitor, understand, and estimate crop production, especially
for corn. Mostly based on vegetation indices [2], [22] or
radiative transfer models [23], biophysical crop parameters are
derived, and then, are implemented into models. However,
optical acquisitions are highly dependent on the weather
conditions (e.g., cloud cover, haze) and the product derived
from their use is non-guaranteed for final users.
In such a context, several studies emphasize the all-weather
capability offered by multi-temporal images provided by SAR
sensors to retrieve biophysical parameters: GAI, biomass, soil
moisture [24][27]. This derived SAR information can be used
directly, or implemented into agro-meteorological models
(coupling weather data and agronomic functioning model of
crop) for estimating the grain mass production of wheat [28],
soybean [19], sunflower [18], and other crops.
In this context, the objective of this study is to estimate grain
corn yield, derived from dry masses (ear, plant, and total
amount), at field scale using a simple agro-meteorological
model controlled by optical and/or SAR satellite images
acquired in the South-West of France. Two approaches are
considered: diagnostic and forecast. The site and the data
collection (in-situ and satellite) are presented in section II. The
methodology is explained in section III (SAR and optical
processing to derive GAI, presentation of the model and the
associated new contribution). Section IV includes the results
and discussion about the model performances in the diagnostic
and forecast approaches. The diagnostic approach aims to
estimate grain yield from all the information acquired
throughout the agricultural season, whereas the forecast
approach consists in estimating grain yield by using an
Fig. 1 Location of the study site and the monitored corn fields (with zooms for the sampled fields: a, b, c, d) in 2015 and 2016, superimposed with the
Landsat-8 and Sentinel-1 swaths.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
3
increasing number of satellite images according to their
acquisition in time.
II. MATERIALS
A. Site description
The study area, centered on the coordinates: 44.06N, 0.42E,
and covering a footprint of 110 km × 140 km, is located in the
South-West of France over two French regions (Nouvelle-
Aquitaine and Occitanie) (Fig. 1). With a tempered climate, the
area is mainly dedicated to the cultivation of seasonal crops
which fills 54% of the land (percentages based on the land cover
map available in [29]. In the West part of the study site, soils
are mostly sandy. They are more heterogeneous in the Eastern
part, depending on the relief and the proximity to large rivers,
and dominated by the fraction of silt or clay (silty clay loam or
silty loam soils as defined in [30]).
B. In situ-data
1) Climatic data: Meteorological stations are implanted in
the vicinity of each group of fields (since the required climatic
variables are not freely available over the area) (Fig. 1). The
weather parameters are standardized at a daily time step and
concern solar radiation, rainfall, air temperature, wind speed
and relative humidity. The reference crop evapotranspiration
(ET0) is then derived, based on the procedure described in [31].
Fig. 2 Temporal evolution of the daily rainfall (blue bar) and
evapotranspiration (ET0, black line) recorded by the station in the South-East
of the study area (Fig. 1, zoom d), during the agricultural season of corn.
Degrees day accumulation (6°C base) is presented for the fields “F1” and
“F5” (associated with phenological stages observations in 2016).
The minimum values of ET0 are observed in spring with 2
mm, they reach 5-6 mm during summer when the climatic
demand is maximum (Fig. 2). In 2016, these values are more
heterogeneous due to the presence of clouds inducing lower
radiative fluxes. The 2015 cultural period (from April to the end
of October) is much drier than that of 2016, as evidence by the
amount of rainfall of 233 mm and 380 mm, respectively. In
2015, two dry periods without rainfall are observed, just after
the sowing around the day 150 and during the flowering and the
tassel emergence from days 170 to 195. In 2016, rainfalls are
regularly distributed except at the end of summer from days 214
to 244. The cumulative degree day (using the 6°C base) allows
presenting the phenological stages (Fig. 2) of corn based from
sowing.
2) Field information: Field information is collected over 122
fields of corn (66 in 2015 and 56 in 2016) managed with
conventional practices (without intercropping) (Table. I). The
sizes of the fields are heterogeneous, with areas ranging from
3.1 ha to 62 ha for the biggest one. They are mostly distributed
over hilly surface (slope of about - or 7% - in average, and
up to 11°). The fields monitored over the two years are
comparable, regarding the sizes (averages of 9.6 and 11.9 ha)
and the slopes (averages of 4.3 and 4°). The corn fields,
cultivated in 2015, show earlier dates of sowing and harvest (in
average, days 103 and 241 days compared to days 121 and 294
in 2016) and a shorter vegetative period (- 35 days). In 2015,
because of a dry summer followed by thundershowers in some
places, many farmers decided to harvest earlier in order to avoid
yield losses.
TABLE I
MAIN FEATURES OF THE CORN FIELD
Year
2015
n
66
Area (ha)
[min max mean]
[3.1 42.3 9.6]
Slope (°)
[min max mean]
[0.3 11.2 4.3]
Sowing day
[min max mean]
[94 127 103]
Harvest day
[min max mean]
[237 288 241]
Yieldmea (q.ha-1)
(15% moisture)
[min max mean]
[31 145 109]
Yieldspa (n)
60
Yield information is generally gathered from the farmer once
the harvest has been weighed. Yet, some combine harvesters,
coupled with positioning technologies, record instantaneously
the yield information. This spatialized knowledge allows
reshaping the field into homogeneous areas, following the
agricultural practices (i.e., by separating the irrigated and non-
irrigated fields, Fig. 3).
Fig. 3 Example of a yield map produced over a heterogeneous area, before (a)
and after been reshaped (b)
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
4
The comparison between the weighed yields and those
obtained from combine harvesters, (Fig. 4) ensures reliability
and guarantees the spatialized yields (yieldspa) to be used as
“ground truth” with a high confidence index (R² > 0.82 and
rRMSE < 11 %). In this case, yield estimated from combine
harvesters have not been reshaped to be consistent, in spatial
scale, with those weighed. When spatialized data are not
available (6 fields in 2015 and 2016), the weighed yield is used
as the reference. Displayed in the trade standard of moisture
(15%), the yield values have the same range of variation (from
31 to 152 q.ha-1) over the two years, with 109 q.ha-1 in average
(Table. I).
Fig. 4 Comparison between yield measured after weighing and yieldspa
measured spatially over a non-reshaped field for the years 2015 and 2016
(15% grain moisture).
3) Ground surveys: Biomass measurements are performed
following an Elementary Sampling Unit protocol (ESU: 20 m ×
20 m - [33]) where 5 plants are weighed in situ before being
dried at 65°C with a minimum of 60 h. The fruit (ear), as yield
component, is separated from the plant (stem and leaves) to
obtain the Ear Dry Mass (EDMmea), the Plant Dry Mass
(PDMmea), and the Total Dry Mass (TDMmea = PDMmea +
EDMmea). The values are expressed in quintal per hectare (q.ha-
1) using the plant density information. In 2015, samples are
collected on one field on two dates at the end of the crop cycle
(days 216 and 236) whereas in 2016 they are collected over 5
fields all along the crop cycle from sowing to harvest (each 3
weeks over 4 fields (Table. II)). The values of TDMmea range
from 0.1 q.ha-1 to a maximum of more than 220 q.ha-1 reached
a few days before the harvest (examples of two fields (“F1” and
“F5”) presented in Fig. 6).
TABLE II
DATES OF BIOMASS MEASUREMENTS DURING THE AGRICULTURAL SEASONS OF
CORN (APRIL TO OCTOBER) FOR YEARS 2015 AND 2016.
C. Satellite data
1) Optical data: Despite the launch of the Sentinel 2A
satellite in June 2015, images are only available for the second
cultural year (from December). To be consistent between the
two years, optical data exclusively comes from the Landsat-8
OLI sensor, launched in 2013. The images are processed by the
level 2A processor named MACCS (Multi-sensor Atmospheric
Correction and Cloud Screening) [34] and are available free of
charge at the Theia land data services [32]. Main features of
images are given in Table. III (only the spectral bands used are
presented).
From April to the end of October, 24 and 22 images are
collected, respectively in 2015 and 2016. This must be
id
n
Date of sampling
(day of year)
Date of sampling
(°C.day)
F1 (2015)
2
216, 236
1572, 1879
F2 (2016)
7
148, 174, 188, 209, 231,
249, 270
202, 508, 709, 1043,
1390, 1693, 1974
F3 (2016)
5
183, 201, 223, 241, 257
679, 944, 1261, 1558,
1801
F4 (2016)
6
147, 172, 194, 215, 236,
267
336, 621, 959, 1280,
1611, 2080
F5 (2016)
7
133, 148, 169, 195, 215,
232, 266
207, 345, 592, 978, 1292,
1567, 2088
F6 (2016)
7
133, 148, 169, 195, 215,
232, 266
207, 345, 592, 978, 1292,
1567, 2088
Fig. 5 Time course of optical and SAR (orbit 30 and 132) acquisition during the cultivation cycle of corn (April to October) for years 2015
and 2016. Cloud cover rate over fields are mentioned by black circle (information not provided for SAR data).
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
5
contrasted assuming that, in average in 2015, 58% of the
information over the fields is worthless because of clouds. It is
even worse in 2016 with 74% of the fields that are masked (Fig.
5). In the following, these images are used to calculate the
NDVI [35].
TABLE III
SATELLITE (SAR AND OPTICAL) FEATURES
Mission
Sentinel-1A
Landsat-8
Swath
250 km
185 km
Repetitivity
12 days
16 days
Resolution
20 x 5
30 x 30 m²
Sensor features
5.405 GHz
Dual
polarization
(VV, VH)
B2 : 0.45-0.52 µm (B)
B3: 0.53-0.60 µm (G)
B4: 0.63-0.68 µm (R)
B5: 0.85-0.89 µm (NIR)
2) Microwave data: SAR images are provided by Sentinel-
1A, freely available from the European Copernicus Services
Data Hub [36]. The satellite features are presented in Table. III.
Over the studied region (Fig. 1) in Ground Range Detected
(GRD) mode with an incidence angle ranged between 29.1° and
46.0°, following two orbits (#30 and #132) [37]. During the
growing period of corn, 28 and 35 images are respectively
acquired in 2015 and 2016. The images are pre-processed
(radiometric calibration, Range Doppler terrain correction and
resampled at 10 m spacing) using the ESA’s SNAP software to
obtain the sigma-naught for the VH and VV polarizations. The
SAR signal is sensitive to the antenna incidence angle, an effect
which can be reinforced by the local incidence angle observed
over hilly landscape. It is necessary to acquire images at regular
visit time using the same incidence angle [38]. Thereby an
angular normalization [19], [39] is applied from sowing to
harvest (1) to take into account the vegetation effect. Thanks to
the swath overlap of the two orbits (#30 and #132), the
monitored fields are finally observed every 6 days, whatever the
weather conditions (Fig. 5).

   (1)
Where 
is the normalized SAR signal, is the local
incidence angle of the acquisition,  is the local incidence
angle reference (37.5°),  is the average angular sensitivity
(0.06 for VH and 0.07 dB.°-1 for VV, calculated from
backscattering coefficient and slope, derived from SAR images
and SRTM digital elevation model, respectively) and
is the
initial SAR signal.
Fig. 6 Times courses of ground data (dry masses), NDVI and the backscattering signals (σ°VH, σ°VV, and σ°VH/VV) observed over all the monitored corn fields
for the years 2015 (a, c, e, g, i) and 2016 (b, d, f, h, j). The crosses (“+”) represent the values observed over the two sampled fields (“F1” and “F5”), the grey
bands represent the standard deviation (SD) of the median values (black line).
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
6
3) SAR and optical signatures: The NDVI, SAR signals and
ground are presented as a function of the degree days (Fig. 2),
which allows aligning the profile to a unique sowing reference
(Fig. 6). The evolution of the NDVI represents the dynamic of
the photosynthetic activity. It increases from sowing to 1000
°C.day, then saturates and finally decreases during senescence
phase (from around 1800 °C.day),. This signal is not affected
by soil moisture changes. At the opposite, The σ°VH and σ°VV
signal are significantly affected by the variations of soil
moisture (due to rains or irrigations) when the vegetative
density is low (i.e., at the beginning of agricultural season and
before the harvest) (Fig. 6e-h) (confirmed by [38]). This
phenomenon is strongly year dependent and is clearly visible in
2016 contrary to 2015. Indeed, rainfalls are more important
during this period (Fig. 6f, h) in 2016 and occur later in 2015
(Fig 6e, g). The impact of irrigations or rainfalls decreases when
the vegetation is fully developed from 900 to 1500 °C.day, as
shown by [38], [40]. The use of ratio σ°VH/VV overcomes this
limitation for vegetation monitoring (compensation of soil
moisture impacts) [38], [41]. From 0 to around 300 °C.day, this
ratio slightly decreases to reach a minimum (Fig. 6i, j) around
the 4 or 5 leaves and corresponds to a low rainfall period and a
sustained evaporative demand (around 5 mm/day) accentuated
by the vegetation demand (Fig. 2). It contradicts the argument
of [40] as no harrowing is practiced after sowing for any of the
fields. Then, the SAR signal (σ°VH/VV) rises till a plateau
(around 1000°C.day) while the corn biomass still grows [39].
At the end of the agricultural season, the SAR signal is
decreasing with disparities depending on the fields, due to
harvest date variations (according to the growth duration of the
variety). In 2015, this decline is less pronounced than in 2016.
This phenomenon is due to the extension of the harvesting
period in 2015 according to favorable climatic conditions (with
some earlier harvests, Table. I). At the opposite, the signal
significantly decreases in 2016, at the end of the growing season
over a shorter period than in 2015. In this case, farmers have
harvested their crops to prevent possible damages due to
predicted heavy rains.
III. METHODOLOGY
As illustrated in Fig. 7, the methodology consists of
assimilating Green Area Index (GAI) derived from optical and
SAR satellite images into the SAFY-WB agro-meteorological
model to simulate GAIsim and dry masses (from which the yield
is derived), following a diagnostic or a forecast approach. The
main steps involved estimating GAI from reflectances or
backscattering coefficients are presented hereinafter, together
with an overview of the SAFY-WB model and the description
of the new proposed module, ending by the calibration and the
validation procedures.
A. From satellite to GAI estimates
The first step consists in deriving the GAI from optical and SAR
satellite data. The time courses of GAI are derived from optical
images using the BVNet tool (Biophysical Variables Neural
NETwork) developed by [23]. The radiative transfer model
PROSAIL is first used to constitute a training dataset, with the
constraint of estimating reflectance (from 400 nm to 2500 nm)
for a wide range of conditions regarding the crop biophysical
variables. Artificial neural network (ANN) are then trained on
Fig. 7 Workflow of the main steps involved to estimate corn yield assimilating satellite (SAR and optical) information in the SAFY-WB model
following two approaches: diagnostic and forecast.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
7
the dataset estimated from PROSAIL (considering reflectance
as input variable, and GAI as output). Regarding to the GAI, the
domain of validity of such approach ranges between 0 to 6
m2.m-2. The trained ANN are finally applied to satellite images,
using the green, red and near-infrared reflectance. Such an
approach has been validated in the South-West of France using
optical images provided by Formosat-2, Spot-4, and Landsat-8,
on independent ground measurements, showing accurate
performance with correlation of 0.92 and RMSE of 0.4 m2.m-2
for corn ([16]).
Fig. 8 Relationship between GAIopt derived from optical images and the ratio
of SAR backscattering coefficients (σ°VH/VV) and the associated residuals
(based on the fit model) from sowing (0 °C.day) to 1000 °C.day.
The sensitivity of the SAR signals to the GAIopt are
established using a non-linear regression, focusing on the
period from the sowing date to the average SAR signal
saturation (1000°C.day, Fig. 6). Analyses are based on images
acquired in 2015, optical data being almost unusable in 2016,
due to cloud cover (Fig. 5). The relationships based on mono-
polarization are limited or not satisfactory to estimate the GAI
(R² = 0.56; RMSE = 0.94 m².m-² for σ°VH and R² = 0.13; RMSE
= 1.13 m².m-² for σ°VV), due to the sensitivity of these signals to
the soil moisture. The best performances are obtained with the
σ°VH/VV, showing high coefficient of determination (R² = 0.72;
RMSE = 0.76 m².m-² - Fig. 8). It confirms that the ratio of
polarization is more sensitive to the volume scattering and is
less affected by rainfall.
Nevertheless, the residuals estimated between the regression
model and GAIopt remain important (> 0.5 m².m-² and RMSE =
1.02 m².m-²) when the ratio is greater than -7.5 dB.
Consequently, equation (2) is used to derive GAIsar inside the
domain of validity [-12 dB, -7.5 dB], given by the analysis of
the residual values.
 

B. The SAFY-WB model
1) Model presentation: The SAFY-WB model (Simple Agro-
meteorological For Yield estimate and Water Balance, [15]
simulates the temporal evolution of GAI (GAIsim) and aerial dry
masses (i.e., PDMsim, EDMsim and TDMsim) from day close of
emergence (D0) until harvest. The model runs at a daily time
steps, and physical processes are controlled by climatic
variables and GAI. The daily dry matter production (3) is
proportional to the PAR (Photosynthetically-Active Radiation),
according to the ELUE (Effective Light-Use Efficiency) (4).
The GAIsim is then derived from the daily total dry matter (5-6).
  ; (3)
with,    ; (4)
  ; (5)
with,     (6)
where ELUE represents the Effective Light-Use-Efficiency;
SLA the Specific Leaf Area; PLI the Partition to Leaf Index;
Sc the Stress Coefficient of water and temperature; doy the
day of year.
For a better description of the vegetation growth, the fixed
parameter SLA is transformed to a dynamic variable, as
proposed by [16], [42]. In this work, the daily values of SLA
are derived from a degrees day reference, following a
relationship estimated from in-situ data (Fig. 9).
Fig. 9 Evolution of the SLA (GAIopt/PDMmea) according to the cumulative
degrees day from the sowing.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
8
2) Corn production module: The daily ear dry mass
production module has been implemented based on ground
measurements of the TDMmea and the EDMmea (Fig. 10).
Fig. 10 Evolution of the ratio TDMmea / EDMmea according to the cumulative
degrees day from the sowing
Contrary to the use of a simple harvest index [16], [17], this
approach allows simulating the evolution of the biomass
allocated to the ear through the ratio between the TDMmea and
the EDMmea. The fraction of the biomass allocated to EDMsim
daily increases following a degrees day reference, the
remaining portion being assigned to the PDMsim. The final yield
is obtained through the maximum of the EDMsim associated
with the fraction of the EDMsim (integrating all the components
of the ear) attributed to the grain part (determined in the
calibration step). At a regional scale, various earliness indices
of corn variety are sown depending on the climate and because
of farmers strategies (e.g., to move forward the market prices)
(Table. IV). This index represents the growth duration of the
crop which is dependent upon variety. It is expressed in degree
days by the seed producers from sowing till 32-35% of grain
moisture. Earlier varieties have shorter growth duration until
maturity. The new module integrates the earliness, announced
by the seed companies (from sowing to the harvested product),
by varying the duration (through coefficients based on the
earliness index of the calibrated field) of the Stt parameter of the
model (value available in Table. V) (7).
TABLE IV
EARLINESS INDICES VARIATION DEPENDING ON THE VARIETY OF CORN SEEDS
ACCORDING TO THE CUMULATIVE DEGREES DAY FROM SOWING TO HARVEST
(USING 6°C BASE) [43]
Earliness
type
Degrees day
(°C.day)
Variation from
Ref. (%)
Number of
fields (2015)
Number of
fields (2016)
Early
1740
-13.65
3
0
Mid-early
1800
-10.70
0
2
Mid-late
1940
-3.72
45
36
Late (ref)
2015
0
15
17
Very late
2090
3.72
3
0
    ) (7)
C. Calibration step
The model is calibrated on the “F5” field of corn (Fig. 1, Fig.
6, and Table. II), selected among the monitored fields, because
it has the advantage of gathering all the ground and satellite
observation. Moreover, no stress or damage has been observed
on this field. Six target parameters (Pla, Plb, Stt, Rs, D0, and
ELUE) are calibrated by minimizing the cost function based on
the difference between the simulated values of the GAIsim and
the TDMsim, and the GAI derived from satellites and the
TDMmea (8-10).
 
 
 (8)
 
 

(9)

  


 (10)
where GAIsim or TDMsim correspond to values simulated by the
model at the time (ti), GAImea and TDMmea the values obtained
from measurements (satellite or ground), N represents the
number of data collected between the sowing (0°C.day) and
the harvest (2500 °C.day).
D. Validation: from diagnostic to forecast
In the validation step, the model is applied to 66 and 55 fields
of corn independent from the calibration step in 2015 and 2016
respectively. D0 and ELUE are optimized field by field by
comparing the GAIsim and the GAI derived from satellite [13].
The values of the others parameters (Pla, Plb, Stt, Rs) stay
constant and come from the calibration step.
In the diagnostic approach, all the available satellite images
are used to derive time series of GAI which finally control the
agro-meteorological model. This approach requires waiting the
end of the crop season to be implemented.
Figure 11: Scheme of the methodology used in the forecast approach,
concerning the estimation of D0 and the number of images assimilated
according to the range of temperature ([0-250], [0-500], …, [0-2500] °C.day).
The aim of the forecast approach is to estimate con yield in
near real time conditions. To this end, the D0 parameters is fixed
as the minimum of the σ°VH/VV (corresponding the 4 or 5 leaves
Figure X: Scheme of the methodology used in the forecast approach, concerning
the estimation of D0 and the number of images assimilated according to the range
of temperature ([0-250], [0-500], …, [0-2500] °C.day).
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
9
stage as observed on the (Fig. 11 top). Then, each 250 °C.day
(from sowing till 2500 °C.day), the forecast is performed by
taking into account all the images (GAIsar+opt) successively
acquired, and considering previous images to adjust the future
prediction (Fig. 11 bottom).
IV. RESULTS AND DISCUSSION
A. Calibration
The results (Fig. 12) show that the GAIsim increases to reach
a plateau around 1000/1200 °C.day (except for the “GAIsar”)
where the plant progressively stops its leaf (and stem)
development and reallocates the energy into the kernel
production. At this step, the biomass is partitioned into EDMsim
and PDMsim thanks to the function of allocation to ear (Fig. 10).
After 2000°C.day, the crop enters finally in its ripening period,
the vegetation dries and the mass production stops till the
harvest. The model simulations, the values of model parameters
and the associated performances are respectively presented in
Fig. 12, in Table. V-VI.
TABLE V
VALUES OF THE MODEL PARAMETERS DEFINED IN THE CALIBRATION STEP ON
THE FIELD “F5” IN 2016
GAIsar
GAIopt
GAIsar+opt
Range [17]
Pla
0.22
0.15
0.10
[0.05 0.5]
Plb
-10-4
0.002
0.002
[10-5 10-2]
Stt (°C.day)
1516
1553
1456
[0 2000]
Rs (°C.day)
12129
6425
8520
[0 105]
D0 (doy)
121
134
137
[90 250]
ELUE (g.MJ-1)
3.72
4.42
3.91
[0.5 6]
The estimations of the parameters are close using “GAIsar+opt
and “GAIopt” to control the model with respectively, emergence
at days 134 and 137 (Table. V) and earlier when using “GAIsar”.
In these cases the D0 corresponds to the 4 or 5 leaves stage (Fig.
2) from an agronomic perspective, few stages after emergence.
The performances are comparable when the model is either
controlled by “GAIoptor “GAIsar+opt” and the simulations well
reproduce the GAIopt with a small error rate (R² > 0.98 and;
rRMSE < 12.18 %) (Fig. 12a, b; Table. VI). The estimations of
the dry masses (PDMsim, EDMsim, TDMsim) are accurate, with
high level of R² (> 0.97) and reasonable rRMSE (inferior to
15.71 %). At the opposite, the GAIsim is strongly overestimated
(rRMSE = 73.62 %), and the simulation exceeds 7 m².m-² (Fig.
12a) when the model is controlled by the “GAIsar” (which is not
realistic for corn [17]). Nevertheless, thanks to the constraint of
TDMmea in the cost function, the dry masses are more fairly
estimated than the GAIsim with rRMSE from 17.60 % (TDM) to
14.43 % (EDM). The grain part (grain yield) of the ears is then
calculated by dividing the measured yieldmea (133 q.ha-1) with
the maximum EDMsim coming from the three type of GAI
assimilated (221.78 q.ha-1, 168.49 q.ha-1, 180.53 q.ha-1
corresponding to assimilation of “GAIsar”, “GAIopt”, and
“GAIsar+opt”). This part represents 60%, 79%, and 74%
respectively of the maximum of EDMsim. The two last values
are more consistent with ground measurements performed in-
situ (86% in mean on 10 samples) whereas the first one is too
small, explained by the overestimation of EDM when the model
is controlled by SAR data only. These coefficients are used in
the following to derive the yieldsim from the maximum of
EDMsim.
TABLE VI
SUMMARY OF STATISTICAL PERFORMANCES OF THE CALIBRATION STEP USING
“GAISAR”, GAIOPT”, GAISAR+OPT”. GAIOPT IS USED AS REFERENCE TO BE
CORRELATED TO THE GAISIM (NO GAIMEA AVAILABLE). A AND B REPRESENT
THE COEFFICIENT OF THE LINEAR REGRESSION USED TO CORRELATE
SIMULATION AND OBSERVATION.
Configuration
Model
output
rRMSE
(%)
n
a
b
GAIsar
GAIsim
0.87
73.62
9
2.61
0.60
PDMsim
0.98
14.71
7
1.27
57.49
EDMsim
0.98
14.43
4
1.09
72.62
TDMsim
0.98
17.60
7
1.16
104.86
GAIopt
GAIsim
0.98
12.18
9
1.05
-0.01
PDMsim
0.98
15.71
7
1.27
29.69
EDMsim
0.98
14.39
4
1.03
81.80
TDMsim
0.97
19.61
7
1.12
93.17
GAIsar+opt
GAIsim
0.99
9.41
9
1.04
-0.01
PDMsim
0.98
11.58
7
1.08
16.56
EDMsim
0.98
11.31
4
0.89
63.54
TDMsim
0.98
14.85
7
0.97
60.66
At the sight of the calibration results, the validation will be
performed by using only the GAI derived from optical (GAIopt)
or both SAR and optical (GAIsar+opt) images.
B. Diagnostic approach
The performances associated with the estimations of the crop
variables (GAI, TDM, PDM, and EDM) are presented in
Fig. 11 Comparison between observed (dots) and simulated (lines) GAI and dry masses (PDM, EDM, and TDM) when the model is controlled by “GAIsar” (a),
“GAIopt” (b), “GAIsar+opt” (c).
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
10
Table.VII. Regardless the considered year, the GAIsim is well
estimated with R² ranging from 0.95 to 0.99 and rRMSE
inferiors to 21.23 % (n > 745 in 2015 and n > 384 in 2016), and
with regression function coefficients (a and b, corresponding to
slope and offset) close to the perfect cases (1 and 0,
respectively). In 2016, the dry masses estimations are better
estimated using the GAIsar+opt. However the precision is more
heterogeneous depending on the considered variable. TDM and
EDM are estimated with similar low errors (about 15% in mean
using GAIsar+opt). Whereas, for both configuration, the model
has some difficulties to estimate fairly the PDMsim with rRMSE
upper to 31.99 %. Moreover, this last variable is strongly
overestimated as pointed out by the high value of the coefficient
“a” of the regression function (1.66 and 1.36 instead of 1.0)
(Table. VII).
TABLE VII
SUMMARY OF STATISTICAL PERFORMANCES DERIVED FROM THE
RELATIONSHIPS BETWEEN SIMULATION AND OBSERVATION USING GAIOPT
AND GAISAR+OPT TO CONTROL THE MODEL (VALIDATION STEP), AND
PERFORMED FROM N POINTS AND LINEAR REGRESSION FUNCTION ARE
CHARACTERIZED BY A AND B COEFFICIENTS.
year
Configuration
Model
output
rRMSE
(%)
n
a
b
2015
GAIopt
GAIsim
0.95
21.23
747
0.93
0.06
GAIsar+opt
GAIsim
0.96
19.00
745
0.94
0.03
2016
GAIopt
GAIsim
0.99
12.42
384
0.99
-0.03
PDMsim
0.85
43.76
19
1.66
81.24
EDMsim
0.92
20.28
11
1.06
55.97
TDMsim
0.87
31.83
19
1.02
238.85
GAIsar+opt
GAIsim
0.98
14.22
384
0.93
0.10
PDMsim
0.88
31.99
19
1.36
37.67
EDMsim
0.95
13.29
11
0.92
15.34
TDMsim
0.95
16.50
19
0.88
114.41
The combined use of SAR and optical data to control the
model (“GAIsar+opt”) improves the accuracy of the TDMsim by
reducing the rRMSE (-15%) and increasing the R² (+0.08). The
best estimates of the TDMsim are associated with accurate
simulated values of the EDMsim (variable of interest in yield
estimation) with rRMSE of 13.29 % for the “GAIsar+opt”.
For both years and configurations, the model is able to
reproduce with confidence the yield heterogeneities as
illustrated in Fig. 13. It reproduces the difference in biomass
production between irrigated and non-irrigated fields,
regardless the level of yield observed on ground. When the
model is controlled by “GAIopt(Fig. 13 a), the performances
of yield estimation are better in 2015 (R² = 0.76 and rRMSE =
15.65 %) than in 2016 (R² = 0.69 and rRMSE = 16.44 %). This
result is explained by the higher number of optical images
exploitable in 2015 (Fig. 5). The performances of yield
estimation varies directly as a function of the number of
available optical data. The results are improved when SAR
information is added as demonstrated over other crop type
(sunflower, soybean, wheat) [18], [19], [28]. In our case, the
GAIsar fills the missing optical information at the beginning of
the vegetative growth and allows estimating the D0 parameter
more accurately (especially in 2016 where cloud cover was
important). In consequence, the performances are more stable
in time or even improved (R² = 0.75 and rRMSE = 12.75 % in
2015, R² = 0.77 and rRMSE = 12.07 % in 2016) when using the
“GAIsar+opt”. This consistency can also be notified by the slope
of the linear regression (similar using the “GAIsar+optwith a =
0.89 and 0.88 in 2015 and 2016).
At field scale, these results are better than those obtained in
[16] (R² = 0.66) by using only optical satellite data.
Nevertheless, the performances are lower than those obtained
at regional scale using low spatial resolution images (R² = 0.92
and rRMSE = 4.6%) [16], or than those obtained using
mechanistic models such as WOFOST (R² = 0.84) [44]. In the
first case, low resolution images are less affected by the
heterogeneities of fields and final performances are smoothed
(as [45] in the US Corn Belt), and final product of yield cannot
Fig. 12 Comparison between yieldmea and the yieldsim simulated by assimilating GAIopt (a) or a combination of SAR and optical GAIsar+opt (b) for the years 2015
and 2016.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
11
be processed at field scale contrary to our approach. In the
second case, the method is strongly limited to few fields as it
requires detailed information as nutrient for example.
The benefit of earliness index is investigated in Table VIII,
to evaluate the robustness of approach when this information is
not available. In 2015, the results show that there is no
improvement on the yield estimate with or without considering
the index. This result is explained by the specific climatic
condition of this year. Indeed, the end of the crop season was
particularly rainy (rain storm) and farmers decided to harvest
earlier the crop to avoid damage and loss on corn production.
Consequently, majority of the field have been harvested earlier,
whatever the crop variety. At the opposite, the result obtained
in 2016 (standard year from a climatic point of view) show that
earliness index have positive effects on yield estimates (R² 0.77
to 0.74). During this year, fields have been harvested according
to their earliness indices. Whatever the considered year, the
rRMSE stays constant using or not the earliness index (≈ 12-
13%). These results demonstrated that the diagnostic approach
stays operational even if no information is provided by the seed
companies.
TABLE VIII
SUMMARY OF STATISTICAL PERFORMANCES DERIVED FROM THE
RELATIONSHIPS BETWEEN SIMULATION AND OBSERVATION USING GAIS AR+OPT
TO CONTROL THE MODEL (WITH OR WITHOUT CONSIDERING THE EARLINESS
INDEX).
2015
2016
With earliness
index
Without
earliness index
With earliness
index
Without
earliness index
0.75
0.75
0.77
0.74
rRMSE
12.75
12.62
12.07
12.61
The impact of angular normalization function on grain yield
retrieval is addressed in Table IX. The results show that, when
backscattering coefficient is not normalized, the final yield
estimation is slightly degraded for the two years (2015 and
2016) based on the coefficient of determination (-0.02 in 2015
and -0.06 in 2016) and the rRMSE (Table IX). It demonstrates
that the yield estimates stays acceptable without angular
normalization, and that performances slightly increase when
SAR signal is angular corrected.
TABLE IX
COMPARISON OF THE COEFFICIENT AND RRMSE ON YIELD ESTIMATES,
ACCORDING TO THE USE OF ANGULAR NORMALIZATION ALGORITHM ON THE
SAR SIGNAL.
With the
angular
normalization
rRMSE
(%)
Without
angular
normalization
rRMSE
(%)
2015
0.75
12.75
0.73
12.78
2016
0.77
12.07
0.71
13.22
In the following, only the GAIsar+opt are used for the forecast
approach because of its lower dependence on cloud cover
(considering the earliness index).
C. Forecast approach
The statistical performances (R² and rRMSE) associated with
the estimates of the yieldsim using the forecast approach are
presented in Fig. 14. Results show that the forecast is not
possible (R² < 0.44 and rRMSE > 21.47 % in 2015; < 0.15
and rRMSE > 50.31 % in 2016) during the first phenological
stages of the crop development, until 1250°C.day. The
maximum of performances (R² > 0.69 and rRMSE < 13.95 %)
are obtained at about 1750/2000 °C.day, just after the fruit
development (Fig. 2). Then, performances are quite stable until
the harvest (in mean, R² = 0.68 and rRMSE = 15 % in 2015 and
2016), where ear/grain dry mass does not evolve anymore and
only the vegetation moisture content decreases.
Fig. 13 Evolution of the statistical performances (R² and rRMSE) associated
with the estimation of corn yield, by updating the assimilation of GAIsar+opt
each 250 °C.day till 2500 °C.day in 2015 and 2016.
In forecasting approach, few studies are available at a field
scale as they mainly use optical information at low resolution
(MODIS, [46] for corn). The accurate estimates of corn yield
are consistent with agronomic studies, where early predictions
are derived from statistical algorithms trained on ground
measurements [47][50]. The proposed approach takes full
advantage of the observation capabilities of satellite images,
offering opportunities for monitoring large areas unlike
empirical approaches often subservient to few surveyed plots.
Moreover, the obtained performances confirm the usefulness of
combining SAR and optical images for early yield estimates,
and extend the promising results presented by [51] to successive
agricultural seasons. It would be interesting, as [52] (using SAR
on rice crop), to mix historical (not available here) and real-time
weather information to minimize uncertainties.
V. CONCLUSION
The aim of this work was to estimate corn dry masses (ear,
plant, and total amount) and grain yield at field scale using an
agro-meteorological model (SAFY-WB) controlled by GAI
derived from optical (GAIopt) and/or Synthetic Aperture Radar
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
12
(SAR) (GAIsar) satellite images. The methodology presented
first the estimation of GAI from SAR and optical satellite
images, with the associated domain of validity. Contrary to
optical, the SAR data (σ°VH/VV) are not able to estimate GAI all
along the crop cycle and saturate early in the crop season
(around -7.5 dB). In parallel, the improvements of the model
were presented with the inclusion of a new “Production
module” which allows simulating the allocation of mass into
the ears from the simulated total dry mass all along the crop
cycle (the yield was then derived from the ear mass) taking into
account the variety (through the earliness index).
The model was calibrated using three configurations 1)
GAIsar 2) GAIopt, and 3) GAIsar+opt. The results show that the use
of GAIopt is well adapted to control the model along the crop
cycle, with accurate estimates of dry masses (rRMSE < 19.61
%). At the opposite, the GAIsar cannot be used alone to
reproduce the temporal behavior of GAI (rRMSE = 73.62 %).
When combined with GAIopt, the GAIsar offers crucial
information to determinate the parameters D0 of the model,
improving the performances of the dry masses and yield
simulations, and its robustness regardless cloud cover rate.
Consequently, only the two last configurations (GAIopt and
GAIsar+opt) were retained for validation. Finally, in the
diagnostic approach, the estimates of dry masses and GAI were
improved especially for the TDMsim (by reducing the rRMSE (-
15%) and increasing the R² (+0.08)), used to derive yield.
Accordingly, yield performances were better and more stable in
time when SAR data were combined to the optical ones (in
mean R² = 0.76 and rRMSE = 12.5 % in 2015/16). The
consideration of the earliness index brought positive effects for
the yield estimation. This impact should be deeper investigated
by considering more contrasted varieties of corn on other
climatic conditions. Therefore, in the forecast approach, only
GAIsar+opt was considered and results demonstrated that the
yield are quite accurate (R² > 0.69 and rRMSE < 13.95%).
The use of only SAR or optical data is not sufficient to
achieve corn monitoring, due to the low sensitivity of SAR
signal to crop development and the uncertainty to obtain enough
optical data because of cloud cover risk. Moreover, in situ dry
masses remained necessary to apply this method year to year,
as only one field (in 2016) was able to calibrate the model for 2
years. In the future, it would be interesting to develop new
approaches to determine a satellite signal more sensitive to dry
masses in order to improve yield forecasting. For this objective
others microwave frequencies could be investigated, especially
those offered by ALOS-2 PALSAR sensors (L-band). At this
frequency, signal penetrates deeper into the vegetation layer,
and saturates later (compared to C-band) according to the
vegetation biomass. Others SAR techniques could be also
investigated as PolInSAR ones.
With the purpose to deliver a service to the people in charge
of forecasting the future yields, the performance of the forecast
around 1500°C.day is acceptable as it means delivering, around
1 month before harvest, a spatialized information that will help
the management of the storage infrastructure and the means to
carry from the fields the grains. Of course weather scourges as
hail, wind, floods, may reduce locally the estimated potential.
ACKNOWLEDGMENT
This work is part of the PRECIEL project, supported by
ACMG, Agralis, Nouvelle-Aquitaine Region, European Union,
and CESBIO and certified by Agri Sud-Ouest Innovation. We
are very grateful to the farmers involved, ACMG (Patrick
Debert, Céline Cazenave …) and the CESBIO (Marjorie
Battude, Florian Helen …) for their help.
REFERENCES
[1] FAO, “FAOSTAT,” 2016. [Online]. Available:
http://www.fao.org/faostat/en/#home. [Accessed: 23-May-2018].
[2] B. Basso, D. Cammarano, and E. Carfagna, “Review of crop yield
forecasting methods and early warning systems,” in Report Presented to
First Meeting of the Scientific Advisory Committee of the Gloal Strategy
to Improve Agricultural and Rural Statistics. FAO, Headquarters, Rome,
Italy, pp. 1819, 2013.
[3] T. Hodges, D. Botner, C. Sakamoto, and J. H. Haug, “Using the CERES-
Maize model to estimate production for the US Cornbelt,” Agricultural
and Forest Meteorology, vol. 40, no. 4, pp. 293303, 1987.
[4] C. O. Stöckle, M. Donatelli, and R. Nelson, “CropSyst, a cropping
systems simulation model,” European journal of agronomy, vol. 18, no.
34, pp. 289307, 2003.
[5] S. J. Maas, “GRAMI: a crop growth model that can use remotely sensed
information,” ARS - U.S. Department of Agriculture, Agricultural
Research Service (USA), p. 78 pp., 1992.
[6] L. Prévot, H. Chauki, D. Troufleau, M. Weiss, F. Baret, and N. Brisson,
“Assimilating optical and radar data into the STICS crop model for
wheat,” Agronomie, vol. 23, no. 4, pp. 297–303, 2003.
[7] I. Supit, A. A. Hoojer, and C. A. Van Diepen, “System description of the
Wofost 6.0 crop simulation model implemented in CGMS. Volume 1:
Theory and Algorithms.” Office for the Official Publications of the
European Communities, 1994.
[8] A. Di Paola, R. Valentini, and M. Santini, “An overview of available crop
growth and yield models for studies and assessments in agriculture:
Overview of crop models for agriculture,” Journal of the Science of Food
and Agriculture, vol. 96, no. 3, pp. 709714, Feb. 2016.
[9] K. O. Rauff and R. Bello, “A Review of Crop Growth Simulation Models
as Tools for Agricultural Meteorology,” Agricultural Sciences, vol. 06,
no. 09, pp. 10981105, 2015.
[10] P. Oteng-Darko, S. Yeboah, S. N. T. Addy, S. Amponsah, and E. O.
Danquah, “Crop modeling: A tool for agricultural research–A,” Journal
of Agricultural Research and Development, vol. 2(1), pp. 001006, 2013.
[11] P. Hoefsloot, A. Ines, J. C. van Dam, G. Duveiller, F. Kayitakire, and J.
Hansen, “Combining crop models and remote sensing for yield
prediction: concepts, applications and challenges for heterogeneous
smallholder environments,” Publications Office of the European Union,
Italy, Jun. 2012.
[12] L. K. Heng, T. Hsiao, S. Evett, T. Howell, and P. Steduto, “Validating the
FAO AquaCrop Model for Irrigated and Water Deficient Field Maize,”
Agronomy Journal, vol. 101, no. 3, p. 488, 2009.
[13] B. Duchemin, P. Maisongrande, G. Boulet, and I. Benhadj, “A simple
algorithm for yield estimates: Evaluation for semi-arid irrigated winter
wheat monitored with green leaf area index,” Environmental Modelling
& Software, vol. 23, no. 7, pp. 876892, Jul. 2008.
[14] Monteith, J. L. 1972. « Solar Radiation and Productivity in Tropical
Ecosystems ». The Journal of Applied Ecology 9 (3): 747.
https://doi.org/10.2307/2401901.
[15] B. Duchemin et al., “Impact of Sowing Date on Yield and Water Use
Efficiency of Wheat Analyzed through Spatial Modeling and
FORMOSAT-2 Images,” Remote Sensing, vol. 7, no. 5, pp. 5951–5979,
2015.
[16] M. Battude et al., “Estimating maize biomass and yield over large areas
using high spatial and temporal resolution Sentinel-2 like remote sensing
data,” Remote Sensing of Environment, Aug. 2016.
[17] M. Claverie et al., “Maize and sunflower biomass estimation in southwest
France using high spatial and temporal resolution remote sensing data,”
Remote Sensing of Environment, vol. 124, pp. 844857, Sep. 2012.
[18] R. Fieuzal, C. Marais Sicre, and F. Baup, “Estimation of Sunflower Yield
Using a Simplified Agrometeorological Model Controlled by Optical and
SAR Satellite Data,” IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing, pp. 111, 2017.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
13
[19] J. Betbeder, R. Fieuzal, and F. Baup, “Assimilation of LAI and Dry
Biomass Data From Optical and SAR Images Into an Agro-
Meteorological Model to Estimate Soybean Yield,” IEEE Journal of
Selected Topics in Applied Earth Observations and Remote Sensing, vol.
9, no. 6, pp. 25402553, Jun. 2016.
[20] P. Silvestro et al., “Estimating Wheat Yield in China at the Field and
District Scale from the Assimilation of Satellite Data into the Aquacrop
and Simple Algorithm for Yield (SAFY) Models,” Remote Sensing, vol.
9, no. 6, p. 509, May 2017.
[21] S. Moulin, A. Bondeau, and R. Delecolle, “Combining agricultural crop
models and satellite observations: from field to regional scales,”
International Journal of Remote Sensing, vol. 19, no. 6, pp. 10211036,
1998.
[22] X. Jin et al., “A review of data assimilation of remote sensing and crop
models,” European Journal of Agronomy, vol. 92, pp. 141–152, Jan.
2018.
[23] F. Baret et al., “LAI, fAPAR and fCover CYCLOPES global products
derived from VEGETATION,” Remote Sensing of Environment, vol.
110, no. 3, pp. 275286, Oct. 2007.
[24] F. Baup, L. Villa, R. Fieuzal, and M. Ameline, “Sensitivity of X-Band
(σ0, γ) and Optical (NDVI) Satellite Data to Corn Biophysical
Parameters,” Advances in Remote Sensing, vol. 05, no. 02, pp. 103117,
2016.
[25] É. Auquière, P. Defourny, V. Baltazart, and A. Guissard, “ERS SAR time
series analysis for maize monitoring using experimental and modeling
approaches,” 2015. [Online]. Available:
https://earth.esa.int/workshops/ers97/papers/auquiere/index-2.html.
[Accessed: 17-Nov-2015]
[26] H. McNairn and B. Brisco, “The application of C-band polarimetric SAR
for agriculture: a review,” Canadian Journal of Remote Sensing, vol. 30,
no. 3, pp. 525542, 2004.
[27] M. El Hajj et al., “Evaluation of SMOS, SMAP, ASCAT and Sentinel-1
Soil Moisture Products at Sites in Southwestern France,” Remote Sensing,
vol. 10, no. 4, p. 569, Apr. 2018.
[28] R. Hadria et al., “Potentiality of optical and radar satellite data at high
spatio-temporal resolutions for the monitoring of irrigated wheat crops in
Morocco,” International Journal of Applied Earth Observation and
Geoinformation, vol. 12, pp. S32S37, Feb. 2010.
[29] J. Inglada « France land cover classification, from Landsat-8 to Sentinel-
2 in 2016 ». Available online: http://osr-cesbio.ups-tlse.fr/~oso/
[30] FAO/IIASA/ISRIC/ISS-CAS/JRC, “Harmonized World Soil Database
(version 1.0),” FAO, Rome, Italy and IIASA, Laxenburg, Austria, 2008.
[31] R. G. Allen, L. S. Pereira, D. Raes, and M. Smith, “FAO Irrigation and
drainage paper No. 56,” Rome: Food and Agriculture Organization of the
United Nations, vol. 56, no. 97, p. 156, 1998.
[32] Theia Land Services. Available online: https://www.theia-land.fr/en
[33] S. Garrigues, D. Allard, M. Weiss, and F. Baret, “Comparing VALERI
sampling schemes to better represent high spatial resolution satellite pixel
from ground measurements: How to characterize an ESU.” Internal
Report, INRA-CSE, Avignon, http://www. avignon. inra. fr/valeri, 2002.
[34] O. Hagolle, M. Huc, D. Villa Pascual, and G. Dedieu, “A Multi-Temporal
and Multi-Spectral Method to Estimate Aerosol Optical Thickness over
Land, for the Atmospheric Correction of FormoSat-2, LandSat, VENμS
and Sentinel-2 Images,” Remote Sensing, vol. 7, no. 12, pp. 2668–2691,
Mar. 2015.
[35] C. J. Tucker, “Red and photographic infrared linear combinations for
monitoring vegetation,” Remote sensing of Environment, vol. 8, no. 2, pp.
127150, 1979.
[36] Copernicus Open Access Hub. Available online:
https://scihub.copernicus.eu/
[37] R. Torres et al., “GMES Sentinel-1 mission,” Remote Sensing of
Environment, vol. 120, pp. 924, May 2012.
[38] X. Blaes, P. Defourny, U. Wegmuller, A. Della Vecchia, L. Guerriero,
and P. Ferrazzoli, “C-band polarimetric indexes for maize monitoring
based on a validated radiative transfer model,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 44, no. 4, pp. 791800, 2006.
[39] R. Fieuzal, “Apports des données radar pour l’estimation des parametres
biophysiques des surfaces agricoles,” Université Toulouse III-Paul
Sabatier, France, 2013.
[40] A. Veloso et al., “Understanding the temporal behavior of crops using
Sentinel-1 and Sentinel-2-like data for agricultural applications,” Remote
Sensing of Environment, vol. 199, pp. 415426, Sep. 2017.
[41] A. Della Vecchia, P. Ferrazzoli, L. Guerriero, L. Ninivaggi, T. Strozzi,
and U. Wegmuller, “Observing and Modeling Multifrequency Scattering
of Maize During the Whole Growth Cycle,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 46, no. 11, pp. 37093718, 2008.
[42] N. G. Danalatos, C. S. Kosmas, P. M. Driessen, and N. Yassoglou, “The
change in the specific leaf area of maize grown under Mediterranean
conditions,” Agronomie, vol. 14, no. 7, pp. 433–443, 1994.
[43] De Sangosse SA, “L’essentiel du maïs.” [Online]. Available:
http://www.desangosse.fr/medias/olds/semences/273-a-
essentiel%20du%20mais.pdf. [Accessed: 08-May-2018].
[44] Z. Cheng, J. Meng, and Y. Wang, “Improving spring maize yield
estimation at field scale by assimilating time-series HJ-1 CCD data into
the WOFOST model using a new method with fast algorithms,” Remote
Sensing, vol. 8, no. 4, p. 303, 2016.
[45] F. J. Morell et al., “Can crop simulation models be used to predict local to
regional maize yields and total production in the U.S. Corn Belt?,” Field
Crops Research, vol. 192, pp. 112, Jun. 2016.
[46] M. S. Mkhabela, P. Bullock, S. Raj, S. Wang, and Y. Yang, “Crop yield
forecasting on the Canadian Prairies using MODIS NDVI data,”
Agricultural and Forest Meteorology, vol. 151, no. 3, pp. 385393, Mar.
2011.
[47] K. Martin, W. Raun, and J. Solie, “By-plant prediction of corn grain yield
using optical sensor readings and mesured plant height,” Journal of Plant
Nutrition, vol. 35, no. 9, pp. 14291439, Jun. 2012.
[48] L. K. Sharma and D. W. Franzen, “Use of corn height to improve the
relationship between active optical sensor readings and yield estimates,”
Precision Agriculture, vol. 15, no. 3, pp. 331345, Jun. 2014.
[49] X. Yin, R. M. Hayes, M. A. McClure, and H. J. Savoy, “Assessment of
plant biomass and nitrogen nutrition with plant height in early-to mid-
season corn,” Journal of the Science of Food and Agriculture, vol. 92, no.
13, pp. 26112617, Oct. 2012.
[50] X. Yin, M. A. McClure, and R. M. Hayes, “Improvement in regression of
corn yield with plant height using relative data,” Journal of the Science of
Food and Agriculture, vol. 91, no. 14, pp. 26062612, Nov. 2011.
[51] R. Fieuzal, C. Marais Sicre, and F. Baup, “Estimation of corn yield using
multi-temporal optical and radar satellite data and artificial neural
networks,” International Journal of Applied Earth Observation and
Geoinformation, vol. 57, pp. 1423, May 2017.
[52] T. Setiyono, A. Nelson, and F. Holecz, “Remote Sensing based Crop
Yield Monitoring and Forecasting,” 2014.
Maël Ameline received the Master’s
degree in Geography with a specialization
in remote sensing, GIS, and environment
from the University of Rennes 2 in
collaboration with the agronomy
engineering school AgroCampus-Ouest,
Rennes, France in 2014. He is currently
working toward the Ph.D. degree in remote
sensing and agro-systems modeling at the University Paul
Sabatier, Toulouse, France.
He had opportunities to work on rice mapping and yield
estimation based on remote sensing in Madagascar within the
Center of Agricultural Research for Development (CIRAD),
Reunion Island, France. He also acquired knowledge in web
mapping, pedology by studying the impact of climate change
on groundwater irrigated agriculture of Southern India, as an
intern in the Indo-French Cell on Water Science (IFCWS),
Bengaluru, India.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
14
Rémy Fieuzal was born in Agen, France,
in 1981. He received the M.S. degree in
ecology, biostatistics, and modeling from
the University Paul Sabatier, Toulouse,
France, in 2007, and the Ph.D. degree in
remote sensing and agronomy from the
University Paul Sabatier, in 2013, for his
work on the contributions of satellite radar
data for estimating biophysical parameters of agricultural land.
He is working on image processing and analysis of satellite data
acquired in optical and microwave domains, with the Centre
d’Etudes Spatiales de la BIOsphère, Toulouse, France. His
research interests include mapping of different surface
parameters (leaf area index, biomass, irrigation, or top soil
moisture) and agrosystems modeling.
Julie Betbeder received the Master’s
degree in environmental engineering from
the Ecole Supérieure d’Agriculture,
Angers, France, in 2010, and the Ph.D.
degree in geography from the University
of Rennes, Rennes, France, in 2015.
She did her Ph.D. work at the Littoral,
Environment, and Remote Sensing (LETG
Rennes COSTEL), on the evaluation of
optical and radar SAR polarimetry data for the characterization
of ecological continuities in European landscapes. She worked
during her Postdoctorate (2015-2016) on the evaluation of the
assimilation of optical and SAR data into agro-meteorological
model for crop yield estimation in the Centre d’Etude Spatiale
de la BIOsphère (CESBIO), Toulouse, France. She is actually a
Researcher in the Research Unit ‘Forests and Societies’ at the
International Center of Agricultural Research for Development
(CIRAD), Montpellier, France. Her research focuses on the
impact of forest degradation on landscape functions in tropical
forest landscapes. She is also working on the evaluation of
remote sensing data for forest degradation quantification and
characterization.
Jean-François Berthoumieu received a
M.S. degree in Habitat and environment
(MST) and later a PH .D in fluid mechanics,
option aérothermie, from the University
Paul Sabatier, Toulouse, France in 1979.
After a Post Doc at the Alberta Research
Council he worked during 4 years on the
climatology study of storms and hail over
the South-West of France. Then he left the
field of fundamental research to work within a farmer
association (ACMG) where he developed technologies based
upon probes and remote sensing for helping famers to better
manage their irrigation. He worked on four Interreg programs
where he found the limitation of optical and thermal images to
provide a continuous service for farmers due to cloud cover. He
met Frédéric Baup in 2011 and they decided to build a new
program called PRECIEL in 2014 base upon SAR images with
the objective for testing their potential for helping farmers on
large scales to better manage agronomy and irrigation. He is the
president of a cluster called Water & Climate Adaptation and
he reviews on atmospheric research for ELSEVIER.
Frédéric Baup received the M.S. degree in
microwaves and optical
telecommunications (M.O.T.O) in 2003,
and the Ph.D. degree in SAR remote
sensing from the University Paul Sabatier,
Toulouse, France, in 2007. Since 2008, he
is a Researcher in microwave remote
sensing with the CESBIO (Centre d’Etudes
Spatiales de la BIOsphère) Laboratory,
Toulouse, France.
He has authored international journals. His research interests
include microwave remote sensing applied to land surfaces and
SAR image analysis to monitor spatio-temporal variations of
soil (moisture and roughness) and vegetation (biomass)
properties over agricultural or natural areas. His education
interests are focused on physics and remote-sensing sciences.
Dr. Baup is a Principal Investigator of projects supported by the
European Space Agency (ESA), Canadian Space Agency
(CSA), Japan Aerospace Exploration Agency (JAXA), and
German Aerospace Center (DLR) agencies. He is a Reviewer
for journals and conference proceedings.
... Therefore, it was discovered that images acquired with large incidence angles can be effective in monitoring the early growth stages of maize. Contrary to McNairn et al. [29,34], the backscatter values of images acquired from three different orbits continued to increase until 19 September (until the average maize height exceeded 250 cm [13]), at which point they reached saturation [10,32,59,70,77,[84][85][86][87], which is a wellknown effect for crops with higher biomass [8,42]. Before harvest, a slight decrease was observed in σ 0 VV and σ 0 VH values of images acquired from three different orbits on 14, 25, and 26 October (when the average maize height was 275-280 cm). ...
... This is most likely due to the green residues that remained in the field and dried out gradually after harvest [13]. Previous studies have also found that backscatter values decrease after maize harvesting [10,13,32,40,70,81,84]. ...
... NDVI, LAI, fCover, and CW values continued to increase until 29 September (until the average maize height exceeded 260 cm), at which point they reached saturation, which is a well-known effect for densely vegetated areas. Similar findings have been reported in previous studies for NDVI [13,[15][16][17]23,26,32,84,[105][106][107][108] and LAI [17,19,85] values. ...
Article
Full-text available
The decrease in water resources due to climate change is expected to have a significant impact on agriculture. On the other hand, as the world population increases so does the demand for food. It is necessary to better manage environmental resources and maintain an adequate level of crop production in a world where the population is constantly increasing. Therefore, agricultural activities must be closely monitored, especially in maize fields since maize is of great importance to both humans and animals. Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical satellite images were used to monitor maize growth in this study. Backscatter and interferometric coherence values derived from Sentinel-1 images, as well as Normalized Difference Vegetation Index (NDVI) and values related to biophysical variables (such as Leaf Area Index (LAI), Fraction of Vegetation Cover (fCover or FVC), and Canopy Water Content (CW)) derived from Sentinel-2 images were investigated. Sentinel-1 images were also used to calculate plant heights. The Interferometric SAR (InSAR) technique was applied to calculate interferometric coherence values and plant heights. For the plant height calculation, two image pairs with the largest possible perpendicular baseline were selected. Backscatter, NDVI, LAI, fCover, and CW values were low before planting, while the interferometric coherence values were generally high. Backscatter, NDVI, LAI, fCover, and CW values increased as the maize grew, while the interferometric coherence values decreased. Among all Sentinel-derived values, fCover had the best correlation with maize height until maize height exceeded 260 cm (R2 = 0.97). After harvest, a decrease in backscatter, NDVI, LAI, fCover, and CW values and an increase in interferometric coherence values were observed. NDVI, LAI, fCover, and CW values remained insensitive to tillage practices, whereas backscatter and interferometric coherence values were found to be sensitive to planting operations. In addition, backscatter values were also sensitive to irrigation operations, even when the average maize height was about 235 cm. Cloud cover and/or fog near the study area were found to affect NDVI, LAI, fCover, and CW values, while precipitation events had a significant impact on backscatter and interferometric coherence values. Furthermore, using Sentinel-1 images, the average plant height was calculated with an error of about 50 cm.
... This problem can significantly reduce the performance and generalization of the machine learning-based yield estimation algorithm Fieuzal et al. 2017a). Owning to the all-weather monitoring capability of synthetic aperture radar (SAR), many studies advocated the use of SAR images for crop yield estimation (Wiseman et al. 2014;Betbeder et al. 2016;Ameline et al. 2018;Mandal and Rao 2020). In the microwave region of the electromagnetic spectrum, the intensity of incident energy scattered by vegetation is primarily a function of the crop geometry and dielectric properties. ...
... Researchers have successfully used a combination of optical and SAR data for crop parameters, soil moisture simulation and used these in traditional agro-meteorological crop yield estimation models (Jin et al. 2017;Ameline et al. 2018;Setiyono et al. 2018;Zhuo et al. 2018). All these studies reported that the addition of SAR images improved the simulation results. ...
Article
Traditional crop cutting experiment-based yield estimation method captures the regional yield variability but lacks field-level information. Satellite images hold enormous crop information at finer spatial resolution. Crop yield mapping with optical images is particularly challenging if cloud-free images are unavailable during the crucial crop developmental stages. All-weather availability and sensitivity to crop structure, dielectric properties make synthetic aperture radar (SAR) images an excellent resource for yield estimation. Both types of data provide complementary information about crop conditions. A random forest regression model with genetic algorithm-based feature selection is developed to exploit the Sentinel-2 optical and Sentinel-1 SAR images for yield estimation. We utilised the crop harvest and quality survey (BEE) yield data set collected by the Hessisches Statistisches Landesamt (HSL), Wiesbaden, Germany, over 490 fields. We prepared 20 m resolution yield maps for winter wheat, winter barley, winter rye, and winter rapeseed. Input features for the yield estimation model are selected based on the prior knowledge of remote sensing of vegetation. Baseline random forest regression models are developed for all the four crop types with optical and SAR input features. An optimised random forest regression model with genetic algorithm-based feature selection results in performance improvement. Dissimilarity in genetic algorithm selected image features highlights the significance of crop-specific feature selection for yield estimation. The optimised models reliably estimate yield by achieving correlation coefficient (r) of 0.65–0.86, mean absolute error (MAE) 0.93 t ha− 1 to 1.16 t ha− 1, and root mean square error (RMSE) 1.12 t ha− 1 to 1.56 t ha− 1 with BEE yield on testing data set. The proposed models could estimate the intra-field yield variation when winter wheat, winter barley, winter rye were in the shooting phase to the beginning of ear-shifting, and winter rapeseed began to flower or was already flowering. These results demonstrate the merits of our model for early-stage crop yield estimation at the field level with mono-temporal image and adaptability for the cropping season with high cloud cover.
... Plant biomass and crop yield are primarily generated and accumulated from GPP. Therefore, this study also provides a physical mechanism for the remote estimation of plant biomass and crop yield from Sentinel-1 SAR images [52]. The SAR-based model was optimized by different variable combinations. ...
Article
Full-text available
Gross primary production (GPP) measures the amount of carbon fixed by plants and thus plays a significant role in the terrestrial carbon cycle and global food security, especially in the context of climate change and carbon neutrality. Currently, all-sky high-resolution (< 100 m) GPP is increasingly needed for better understanding the food-carbon-water-energy nexus. However, previous studies usually used optical satellites to estimate clear-sky GPP at kilometer-scale resolution. Due to missing estimates under cloudy-sky conditions, monitoring spatio-temporal changes in GPP from optical satellites would suffer from some uncertainties. Moreover, one issue of some previous studies is that they only used optical satellite images or environmental data to estimate GPP rather than jointly integrating them and biome types. To address these challenges, this study attempts to use active microwave Sentinel-1 synthetic aperture radar (SAR) images at 10 m resolution to estimate all-sky GPP. GPP measurements across nine biome types in North America and Sentinel-1 images were employed to develop the SAR-based all-sky model. Meanwhile, an optical-based clear-sky model with Landsat-8 images was also proposed for comparison. The results revealed that (1) Sentinel-1 SAR images can be utilized to estimate all-sky GPP. By integrating Sentinel-1 SAR images, environmental data, and biome types, the optimal SAR-based model showed high accuracy in estimating all-sky daily GPP that coefficient of determination (R 2 ) = 0.764, root mean square error (RMSE) = 1.976 gC/m 2 /d, and mean absolute error (MAE) = 1.308 gC/m 2 /d. (2) The optimal optical-based model had reasonable validation results in estimating clear-sky daily GPP (R 2 = 0.809, RMSE = 1.762 gC/m 2 /d, and MAE = 1.165 gC/m 2 /d). (3) Landsat-8 optical images contributed more than environmental data in the optical-based model, while the contribution of environmental data was higher than Sentinel-1 SAR images in the SAR-based model. (4) The optical-based model had better performance than the SAR-based model in estimating clear-sky daily GPP, and these two models showed reasonable consistency (R 2 = 0.730 and RMSE = 1.858 gC/m 2 /d) and can be utilized together. Therefore, this study demonstrated that active microwave provides an important data source to estimate all-sky high-resolution GPP, advancing our understanding of the carbon cycle, food security, and environmental change.
... The complementarity of data from different sensors was crucial for corn Nc monitoring. Optical sensors are sensitive to chlorophyll activity and C-SAR signals to plant structure (leaves, stems, and cobs) and soil surface [21,31,34,40,55]. Therefore, sensor data fusion predicted Nc accurately because it integrated information from different biophysical parameters associated with corn N status. ...
Article
Full-text available
Corn (Zea mays L.) nitrogen (N) management requires monitoring plant N concentration (Nc) with remote sensing tools to improve N use, increasing both profitability and sustainability. This work aims to predict the corn Nc during the growing cycle from Sentinel-2 and Sentinel-1 (C-SAR) sensor data fusion. Eleven experiments using five fertilizer N rates (0, 60, 120, 180, and 240 kg N ha−1) were conducted in the Pampas region of Argentina. Plant samples were collected at four stages of vegetative and reproductive periods. Vegetation indices were calculated with new combinations of spectral bands, C-SAR backscatters, and sensor data fusion derived from Sentinel-1 and Sentinel-2. Predictive models of Nc with the best fit (R2 = 0.91) were calibrated with spectral band combinations and sensor data fusion in six experiments. During validation of the models in five experiments, sensor data fusion predicted corn Nc with lower error (MAPE: 14%, RMSE: 0.31 %Nc) than spectral band combination (MAPE: 20%, RMSE: 0.44 %Nc). The red-edge (704, 740, 740 nm), short-wave infrared (1375 nm) bands, and VV backscatter were all necessary to monitor corn Nc. Thus, satellite remote sensing via sensor data fusion is a critical data source for predicting changes in plant N status.
... In recent years, there has been an increasing amount of studies on integrating optical and SAR remote sensing data and combining the advantages of both for crop growth monitoring [12], [13]. Pipia et al. [14] proposed a method of Multi-Output Gaussian Processes (MOGPs) for filling the missing Sentinel-2 LAI caused by clouds by using Sentinel-1 time series Radar Vegetation Index (RVI), which can be applied to multiple types of crops and had different LAI prediction capabilities at different growth stages of the crops. ...
Article
Full-text available
Regions with excessive cloud cover lead to limited feasibility of applying optical images to monitor crop growth. In this study, we built an upsampling moving window network for regional crop growth monitoring (UMRCGM) model to estimate the two key biophysical parameters (BPs), leaf area index (LAI) and canopy chlorophyll content (CCC) during the main growth period of winter wheat by using Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-3 optical images. Sentinel-1 imagery is unaffected by cloudy weather and Sentinel-3 imagery has a wide width and short revisit period, the organic combination of the two will greatly improve the ability to monitor crop growth at a regional scale. The impact of two different types of SAR information (intensity and polarization) on the estimation of the two BPs was further analyzed. The UMRCGM model optimized the correspondence between inputs and outputs, it had more accurate LAI and CCC estimates compared with the three classical machine learning models, and had the highest accuracy at the green-up stage of winter wheat, followed by the jointing stage and the heading-filling stage, and the lowest accuracy was found at the milk maturity stage. The estimation accuracies of CCC were slightly higher than that of LAI for the first three growth stages of winter wheat, while lower than that of LAI for the milk maturity stage. This study proposes a new method for regional BPs (especially for CCC) estimation by combining SAR and optical imagery with large differences in spatial resolution under a deep learning framework.
... K gs simulated by hyperspectral imagery were also effective in identifying the extent of natural gas leakage. This was consistent with previous research on the sensitivity of SAFY model parameters to environmental stress [47], [48]. Although results were overall encouraging, the simulation results of some plots did not completely match the experimental design, possibly due to a few weed disturbances in those sample plots. ...
Article
Full-text available
Natural gas leakage occurs frequently due to aging pipes and other factors, but is challenging to detect. In this article, a new, robust method for nondestructive natural gas microleakage detection was proposed. It combines a crop growth model with a convolutional neural network (CNN) approach to quantitatively detect underground natural gas leakage using unmanned aerial vehicle (UAV) hyperspectral imagery. The environmental stress on wheat was used as an indicator to reflect the intensity of natural gas leakage. First, a crop growth model (simple algorithm for yield, SAFY) was used to simulate the growth of wheat, and the environmental stress factor in the model was used to construct the natural gas stress index (Kgs). Subsequently, CNN models were used to estimate the Kgs value with a hyperspectral image as the input. Finally, the CNN estimated Kgs was used to detect the natural gas leakage in the study area. Results showed that the SAFY model Kgs value could effectively identify natural gas leakage, with statistically significant differences (p-value < 0.05) among three leakage levels. Furthermore, compared to a single spectral index, Kgs had superior robustness throughout the wheat growth period. The CNN-1D model with InceptionV2 architecture exhibited the best accuracy in estimating Kgs, with a robust nRMSE of 6.9%. Overall, the combined CNN and SAFY models could accurately detect natural gas leakage, and this approach is more robust than traditional spectral index-based methods. This article provides a new method for nondestructive detecting of natural gas microleakage.
... Los modelos dinámicos que consiguen un menor error en la estimación requieren variables tomadas en campo, e.g. el total de biomasa seca (Ameline et al., 2018), difíciles de obtener sobre explotaciones no experimentales. ...
Article
Full-text available
A precise estimation of agricultural production provides relevant information for upcoming seasons, and helps in the assessment of crop losses before harvest in case of adverse situations. The objective of this work is to explore the development of a model capable of estimating barley production of a small agricultural production (127 ha) in Belchite, Spain. Variables adapted to the crop calendar of the growing barley are used to achieve that purpose. The variables have been created with weather data and remote sensing images. These images are acquired in two ranges of the electromagnetic spectrum, i.e., microwaves and optical spectral range, obtained from Sentinel-1 and Sentinel-2, respectively. Models are defined with a multiple linear regression method using all combinations of the independent variables correlated with production. The best linear regression model has a prediction error of 57.38 kg/ha (4%). The use of spectral variables, derived from radar vegetation index Cross Ratio (CR) and optical Inverted Red Edge Chlorophyll Index (IRECI), and climatic variables adapted to the crop calendar and climatic conditioning is revealed as an adequate strategy to obtain adjusted models.
Article
Full-text available
Nitrogen (N) nutrition index (NNI) is a reliable indicator of plant N status for field crops, but its determination is both labor- and cost-intensive. The utilization of remote sensing approaches for monitoring N, mainly in relevant crops such as of corn (Zea mays L.), will be critical for enhancing effective use of this nutrient. Therefore, the aim of this study was to assess NNI predicted from optical and C-band Synthetic Aperture Radar (C-SAR) satellite data and available soil N (Nav) at different vegetative growth stages for corn crop. Eleven field studies were conducted in the Pampas region (Argentina), applying five fertilizer N rates (0, 60, 120, 180, and 240 kg N ha-1), all at sowing time. Plant samples were collected at sixth-leaf (V6), tenth-leaf (V10), fourteen-leaf (V14), and flowering (R1). Using linear regression models, NNI was best predicted using only optical satellite data from V6 to V14, and integrating optical with C-SAR plus Nav at R1. The best monitoring model integrated vegetation spectral indices, C-SAR and Nav data at V10 with an adjusted R2 of 0.75 achieved during calibration in the northern Pampa. During validation, it predicted NNI with an RMSE of 0.14 and a MAPE of 12% in the southeastern Pampa. The red-edge spectrum and Local Incidence Angle of C-SAR were necessary to monitor the corn N status via prediction of NNI. Thus, this study provided empirical models to remotely sensed corn N status within fields during vegetative period, serving as a foundational data for guiding future N management.
Article
Full-text available
This study evaluates the accuracy of several recent remote sensing Surface Soil Moisture (SSM) products at sites in southwestern France. The products used are Soil Moisture Active Passive “SMAP” (level 3: 36 km × 36 km, level 3 enhanced: 9 km × 9 km, and Level 2 SMAP/Sentinel-1: 1 km × 1km), Advanced Scatterometer “ASCAT” (level 2 with three spatial resolution 25 km × 25 km, 12.5 km × 12.5 km, and 1 km × 1 km), Soil Moisture and Ocean Salinity “SMOS” (SMOS INRA-CESBIO “SMOS-IC”, SMOS Near-Real-Time “SMOS-NRT”, SMOS Centre Aval de Traitement des Données SMOS level 3 “SMOS-CATDS”, 25 km × 25 km) and Sentinel-1(S1) (25 km × 25 km, 9 km × 9 km, and 1 km × 1 km). The accuracy of SSM products was computed using in situ measurements of SSM observed at a depth of 5 cm. In situ measurements were obtained from the SMOSMANIA ThetaProbe (Time Domaine reflectometry) network (7 stations between 1 January 2016 and 30 June 2017) and additional field campaigns (near Montpellier city in France, between 1 January 2017 and 31 May 2017) in southwestern France. For our study sites, results showed that (i) the accuracy of the Level 2 SMAP/Sentinel-1 was lower than that of SMAP-36 km and SMAP-9 km; (ii) the SMAP-36 km and SMAP-9 km products provide more precise SSM estimates than SMOS products (SMOS-IC, SMOS-NRT, and SMOS-CATDS), mainly due to higher sensitivity of SMOS to RFI (Radio Frequency Interference) noise; and (iii) the accuracy of SMAP-36 km and SMAP-9 km products was similar to that of ASCAT (ASCAT-25 km, ASCAT-12.5 km and ASCAT-1 km) and S1 (S1-25 km, S1-9 km, and S1-1 km) products. The accuracy of SMAP, Sentinel-1 and ASCAT SSM products calculated using the average of statistics obtained on each site is defined by a bias of about −3.2 vol. %, RMSD (Root Mean Square Difference) about 7.6 vol. %, ubRMSD (unbiased Root Mean Square Difference)about 5.6 vol. %, and R coefficient about 0.57. For SMOS products, the station average bias, RMSD, ubRMSD, and R coefficient were about −10.6 vol. %, 12.7 vol. %, 5.9 vol. %, and 0.49, respectively.
Article
Full-text available
Accurate yield estimation at the field scale is essential for the development of precision agriculture management, whereas at the district level it can provide valuable information for supply chain management. In this paper, Huan Jing (HJ) satellite HJ1A/B and Landsat 8 Operational Land Imager (OLI) images were employed to retrieve leaf area index (LAI) and canopy cover (CC) in the Yangling area (Central China). These variables were then assimilated into two crop models, Aquacrop and simple algorithm for yield (SAFY), in order to compare their performances and practicalities. Due to the models' specificities and computational constraints, different assimilation methods were used. For SAFY, the ensemble Kalman filter (EnKF) was applied using LAI as the observed variable, while for Aquacrop, particle swarm optimization (PSO) was used, using canopy cover (CC). These techniques were applied and validated both at the field and at the district scale. In the field application, the lowest relative root-mean-square error (RRMSE) value of 18% was obtained using EnKF with SAFY. On a district scale, both methods were able to provide production estimates in agreement with data provided by the official statistical offices. From an operational point of view, SAFY with the EnKF method was more suitable than Aquacrop with PSO, in a data assimilation context.
Article
Full-text available
The objective of this work was to evaluate the sensitivity of three different satellite signals (interferometric coherence (γ), backscattering coefficient (σ0) and NDVI) to corn biophysical parameters (leaf area index, height, biomass and water content) throughout its entire vegetation cycle. All of the satellite and in situ data were collected during the Multi-spectral Crop Monitoring (MCM’10) experiment conducted in 2010 by the CESBIO Laboratory over eight different agricultural sites located in southwestern France. The results demonstrated that the NDVI is well adapted for leaf area index monitoring, whereas γ27.3° is much more suited to the estimation of the three other Biophysical Parameters throughout the entire crop cycle, with a coefficient of determination ranging from 0.83 to 0.99, using non-linear relationships. Moreover, contrary to the use of the NDVI or backscattering coefficients, the use of coherence exhibited a low sensitivity to the changes in vegetation and soil moisture occurring during senescence, offering interesting perspectives in the domain of applied remote sensing.
Article
Full-text available
Crop monitoring at a fine scale and crop yield estimation are critical from an environmental perspective because they provide essential information to combine increased food production and sustainable management of agricultural landscapes. The aim of this article is to estimate soybean yield using an agro-meteorological model controlled by optical and/or synthetic aperture radar (SAR) multipolarized satellite images. Satellite and ground data were collected over seven working farms. Optical and SAR images were acquired by Formosat-2, Spot-4, Spot-5, and Radarsat-2 satellites during the soybean vegetation cycle. A vegetation index (NDVI) was derived from the optical images, and backscattering coefficients and polarimetric indicators were computed from full quad-pol Radarsat-2 images. An angular normalization of SAR data was performed to minimize the incidence angle effects on SAR signals by using the complementarities provided by SAR and optical data. The best results are obtained when the model is controlled by both the leaf area index (LAI) derived from the optical vegetation index modified triangular vegetation index (MTVI2) or from the SAR backscattering coefficient σ° VV [Formula: see text] or ([Formula: see text]) and the dry biomass (DB) derived from the SAR Pauli matrix T33 [Formula: see text], demonstrating the complementary of optical and SAR data.
Article
Timely and accurate estimation of crop yield before harvest to allow crop yields management decision-making at a regional scale is crucial for national food policy and security assessments. Modeling dynamic change of crop growth is of great help because it allows researchers to determine crop management strategies for maximizing crop yield. Remote sensing is often used to provide information about important canopy state variables for crop models of large regions. Crop models and remote sensing techniques have been combined and applied in crop yield estimation on a regional scale or worldwide based on the simultaneous development of crop models and remote sensing. Many studies have proposed models for estimating canopy state variables and soil properties based on remote sensing data and assimilating these estimated canopy state variables into crop models. This paper, firstly, summarizes recent developments of crop models, remote sensing technology, and data assimilation methods. Secondly, it compares the advantages and disadvantages of different data assimilation methods (ca-libration method, forcing method, and updating method) for assimilating remote sensing data into crop models and analyzes the impacts of different error sources on the different parts of the data assimilation chain in detail. Finally, it provides some methods that can be used to reduce the different errors of data assimilation and presents further opportunities and development direction of data assimilation for future studies. This paper presents a detailed overview of the comparative introduction, latest developments and applications of crop models, remote sensing techniques, and data assimilation methods in the growth status monitoring and yield estimation of crops. In particular, it discusses the impacts of different error sources on the different portions of the data assimilation chain in detail and analyzes how to reduce the different errors of data assimilation chain. The literature shows that many new satellite sensors and valuable methods have been developed for the retrieval of canopy state variables and soil properties from remote sensing data for assimilating the retrieved variables into crop models. Additionally, new proposed or modified crop models have been reported for improving the simulated canopy state variables and soil properties of crop models. In short, the data assimilation of remote sensing and crop models have the potential to improve the estimation accuracy of canopy state variables, soil properties and yield based on these new technologies and methods in the future.
Article
This paper aims at estimating the yield of sunflower at field scale by testing different ways to control an agrometeorological model, using the leaf area index (LAI) and/or the dry mass (DM) of the crop. The LAI is derived from multitemporal satellite images acquired in the visible, near-infrared, and microwave wavelengths by five sensors (Formosat-2, Spot-4/5, TerraSAR-X, Radarsat-2), over a region located in southwestern France, while the DM is measured in situ. Results show that the temporal dynamics of LAI and DM can be accurately simulated over the five studied working farms (R DM 2 > 0.85 and R LAI 2 > 0.94 with a relative root-mean-square error (RMSE) <; 30%), when the model is controlled by both the DM (measured over only one field) and the LAI (estimated from the modified triangular vegetation index (MTVI2) or σ° C -VV ). The seed yield is then accurately estimated (R 2 > 0.85) over 140 ha, with low errors ranging from 0.20 to 0.54 t·ha -1 by using both the LAI derived from microwave and optical data (LAI MTVI2 and LAI σ°C-VV ), in order to reduce the overestimations related to the excessive development of the LAI during the first phenological stages, or to the presence of dew on the vegetation, respectively, affecting the LAI derived from the MTVI2 or σ° C -VV .
Article
Crop monitoring information is essential for food security and to improve our understanding of the role of agriculture on climate change, among others. Remotely sensing optical and radar data can help to map crop types and to estimate biophysical parameters, especially with the availability of an unprecedented amount of free Sentinel data within the Copernicus programme. These datasets, whose continuity is guaranteed up to decades, offer a unique opportunity to monitor crops systematically every 5 to 10 days. Before developing operational monitoring methods, it is important to understand the temporal variations of the remote sensing signal of different crop types in a given region. In this study, we analyse the temporal trajectory of remote sensing data for a variety of winter and summer crops that are widely cultivated in the world (wheat, rapeseed, maize, soybean and sunflower). The test region is in southwest France, where Sentinel-1 data have been acquired since 2014. Because Sentinel-2 data were not available for this study, optical satellites similar to Sentinel-2 are used, mainly to derive NDVI, for a comparison between the temporal behaviors with radar data. The SAR backscatter and NDVI temporal profiles of fields with varied management practices and environmental conditions are interpreted physically. Key findings from this analysis, leading to possible applications of Sentinel-1 data, with or without the conjunction of Sentinel-2, are then described. This study points out the interest of SAR data and particularly the VH/VV ratio, which is poorly documented in previous studies.
Article
The yield forecasting of corn constitutes a key issue in agricultural management, particularly in the context of demographic pressure and climate change. This study presents two methods to estimate yields using artificial neural networks: a diagnostic approach based on all the satellite data acquired throughout the agricultural season, and a real-time approach, where estimates are updated after each image was acquired in the microwave and optical domains (Formosat-2, Spot-4/5, TerraSAR-X, and Radarsat-2) throughout the crop cycle. The results are based on the Multispectral Crop Monitoring experimental campaign conducted by the CESBIO (Centre d’Études de la BIOsphère) laboratory in 2010 over an agricultural region in southwestern France. Among the tested sensor configurations (multi-frequency, multi-polarization or multi-source data), the best yield estimation performance (using the diagnostic approach) is obtained with reflectance acquired in the red wavelength region, with a coefficient of determination of 0.77 and an RMSE of 6.6 q ha⁻¹. In the real-time approach the combination of red reflectance and CHH backscattering coefficients provides the best compromise between the accuracy and earliness of the yield estimate (more than 3 months before the harvest), with an R² of 0.69 and an RMSE of 7.0 q ha⁻¹ during the development of the central stem. The two best yield estimates are similar in most cases (for more than 80% of the monitored fields), and the differences are related to discrepancies in the crop growth cycle and/or the consequences of pests.
Article
Crop simulation models are used at the field scale to estimate crop yield potential, optimize current management, and benchmark input-use efficiency. At issue is the ability of crop models to predict local and regional actual yield and total production without need of site-year specific calibration of internal parameters associated with fundamental physiological processes. In this study, a well-validated maize simulation model was used to estimate yield potential for 45 locations across the U.S. Corn Belt, including both irrigated and rainfed environments, during four years (2011-2014) that encompassed diverse weather conditions. Simulations were based on measured weather data, dominant soil properties, and key management practices at each location (including sowing date, hybrid maturity, and plant density). The same set of internal model parameters were used across all site-years. Simulated yields were upscaled from locations to larger spatial domains (county, agricultural district, state, and region), following a bottom-up approach based on a climate zone scheme and distribution of maize harvested area. Simulated yields were compared against actual yields reported at each spatial level, both in absolute terms as well as deviations from long-term averages. Similar comparisons were performed for total maize production, estimated as the product of simulated yields and official statistics on maize harvested area in each year. At county-level, the relationship between simulated and actual yield was better described by a curvilinear model, with decreasing agreement at higher yields (>12Mgha-1). Comparison of actual and simulated yield anomalies, as estimated from the yearly yield deviations from the long-term actual and simulated average yield, indicated a linear relationship at county-level. In both cases (absolute yields and yield anomalies comparisons), the agreement increased with increasing spatial aggregation (from county to region). An approach based on long-term actual and simulated yields and year-specific simulated yield allowed estimation of actual yield with a high degree of accuracy at county level (RMSE≤18%), even in years with highly favorable weather or severe drought. Estimates of total production, which are of greatest interest to buyers and sellers in the market, were also in close agreement with actual production (RMSE≤22%). The approach proposed here to estimate yield and production can complement other approaches that rely on surveys, field crop cuttings, and empirical statistical methods and serve as basis for in-season yield and production forecasts.