IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 47, NO. 2, FEBRUARY 2009 527
Wheat Crop Mapping by Using ASAR AP Data
Giuseppe Satalino, Francesco Mattia, Senior Member, IEEE, Thuy Le Toan, and Michele Rinaldi
Abstract—The purpose of this paper is to assess the use of
C-band HH/VV backscatter ratio for mapping winter wheat. This
paper analyzes two temporal series of images acquired in 2006
and 2007 by the Advanced Synthetic Aperture Radar (ASAR)
system in alternating polarization (AP) mode, over an agricultural
site located in southern Italy. Results on test data show that
classification accuracies between 75% and 80% can be achieved
by using a single ASAR image, acquired during the peak of
the wheat-growing season. To achieve accuracies close to 90%, a
spatial averaging at field scale is necessary.
Index Terms—Advanced Synthetic Aperture Radar (ASAR),
polarization ratio, SAR data classification, wheat mapping.
land cover classification (see  for a topical review). For
instance,  and  showed that, at C-band, the H- and
V-polarized radar signal is differentially attenuated by wheat
crops and that the attenuation increases with the incidence
angle, from 20◦to 40◦. The phenomenon is due to the pre-
dominant vertical structure of the wheat canopy, and it is
not observed for crops with a more branching structure or
with broader leaves, such as tomato or sugar beet. Advanced
Synthetic Aperture Radar (ASAR) data acquired in alternating
polarization (AP) mode are therefore well suited for developing
algorithms for wheat crop mapping. Despite this favorable con-
text, the interpretation of SAR images is not always unambigu-
ous because SAR data are sensitive not only to vegetation cover
but alsotosoilroughness, soilmoisture,field orientation, etc. In
addition, the presence of speckle noise limits the discrimination
among different targets. In this context, the objective of this
paper is to investigate the extension to which a simple rule-
based approach using the HH/VV backscatter ratio is able
to discriminate wheat from nonwheat classes. The data set
EVERAL past studies have demonstrated that C-band SAR
data hold an important potential for crop monitoring and
Manuscript received March 7, 2008; revised July 31, 2008. Current version
published January 28, 2009. This work was supported in part by the Italian
Ministry of Agriculture, Food and Forestry Policies (AQUATER Project, under
Contract 209/7303/05) and in part by ESA–ESTEC (ELASIM Project, under
G. Satalino and F. Mattia are with the Consiglio Nazionale delle Ricerche,
Istituto di Studi sui Sistemi Intelligenti per l’Automazione (CNR-ISSIA), CNR,
70126 Bari, Italy (e-mail: firstname.lastname@example.org; email@example.com).
T. Le Toan is with the Centre d’Etudes Spatiales de la Biosphère (CESBIO),
31401 Toulouse Cedex 9, France (e-mail: firstname.lastname@example.org).
M. Rinaldi is with the Consiglio per la Ricerca e Sperimentazione in
Agricoltura, Unità di Ricerca per i sistemi colturali degli ambienti caldo-aridi
(CRA-SCA), 70125 Bari, Italy (e-mail: email@example.com).
Digital Object Identifier 10.1109/TGRS.2008.2008026
ASAR AP DATA ACQUIRED IN 2006 AND 2007 OVER THE
AREA OF INTEREST (IMAGES LABELED WITH∗
WERE USED FOR TEMPORAL FILTERING)
consists of two temporal series of ASAR AP images acquired
in 2006 and 2007, over an agricultural area located close to the
town of Foggia (southern Italy). First, the temporal behavior
of the HH/VV backscatter ratio (hereafter referred to as γ)
estimated over wheat and nonwheat crops was investigated.
Then, the classification scheme was tested over the entire study
area. In the next section, the experimental data are described.
Then, the classification approach is illustrated, and the results
II. EXPERIMENTAL DATA AND ANALYSIS
The study area lies in the Capitanata plain, in the Puglia
region, southern Italy. The selected agricultural area, covering
approximately 736 km2, is mainly devoted to the cultiva-
tion of durum wheat (up to 50% of the area), characterized
by significant lower values of fresh/dry biomass and yield,
with respect to other wheat species. Other annual main crops
of the region are sugar beet and tomato. Permanent crops
with a significant presence in the area are grapes and olives.
Accurate and updated land use maps of the study area were
obtained by classifying multitemporal SPOT images acquired
in 2006 and 2007. In the framework of a three-year study
project , the area was periodically surveyed, and land cover
was monitored over a large number of training fields covering
an area of 9.4 km2. A temporal shift of a few weeks between
the peaks of the 2006 and 2007 wheat-growing seasons was
observed. The former was approximately in May, whereas
the latter fell in April. The values of fresh biomass, sam-
pled at peak stage, approximately ranged between 2.0 and
ASAR Alternating Polarization Mode Single Look Complex
0196-2892/$25.00 © 2009 IEEE
528 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 47, NO. 2, FEBRUARY 2009
training fields in 2006 and 2007. An error bar equal to the root mean square of
backscatter variability over the fields is also shown.
Temporal behavior of the HH/VV backscatter ratio estimated over the
OA OF WHEAT CLASSIFICATION FROM SAR DATA
(A) SPATIALLY AVERAGED OVER A 7 × 7 WINDOW AND
(B) TEMPORALLY FILTERED AND SPATIALLY AVERAGED
OVER A 5 × 5 WINDOW, RESPECTIVELY
between 37◦and 44◦(i.e., swaths I5–I7) were selected because
 showed that, at approximately 40◦of incidence, the HH/VV
backscatter is highly sensitive to wheat canopy and almost
insensitive to soil status (the slightly different incidence angle,
between images acquired at swaths I5 and I7, has a reduced
impact on the HH/VV backscatter and will be disregarded
in the following analysis). The ASAR APS products were
calibrated, coregistered, and geocoded. The pixel size of the
obtained images is 12.5 m, and the equivalent number of looks
is approximately one.
Fig. 1 shows the temporal behavior of γ estimated over the
wheat and nonwheat training fields, whereas the differences
between wheat and nonwheat γ values (i.e., Δγ), as a function
of day of the year (DoY), are reported in Table II. Both in
2006 and 2007, the temporal series of nonwheat γ values
show an almost constant level (below 0 dB). On the contrary,
the γ of wheat fields, in 2006, increases from March to May
and then decreases from May to June; in 2007, it steadily
decreases from April to June. In both years, the maximum
γ values for wheat fields are reached during the peak of
biomass-production growing stage (usually before heading).
Concerning the possibility of using the HH/VV backscatter
ratio not only for discriminating wheat from nonwheat classes
but also for retrieving wheat biomass, as suggested in , the
analysis of the entire data set collected over the Foggia site
uous line) corresponding pdf, referring to data acquired at DoY 136 (2006).
Data were averaged over a 7 × 7 window.
Histogram of (dotted and dashed lines) R coefficients and (contin-
showed that a linear relationship between γ and the wheat fresh
biomass holds for wheat biomass lower than 2.0–2.5 kg/m2.
Whereas for larger values of wheat biomass, the linear relation-
ficient during the first stages of wheat-growing season is always
shown by wheat crops, the phenological stage (or the value of
fresh biomass) at which γ starts to decrease may depend on
the particular wheat species (or on the wheat plant density). In
particular, for wheat species that can reach values of fresh bio-
mass as high as 6–7 kg/m2, the γ coefficient may start declining
well before the peak of biomass production. As a consequence,
γ can be considered as a robust feature to discriminate wheat
from nonwheat in the appropriate temporal window, but it may
not be reliable for retrieving wheat biomass during the entire
III. WHEAT CLASSIFICATION
Under the hypothesis of fully developed speckle, the ratio of
the n-look HH and VV sample estimate of the backscattering
coefficients (hereafter referred to as R) is a random variable
characterized by the probability density function (pdf) reported
in (1) 
(γ + R)2n.
For n > 2, the mean and the variance of R are 2γn/(2n − 1)
and γ2n(2n − 1)/((n − 2)(n − 1)2), respectively. As an ex-
ample, Fig. 2 shows the histogram and the corresponding pdf
of R for the two classes of wheat and nonwheat estimated from
ASAR data, acquired on DoY 136 of 2006 and averaged over
a 7 × 7 (pixels) window. Both the histogram and the pdf are
weighed for the a priori probabilities of the classes. Based on
the two estimated pdfs, the optimal threshold separating the two
classes can be computed by applying the Bayesian decision
rule that minimizes the classification errors . As a result,
the optimal threshold is identified as the intersection between
the pdf of the two classes (Fig. 2). For each acquisition date,
a slightly different threshold, ranging from 0.97 to 1.17, was
SATALINO et al.: WHEAT CROP MAPPING BY USING ASAR AP DATA 529
CONFUSION MATRIX OBTAINED FOR THE CLASSIFICATION OF
WHEAT AND NONWHEAT CLASSES, COMPUTED OVER THE
TEST AREA FROM TEMPORALLY AND SPATIALLY
FILTERED DATA AT DoY 136, YEAR 2006
obtained. The classification error decreases with the increase of
the distance Δγ and with the decrease of the associated stan-
dard deviations (or, equivalently, with the number of looks n).
For instance, when Δγ equals 2 dB, a probability of classifi-
cation error less than or equal to 15% can be obtained with a
number of looks greater than 40 . To obtain such a number
of looks, two filtering approaches were implemented. The first
solely consists in a spatial filtering and uses a single ASAR
AP image on which a spatial averaging over a 7 × 7 window
is performed. The second is based on a combined temporal
and spatial filtering. It exploits a stack of seven ASAR AP
images (see Table I) and, according to [8, eq. 22], estimates
the coefficients of the temporal filtering over a 5 × 5 window.
In addition to the temporal filtering, a spatial averaging over a
5 × 5 window is also applied.
Table II reports the overall accuracy (OA) of the wheat
classification from SAR data acquired in 2006 and 2007 over
the testing area during the peak growing stage. For both years,
an area of roughly 560 km2, containing a percentage of wheat
fields of approximately 60%, was employed. Results refer
to SAR data to which (A) solely a spatial and (B) a com-
bined temporal and spatial filtering were applied, respectively.
In both cases, the OA values estimated on test data are signif-
icantly lower compared to the expected accuracies estimated
on training fields (i.e., 75% against 85%). This is mostly due
to the fact that the test area includes a larger number of wheat
fields at different phenological stages. Table III shows the con-
fusion matrix (expressed in %) obtained for the classification
of wheat and nonwheat classes (case B, SAR data temporally
and spatially filtered at DoY 136, 2006). The OA is 77.6%,
whereas the percentages of wheat classified as nonwheat and
nonwheat classified as wheat are 14.4% and 8.0%, respec-
tively. The results reported in Table II were obtained using
the optimal thresholds for each date. However, the robustness
of the method was tested by simulating the unavailability of
training fields. For instance, by applying the threshold of 1.17
(detected on the data acquired in May 2006) to the data
acquired in April 2007, the results deteriorate at the most
Moreover, a number of tests were carried out by filtering
the SAR data with larger windows (e.g., 9 × 9, 11 × 11, and
13 × 13 pixels). As a result, it was observed that, although
the OA increases (particularly when the combined temporal
and spatial filtering is used), it never exceeds 80%. Conversely,
by performing solely a spatial filtering at field scale, which
maximizes the number of looks while avoiding average pixels
belonging to nonhomogeneous areas, it is feasible to obtain in
the test OA values approximately equal to 90%. Of course, the
possibility of applying this method relies on the availability of
a priori information for the field boundaries, for instance as a
layer of a Geographical Information System.
In this paper, a physically based method for mapping winter
wheat using ASAR AP data, acquired at HH and VV polar-
izations and at high incidence angles (i.e., between 37◦and
44◦), was assessed. This paper analyzed two temporal series
of ASAR AP images acquired in 2006 and 2007 over an
The wheat classification was obtained by applying an optimal
threshold to the HH/VV-copolarized backscatter ratio acquired
during the peak growing stage. Classification accuracies on test
data ranging between 75% and 80%, depending on the amount
of spatial and temporal filtering performed, were obtained. In
order to attain classification accuracies approximately equal
to 90%, the spatial averaging needs to be carried out at field
scale. The obtained accuracies were not critically sensitive to
the adopted threshold or to the specific acquisition date during
the peak growing stage.
One limitation of this paper is that the analyzed data refer
to an agricultural site mainly devoted to wheat cultivation and,
more specifically, to the durum wheat. As a consequence, the
species remains to be assessed.
ASAR and SPOT data were supplied in the framework of
ENVISAT AO 662 and CNES (2006) Distribution Spot Image
S.A./OASIS program, respectively.
 H. McNairn and B. Brisco, “The application of C-band polarimetric SAR
for agriculture: A review,” Can. J. Remote Sens., vol. 30, no. 3, pp. 525–
542, Jun. 2004.
 S. C. M. Brown, S. Quegan, K. Morrison, J. C. Bennett, and
wheat canopies—Implications for crop parameter retrieval,” IEEE
Trans. Geosci. Remote Sens., vol. 41, no. 7, pp. 1602–1610, Jul. 2003.
 F. Mattia, T. Le Toan, G. Picard, F. Posa, A. D’Alessio, C. Notarnicola,
A. M. Gatti, M. Rinaldi, G. Satalino, and G. Pasquariello, “Multitemporal
C-band radar measurements on wheat fields,” IEEE Trans. Geosci. Remote
Sens., vol. 41, no. 7, pp. 1551–1560, Jul. 2003.
 M. Rinaldi, P. Campi, M. Castellini, R. M. Ferrara, C. Fiorentino,
P. Garofalo, and G. Satalino, “AQUATER project: The measurement cam-
paign in Capitanata plain of soil–plant–atmosphere continuum,” in Proc.
5th Int. Symp. Retrieval Bio- Geophys. Parameters SAR Data Land Appl.,
Bari, Italy, Sep. 25–28, 2007.
 C. Oliver and S. Quegan, Understanding Synthetic Aperture Radar Images.
Norwood, MA: Artech House, 1998.
 O. R. Duda and P. E. Hart, Pattern Classification, 2nd ed. Hoboken, NJ:
 E. J. M. Rignot and J. van Zyl, “Change detection techniques for ERS-1
SAR data,” IEEE Trans. Geosci. Remote Sens., vol. 31, no. 4, pp. 896–906,
 S. Quegan and J. J. Yu, “Filtering of multichannel SAR images,”
IEEE Trans. Geosci. Remote Sens., vol. 39, no. 11, pp. 2372–2379,
measurements ofscattering in
530 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 47, NO. 2, FEBRUARY 2009 Download full-text
Giuseppe Satalino received the Laurea degree
(cum laude) in computer science from the University
of Bari, Bari, Italy, in 1991.
He was a Summer Student with the European
Switzerland, in 1991, where he worked on appli-
cations of neural networks to high-energy physics.
From 1993 to 1996, he was Grant Holder of Alenia
and Consiglio Nazionale delle Ricerche (CNR).
Since 1996, he has been with the Institute of Intel-
ligent Systems for Automation (ISSIA), CNR, Bari.
He worked in several national and international research projects concerning
image processing, data classification, and remote sensing applications. He
participated in conducting several SAR and ground radar experiments for
studies concerning the use of remote sensing for agricultural and hydrologic
applications. His main research field includes data classification techniques
and methods for the retrieval of geophysical parameters from SAR and optical
Francesco Mattia (M’99–SM’08) received the Lau-
rea degree in physics and the M.S. degree in signal
processing from the University of Bari, Bari, Italy,
in 1990 and 1994, respectively, and the Ph.D. degree
from the University Paul Sabatier, Toulouse, France,
From 1991 to 1994, he was a grant holder of the
Italian National Council of Research (CNR) and the
European Commission at the Institute for Remote
Sensing Applications of the Joint Research Centre,
Ispra, Italy. Since 1995, he has been with CNR
where he is currently Senior Research Scientist with the Institute of Intel-
ligent Systems for Automation (ISSIA), CNR, Bari. During 1996–1999, he
was a Visiting Scientist with the Centre d’Etudes Spatiales de la Biosphère
(CESBIO), Toulouse. In 2007, he was the Coorganizer of the 5th International
Symposium on Retrieval of Bio- and Geophysical Parameters from SAR Data
for Land Applications, held in Bari. His scientific interests include the direct
and inverse modeling of microwave scattering from natural surfaces and the use
of information derived from Earth observation sensors to improve land surface
process models (e.g., hydrologic or crop growth models).
Thuy Le Toan, photograph and biography not available at the time of
Michele Rinaldi received the Laurea degree
(summa cum laude) in agricultural science from the
University of Bari, Bari, Italy.
Since 1988, he has been a Researcher with the
Agricultural Research Council—Research Unit for
cropping systems in dry environments in Bari. He
attended several courses about classical statistic and
technical software, experimental design applied to
biology, precision farming, and simulation models
(CropSyst and DSSAT). His main research interests
include crop simulation models, cropping systems,
and crop growth.Hecoordinatesa nationalproject(AQUATER) aboutcropping
scale. He is responsible for several national research projects about water and
cropping systems and cooperates with foreign partners in European projects.
He is the author of more than 150 scientific publications. He is in the lists of
experts of the Ministry of Agriculture and the Ministry of University.
Mr. Rinaldi was awarded an OECD Fellowship (2002) “Evaluating
Simulation Modules for the Assessment of Crop Emergence Under Extreme/
Variable Climate and Soil Conditions,” carried out in Rothamsted Research,
Hertfordshire, U.K. He is a member of the Italian Society of Agronomy and the
Italian Society of Agrometeorology.