IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 58, NO.

9, SEPTEMBER 2020 6321

A Radar Vegetation Index for Crop Monitoring

Using Compact Polarimetric SAR Data
Dipankar Mandal , Student Member, IEEE, Debanshu Ratha , Student Member, IEEE,
Avik Bhattacharya , Senior Member, IEEE, Vineet Kumar , Member, IEEE, Heather McNairn ,
Yalamanchili S. Rao, Member, IEEE, and Alejandro C. Frery , Senior Member, IEEE

Abstract— Crop growth monitoring using compact-pol syn- I. I NTRODUCTION

thetic aperture radar (CP-SAR) data is gaining attention with
the rapid advancements toward operational applications. In this
article, we propose a vegetation index for compact polarimetric
(CP) SAR data [compact-pol radar vegetation index (CpRVI)].
P OLARIMETRIC SAR data have provided exceptional
opportunities to track temporal changes in vegetation
conditions and to map crop characteristics over large spatial
The CpRVI is derived using the concept of a geodesic distance
between the Kennaugh matrices projected on a unit sphere. This extents [1], [2]. With the advent of full-pol SAR sensors,
distance is utilized to compute a similarity measure between the it is now possible to monitor crop growth by relating the
observed Kennaugh matrix and the Kennaugh matrix of an ideal physical scattering mechanisms extracted from the data with
depolarizer (a realization of vegetation canopy). The similarity crop phenology [3]. Several studies have reported the use
measure is then modulated with a scaled quantity derived from of backscatter intensity ratios and polarimetric decomposition
the scattering power ratio of the same and opposite sense
polarization with respect to the transmitted circular polarization. parameters as a proxy for crop growth [4]–[7].
In this article, we utilize time-series-simulated RADARSAT Similar to the spectral indices that are well established in
Constellation Mission (RCM) compact-pol SAR data (RH-RV) optical remote sensing, a vegetation index derived from SAR
obtained from the full-pol RADARSAT-2 observations during the data could be an alternative for crop growth monitoring. In this
soil moisture active passive (SMAP) validation experiment 2016 direction, Kim and van Zyl [8] made the first attempt by
(SMAPVEX16-MB) campaign in Manitoba, Canada, to assess the
proposed vegetation index. Among the various crops grown in introducing the radar vegetation index (RVI) using a measure
this region, in particular, we analyze the growth stages of wheat of scattering randomness from vegetation targets. The RVI was
and soybean due to their different canopy structures. A temporal formulated by modeling the vegetation canopy as a collection
analysis of the proposed CpRVI with crop biophysical parameters of randomly oriented dipoles. Several studies have utilized the
[the plant area index (PAI) and vegetation water content (VWC)] RVI for crop growth monitoring and biophysical parameter
at different phenological stages confirms the trend of CpRVI with
the plant growth. Nevertheless, variations of CpRVI values are estimation [9]–[11].
apparent with different plant densities for both the crop types. Even though full polarimetric SAR sensors offer unique
Also, the linear regression analysis confirms that the CpRVI performances in terms of the retrieval of rich quantitative
values significantly correlate with PAI (r = 0.72 and 0.85) and information, the advantages of compact polarimetric (CP)
VWC (r = 0.62 and 0.75) for both wheat and soybean. We SAR [12] systems for several Earth and planetary appli-
observed good retrieval of PAI and VWC for both wheat and
soybean. cations are gaining attention [13]–[17]. Compact-pol syn-
thetic aperture radar (CP-SAR) modes have advantages
Index Terms— Compact-pol, crop, geodesic distance (GD), over full-pol measurements in terms of larger swath width
RADARSAT Constellation Mission (RCM), vegetation index.
and less data volume at the expense of less polarimetric
Manuscript received May 16, 2019; revised November 3, 2019; accepted information.
February 22, 2020. Date of publication March 16, 2020; date of current version In the literature, the utility of CP-SAR data is
August 28, 2020. This work was supported in part by the Shastri Indo-Candian
Institute, New Delhi, India. The work of Debanshu Ratha was supported by the explored by comparing its performance with dual- and
Council of Scientific and Industrial Research (CSIR). (Corresponding author: full-pol SAR data for the analysis of different land
Dipankar Mandal.) features [16], [18]–[25]. In particular, the potential of CP-SAR
Dipankar Mandal, Debanshu Ratha, Avik Bhattacharya, Vineet Kumar, and
Yalamanchili S. Rao are with the Microwave Remote Sensing Laboratory, data has also been investigated for agricultural applications,
Centre of Studies in Resources Engineering, IIT Bombay, Mumbai 400076, such as crop growth monitoring [26]–[31], biophysical para-
India (e-mail: [email protected]; [email protected]; meter retrieval [32], [33], and soil moisture retrieval [34]–[37],
[email protected]; [email protected]; [email protected]).
Heather McNairn is with the Agriculture and Agri-Food Canada, Ottawa, using real and simulated CP-SAR data from full-pol
ON K1A0C6, Canada (e-mail: [email protected]). observations.
Alejandro C. Frery is with the Laboratório de Computação Cientí- The potential of CP-SAR modes might hold significant
fica e Análise Numérica (LaCCAN), Universidade Federal de Alagoas,
Maceió 57072-900, Brazil, and with the Key Laboratory of Intelligent interest from an operational perspective for agricultural appli-
Perception and Image Understanding of the Ministry of Education, Xidian cations with planned time-series data sets. These could be
University, Xi’an 710071, China (e-mail: [email protected]). globally obtained from multiple SAR satellites considering
Color versions of one or more of the figures in this article are available
online at http://ieeexplore.ieee.org. the rapid expansion of constellations of satellites, such as
Digital Object Identifier 10.1109/TGRS.2020.2976661 the Canadian RADARSAT Constellation Mission (RCM),
0196-2892 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
6322 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 58, NO. 9, SEPTEMBER 2020

SAOCOM (TOPSAR with experimental CP-mode), and the the study area and data sets. Section IV discusses the results
NASA-ISRO SAR (NISAR) L- and S-band mission. in detail by relating the CpRVI with two biophysical parame-
In a recent study, McNairn et al. [38] showed that the ters: the plant area index (PAI) and vegetation water content
simulated CP parameters correlate with the normalized differ- (VWC) for wheat and soybean. Finally, the work is succinctly
ence vegetation index (NDVI) derived from optical sensors. summarized and concluded in Section V.
Homanouyi et al. [39] assessed the volume to odd-bounce
scattering power ratio (Pv /Ps ) derived from a scattering power II. M ETHODOLOGY
decomposition method from the simulated CP-SAR data for
A. Kennaugh Matrix
vegetation condition monitoring. The Pv /Ps ratio shows a
good correlation with NDVI and crop biomass for several In PolSAR, the 2 × 2 complex scattering matrix S encom-
crop types. However, as the crop canopy attained its peak passes complete polarimetric information about backscattering
growth during the end of the vegetative stage and fruit devel- from targets
opment, the magnitude of Pv /Ps was overestimated relative  
SHH SHV
to that obtained from full-pol data. Nevertheless, these studies S= (1)
SVH SVV
confirm that a derived CP-SAR metric similar to NDVI is
essential for agricultural applications. Even though the RVI is where the subscripts H and V denote the horizontal and
a good alternative, it is restricted to the use of full-polarimetric vertical polarizations, respectively. Following the reciprocity
SAR data. Thus, there is a need for an RVI for CP-SAR data. theorem in the case of a monostatic radar, SHV = SVH .
It is noteworthy that CP-SAR mode will become an important A suitable representation of PolSAR data in terms of power
mode for the operational monitoring of Canada’s agricultural is given by the 4 × 4 real Kennaugh matrix K defined as
extent. ⎡ ⎤
1 0 0 1
In this article, we have proposed a new vegetation index, 1   ⎢1 0 0 −1 ⎥
K = A∗ S ⊗ S∗ A−1 , A = ⎢ ⎣
⎥ (2)
namely, Compact-pol RVI (CpRVI), for CP-SAR data using 2 0 1 1 0 ⎦
the concept of a geodesic distance (GD) between two 0 j −j 0
Kennaugh matrices projected on unit sphere [40], [41]. In √
recent work, the authors have introduced a novel vegetation where ⊗ is the Kronecker product, and j = −1. The
index [42] for full-pol SAR data using the generalized volume Kennaugh matrix for the incoherent case can also be obtained
scattering model (GVSM) proposed by Antropov et al. [43]. from the elements of the coherency matrix T in the following
This model is a function of two parameters that correspond to manner in (3), as shown at the bottom of this page, where Ti j
the copolarized ratio and cross-polarized correlation coefficient is the (i, j )th entry of T, and  and  denote the real and
obtained from full-pol SAR data. However, for CP-SAR data, imaginary parts of a complex number.
we have utilized the ideal depolarizer (ID) as a volume model Thus, the coherent elementary targets represented by S and
that converts all Stokes vectors into a randomly polarized the distributed targets usually represented by T (or the covari-
wave [44]. We compute a similarity measure between the ance matrix C) can be suitably transformed into their K matrix
observed Kennaugh matrix and a Kennaugh matrix corre- representation using (2) and (3), respectively. Table I presents
sponding to the ID. The formulation of CpRVI is direct the Kennaugh matrix forms of some well-known elementary
and intuitive, as well as easy to implement. We tested the scatterers and the ID from PolSAR literature. These include
CpRVI with simulated RCM CP-SAR mode data from the symmetric scatterers such as trihedral (t), cylinder (c), dipole
full-pol RADARSAT-2 observations over the Joint Experiment (dp), quarter waves (±1/4), narrow dihedral (nd), dihedral (d),
for Crop Assessment and Monitoring (JECAM) test site in asymmetric scatterers, such as left and right helices (lh and rh),
Carman (Manitoba), Canada. Also, we perform comparative and ID.
analysis for the full crop season between CpRVI and full-pol It is noteworthy that different scatterers pertain to the
vegetation indices. description of features of interest in a PolSAR image. Thus,
The rest of the manuscript is organized in the following using a suitable distance measure to compare observations
order. Section II describes the formulation of the Kennaugh with the scattering models provides a good starting point
matrix for CP-SAR data, the similarity measure using GD, for several applications. From this perspective, GD to com-
and the formulation of CpRVI. Section III briefly describes pare two Kennaugh matrices was first proposed in [40] and,

⎡ ⎤
T11 + T22 + T33
⎢ (T12 ) (T13 ) (T23 ) ⎥
⎢ 2 ⎥
⎢ T11 + T22 − T33 ⎥
⎢ (T12 ) (T23 ) (T13 ) ⎥
K=⎢
⎢
2
T11 − T22 + T33
⎥
⎥
⎢ (T13 ) (T23 ) −(T12 ) ⎥
⎢ 2 ⎥
⎣ −T11 + T22 + T33 ⎦
(T23 ) (T13 ) −(T12 )
2
(3)

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
MANDAL et al.: RVI FOR CROP MONITORING USING CP-SAR DATA 6323

TABLE I scattering model to study backscattering from vegetation. Most

K ENNAUGH M ATRICES FOR E LEMENTARY TARGETS AND THE ID of the volume models in the PolSAR literature follow this
generalized volume model as a particular case by fixing its
two parameters: γ and ρ. The parameters γ and ρ correspond
to the copolarized channel ratio and the copolarized channel
correlation coefficient, respectively.
Ratha et al. [42] synergistically combine GD of the observa-
tion with the generalized volume model, as well as GDs from
even and odd bounce scatterers to propose an RVI [generalized
RVI (GRVI)] for full-pol SAR data. The GRVI is compared
with RVI proposed by Kim and Van Zyl [8] in terms of its
correlation with biophysical parameters, such as the PAI and
volumetric water content. We obtain GD-based indices for full
since then, is used as an adaptable measure for several polarimetric SAR data represented by the Kennaugh matrices
applications [40]–[42], [45]–[47]. without any assumption on the polarization mode, so they can
also be utilized for compact-pol SAR data.
B. GD and Its Applications
GD is motivated by the task of measuring the distance over C. Compact-Pol RVI
a manifold (i.e., higher dimensional surface) intrinsically with- The hybrid CP mode measures a projection of the 2 × 2
out leaving the surface along the shortest path (or geodesic). complex scattering matrix S as
For example, the unit sphere is one of the simplest manifolds     
that are often encountered in several applications. E CH 1 SHH SHV 1
= √
By definition, a great circle is the circle of radius equal to E CV 2  SVH SVV  ±i
that of the sphere, and the shortest path between two points on 1 SHH ± i SHV
= √ (5)
the unit sphere is along the great circle passing through the two 2 SVH ± i SVV
points. The distance measured along this circle is GD on the
unit sphere. In general, this notion can be easily generalized where the subscript C can be either the left-hand circular
to a higher dimensional real unit sphere. (LHC) transmit with a + sign or the right-hand circular (RHC)
In the context of PolSAR data, the space of 4 × 4 real matri- transmit with a − sign. The 2 × 2 covariance matrix is then
ces where all the Kennaugh matrices reside forms a convenient obtained from the elements of the scattering vector as
space to study GD on the unit sphere centered at the origin.  ∗

|E CH |2  E CH E CV 
Ratha et al. [40] proposed GD for Kennaugh matrices on this [C] = ∗ . (6)
E CV E CH  |E CV |2 
unit sphere. GD between two arbitrary Kennaugh matrices (K1
and K2 ) is defined as For CP-SAR data, the 4 × 1 Stokes vector g can be written
  in terms of the elements of the 2 × 2 covariance matrix C as
2 −1 Tr K1T K2 ⎡ ⎤ ⎡ ⎤
GD(K1 , K2 ) = cos     (4) g0 C11 + C22
π Tr K T K Tr K T K ⎢ g1 ⎥ ⎢ C11 − C22 ⎥
1 2
g=⎢ ⎥ ⎢ ⎥
1 2
⎣ g2 ⎦ = ⎣ C12 + C21 ⎦ (7)
where Tr is the trace operator, and the superscript T
denotes transpose. The 2/π factor is used to normalize g3 ± j (C12 − C21 )
its range to [0, 1]. GD is adaptively utilized in several where ± corresponds to left and right circular polarization
applications [40]–[42], [45]–[48] for full-polarimetric SAR respectively. The Kennaugh matrix (KCP ) for CP-data can be
data. The versatility of GD is due to its elegant properties, such written in terms of the elements of the Stokes vector as
as boundedness, scale invariance, and invariance under the ⎡ ⎤
g0 0 g2 /2 0
orthogonal transformation of the polarization basis. A detailed ⎢ 0 0 0 g1 ⎥
treatment of GD with rigorous mathematical derivations of KCP = ⎢⎣ g2 /2 0
⎥. (8)
0 0 ⎦
its useful properties and new roll invariant parameters using
0 g1 0 g3 /2
elementary models are detailed in [48].
Using specific scattering models in GD for particular appli- In the absence of any volume scattering models to suit-
cations, Ratha et al. [45] proposed a novel radar built-up index ably describe scattering from vegetation in CP-SAR data, ID
(RBUI) that utilizes GD for observations from even bounce is utilized to measure randomness from distributed targets.
scatterers (nd, d) and helices (lh, rh) for extracting built-up Moreover, GDs from the odd (i.e., trihedral) and even bounce
area from PolSAR data. However, the standard elementary (i.e., dihedral) scatterers that appear in GRVI [42] are suitably
models do not approximate well the scattering from vegetation. replaced by the received echo powers from the same circular
Hence, the concept of working with distributed targets in the (SC) and the opposite circular (OC) sense to the transmitted
form of volume scattering models for vegetation monitoring polarization.
is prevalent in PolSAR literature. In this respect, Antropov In this light, we propose a novel CpRVI (9) by utilizing
et al. [43] proposed a generalized two-parameter volume the measure of similarity f ID = 1 − (3/2)GDID between a

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
6324 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 58, NO. 9, SEPTEMBER 2020

Kennaugh matrix for an ID KID and the observed Kennaugh

matrix K using GD: GDID = GD(K, KID ).
The quantity 2/3 is the maximum GD between an ID (KID )
and an observed Kennaugh matrix K. Moreover, the similarity
measure f ID is zero for any other elementary target, and
therefore, CpRVI = 0.
The parameter β in (10) is proportional to the ratio of the
minimum to the maximum backscattering power in the same
(SC) and the opposite (OC) sense to the transmitted circular
polarization

CpRVI = β f ID , 0 ≤ CpRVI ≤ 1 (9)


p 2 (3/2 GDID )
β = (10)
q
p = min{SC, OC}, q = max{SC, OC} (11)
g0 − g3 g0 + g3
SC = , OC = (12)
2 2
Fig. 1. Dependence of CpRVI on f ID and β.
where g0 (the first element of the Stokes vector) is a measure
of the total average received power and g3 (the fourth element
of the Stokes vector) is a measure of the average received III. S TUDY A REA AND DATA S ET
power in circular polarization. This article was conducted over the JECAM test site in
The proportion of the power that is received by the radar Carman, Manitoba (Canada), shown in Fig. 2. The site covers
in opposite-sense circular polarization to that transmitted is an area of 26 km × 48 km and is characterized by various
(g0 + g3 )/2. For example, OC = (g0 + g3 )/2 = |E R |2 for crops and soil conditions. The major crops grown in this test
left-circular (L) polarization on transmit, where |E R |2 is the site are wheat, soybean, canola, and corn. According to the
intensity of the right-circular component. Such a change of 2016 Province Report [49], the seeding of annual crops in
polarity occurs when an electromagnetic (EM) wave undergoes Manitoba started at the end of April to mid-May with harvest-
an odd number of reflections from a target. ing commencing in August. The in situ measurements were
Similarly, (g0 − g3 )/2 is a measure of the power received collected over the area with near coincident satellite passes as
by the radar in the same-sense circular polarization as it was a part of the Soil Moisture Active Passive (SMAP) Validation
transmitted, which represents an even number of reflections Experiment 2016 (SMAPVEX16-MB) campaign [50].
from a target. For example, SC = (g0 − g3 )/2 = |E L |2 for
left-circular (L) polarization on transmit, where |E L |2 is the
intensity of the left-circular component. A. In Situ Sampling Strategy
The role of β is similar to that of the damping factor used During the campaign, in situ measurements of crops and soil
in GRVI. It ensures the imbalance between the SC and OC were collected in two separate periods (8 June to 22 June, and
returns, which is identical to the case for even and odd bounce 9 July to 22 July 2016) over 50 agricultural fields. The nominal
returns for GRVI. For purer targets with either an odd or an size of each field is around 800 m × 800 m. In each field, soil
even bounce-dominant scattering mechanism, there will be a moisture was measured at sixteen sampling points, arranged
substantial imbalance between the SC and OC returns. This in two parallel transects along the row direction (see Fig. 2).
imbalance leads to a smaller value of β, thereby resulting in Sampling points were separated by approximately 75 m, and
a lower value of CpRVI. While for random scattering media each transect was separated by 200 m. Soil moisture was
(namely, vegetation), which do not have an overall preference measured at each point using hydra probes.
for SC or OC returns, the value of β is high (i.e., close to 1), The vegetation was sampled in each field at three points
which leads to a higher value of CpRVI. (i.e., points 2, 11, and 14 in the first week, and 3, 10,
Theoretically, for a smooth bare surface, the similarity and 13 in the second week). Vegetation sampling included
measure, f ID ≈ 0, between the observed Kennaugh matrix measurement of PAI, wet biomass, plant height, density, and
K and KID , hence, for this case, β 1, as SC OC. phenology through destructive and nondestructive methods.
This scenario also indicates a lower circular polarization ratio The SMAPVEX16-MB report [50] and [51] provides detailed
(CPR) with CpRVI ≈ 0. In the case of a random volume descriptions of vegetation and soil sampling strategies.
scattering (with little to no polarization structure), f ID ≈ 1
and β ≈ 1. The SC and OC components would then be
B. SAR Data and Preprocessing
similar (equal to g0 /2) for such random volume cases with
CpRVI ≈ 1. Fig. 1 shows the variation of f ID versus β with Several RADARSAT-2 images were acquired during the
CpRVI. campaign in Fine Wide quad-pol mode (FQW); see Table II

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
MANDAL et al.: RVI FOR CROP MONITORING USING CP-SAR DATA 6325

TABLE II
S PECIFICATION OF RADARSAT-2 D ATA A CQUISITIONS
U SED IN T HIS A RTICLE

Research over Canadian JECAM test sites has confirmed that

despite its higher NESZ, simulated RCM CP-SAR data can
accurately classify crops [54], [55]. As the NESZ of RCM
is higher than the RADARSAT-2 system, there could be a
decreased sensitivity to low backscatter values. Hence, care
should be taken when deriving the polarimetric parameters
from simulated RCM products.
In the simulation process, the complex scattering vector
in RH and RV polarizations are then calculated from the
SHH , SHV , SVH , and SVV elements, as given in (5). Here,
Fig. 2. Study area and sampling locations over the JECAM-Carman (Canada) it is important to note that the ellipticity, χ = −45◦ , and
test site. The layout of the 16 sampling locations within each field is orientation angle ψ = 0◦ are considered constant values in
highlighted. In each field, three points (i.e., points {2, 11, 14} or {3, 10, 13}) the RCM simulator to generate a perfect circular transmit
were selected for vegetation sampling.
condition. Based on the RCM product specifications, the RH
and RV imagery are then downsampled using a 2-D fast
Fourier transformation technique to obtain the spatial resolu-
for details of the satellite data used in this article. The selection tion defined for the specific beam mode. The elements of the
of acquisition dates was based on in situ measurement periods. scattering vector in RH and RV polarizations are then stored
Several polarization combinations at varying spatial resolu- in a product file along with a metadata information file as an
tions and noise floors (ranging from −25 to −17 dB) would RCM simulator output for each acquisition date (see Table II).
be available for RCM [52] that includes single polarization These simulated outputs are utilized further to generate the
(i.e., HH, VV, HV, and VH), dual-pol (i.e., HH-VV, HH-HV, CpRVI. First, the 2 × 2 covariance matrix is obtained from
and VV-VH), full-pol (i.e., HH-HV-VH-VV), and compact-pol the elements of the scattering vector, as given in (6). During
(i.e., RH-RV) modes. the [C] generation, a 2 × 2 multilook is applied in the
In this article, we simulated the compact-pol (RH-RV) data range and azimuth directions. The Stokes vectors are then
from full-pol RADARSAT-2 data using the RCM Compact used to calculate the Kennaugh matrix K. Finally, CpRVI
Polarimetry Simulator developed at the Canada Centre for images are obtained from the derived K for the simulated
Remote Sensing (CCRS) [16], [53]. The simulator produces CP-SAR data. To reduce the effect of speckle, the CpRVI
the simulated SAR products based on the prelaunch speci- is calculated over a 5 × 5 averaging window. The CpRVI
fications of RCM provided by the Canadian Space Agency. images are coregistered using ground control points (GCPs)
Specifications such as the spatial resolution and noise equiva- with an RMSE ≤ 0.33 m and georeferenced. The sampling
lent sigma zero (NESZ) of each mode are considered during points are overlayed on the CpRVI images, and CpRVI values
the simulation process. The RCM simulator uses full-pol for each site (point measurement) are extracted from the
RADARSAT-2 images as input to simulate RCM data products coregistered CpRVI images over wheat and soybean fields with
of any of the RCM modes. 3 × 3 window.
The RCM simulator first converts the RADARSAT-2 16-bit Also, a comparative analysis is performed for the full crop
complex products (SHH , SHV , SVH , and SVV ) to 32-bit float season between CpRVI and full-pol vegetation indices. The
complex by applying the calibration coefficients specified for full-pol indices, namely, GRVI, and RVI are obtained from the
a particular product of RCM. Random noise is added to the seven full-pol RADARSAT-2 images (see Table II). The full-
complex imagery. In this article, the compact-pol (RH-RV) polarimetric C-band RADARSAT-2 images are multilooked by
data are simulated with −19-dB NESZ for the high-resolution 2 × 2 (in the range and azimuth) to generate the coherency
(5-m pixel size) imaging mode of RCM [52]. It is noteworthy matrix T. The elements of the T matrix are then used to
that CP-SAR mode will become an important mode for calculate the Kennaugh matrix K, as discussed in Section II-A.
the operational monitoring of Canada’s agricultural extent. The GRVI images are generated from the derived K for each

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
6326 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 58, NO. 9, SEPTEMBER 2020

Fig. 3. Temporal pattern of CpRVI for several wheat fields at different crop growth stages. Wheat fields are selected based on variations in PD (number of
plants per m2 area). The in situ measurements of PAI (m2 m−2 ) and VWC (kg m−2 ) are plotted in secondary axis for each field.

acquisition over 5 × 5 moving overlapping window [42]. measurements are compared among three groups of which are
In a separate process flow, the RVI [8] images are obtained separated based on plant density (PD: number of plants per
from the eigenvalue spectrum obtained from the T matrix [56]. square meter). The growth trends are similar irrespective of
Both the GRVI and RVI images are then coregistered using PD among several fields, i.e., an increase of PAI and VWC
GCPs with an RMSE ≤ 0.54 m. The GRVI and RVI values with crop development.
for each sampling location (point measurements) are extracted At the early stage of development (on 30 May), the mag-
over 3 × 3 moving overlapping window. nitude of CpRVI is low (0.2) in the majority of the wheat
plots. Alternatively, a high degree of polarization (DOP) values
IV. R ESULTS AND D ISCUSSION (≈0.75 ± 0.03) suggests low roughness and randomness from
The CpRVI for different sampling sites are generated from the fields (see Fig. 5). The low CPR values (≈0.25 ± 0.02)
the simulated RCM data set, and we perform a seasonal indicate high opposite sense (OC) received power, which is
analysis with different growth stages for wheat and soybean. likely due to odd-bounce scattering phenomena from the bare
Furthermore, the correlation of CpRVI with PAI (m2 m−2 ) soil with less vegetation cover at the early stage of leaf
and VWC (kg m−2) are analyzed for individual crops. Figs. 3 development [28], [31]. These low values of CpRVI indicate
and 7 show the temporal trends of CpRVI averaged for three that the observed Kennaugh matrix would be similar to an
sampling points in each field. In total, 96 samples from 13 elementary target (i.e., trihedral in this case).
wheat fields and 43 samples from 14 soybean fields were Separation of CpRVI values among several fields becomes
used in the regression model (see Fig. 4 and Table IV). apparent when wheat has advanced from the leaf development
Consequently, the PAI and VWC are estimated using these to the tillering stage on 15 June. Fields with PD of ≈100 m−2
models for both wheat and soybean crops. (Fields no. 220 and 32) have low CpRVI values (≈0.37±0.05)
that are comparatively lower than wheat fields with high
A. Wheat (Triticum Aestivum) PD (125 and 190 m−2 ). The in situ measurements of PAI
Fig. 3 shows the temporal response of CpRVI over the and VWC are also comparatively higher (>3.5 m2 m−2 and
phenological stages of wheat. For the analysis of CpRVI, field ≈1.0 kg m−2 ) for wheat fields with high PD.

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
MANDAL et al.: RVI FOR CROP MONITORING USING CP-SAR DATA 6327

Fig. 4. Linear model correlations (r) between CpRVI with PAI (m2 m−2 ) and VWC (kg m−2 ) for wheat and soybean.

Fig. 5. Temporal variations of DOP and CPR for wheat fields (220, 32, 104, 55, and 233).

It is important to note that CpRVI values reached their values become low (≈0.35), and CPR saturates to ≈1. Also,
maximum when crop advanced from stem elongation to early the SC and OC components are quite similar, which indicates
dough or fruit development stages on 17 July. CpRVI values high volume scattering. However, an ID does not exist in
reach up to 0.6 for low PD fields, while these values peak at nature, and hence, as mentioned earlier, GD(K, KID ) varies in
≈0.8 for high PD wheat fields. Such difference may be due the range [0, 2/3]. Also, the spatiotemporal changes in several
to the high degree of randomness in scattering (DOP ≈ 0.4 wheat fields are apparent in Fig. 6.
on 17 July) from the canopy elements during the flowering to The correlation analysis of CpRVI with PAI and VWC along
fruit development stages. with the confidence interval (CI) is shown in Fig. 4, which
The in situ measurements of PAI and VWC at the flow- indicates the variation between these two biophysical parame-
ering to fruit development stages confirm their increase up ters. The correlation coefficient (r ) of CpRVI with PAI is 0.72
to approximately 6.5 m2 m−2 and 3.0 kg m−2 , respectively. ( p = 0.005, 95% CI: [0.60, 0.80]), which is slightly higher
This behavior indicates the dominance of volume scattering than VWC with r = 0.62 ( p = 0.006, 95% CI: [0.48, 0.73]).
in the fields, which might increase the similarity between the Variations in CpRVI with PAI and VWC at later growth
observed K and the ID KID . Also, the presence of even bounce stages are likely due to scattering from the upper canopy
scattering is often due to the interaction of the EM wave with layer (i.e., wheat heads). Similarly, Wu et al. [57] reported
the vertical stems of the wheat canopy and the underlying soil. using a ground-based scatterometer experiment that the wheat
This increases the SC component [31], thereby increasing the heads dominate the total scattering power at the heading stage.
β factor in the CpRVI formulation. Moreover, CPR values However, during the ripening stage (when the heads become
>1.2 also confirm that the same sense (SC) power is higher drier), the backscatter from the ground becomes dominant,
than OC during this period (see Fig. 5). and the backscatter power from the heads is insensitive to the
The differential increase in CpRVI among the wheat fields moisture content. Also, variations in backscatter power are less
is apparent in Fig. 3. Variations in the PD could cause prominent with changes in leaf area or biomass [58]. In fact,
dissimilarity with CpRVI in several fields, even if they are during the ripening stage of wheat, the backscatter power is
in the same phenological stage. Multiple interactions of the insensitive to plant biomass [59] at low incidence angle (as
SAR signal cause high volume scattering in the dense canopy, used in the present case of RADARSAT-2 FQ7W mode of
which is observed for fields with high PD. ≈25◦ incidence angle).
The rate of increase in CpRVI values slows down at the end
of July after the stagnation of the vegetative growth and crops
begin seed development. The ripening of wheat started during B. Soybean (Glycine Max)
the first two weeks of August, followed by swathing and Compared to wheat, the soybean canopy has more ran-
harvest in the last week of August. During this period, the DOP dom structure [60] with secondary stems and randomly

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
6328 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 58, NO. 9, SEPTEMBER 2020

Fig. 6. Spatiotemporal variations of CpRVI within the wheat and soybean fields. The black and white polygons show the sampling fields of wheat and
soybean, respectively. The field photographs during the campaign at specific instances are presented for wheat and soybean.

oriented leaves. Hence, unlike erectophiles (canopy ele- component is much higher than SC (see Fig. 8). Bare soil with
ments have predominant vertical distribution), the response shallow vegetation likely produces higher surface scattering.
of CpRVI would be different for the random structure of The dominance of surface scattering during the early stage was
the soybean canopy. The Manitoba weekly crop reports [49] also reported in [39], where the volume to surface scattering
suggest that soybean seeding was completed by the third week power ratio in CP-SAR data was below 0.7.
of May. Thus, crop development during the beginning of the With the increase in vegetation components, the varia-
SMAPVEX-16 campaign in June was primarily limited to veg- tions in CpRVI values among several fields are apparent.
etative growth. Subsequently, inflorescence emergence, flower- The soybean fields are separated into three groups: group A
ing, and pod development started during the last week of July. (RS < 40 cm and PC < 40 per meter length of row) includes
The development of pods, ripening of seeds, and senescence fields 72, 113, and 221; group B (RS > 75 cm and PC >
are followed in August until the second week of September. 65 per meter length of row) includes fields 65, 82, and 228;
Here, it is worth mentioning that the seeding pattern and the and group C (RS ≈ 50 cm and PC ≈ 50 per meter length
choice of row-plant spacing (single- and double-row plantings) of row) includes fields 223, 231, and 232. During the pod
varied among the fields (see Fig. 7). development stage, the CpRVI values reach up to 0.75 ± 0.06
Fig. 7 presents the temporal trends of CpRVI for several for group C.
fields with different row spacing (RS) and plant count (PC) per Here, it is essential to note that despite having PC > 65 m−1
meter length. The CpRVI values for each field increase rapidly for group B, the CpRVI values are comparatively lower
as the vegetation growth increases from the early vegetative (approximately 10%–15%) than fields of group C. The in situ
growth stages to the beginning of pod development. The measurements confirmed that fields in group B had wider RS
CpRVI value is approximately 0.25 ± 0.06 at the end of the (approximately 77 cm) and less canopy closure (see Fig. 6),
early leaf development stage (15 June). The high value of DOP while the number of plants along the row was higher (a double-
(0.75 ± 0.05 in Fig. 8) also suggests less randomness from the row-seeded soybean). Therefore, the PAI of group B fields is
soybean fields, which is primarily affected by the underlying lower than the fields of group C. In contrast to the fields in
soil as a result of less plant cover (PAI < 0.1 m2 m−2 ). The groups B and C, there are narrower RS (approximately 38 cm)
CPR value is approximately 0.3, which implies that the OC and low PC (<40 per meter row) for fields in group A. Hence,

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
MANDAL et al.: RVI FOR CROP MONITORING USING CP-SAR DATA 6329

Fig. 7. Temporal pattern of CpRVI for several soybean fields at different crop growth stages. Soybean fields are selected based on variations in plant RS
(cm) and PC (number of plants per meter of row length). The in situ measurements of PAI (m2 m−2 ) and VWC (kg m−2 ) are plotted in secondary axis for
each field.

Fig. 8. Temporal variations of DOP and CPR for soybean fields (72, 113, 65, 228, 231, and 232).

more volume scattering is possible from the fields of group B, case of group A fields, the CpRVI values are restricted within
which, in turn, results in higher CpRVI. Wiseman et al. [61] the range of 0.65 ± 0.05, which is comparatively lower than
also reported differences in volume scattering with such var- group C. Also, PAI and VWC are high during this stage.
ied PD. The correlation plots shown in Fig. 4 indicate that the
It is also apparent from Fig. 8 that low DOP and CPR CpRVI values are correlated better with the PAI than VWC.
values (approximately 0.85) during the pod development stage It is observed that the correlation coefficients of the PAI
indicate an increase in the volume scattering component. In the and VWC with CpRVI are r = 0.85 ( p = 0.002,

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
6330 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 58, NO. 9, SEPTEMBER 2020

Fig. 9. Temporal pattern of vegetation indices (CpRVI, RVI, and GRVI) for wheat and soybean fields at different phenological stages. Wheat fields are
selected based on variations in PD (number of plants per m2 ). Soybean fields are selected based on variations in plant RS (cm) and PC (number of plants per
meter of row length). The in situ measurements of PAI (m2 m−2 ) and VWC (kg m−2 ) are plotted in the secondary axis for each field.

95% CI: [0.75, 0.91]) and r = 0.75 ( p = 0.003, TABLE III

95% CI: [0.59, 0.85]), respectively. It is evident from Fig. 4 C ORRELATION OF V EGETATION I NDICES W ITH PAI AND
VWC FOR W HEAT AND S OYBEAN
that for both PAI and VWC, the CpRVI values are low at
the early vegetative stages (PAI < 1.0 m2 m−2 and VWC >
0.35 kg m−2 ). This phenomenon is likely because the CpRVI
derived at the early growth stage is affected to a higher degree
by the underlying soil rather than the vegetation canopy.
Fig. 6 shows the CpRVI maps for the four acquisition dates
in over a subset of the test site. Both spatial and temporal
variabilities are observed in these maps, which are similar to
the spatiotemporal behavior of soybean, as shown in Fig. 7.
During the second week of June, the soybean plants advanced
from the unifoliate to the third trifoliate stage, which is well
represented by the low values (<0.25) of CpRVI. The plant soybean, as shown in Fig. 9. From several sampling sites,
growth in subsequent weeks of June and the first week of three representative fields (field numbers: 220, 105, 233 for
July is apparent (increased to ≈0.55) in the CpRVI images for wheat, and 113, 82, 232 for soybean) for each crop are used
almost all soybean fields. The Manitoba weekly reports [49] for the temporal analysis of vegetation indices. The temporal
inform the seventh trifoliate stage in the majority of the fields. trend follows PAI and VWC for both wheat and soybean.
Rapid growth is also evident in the subsequent weeks of Similar to the analysis reported in Section IV-A for wheat,
July with soybean at its full vegetative growth. Following the field measurements are compared among the three groups
vegetative growth, soybeans continued to flower and pods that are separated based on PD (PD: number of plants per
developed in August [49] with high CpRVI values. square meter). The growth trends of the vegetation indices
are similar irrespective of PD among several fields, i.e., an
increase of PAI and VWC with crop development. However,
C. Comparison of CpRVI With Full-Pol Indices the RVI, GRVI, and CpRVI values for field no. 105 are
We compare three vegetation indices (i.e., RVI, GRVI, comparatively higher for all phenological observations that
and CpRVI) for different phenological stages of wheat and have the highest PD (approximately 230 plants/m2 ). In the

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
MANDAL et al.: RVI FOR CROP MONITORING USING CP-SAR DATA 6331

Fig. 10. Validation plots of PAI (m2 m−2 ) and VWC (kg m−2 ) retrieval for wheat and soybean.

other two fields, separation of CpRVI, GRVI, and RVI values of CpRVI with PAI is 0.72 and 0.85, which is higher than that
within a field becomes apparent (see Fig. 9) when wheat has of RVI (r = 0.68 and 0.76) for wheat and soybean. A similar
advanced from the leaf development to booting and start of improvement of r is observed for VWC for both the crops.
the flowering stage. The correlation analysis shows marginally better performance
CpRVI values reached their maximum when crop advanced of CpRVI compared with RVI, while it is inferior to GRVI for
from stem elongation to early dough or fruit development characterizing vegetation growth.
stages on 17 July. CpRVI values reach up to 0.65 for low
PD fields, while these values peak at ≈0.8 for high PD
wheat fields. This variation may be due to the high degree D. PAI and VWC Estimation
of randomness in scattering from the canopy elements during We fit linear regression models for explaining the bio-
the flowering to fruit development stages. Unlike CpRVI, physical parameters PAI and VWC using CpRVI for several
which follows almost a monotonic increase along phenological sampling locations (see Fig. 4 and Table IV). The retrieval
stages, the fluctuation of RVI values is more apparent. It is results are presented in Fig. 10 with a 1:1 plot for wheat and
possibly due to the inherent noise that arises in the numerator soybean. We measure the accuracy of PAI and VWC retrieval
term of RVI formulation, i.e., the third eigenvalue, which by the correlation coefficient (r ) and rms error.
is more affected by noise rather than changes in vegetation For wheat, the estimated PAI closely follows the 1:1 line
randomness. However, the GRVI values are more stable than with r = 0.79 and RMSE = 0.711 m2 m−2 . Higher disper-
the RVI and CpRVI. sion of estimates is observed throughout the entire range of
Unlike the RVI, which models the vegetation layer as an VWC with r = 0.75 and RMSE = 0.567 kg m−2 . The error
aggregation of randomly oriented dipoles, the GRVI offers may have propagated from the CpRVI-VWC linear regression
the flexibility to choose the parameters of the GVSM from the model (see Fig. 4). However, for both PAI and VWC retrieval,
measured data to describe the volume scattering model [42]. In the overestimations are apparent in the early development
this way, the GVSM might capture the phenological changes stage of wheat. At this early stage, backscatter response is
within a volume model, whereas both CpRVI and RVI lack dominated by soil moisture contributions due to low leaf
this aspect. On the contrary, in the absence of any volume area [62] as reported for PAI and biomass estimation for wheat
scattering models to suitably describe scattering from vegeta- using this data set [42], [63]. Moreover, the assumption of ran-
tion in CP-SAR data, the ID is utilized to measure randomness dom volume scattering from wheat fields is quite challenging
from distributed targets in CpRVI. Moreover, GDs from the to satisfy. It is because wheat has erectophile leaf orientation,
odd (i.e., trihedral) and even bounce (i.e., dihedral) scatterers satisfying a predominant vertical structure instead of a random
that appear in GRVI [42] are suitably replaced by the received volume structure. However, volume scattering can occur due
echo powers from the SC and the OC sense to the transmitted to multiple interactions of the EM wave with stems, leaves,
polarization. and the underlying soil [64] when the canopy density is high.
The variations of CpRVI and RVI values are more apparent On the other hand, the structure of the soybean canopy
in the low-biomass soybean crop. Fig. 9 shows the temporal is random with secondary and tertiary stems and randomly
trends of CpRVI and RVI for three representative soybean oriented leaves. During the full vegetative stage, the soybean
fields with different RS and PC per meter length. Both the canopy forms a very lossy volume medium, as indicated
CpRVI and full-pol indices values for each field increase in [57]. Hence, the observed K matrix might be similar to
as the vegetation growth increases from the early vegetative KID under the GD framework. The estimated PAI follows the
growth stages to the beginning of pod development. With the 1:1 line, with r = 0.87 and RMSE = 0.621 m2 m−2 , while the
increase in vegetation components, the differential increase in VWC estimation shows r = 0.83 and RMSE = 0.255 kg m−2 .
vegetation indices values among several fields is apparent. The It is evident from Fig. 10 that the dispersion of the estimates
correlation analysis of vegetation indices with crop biophysical throughout the entire range of PAI and VWC is lower for
parameters also indicates a higher correlation (r ) of CpRVI soybean compared with wheat. Nonetheless, errors might be
than RVI, as given in Table III. The correlation coefficient (r ) propagating from the CpRVI-PAI and CpRVI-VWC models

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
6332 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 58, NO. 9, SEPTEMBER 2020

TABLE IV
R EGRESSION M ODEL C OEFFICIENTS AND S TATISTICS OF C P RVI V ERSUS PAI (m2 m−2 ) AND VWC (kg m−2 ) FOR W HEAT AND S OYBEAN

(see Fig. 4) into the linear regression. In both PAI and VWC providing RADARSAT-2 data to the SMAPVEX16-MB
retrievals, overestimation is apparent in the early development research team. RADARSAT-2 Data and Products Maxar Tech-
stages that are likely due to the underlying soil contribution nologies Ltd. (2018)-All Rights Reserved. RADARSAT is an
to total backscatter power [4]. official mark of the Canadian Space Agency. In addition,
they would also like to thank the Canada Centre for Remote
V. C ONCLUSION Sensing (CCRS) for providing the RCM Compact Polarimetry
Simulator.
We have proposed a novel RVI (CpRVI) for CP-SAR data
using the concept of a GD between the observed Kennaugh
matrix and the Kennaugh matrix for an ID. The temporal R EFERENCES
analysis of the proposed CpRVI derived from the simulated
RCM data suggests that it follows crop growth development, [1] H. McNairn and B. Brisco, “The application of C-band polarimetric
i.e., increasing with an increase in PAI and VWC. Also, SAR for agriculture: A review,” Can. J. Remote Sens., vol. 30, no. 3,
pp. 525–542, Jun. 2014.
a positive correlation with these crop biophysical parameters [2] S. C. Steele-Dunne, H. McNairn, A. Monsivais-Huertero, J. Judge,
confirms the potential of CpRVI. P.-W. Liu, and K. Papathanassiou, “Radar remote sensing of agricultural
CpRVI followed the advancement of plant growth until canopies: A review,” IEEE J. Sel. Topics Appl. Earth Observ. Remote
Sens., vol. 10, no. 5, pp. 2249–2273, May 2017.
full canopy development with the accumulation of PAI and [3] F. Vicente-Guijalba, T. Martinez-Marin, and J. M. Lopez-Sanchez,
biomass, which is evident from the high correlation with these “Crop phenology estimation using a multitemporal model and a Kalman
parameters. However, with the maturity of crops, the reduction filtering strategy,” IEEE Geosci. Remote Sens. Lett., vol. 11, no. 6,
pp. 1081–1085, Jun. 2014.
of CpRVI is not apparent. We hypothesize that CpRVI is [4] X. Jiao, H. McNairn, J. Shang, E. Pattey, J. Liu, and C. Champagne,
related to the similarity between the observed Kennaugh “The sensitivity of RADARSAT-2 polarimetric SAR data to corn and
matrix K and that of the ID, which depends on both the plant soybean leaf area index,” Can. J. Remote Sens., vol. 37, no. 1, pp. 69–81,
Jun. 2014.
dielectric constant (water content) and structure of the canopy. [5] X. Blaes, P. Defourny, U. Wegmuller, A. D. Vecchia, L. Guerriero, and
Thus, the decrease in CpRVI values would be evident for late P. Ferrazzoli, “C-band polarimetric indexes for maize monitoring based
senescence (before harvest) when the plant VWC reduces to on a validated radiative transfer model,” IEEE Trans. Geosci. Remote
Sens., vol. 44, no. 4, pp. 791–800, Apr. 2006.
a minimum. Both PAI and VWC were accurately estimated [6] J. Cable, J. Kovacs, X. Jiao, and J. Shang, “Agricultural monitoring
using an empirical model with CpRVI. in Northeastern Ontario, Canada, using multi-temporal polarimetric
Notably, the proposed CpRVI for CP-SAR data holds sig- RADARSAT-2 data,” Remote Sens., vol. 6, no. 3, pp. 2343–2371,
Mar. 2014.
nificant interest from an operational perspective for upcoming
[7] H. McNairn and J. Shang, “A review of multitemporal synthetic aperture
SAR missions, e.g., the RCM, RISAT-1A, and NISAR; all of radar (SAR) for crop monitoring,” in Multitemporal Remote Sensing.
them will be collecting data in the CP mode. In particular, Cham, Switzerland: Springer, 2016, pp. 317–340.
the RCM will be capable of acquiring CP data at swaths of [8] Y. Kim and J. J. van Zyl, “A time-series approach to estimate soil
moisture using polarimetric radar data,” IEEE Trans. Geosci. Remote
up to 500 km with different NESZ values. Hence, monitoring Sens., vol. 47, no. 8, pp. 2519–2527, Aug. 2009.
of crop growth over national scales with compact-pol indices [9] Y. Kim, T. Jackson, R. Bindlish, H. Lee, and S. Hong, “Radar vege-
would be a good proxy. However, implications with real tation index for estimating the vegetation water content of rice and
soybean,” IEEE Geosci. Remote Sens. Lett., vol. 9, no. 4, pp. 564–568,
compact-pol data in terms of transmitted circularity, NESZ, Jul. 2012.
and hardware limitations need to be further explored, as this [10] Y. Kim, T. Jackson, R. Bindlish, S. Hong, G. Jung, and K. Lee,
may affect the scattering response from targets. Moreover, “Retrieval of wheat growth parameters with radar vegetation indices,”
IEEE Geosci. Remote Sens. Lett., vol. 11, no. 4, pp. 808–812, Apr. 2014.
experimental validation of vegetation indices on the incidence [11] Y. Huang, J. P. Walker, Y. Gao, X. Wu, and A. Monerris, “Estima-
angle variations is essential for wide swath RCM products. tion of vegetation water content from the radar vegetation index at
Nevertheless, the effectiveness of the proposed CpRVI for L-band,” IEEE Trans. Geosci. Remote Sens., vol. 54, no. 2, pp. 981–989,
Feb. 2016.
large swath observation needs further studies with experiment [12] R. K. Raney, “Hybrid-polarity SAR architecture,” IEEE Trans. Geosci.
setups (in situ measurements and satellite passes) to provide Remote Sens., vol. 45, no. 11, pp. 3397–3404, Nov. 2007.
a conclusive exposition. [13] J.-C. Souyris, N. Stacy, T. Ainsworth, J.-S. Lee, and
P. Dubois-Fernandez, “SAR compact polarimetry (CP) for Earth
observation and planetology: Concept and challenges,” in Proc.
ACKNOWLEDGMENT PolInSAR, 2007, pp. 1–7.
[14] L. M. Carter, “Impact crater related surficial deposits on Venus: Multipo-
The authors would like to thank the Canadian Space larization radar observations with Arecibo,” J. Geophys. Res., vol. 109,
Agency and MAXAR Technologies Ltd. (formerly MDA) for no. E6, pp. 1–13, 2004.

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
MANDAL et al.: RVI FOR CROP MONITORING USING CP-SAR DATA 6333

[15] P. D. Spudis et al., “Initial results for the North Pole of the Moon from [36] G. G. Ponnurangam, T. Jagdhuber, I. Hajnsek, and Y. S. Rao, “Soil
Mini-SAR, Chandrayaan-1 mission,” Geophys. Res. Lett., vol. 37, no. 6, moisture estimation using hybrid polarimetric SAR data of RISAT-1,”
pp. 1–6, Mar. 2010. IEEE Trans. Geosci. Remote Sens., vol. 54, no. 4, pp. 2033–2049,
[16] F. J. Charbonneau et al., “Compact polarimetry overview and applica- Apr. 2016.
tions assessment,” Can. J. Remote Sens., vol. 36, no. 2, pp. S298–S315, [37] G. G. Ponnurangam and Y. S. Rao, “The application of compact polari-
Jun. 2014. metric decomposition algorithms to L-band PolSAR data in agricultural
[17] R. K. Raney, J. T. S. Cahill, G. W. Patterson, and D. B. J. Bussey, areas,” Int. J. Remote Sens., vol. 39, no. 22, pp. 8337–8360, Jul. 2018.
“Them-chidecomposition of hybrid dual-polarimetric radar data with [38] H. McNairn, S. Homayouni, M. Hosseini, J. Powers, K. Beckett, and
application to lunar craters,” J. Geophys. Res., Planets, vol. 117, no. E12, W. Parkinson, “Compact polarimetric synthetic aperture radar for mon-
p. 8, May 2012. itoring crop condition,” in Proc. IEEE Int. Geosci. Remote Sens. Symp.
[18] T. L. Ainsworth, J. P. Kelly, and J.-S. Lee, “Classification comparisons (IGARSS), Jul. 2017, pp. 4358–4361.
between dual-pol, compact polarimetric and quad-pol SAR imagery,” [39] S. Homayouni, H. McNairn, M. Hosseini, X. Jiao, and J. Powers, “Quad
ISPRS J. Photogramm. Remote Sens., vol. 64, no. 5, pp. 464–471, and compact multitemporal C-band PolSAR observations for crop char-
Sep. 2009. acterization and monitoring,” Int. J. Appl. Earth Observ. Geoinf., vol. 74,
[19] B. Brisco, K. Li, B. Tedford, F. Charbonneau, S. Yun, and pp. 78–87, Feb. 2019.
K. Murnaghan, “Compact polarimetry assessment for rice and wetland [40] D. Ratha, S. De, T. Celik, and A. Bhattacharya, “Change detection in
mapping,” Int. J. Remote Sens., vol. 34, no. 6, pp. 1949–1964, polarimetric SAR images using a geodesic distance between scatter-
2013. ing mechanisms,” IEEE Geosci. Remote Sens. Lett., vol. 14, no. 7,
[20] D. Haldar, A. Das, S. Mohan, O. Pal, R. S. Hooda, and M. Chakraborty, pp. 1066–1070, Jul. 2017.
“Assessment of L-band SAR data at different polarization combinations [41] D. Ratha, A. Bhattacharya, and A. C. Frery, “Unsupervised classifica-
for crop and other landuse classification,” Prog. Electromagn. Res. B, tion of PolSAR data using a scattering similarity measure derived from
vol. 36, pp. 303–321, Jan. 2012. a geodesic distance,” IEEE Geosci. Remote Sens. Lett., vol. 15, no. 1,
[21] M. E. Nord, M. E. Nord, T. L. Ainsworth, J.-S. Lee, and N. J. S. Stacy, pp. 151–155, Jan. 2018.
“Comparison of compact polarimetric synthetic aperture radar modes,” [42] D. Ratha, D. Mandal, V. Kumar, H. McNairn, A. Bhattacharya, and
IEEE Trans. Geosci. Remote Sens., vol. 47, no. 1, pp. 174–188, A. C. Frery, “A generalized volume scattering model-based vegetation
Jan. 2009. index from polarimetric SAR data,” IEEE Geosci. Remote Sens. Lett.,
[22] L. Xie, H. Zhang, H. Li, and C. Wang, “A unified framework for vol. 16, no. 11, pp. 1791–1795, Nov. 2019.
crop classification in Southern China using fully polarimetric, dual [43] O. Antropov, Y. Rauste, and T. Hame, “Volume scattering modeling in
polarimetric, and compact polarimetric SAR data,” Int. J. Remote Sens., PolSAR decompositions: Study of ALOS PALSAR data over boreal for-
vol. 36, no. 14, pp. 3798–3818, Jul. 2015. est,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 10, pp. 3838–3848,
[23] R. K. Raney, “Comparing compact and quadrature polarimetric SAR per- Oct. 2011.
formance,” IEEE Geosci. Remote Sens. Lett., vol. 13, no. 6, pp. 861–864, [44] S. Cloude, Polarisation: Applications in Remote Sensing. Oxford, U.K.:
Jun. 2016. Oxford Univ. Press, 2010, pp. 83–84.
[24] M. Ohki and M. Shimada, “Large-area land use and land cover classifi- [45] D. Ratha, P. Gamba, A. Bhattacharya, and A. C. Frery, “Novel tech-
cation with quad, compact, and dual polarization SAR data by PALSAR- niques for built-up area extraction from polarimetric SAR images,” IEEE
2,” IEEE Trans. Geosci. Remote Sens., vol. 56, no. 9, pp. 5550–5557, Geosci. Remote Sens. Lett., vol. 17, no. 1, pp. 177–181, Jan. 2020.
Sep. 2018. [46] D. Ratha, A. Bhattacharya, and A. C. Frery, “A scattering power fac-
[25] V. Kumar, Y. S. Rao, A. Bhattacharya, and S. R. Cloude, “Classifi- torization framework using a geodesic distance in radar polarimetry,”
cation assessment of real versus simulated compact and quad-pol in Proc. IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), Jul. 2018,
modes of ALOS-2,” IEEE Geosci. Remote Sens. Lett., vol. 16, no. 9, pp. 6075–6078.
pp. 1497–1501, Sep. 2019. [47] D. Ratha, A. Bhattacharya, A. C. Frery, and E. Pottier, “A scattering
[26] J. D. Ballester-Berman and J. M. Lopez-Sanchez, “Time series of power factorization framework using a geodesic distance for multi-
hybrid-polarity parameters over agricultural crops,” IEEE Geosci. looked polsar data,” in Proc. IEEE Int. Geosci. Remote Sens. Symp.
Remote Sens. Lett., vol. 9, no. 1, pp. 139–143, Jan. 2012. (IGARSS), Jul. 2019, pp. 5125–5128.
[27] J. Shang, H. McNairn, F. Charbonneau, Z. Chen, and X. Jiao, “Sensi- [48] D. Ratha, E. Pottier, A. Bhattacharya, and A. C. Frery, “A PolSAR scat-
tivity analysis of compact polarimetry parameters to crop growth using tering power factorization framework and novel roll-invariant parameter-
simulated RADARSAT-2 SAR data,” in Proc. IEEE Int. Geosci. Remote based unsupervised classification scheme using a geodesic distance,”
Sens. Symp., Jul. 2012, pp. 1825–1828. IEEE Trans. Geosci. Remote Sens., to be published, doi: 10.1109/
[28] J. M. Lopez-Sanchez, F. Vicente-Guijalba, J. D. Ballester-Berman, and TGRS.2019.2957514.
S. R. Cloude, “Polarimetric response of rice fields at C-band: Analysis [49] Manitoba Agriculture. (2016). Agriculture|Province of Manitoba.
and phenology retrieval,” IEEE Trans. Geosci. Remote Sens., vol. 52, [Online]. Available: http://www.gov.mb.ca/agriculture/crops/seasonal-
no. 5, pp. 2977–2993, May 2014. reports/crop-report-archive/index.html
[29] Z. Yang, K. Li, L. Liu, Y. Shao, B. Brisco, and W. Li, “Rice growth [50] H. A. K. M. Bhuiyan et al., “Assessing SMAP soil moisture scaling and
monitoring using simulated compact polarimetric C-band SAR,” Radio retrieval in the Carman (Canada) study site,” Vadose Zone J., vol. 17,
Sci., vol. 49, no. 12, pp. 1300–1315, Dec. 2014. no. 1, Dec. 2018, Art. no. 180132, doi: 10.2136/vzj2018.07.0132.
[30] R. K. Venkata et al., “Multi-incidence angle RISAT-1 Hybrid Polari- [51] H. Mcnairn et al. (2016). Experimental Plan SMAP Validation Exper-
metric SAR data for large area mapping of maize crop—A case study iment 2016 in Manitoba, Canada (SMAPVEX16-MB). [Online]. Avail-
in Khagaria district, Bihar, India,” Int. J. Remote Sens., vol. 38, no. 20, able: https://smap.jpl.nasa.gov/internal_resources/390/
pp. 5487–5501, Jun. 2017. [52] A. A. Thompson, “Overview of the RADARSAT constellation mission,”
[31] V. Kumar, H. McNairn, A. Bhattacharya, and Y. S. Rao, “Temporal Can. J. Remote Sens., vol. 41, no. 5, pp. 401–407, Dec. 2015.
response of scattering from crops for transmitted ellipticity variation [53] T. Geldsetzer, F. Charbonneau, M. Arkett, and T. Zagon, “Ocean wind
in simulated compact-pol SAR data,” IEEE J. Sel. Topics Appl. Earth study using simulated RCM compact-polarimetry SAR,” Can. J. Remote
Observ. Remote Sens., vol. 10, no. 12, pp. 5163–5174, Dec. 2017. Sens., vol. 41, no. 5, pp. 418–430, Oct. 2015.
[32] W. Zhang et al., “Compact polarimetric response of rape (Brassica napus [54] L. D. Robertson, A. Davidson, H. McNairn, M. Hosseini, and
L.) at C-band: Analysis and growth parameters inversion,” Remote Sens., S. Mitchell, “Compact polarimetry for agricultural mapping and
vol. 9, no. 6, p. 591, Jun. 2017. inventory: Preparation for radarsat constellation mission,” in Proc.
[33] S. Chauhan, H. S. Srivastava, and P. Patel, “Wheat crop biophysical IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), Jul. 2019,
parameters retrieval using hybrid-polarized RISAT-1 SAR data,” Remote pp. 5851–5854.
Sens. Environ., vol. 216, pp. 28–43, Oct. 2018. [55] M. Mahdianpari et al., “Mid-season crop classification using Dual-,
[34] M.-L. Truong-Loi, A. Freeman, P. C. Dubois-Fernandez, and E. Pottier, Compact-, and full-polarization in preparation for the radarsat constella-
“Estimation of soil moisture and Faraday rotation from bare surfaces tion mission (RCM),” Remote Sens., vol. 11, no. 13, p. 1582, Jul. 2019.
using compact polarimetry,” IEEE Trans. Geosci. Remote Sens., vol. 47, [56] J. J. van Zyl, Synthetic Aperture Radar Polarimetry, vol. 2. Hoboken,
no. 11, pp. 3608–3615, Nov. 2009. NJ, USA: Wiley, 2011.
[35] J. D. Ouellette et al., “A simulation study of compact polarimetry for [57] L.-K. Wu, R. Moore, and R. Zoughi, “Sources of scattering from
radar retrieval of soil moisture,” IEEE Trans. Geosci. Remote Sens., vegetation canopies at 10 GHz,” IEEE Trans. Geosci. Remote Sens.,
vol. 52, no. 9, pp. 5966–5973, Sep. 2014. vols. GE-23, no. 5, pp. 737–745, Sep. 1985.

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
6334 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 58, NO. 9, SEPTEMBER 2020

[58] M. Jia, L. Tong, Y. Zhang, and Y. Chen, “Multitemporal radar backscat- Avik Bhattacharya (Senior Member, IEEE)
tering measurement of wheat fields using multifrequency (L, S, C, received the integrated M.Sc. degree in mathematics
and X) and full-polarization,” Radio Sci., vol. 48, no. 5, pp. 471–481, from IIT Kharagpur, Kharagpur, India, in 2000,
Sep. 2013. and the Ph.D. degree in remote sensing image
[59] F. Mattia et al., “Multitemporal C-band radar measurements on processing and analysis from Télécom ParisTech,
wheat fields,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 7, Paris, France, and the Ariana Research Group,
pp. 1551–1560, Jul. 2003. Institut National de Recherche en Informatique et
[60] H. Wang, R. Magagi, and K. Goita, “Polarimetric decomposition for en Automatique (INRIA), Nice, France, in 2007.
monitoring crop growth status,” IEEE Geosci. Remote Sens. Lett., He was a Canadian Government Research Fellow
vol. 13, no. 6, pp. 870–874, Jun. 2016. with the Canadian Centre for Remote Sensing
[61] G. Wiseman, H. McNairn, S. Homayouni, and J. Shang, “RADARSAT- (CCRS), Ottawa, ON, Canada. He is currently an
2 polarimetric SAR response to crop biomass for agricultural production Associate Professor with the Centre of Studies in Resources Engineering,
monitoring,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., IIT Bombay (CSRE, IITB), Mumbai, India, where he is also leading the
vol. 7, no. 11, pp. 4461–4471, Nov. 2014. Microwave Remote Sensing Laboratory. His research interests include
[62] F. T. Ulaby, C. T. Allen, G. Eger, and E. Kanemasu, “Relating the synthetic aperture radar (SAR) polarimetry, statistical analysis of polarimetric
microwave backscattering coefficient to leaf area index,” Remote Sens. SAR images, and applications of radar remote sensing in agriculture,
Environ., vol. 14, nos. 1–3, pp. 113–133, Jan. 1984. cryosphere, urban, and planetary studies.
[63] D. Mandal, V. Kumar, H. McNairn, A. Bhattacharya, and Y. S. Rao, Dr. Bhattacharya received the Natural Sciences and Engineering Research
“Joint estimation of plant area index (PAI) and wet biomass in wheat Council of Canada Visiting Scientist Fellowship at the Canadian national
and soybean from C-band polarimetric SAR data,” Int. J. Appl. Earth laboratories, from 2008 to 2011. He is the Founding Chairperson of the
Observ. Geoinf., vol. 79, pp. 24–34, Jul. 2019. IEEE Geoscience and Remote Sensing Society (GRSS) Chapter of the
[64] S. C. M. Brown, S. Quegan, K. Morrison, J. C. Bennett, and Bombay Section. He was an Associate Editor of the IEEE G EOSCIENCE
G. Cookmartin, “High-resolution measurements of scattering in wheat AND R EMOTE S ENSING L ETTERS (GRSL). He has been a Guest Editor of
canopies-implications for crop parameter retrieval,” IEEE Trans. Geosci. the Special Issue on Applied Earth Observations and Remote Sensing in
Remote Sens., vol. 41, no. 7, pp. 1602–1610, Jul. 2003. India in the IEEE J OURNAL OF S ELECTED T OPICS IN A PPLIED E ARTH
O BSERVATIONS AND R EMOTE S ENSING (J-STARS) in 2017. He was one of
the guest editors of the special stream on Advanced Statistical Techniques
in SAR Image Processing and Analysis in the IEEE G EOSCIENCE AND
R EMOTE S ENSING L ETTERS in 2018. He is also the Editor-in-Chief of the
IEEE GRSL.
Dipankar Mandal (Student Member, IEEE)
received the B.Tech. degree in agricultural engineer-
ing from Bidhan Chandra Krishi Viswavidyalaya,
Mohanpur, India, in 2015. He is currently pursuing
the joint M.Tech./Ph.D. degree in geoinformatics
and natural resources engineering with the Centre Vineet Kumar (Member, IEEE) received the
of Studies in Resources Engineering, IIT Bombay, B.Tech. degree in electronics and communication
Mumbai, India. engineering from Uttar Pradesh Technical Univer-
He was a Visiting Researcher with the Agri- sity, Lucknow, India, in 2010, the M.Tech. degree in
culture and Agri-Food Canada (AAFC), Ottawa, remote sensing and geographic information system
ON, Canada, and Carleton University, Ottawa, from (GIS) from the Indian Institute of Remote Sensing,
October 2018 to February 2019. As a Visiting Researcher, he contributed to Dehradun, India, in 2013, and the Ph.D. degree in
synthetic aperture radar (SAR) intercomparison experiment for crop biophys- synthetic aperture radar (SAR) remote sensing from
ical parameter estimation within the Joint Experiment for Crop Assessment IIT Bombay, Mumbai, India, in 2019.
and Monitoring (JECAM) network of GEO Global Agricultural Monitoring. He was a Visiting Researcher with Carleton Uni-
His research interests include applications of SAR polarimetry for crop versity, Ottawa, ON, Canada, and the Agriculture
classification, vegetation biophysical parameter estimation, deriving vegetation and Agri-Food Canada (AAFC), Ottawa, from April to July 2016. His
indices, and yield forecasting. research interests include applications of full and compact SAR polarimetry,
Mr. Mandal was a recipient of the Shastri Research Student Fellowship for classification and characterization of vegetation, and relating biophysical
Award by the Shastri Indo-Canadian Institute, India, from 2018 to 2019. variables of crops to SAR observables.
Dr. Kumar was a recipient of the Shastri Research Student Fellowship
2016 Award by the Shastri Indo-Canadian Research Institute, India.

Debanshu Ratha (Student Member, IEEE) received

the five-year integrated M.Sc. degree in mathematics
from the National Institute of Science Education and
Research (NISER), Bhubaneswar, Odisha, in 2013.
He is currently pursuing the Ph.D. degree with the
Centre of Studies in Resources Engineering (CSRE),
IIT Bombay, Mumbai, India. Heather McNairn received the bachelor’s degree in
Since then, he worked in full-time research posi- environmental studies from the University of Water-
tions at Delft Institute of Applied Mathematics loo, Waterloo, ON, Canada, in 1987, the master’s
(DIAM), Delft University of Technology (TU Delft), degree in soil science from the University of Guelph,
Delft, The Netherlands, and CSRE, IIT Bombay. Guelph, ON, Canada, in 1991, and the Ph.D. degree
He has been a Visiting Researcher with the Institut d’Électronique et de in geography from Université Laval, Quebec, QC,
Télécommunications de Rennes (IETR) Laboratory, University of Rennes Canada, in 1999.
1, Rennes, France, through the Raman Charpak Fellowship Program. His She is currently a Senior Scientist with the Agri-
research interests include the theory of remote sensing and its applications, culture and Agri-Food Canada, Ottawa, ON, Canada.
radar polarimetry, variational analysis, information geometry, graph theory, She has 25 years of experience researching meth-
machine learning, and data fusion. ods to monitor crops and soil using multispectral,
Mr. Ratha was a recipient of many prestigious scholarships and visiting hyperspectral, and synthetic aperture radar sensors. She is also an Adjunct
fellowships in the past. He received the CSIR Fellowship Award by the CSRE, Professor with the University of Manitoba, Winnipeg, MB, Canada, and
IIT Bombay. Carleton University, Ottawa.

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.
MANDAL et al.: RVI FOR CROP MONITORING USING CP-SAR DATA 6335

Yalamanchili S. Rao (Member, IEEE) received the Alejandro C. Frery (Senior Member, IEEE)
M.Sc. degree in physics from Andhra University, received the B.Sc. degree in electronic and elec-
Visakhapatnam, India, in 1982, and the Ph.D. degree trical engineering from Universidad de Mendoza,
in passive microwave remote sensing of soil mois- Mendoza, Argentina, in 1985, the M.Sc. degree in
ture from IIT Bombay, Mumbai, India, in 1992. applied mathematics (Statistics) from the Instituto
From 2005 to 2009, he was a Senior Research Nacional de Matemática Pura e Aplicada (IMPA),
Scientist and then Associate Professor from 2009 to Rio de Janeiro, Brazil, in 1990, and the Ph.D. degree
2014. He was engaged in both passive and active in applied computing from the Instituto Nacional de
microwave remote sensing for several applications Pesquisas Espaciais (INPE), São José dos Campos,
viz., soil moisture, crop monitoring, flood mapping, Brazil, in 1993.
and land use/land cover. He joined the Centre of He is currently the Leader of the Laboratório de
Studies Resources Engineering, IIT Bombay, in 1985, as a Senior Research Computação Científica e Análise Numérica (LaCCAN), Universidade Federal
Assistant, and then became a Research Scientist in 1999. He is currently de Alagoas, Maceió, Brazil, and also holds a Huashan Scholar position
a Professor with IIT Bombay. He has participated in several spaceborne at the Key Laboratory of Intelligent Perception and Image Understanding
campaigns for collecting synchronous ground-truth data and has experience of the Ministry of Education, Xidian University, Xi’an, China, for the
in handling various data sets for several applications. His research interests period of 2019–2021. His research interests include statistical computing and
include the application of polarimetry for bio-geophysical parameter retrieval stochastic modeling.
and synthetic aperture radar (SAR) interferometry for digital elevation model
(DEM) and displacement map generation.

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on May 17,2022 at 01:16:27 UTC from IEEE Xplore. Restrictions apply.