Remote Sensing Applications: Society and Environment 18 (2020) 100290

Contents lists available at ScienceDirect

Remote Sensing Applications: Society and Environment

journal homepage: http://www.elsevier.com/locate/rsase

Mapping rice crop using sentinels (1 SAR and 2 MSI) images in tropical
area: A case study in Fogera wereda, Ethiopia
Teshome Talema a, Binyam Tesfaw Hailu a, b, *
a
School of Earth Science, College of Natural and Computational Sciences, Addis Ababa University, Ethiopia
b
Philipps-Universit€
at Marburg, Department of Geography, Deutschhausstraße 12, 35032, Marburg, Germany

A R T I C L E I N F O A B S T R A C T

Keywords: Though optical remote sensing has various importance for land-cover mapping and monitoring, it is very difficult
Sentinel-1A to assess and monitor rice agriculture over large areas due to its cloud cover and the nature of rice agriculture.
Sentinel-2A Therefore, we integrate Sentinel-1A Synthetic Aperture Radar (SAR) with Sentinel-2 multispectral sensor (MSI)
Fogera wereda
images to map rice field extent in a tropical area between 1774 and 2494 m above sea level (Fogera wereda,
CART
Rice mapping
Ethiopia). First, we extract the temporal backscatter (TB) value of rice fields and background land-cover types at
Ethiopia the Vertical transmitted and Vertically received (VV) and Vertically transmitted and Horizontal received (VH)
polarizations. Second, Classification And Regression Trees (CART) model was applied to identify the optimal
node and map the rice field and other Land Use Land Covers. Third, the map from Sentinel-1A image was in­
tegrated with Sentinel-2A image products (Normalized Difference Vegetation Index (NDVI) and the Modified
Normalized Difference Water Index (MNDWI)) to improve the classification accuracy. The result shows that the
TB value of rice increased sharply at the early planting stage and decreased during the high flooding stages. The
increase in rice backscatter is more sustained at the Sentinel-1A VH polarization, and two-class separability
measure further shows that VH is better than VV in discriminating rice fields. The CART model result shows
August 01, 2017 is the best imaging date to map rice fields that account 19,892.5 ha of the study area with F1-
macro accuracy 0.71. The integration of NDVI and MNDWI from Sentinel-2A image with TB increased the ac­
curacy by 0.08. The refined land use land-cover map shows 19,157.8 ha of rice field.

1. Introduction
One of the major agricultural crops in the world is rice (Oryza sativa)
that is the staple food for more than half of the world population (Tor­
bick et al., 2017; Dawe et al., 2010). According to Reddy (2004), the two
most cultivated species are the Asian rice (Oryza sativa) and the African
rice (Oryza glaberrima). In Ethiopia, rice was introduced based on the
adaptation of wild rice in the north-western part of the country near
Lake Tana (the largest lake in Ethiopia) in the early 1970s (Takele,
2010). Rice production has been expanding in the low-lying areas with
the unceasing engagement of extension systems. Rice is used for food,
local alcoholic drinks and as a cash crop to the community. Under
normal circumstances, an expansion of rice fields should increase rice
productivity and food supply (Inoue et al., 2014; Mansaray et al., 2017;
Torbick et al., 2017). A large amount of money and labor are required to
monitor rice acreage at a large scale based on the traditional way of a
survey (Shen et al., 2009; Bouvet and Le Toan, 2011). Thus, both optical
and RAdio Detection And Ranging (RADAR) remote sensing data have
been exploited to map and monitor rice growth (Mansaray et al., 2017)
because these data on rice planted areas are available at a variety of
spatial scales. In the last two decades, the advances in Synthetic Aper­
ture Radar (SAR) technologies brought about a remote sensing capa­
bility of an entirely different kind. This coherent microwave sensor can
penetrate foliage and clouds, and can be operated during day or night,
because it provides its own illumination. The SAR synthesizes a large
antenna by the motion of the radar platform. The resolution in the along
track (azimuth) direction is independent of the range (Elachi, 1988).
Consequently, such a sensor has numerous advantages for remote
sensing, especially from space. The application of RADAR data
compared to optical data has not been well utilized in agricultural ap­
plications due to i) the complexity, diversity, and availability, and ii) the
difficulty of data interpretation (Veloso et al., 2017).
RADAR images are currently preferable to estimate the area of rice
field due to i) the RADAR sensor’s all-weather imaging ability and ii) the

* Corresponding author. School of Earth Science, College of Natural and Computational Sciences, Addis Ababa University, Ethiopia.
E-mail addresses: binyam.tesfaw@aau.edu.et, binyam.hailu@geo.uni-marburg.de (B.T. Hailu).

https://doi.org/10.1016/j.rsase.2020.100290
Received 8 February 2019; Received in revised form 8 November 2019; Accepted 26 January 2020
Available online 31 January 2020
2352-9385/© 2020 Elsevier B.V. All rights reserved.
T. Talema and B.T. Hailu
presence of clouds and its effect on optical satellite images in tropical
areas (Koppe et al., 2013). RADAR data could be the preeminent data for
monitoring paddy rice fields in cloud-prone tropical regions and com­
plementary to optical data. SAR data have a proven ability to detect
lowland rice systems (both irrigated and rainfed) through the unique
temporal signature of the backscatter coefficient (also termed sigma
nought or σ� ) exhibited by the crop. Mapping of rice crop is primarily
based on the identification and processing of its reflectance character­
istics in the visible near-infrared (NIR) bands of optical satellite images.
The investigated spectral reflectance showed by paddy rice have been
mostly applied to differentiate the rice crop from other land cover
classes during growth. On the other hand, mapping paddy rice using
microwave data is largely a function of growth stage and environmental
conditions, for example, RADAR SAR satellite data is based on the
backscatter response (Yonezawa et al., 2012).
Surface and volume scatterings account in radar response with a
different time period due to an interaction between electromagnetic
radiation with the vegetation (Zhang et al., 2009). Before rice
sowing/planting, fields are flooded and smooth or specular reflection,
which has a low backscatter characteristic is observed from the flooded
area (Kuenzer and Knauer, 2013; Inoue et al., 2014). According to Zhang
et al. (2009), when rice crops grow and advance in biomass, the back­
scatter response increases proportionally with more interaction from
direct volume scattering, and double-bounce mechanisms until crops
head. After ripening and close to harvest time, decrease in saturation
and moisture due to mature by stem, and if still flooded, paddies are
usually drained and leaf area, which causes a shift toward the decline in
backscatter (Nguyen Lam-Dao, 2009; Inoue et al., 2014; Mansaray et al.,
2017). The temporal backscatter (TB) change helps to differentiate the
rice crop area from other surfaces using microwave remote sensing data
such as Sentinel-1A SAR images (Torbick et al., 2017). A number of
studies have been conducted on rice mapping (Inglada et al., 2016;
Mansaray et al., 2017; Torbick et al., 2017; Sakamoto et al., 2018;
Nelson et al., 2014). There are also some researches on using both op­
tical and microwave sensors, for example, Campos-Taberner et al.
(2017). However, gaps remain on the combined use of microwave and
optical data for rice area estimation in tropical areas between 1774 and
2494 m above sea level (a.s.l.) areas. Therefore, this research work has
Fig. 1. Location map of study area A) with administrative and river network and B) with Sentinel-2A natural color composite image (Red: Band 4, Green: Band 3 and
Blue: Band 2) acquired on October 09, 2017. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of
this article.)
2
Remote Sensing Applications: Society and Environment 18 (2020) 100290
used both optical specifically, Normalized Difference Vegetation Index
(NDVI) and the Modified Normalized Difference Water Index (MNDWI)
and microwave remote sensing data sets to map paddy rice using Clas­
sification and Regression Trees (CART) using images from the new
generation Sentinel-1A and Sentinel-2A to discriminate paddy rice field
in Fogera wereda, tropical area during the 2017 rice-growing season.
2. Materials
2.1. Study area
Fig. 1 shows the study area that is Fogera wereda overlaid with the
Sentinel-2A image footprints. Geographically, it is bounded by latitude
11� 380 4100 –12� 00 3700 N and longitude 37� 310 2800 –38� 00 2900 E, covering a
total area of 111,140.4 ha with an altitude between 1774 and 2494 m a.
s.l. According to Takele (2010), Fogera wereda, is classified into Moist
Weyena Dega and Wet Weyena Dega zones based on the agro-ecological
classification method (altitude and rainfall). The Moist Weyena Dega
and Wet Weyena Dega cover 43.6% and 56.4% land of the wereda,
respectively. The mean annual rainfall is 1216 mm both the short
(March and April) and long (June to September) rainy seasons and
farmers depend on long rainy season (“Kremt”) for crop production.
The farming system in the north-western and western part of the
wereda, encompassing the area known as the Fogera plains and gets high
flood during the long rainy season. The people cultivate mainly rice and
experience fishing, horticulture, and livestock (Takele, 2010). The
farming system in the south and the eastern section, which varies from
lowlands to highlands is cereals, oil crops, horticulture, livestock, and
apiculture. Cattle-farming is still a major activity, but crop cultivation
has become highly important.
2.2. Sentinel-1A SAR image
Sentinel-1 collects data in four modes, namely Strip Map (SM),
Interferometric Wide Swath (IW), Extra Wide (EW) and Wave Mode
(WM), with varying resolutions, extents, incidence angles, and polari­
zations. The IW mode is the default operational mode for the applica­
tions of forestry, agriculture, and other natural resources (Skolnik,
T. Talema and B.T. Hailu
2008). Hence, the IW mode is selected in this study to map rice fields in
Fogera wereda, Ethiopia. IW mode observes in two polarizations: Ver­
tical transmitted and Vertically received (VV), and Vertically trans­
mitted and Horizontally received (VH) with a 250 km footprint in the
range direction. The European Space Agency (ESA) has made these data
available from its hub and mirrored at the Alaska Satellite Facility (ASF).
Data were obtained as Standard Archive Format for Europe (SAFE),
which contains general product information in XML, for the entire re­
gion, using the open-access data archives. Annotated data sets hold
metadata on the main characteristics, including acquisition, image
properties, polarization, Doppler information, swath merging, calibra­
tion, and geographic location (Torbick et al., 2017). SAFE products were
obtained as Ground Range Detected (GRD), which have already been
focused, calibrated and projected in ground range.
In this study, we have acquired six Sentinel-1A GRD images in IW
mode is used during the 2017 rice growing season in Fogera wereda. The
imaging dates are June 14, 2017 (B1), July 20, 2017 (B2), August 1,
2017 (B3), August 25, 2017 (B4), September 30, 2017 (B5) and October
12, 2017 (B6). These images span from the flooding/transplanting stage
to the heading/milking stage of paddy rice, a period at which changes in
TB of the rice crop are most diagnostic. The acquired images are both in
the single and dual polarization of VV, and VH channels in a range and
azimuth resolution of 5 m and 20 m, respectively. The VV and VH
Sentinel-1A with incidence angle near and far were 31.01� and 46.19� ,
respectively. The six images have descending pass direction (PD), 152 in
Relative Orbit Number (RON), and normal acquisition type (AT).
2.3. Sentinel-2A multispectral sensor (MSI) image
The Sentinel-2A MSI image has 13 bands in total, in which four bands
(blue, green, red and NIR) have a spatial resolution of 10 m and six
bands (including SWIR band) have a spatial resolution of 20 m. The ESA
launched a new optical Sentinel-2A on June 23, 2015. Sentinel-2A op­
tical images are freely available to the user community. Due to persistent
cloudy weather conditions in the study area during the rice growing
season, only a single cloud masked Sentinel-2A Level 1C image acquired
on October 09, 2017 was used in this study.
Fig. 2. A) GPS data points collected from field work with Sentinel 1 A SAR images of 2017, B) Rice field at high flooding C) Field at maturity stages.
3
Remote Sensing Applications: Society and Environment 18 (2020) 100290
2.4. Field data
Field data collection was conducted from September 5 to September
15, 2017, to obtain training and validation datasets. Fig. 2 shows photos
taken during the 2017 field campaigns. Based on these data, the wereda
land-cover classification was divided into six broad classes: these are
rice (paddy rice fields), non-rice crops (wheat, sorghum, maize, teff,
horticulture etc.), water (rivers, lakes, streams, and ponds), built up
(residential and commercial houses), trees (plants with considerable
height and density) and others (roads, bare land surfaces, grassland,
other low biomass species). This broad classification scheme was
adopted to help in identification of optimal polarization (VV/VH) of
Sentinel-1A images.
A total of 611 training and validation data were randomly collected
from the paddy rice field and other land-cover using a portable Garmin
Global Positioning System (GPS) device, and an average horizontal ac­
curacy of 3.0 m was recorded (Fig. 2). For paddy rice field 93 training
data were collected. The remaining 76 points were reserved for rice field
accuracy assessment. Other land cover types (non-rice crops, water,
built up and trees) account for a total of 226 training points which were
collected and exported to Sentinel-1A data for the extraction of TB
profiles.
3. Methods
This research has five stages for rice mapping: i) Image pre-
processing, ii) S Sentinel-1A image analysis and identifying the
optimal polarization for mapping, iii) Land Use Land Cover (LULC)
processing using CART Model, iv) Classification improvement using
indices from Sentinel-1A images and iv) validation.
3.1. Image pre-processing
Sentinel-1A SAR images of VV/VH polarizations were first calibrated
to sigma nought (σ0 ) to provide imagery in which the pixel values can be
directly related to the radar backscatter of the scene. The Sentinel-1A
GRD product was radiometrically calibrated using SNAP toolbox
T. Talema and B.T. Hailu
automatically with the following equation:
2
DN
σ0 ​ ¼ ​
Aσ2
where DN: is the pixel digital Number of GRD pixel amplitude directly
taken from the measurement file and A is the value of sigma nought (σ0 ).
The calibrated Sentinel-1A images were then automatically terrain
corrected using range-Doppler terrain correction to produce 10 m square
pixel resolution images using Universal Transverse Mercator (UTM)
zone 37 North.
The speckle filtering suppresses the noise that eases the better
backscatter analysis and thus, it turns out to be a critical pre-processing
step for detection or classification optimization. In this study speckle
filtering using a 7 � 7 gamma map filter with 3 look techniques was
applied (Wu et al., 2011; Mansaray et al., 2017). A good filter shows
lower Mean Square Error (MSE), higher Signal to Noise Ratio (SNR), and
a β closer to one. The Gamma-MAP filter with a 7 � 7 kernel has the
lowest Mean Square Error and highest Signal-to-Noise Ratio and β in
compare to other filters with the same kernel. Using the 14 June image
as master, all the Sentinel-1A images were co-registered on-the-fly in
SNAP based on bilinear interpolation, using ground control points
(GCPs). Backscatter is usually expressed in decibel (backscattering co­
efficient). Each Sentinel-1A (both VV and VH) linear to decibel con­
version was applied using equation (2).
σ� (dB) ¼ 10*log10σ� (energy ratio),
Received ​ energy ​ from ​ the ​ sensor ​
where by, Energy Ratio ¼ Energy ​ reflected ​ in ​ an ​ isotropic ​ way ​
The backscattered coefficient can be a positive number if there is a
focusing of backscattered energy towards the radar or a negative num­
ber if there is a focusing of backscattered energy away from the radar
(Bartsch et al., 2012). By performing this operation, the master and slave
images share the same geo-positioning information and have the similar
dimensions.
The pre-processing of Sentinel-2A Level 1C multispectral images
were atmospheric correction using the Fast Line-of-sight Atmospheric
Analysis of Hypercube (FLAASH) in Environment for Visualizing Images
(ENVI) software (version 5.1). This product is composed of 100 km �
100 km tiles in the UTM/World Geodetic System (WGS) 84 projection.
Geometric pixel-to-pixel precision correction was implemented through
the georeferencing model of IMPACT tool v1.3 b (Simonetti et al., 2015).
Pan-sharpening was applied to Sentinel-2A, to downscale the 20-m
bands to 10 m spatial resolution. Pan-sharpening (PANSHARP algo­
rithm) resampling method chosen in our study because it has no dif­
ference with others like Gram-Schmidt (GS), nearest neighbour diffusion
(NNDiffusion), wavelet transformation fusion (WTF) and high-pass filter
fusion (HPF) (Zheng et al., 2017). The output radiometric resolution will
be converted into Byte (8 bit) by using a linear transformation (equation
(3)). Image DN values [0–4095] are converted to TOA Reflectance [0,1]
using the provided conversion factor (10000) and multiplied by 255 to
obtain a byte TOA Reflectance [0,255].
TOA ​ RefByte ​ ¼ ðDN � 0:0255Þ
where DN is raw Digital Number.
3.2. Optimal polarization for mapping
As the Sentinel-1A data are in both single and dual polarization,
proposing the most optimal polarization for mapping the targeted land-
cover type is important in order to reduce data redundancy and
dimensionality. The temporal increase of paddy rice backscatter and a
two-class separability measure were considered to select optimal po­
larization. The optimal polarization is assumed to be the one which
exhibits the most sustained increase in paddy rice backscatter, while at
the same time producing the best results with the two-class separability
Remote Sensing Applications: Society and Environment 18 (2020) 100290
measure. In other words, a polarization has to satisfy both conditions to
be considered optimal or else both polarizations will be used as inputs in
(1) the decision tree classifier.
Two classes are said to be well separated if the distance between their
respective class means is larger than that of their standard deviations
(Choudhury and Chakraborty, 2005). The separability criteria between
classes a and b are measured by the Bahattacharrya distance as equation
(4).
jμa μb j
Sab ¼ (4)
Sa þ Sb
where μ and S are the mean and standard deviation of the features to be
separated, respectively.
The Bahattacharya distance is directly related to the upper bound of
the probabilities of the classification errors (Choudhury and Chakra­
borty, 2005). It measures the real values between 0 and 2, where 0 in­
dicates complete overlap between the signatures of the two classes and 2
indicates the complete separation between the two classes (Lin et al.,
2009). The larger the separability values among the class pairs, the
better the final classification results (Wang et al., 2010). A separability
value below 0.8 is regarded not large enough to separate two classes.
Between 0.8 and 1.5, the quality of the separation between the two
classes is regarded as average. Values above 1.5 indicate almost com­
plete separability between classes (Shi et al., 1994).
(2)
3.3. CART model
3.3.1. Model description
This algorithm identifies optimal polarization and generates Land
Use Land Cover (LULC) map using a stacked Sentinel-1A image. CART is
inbuilt method, often described in graphical terms, a node is typically
shown growing upside down, beginning at its root. Nodes (trees)
partition the explanatory variables in a series of leaves that contain the
most homogeneous collection of outcomes possible. Trees are typically
fitted via binary recursive partitioning. One explanatory variable makes
a single split. The location of the split and the variable are chosen to
reduce the impurity of the node at that point. The learning and test
samples are formed by collecting two independent data sets, which are
319 and 292, respectively. CART divided the data of major land-cover
types in to learn and test data as shown in Table 1 in number and per­
centage of collected learn and test data.
3.3.2. Identification of optimal node of sentinel 1 RADAR images
The optimal tree (node) is a tree in the pruned sequence that achieves
minimum cost of test data in CART model. This is because test mis-
classification cost measurement is subject to sampling error, uncer­
tainty always remains regarding which tree in the pruning sequence is
optimal (Batterham et al., 2009). The sequence starts at the root node,
which is CART evaluates all possible splits of all predictor variables and
choices the “best” single split overall. The data is then partitioned ac­
cording to that best split. To all cases placed in a terminal node, all
(3) subsequent nodes apply the same process. Because the final partitions do
not overlap, each case can only be in one terminal node. The “best” split
is the one that reduces impurity the most (Li et al., 2012).
Identification optimal nodes are important to overcome the problem
of overlapping and usually increased the overall accuracy in the class of
interest. To identify the optimal node, the collected data divided into 10
nodes, which are the best nodes after using Gigi index, based on 6-day
imaging dates. The Gini index impurity measure at a node t is defined
as shown in equation (5).
X �i�
iðtÞ ​ ¼ C p ​ ði = tÞ ​ pðj = tÞ (5)
i;j
j
where C (i/j) is the cost of a mis-classifying a class j case as a class i case
and p (j/t) the probability of a case in class j given that it falls into node t.
4
T. Talema and B.T. Hailu Remote Sensing Applications: Society and Environment 18 (2020) 100290

Table 1
Learn and test data derived from CART model.
Classes Rice Water Built Trees Non-Rice crops

Sample Learn Test Learn Test Learn Test Learn Test Learn Test

No. 93 76 38 40 70 70 66 60 52 46
in % 55 45 48.7 51.3 50 50 52.4 47.6 53.1 46.9

The Gini splitting criterion is the decrease of impurity defined as
stated in equation (6).
Δiðs; tÞ ​ ¼ iðtÞ ​ ​ PL i ​ ðtL Þ ​ PR i ​ ðtR Þ
where PL and PR are probabilities of sending a case to the left child node
tL and to the right child node, respectively (equation (7)).
PL ¼ PðtL Þ = PðtÞ and PR ¼ P ðtR Þ = PðtÞ
where p(t) is the probability of a case in node t.
Based on the optimal node threshold number, a decision tree clas­
sification algorithm was applied in ENVI 5.1 software because it is less
sensitive to non-linearities in the input data than classification methods
that require an assumption of Gaussian distributions. A decision tree
employs tree-structured rules which recursively divide the initial input
dataset into increasingly homogeneous subsets based on well-defined
splitting criteria (Wang et al., 2010). At each split or node, the values
of each explanatory variable are examined and a particular threshold of
a single variable that produces the largest reduction in a deviance
measure is chosen to partition the data (Rogan et al., 2003). Variable
importance analysis is used to know which imaging date better separate
rice from other land-cover types and to rank the six-imaging date based
on their separating importance. The variable importance of each
six-imaging date of Sentinel 1 RADAR data Xk is measured by equation
(8).
X other land-cover. All 8 images in the co-registered layer stack is used in
M ​ ðXk Þ ​ ¼ Δið ​ s�k ; tÞ (8)
t ​ 2T
where s�k is a surrogate split on Xk for s� (the best split of t into the left
child node (tL ) and the right child node (tR )).
3.4. Integrating optical indices with Sentinel-1A images
From Sentinel-2A optical cloud masked image NDVI and MNDWI
were computed to integrate with the temporal Sentinel-1A images. NDVI
has been widely applied in monitoring density and intensity of green
vegetation growth using the spectral reflectivity of solar radiation.
Green vegetation leaves demonstrate better reflection in the NIR band
than in visible. When leaves are water stressed, diseased, or dead, they
become more yellow and reflect significantly less in the near-infrared
range (Mansaray et al., 2017). NDVI from the Sentinel-2A data was
derived using equation (9).
ρnir –ρred
NDVI ​ ¼ ​
ρnir þ ρred
where ρnir and ρred correspond to near-infrared and red spectral bands of
Sentinel 2, respectively.
The MNDWI (equation (10)) is a modified NDWI, which is an index
to extract water bodies and wetlands (Ahmed and Akter, 2017). MNDWI
can not only enhance information about water and restricting the one
from vegetation and soil but can also significantly distinguish built-up
features from water bodies (Mansaray et al., 2017).
ρgreen –ρswir
MNDWI ​ ¼ ​ (10)
ρgreen þ ρswir
where ρgreen and ρswir are the green and shortwave infrared bands,
respectively.
In the current study, it is proposed that the integration of the
aforementioned optical indices with temporal Sentinel-1A images could
provide additional information that would improve overall classification
(6)
accuracy. The six Sentinel-1A images from the optimal polarization and
the NDVI and MNDWI images were finally layer stacked via bilinear
interpolation to produce an 8-band dataset as follows: 14 June image
(b1), 20 July Image (b2), 1 August image (b3), 25 August image (b4), 30
(7) September image (b5), 12 October image (b6), 09 October NDVI image
(b7) and 09 October MNDWI image (b8). Based on the TB of the
investigated land-cover classes at the optimal polarization and the
spatial distribution of NDVI and MNDWI values, a decision rule mapping
algorithm is proposed and validated in this study.
3.5. Mapping algorithm
The classification rule was based on expert knowledge of class sig­
natures and data characteristics using the per-pixel information in each
image, with its temporal behavior, and the corresponding TB co­
efficients (dynamic range, minimum and maximum) and the single-date
optical indices in each of the land-cover types (Mansaray et al., 2017).
NDVI threshold values of rice, trees, and non-rice crops were developed
using CART model based on training data collected.
MNDWI and NDVI threshold values for water and built was adopted
from (Mansaray et al., 2017) to discriminate paddy rice fields from the
the decision rule. Finally, all pixels which do not meet the classification
criteria were classified as others. Finally, a decision tree classification
algorithm is implemented in this study using ENVI 5.1 software to map
land-cover classes by integrating Sentinel-1A and Sentinel-2A optical
indices. Results obtained from this algorithm were validated via
F1-macro accuracy tests.
4. Results
4.1. Temporal backscatter of rice and other land-cover classes
Table 2 show the average temporal radar backscatter coefficients and
their respective standard deviations at VH polarizations for the rice class
and other LULC classes. The VH polarization shows the most consistent
temporal increase in paddy rice backscatter apart from August 01, 2017
image date. There is an increase in backscatter until July 20, 2017 and a
sharp decline on the August 1, 2017 at VH polarization due to high
flooding condition (Fig. 3). The VV polarization shows relatively linear
(9)
backscatter value in all imaging dates. After August 01, 2017 imaging
date, VH polarization shows an increase in backscatter. The VH polari­
zation is to provide a better discrimination of rice from other land-cover
types because TB profiles are generally stable during the paddy rice
growing season in the area under investigation.
4.2. Imaging date importance
The VH polarization image of 01 August has high importance in
separating the different land-cover classes (100%), which is developed
using the CART model, followed by 25 August and July 20, 2017 with
89.87% and 85.63%, respectively (Fig. 4F). The Sentinel 1 RADAR VH
polarization image of 14 June, 30 September and October 12, 2017 has
relatively low importance to separate land-cover classes. Based on the

5
T. Talema and B.T. Hailu Remote Sensing Applications: Society and Environment 18 (2020) 100290

Table 2
Temporal average backscatter coefficients for Rice and other land-cover classes at VH polarization.
Rice Non-Rice crops Water Built up Trees

Image μ s μ s μ s μ s μ s

June 14, 2017 20.3 18.7 1.5 19.0 3.8 15.3 3.0 15.6 2.3
July 20, 2017 19.3 2.5 16.0 1.5 24.8 4.1 14.0 2.0 15.0 2.0
August 1, 2017 49.0 3.0 43.6 1.3 53.2 3.4 42.0 1.8 42.4 2.1
August 25, 2017 18.1 2.0 15.5 1.4 25.8 3.9 14.1 2.1 15.0 1.9
September 30, 2017 16.6 1.4 16.0 1.5 24.6 3.7 14.6 1.7 15.0 1.9
October 12, 2017 16.8 1.9 15.4 1.3 22.2 5.4 14.5 2.1 15.4 1.8

μ and s are the mean and standard deviation, respectively, of backscatter (dB) at corresponding class and date.

Fig. 3. The average temporal backscatter profiles of A) rice class in VV and VH polarizations and B) rice class with the other land-cover at VH polarization.

variable importance identified, the land-cover type mapped using of rice class with non-rice crops, water, built up and trees were measured
Sentinel 1 RADAR VH polarization images in the subsequent imaging using Bahattacharya distance, as shown in Table 3. The minimum
date have a different degree of mis-classification errors. separability has been observed in VV polarization. Overall, the separa­
bility among various rice and other non-rice classes is low to very high.
The class pair “rice and non-rice crops” has 0.5 Bahattacharya distance,
4.3. Optimal node of sentinel 1 RADAR VH images
which is a very low separability as the non-rice crops class is a combi­
nation of wheat, maize, sorghum, teff, horticulture, etc., which generally
In Fig. 4, Nodes 3 and 2 are the 14 June and 01 August, respectively,
exhibit similar (low) backscatter values especially at the start of the rice
of Sentinel 1 RADAR VH image optimal nodes that differentiate rice
growing season. The rice class with trees and built up classes has low
from the other land-cover classes, which is in yellow boxes. Node1 (25
separability, 0.7 and 0.8, respectively (Table 3).
August) was for water and Node 8 (12 October) and Node 10 (25 August)
In Table 3, it can also be observed that with exception of the class
was used for the mapping built up. Node 4 (14 June), Node 6 (25
pairs “built up and trees”, the VH polarization produces better inter-class
August) and Node 9 (25 August) were used for the tree mapping and
signature separability. Based on the most sustained increase in paddy
Node 7 (25 August) and Node 5 (14 June) were used for the mapping of
rice backscatter at the VH polarization, and the superiority of the VH
non-rice crops.
over the VV channel in the two-class separability measure, the VH po­
Fig. 5 shows the optimal nodes that separate rice and other land-
larization as the optimal polarization and with which the optical indices
cover types. The optimal node values identified for the rice mapping
are combined for the mapping of rice fields and the other land-cover
are all pixels less than or equal to -20.14 dB b1 (14 June) and -46.3 dB b3
categories in this study. It has also been observed that C-band, cross
(01 August). Water the identified optimal node values are all pixels less
(VH) polarization produces better crop classification results than the like
than or equal to -24.63 dB at 25 August dates of imaging (b4). All pixels
(VV) polarization, and the use of dual polarization does not further
greater than -13.63 dB b4 (25 August) and -14.26 dB b6 (12 October)
improve the classification results.
were identified for the mapping of built up. This class exhibited the
highest TB values because of corner reflections from buildings and other
vertical objects. For the mapping of trees, all pixels greater than -20.14
4.5. Rice and the other land-cover map from Sentinel-1A
dB b1 (14 June), -15.37 dB b4 (25 August) and -14.26 dB b6 (12
October) were the optimal node values. All pixels less than or equal to
Fig. 6 shows rice and other LULC map of the study area based on 01
-17.52 dB b1 (14 June) and -15.37 dB b4 (25 August) were identified as
August Sentinel 1 RADAR VH images. Based on the optimal node
optimal node values for non-rice crops. Finally, pixels that do not meet
identified in Fig. 4, paddy rice identified as more accurate in this im­
the classification criteria outlined above, were classified as others like
aging date. In addition, the variable importance in Fig. 4F shows the 1st
roads, barren, grasslands, and other low biomass species.
of August VH image has a higher percentage in separating the different
land-cover types. From the total area of the wereda, rice field accounts
4.4. Separability measures using Bahattacharya distance 19,846.9 ha. CART model in 01 August Sentinel 1 RADAR VH image.
The accuracy (F1-macro) of mapping using Sentinel 1 RADAR VH im­
The determination of the optimal polarization channel based on two- ages developed by CART is 0.73 and 0.71 of learning and test data,
class signature separability measures are shown in Table 3. Separability respectively as shown in Table 4.

6
T. Talema and B.T. Hailu Remote Sensing Applications: Society and Environment 18 (2020) 100290

Fig. 4. Optimal nodes identified in CART model (yellow boxes) for A) Rice, B) Water, C) Built up, D) Trees and E) Non-Rice crops, F) The optimal variable importance
in terms of percentage. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

4.6. Mapping rice and non-rice from Sentinel-1A and Sentinel-2A
Results from both the temporal VH images of Sentinel-1A and the
single-date NDVI and MNDWI from Sentinel-2A show greater class ho­
mogeneity as much of the mis-classified, island or mixed pixels have
been eliminated compared to maps using Sentinel-1A VH images
(Fig. 7.). Rice is more clearly delineated. The area covered by trees has
significantly reduced as a consequence of the increase in the area
covered by the class others. In addition, the area covered by built up and
non-rice crops have significantly increased as a consequence of the
decreased in the area covered by the class trees. There is an increase in
the precision of built up and trees see Table 5. The accuracy report
developed from mapping results from integrated Sentinel 1 RADAR VH
and Sentinel 2 optical indices increased F1-macro from 0.71 to 0.79.
From the total area of the wereda, paddy rice field covers 19,157.8 ha.
Paddy rice cultivation is mostly concentrated in the northwestern part of
the study area, which mainly includes the lowland paddy rice.

7
T. Talema and B.T. Hailu Remote Sensing Applications: Society and Environment 18 (2020) 100290

Fig. 5. CART decision tree classifier.

land-cover at the VH polarization. It can be clearly seen that, apart from

Table 3
the rice class, the other land-cover classes, especially built up and trees
Comparison of two-class separability results using Bahattacharya distance at VV
show higher overlap TB profiles to all paddy rice growing season. In
and VH polarizations.
addition, mapping and monitoring of rice after 01 August VH images
LULC VV Polarization
was not possible, due mainly to high overlap of rice backscatter with the
Rice Water Built Trees Non-Rice other land-cover classes.
up crops The first paddy rice mapping results presented are only based on 14
VH Rice 1 0.8 0.7 0.5 June VH image from Sentinel-1A data, which is during the rice sowing
Polarization Water 1.6 1.3 1.4 1.4 period. This mapping was employed using the CART classification
Built up 1.5 2.2 0.4 0.4
method, the paddy rice mapping during sowing exclude late planting
Trees 1.3 2 0.3 0.3
Non-Rice 1.3 2 0.8 0.8 rice. The resulting land use map from this approach shows high mis-
crops classifications of rice and the other land-cover types. The second
paddy rice mapping presented in this study is only based on 01 August
VH images from Sentinel-1A data. Mapping paddy rice using 20 July, 25
5. Discussion August, 30 September and 12 October was not possible, mainly due to
high overlap between rice and the other land-cover types in the study
In this study, six temporal images of Sentinel-1A data have area. The paddy rice mapping during high flooding stage includes late-
employed, covering a part of the rice-growing season (sowing to planted rice. Even though most validation rice crop pixels were correctly
ripening) in which a marked difference in backscatter profiles are most classified using 01 August VH images, there were some commission
pronounced. In this study, the observed TB values at the VV Sentinel-1A errors in this land use type. Several water pixels have been mis-classified
polarization are higher for all land use types, this work is comparable as rice. An increase in surface roughness, which could be brought by the
with investigations done by (Wang et al., 2010; Wu et al., 2011). moving wind, could increase the backscatter profile of a given pixel
However, the VH Sentinel-1A polarization showed the most sustained between Sentinel-1A acquisitions. High confusions also exist between
increase in backscatter profile of rice compared to VV except 01 August built up and trees, this indicates that the application of Sentinel-1A
high flooding imaging date. VH backscatter of rice sharply decreased images for the identification of built up and trees is poor in this study
during the 1st of August imaging date, increasing sharply on 25 August area. This is due to the material used for the construction of built up in
and decreasing slightly at the ripening growth stage. The VV channel the study area, which is mostly mud and may result in similar back­
shows a relatively linear backscatter profile of rice at periods equivalent scatter value like trees.
to the rice growing stage. The application of Sentinel-1A images for mapping and monitoring
In determining the optimal polarization based on the temporal in­ of built up, trees and non-rice crops is not good in the result found in this
crease in paddy rice backscatter, backscatter profiles of points were study area. On maps prepared by Sentinel-1A shows the large area
critically examined at both VV and VH polarizations. In any rice field, covered by trees, a condition that is very unlikely in the study area. To
lowest backscatter profiles are recorded in the sowing and flooding reduce some of these confusions, we combined the temporal VH
stages of rice. From Table 3, it can be observed that the TB values at the Sentinel-1A data with the NDVI and MNDWI indices. The introduction of
VV polarization is higher than that at VH polarization at all imaging these indices into the mapping algorithm brought an increase of the
dates, in agreement with investigations done by (Nguyen et al., 2015). accuracy from 0.71 to 0.79. In the rice land use types, these optical
Fig. 4B shows the average TB profiles of rice class and the other broad indices increase the TP from 62 to 70. It is assumed that on October 09,

8
T. Talema and B.T. Hailu Remote Sensing Applications: Society and Environment 18 (2020) 100290

Fig. 6. Rice and the other land-cover map derived only from VH images of Sentinel 1RADAR data.

Table 4
Confusion Matrix - Learn and Test classification with Sentinel-1A VH images.
Rice Water Built Trees Non-Rice crops Row Total

Classes Learn Test Learn Test Learn Test Learn Test Learn Test Learn Test

Rice 74 62 3 3 0 1 1 0 0 2 78 68
Water 0 0 33 37 0 0 0 0 0 0 33 37
Built up 0 0 0 0 36 34 9 14 1 3 46 51
Trees 1 2 0 0 24 23 49 35 6 5 83 68
Non-Rice crops 18 12 2 0 10 12 7 11 45 36 82 71
CT 93 76 38 40 70 70 66 60 52 46 319 292
TP 74 62 33 37 36 34 49 35 45 36
FP 19 14 5 3 34 36 17 25 7 10
FN 4 6 0 0 10 17 34 33 37 35
Recall 0.95 0.91 1 1 0.78 0.67 0.59 0.51 0.55 0.5
Precision 0.80 0.82 0.87 0.93 0.49 0.49 0.74 0.58 0.86 0.78
F1 Score 0.87 0.86 0.93 0.96 0.6 0.57 0.66 0.54 0.67 0.6

Learn: MAP, MAR and F1-macro are 0.70, 0.77 and 0.73, respectively.
Test: MAP, MAR and F1-macro are 0.72, 0.71 and 0.71, respectively.
CT, TP, FN, FP, MAP and MAR are Column Total, Macro Average Precision, Macro Average Recall, True Positive, False Negative, and False Positive, respectively.

rice plants have acquired the surface area for NDVI to be greater than or
equal to 0.1, but less than 0.3. Therefore, we used NDVI greater than or
equal 0.5 that slightly reduces the precision errors in trees and hence, an
overall slight increase in the classification accuracy of trees and built up
is shown in Table 5. With the integration of optical indices, there is an
increase in the accuracy of rice compared to test data on optimal
Sentinel-1A accuracy. Thus, this research study proposed that if a pixel
meets the TB value criterion, that could be classified as rice, while
having NDVI values greater than or equal to 0.1, but less than 0.3. The
integration of Sentinel-1A VH images with optical image products (NDVI
and MNDWI) creates a considerable increase in the accuracy of rice land
use class and the decrease of errors in non-rice land use classes. This also
brings a high increase in F1-macro land-cover classification accuracy.
6. Conclusion
In this study, we showed the operational applicability of integrating
the new Sentinel-1A data with Sentinel-2A optical indices for mapping
rice fields in tropical area, Fogera wereda, Ethiopia. The VH polarization
of Sentinel-1A is suitable for rice field mapping as it exhibited the most
sustained increase in the TB of rice, relative to the nearly uniform TB
profiles of the other land-cover classes. It is also better for inter-class
discrimination based on a two-class separability measure.
CART model is robust in identifying the optimal node, which B5,
extracting paddy fields that covers an area of 19,892.5 ha, and LULC
classification with an F1-macro of 0.71 in test data validation. The
introduction of NDVI and MNDWI from Sentinel 2 A satellite image
increases the F1-macro accuracy from CART test validation of Sentinel-
1A VH image data to 0.79. The increase in rice accuracy compared to

9
T. Talema and B.T. Hailu Remote Sensing Applications: Society and Environment 18 (2020) 100290

Fig. 7. Simplified rice and non-rice map of study area derived from Sentinel-1A VH and Sentinel-2A optical images.

Declaration of competing interest

Table 5
Error matrix of accuracy assessment of classification with Sentinel-1A VH and
The authors have no any conflict of interest.
Sentinel-2A optical.
Classes Rice Water Built Trees Non-Rice Others Row
crops Total
Acknowledgment

Rice 70 0 0 0 0 0 70
Authors acknowledge to the European Space Agency (ESA) for
Water 1 38 0 0 0 2 41
Built up 0 0 51 8 1 5 65 providing Sentinel 1 SAR and Sentinel 2 MSI images free of charge in
Trees 0 0 8 42 2 6 58 their data hub. Dr. Binyam Tesfaw Hailu would like to acknowledge
Non-Rice 3 0 5 4 39 5 56 Georg Fo €rster Post-doctoral Fellowship program in Alexander Von
crops
Humboldt Foundation.
Others 2 2 6 6 4 32 52
CT 76 40 70 60 46 50 342
TP 70 38 51 42 39 32 Appendix A. Supplementary data
FP 6 2 19 18 7 18
FN 0 3 14 16 17 20
Supplementary data to this article can be found online at https://doi.
Recall 1 0.93 0.78 0.72 0.70 0.62
Precision 0.92 0.95 0.73 0.70 0.85 0.64 org/10.1016/j.rsase.2020.100290.
F1 Score 0.96 0.94 0.75 0.71 0.77 0.63

MAP, MAR and F1-macro are 0.79, 0.80 and 0.79, respectively.
CT, TP, FN, FP, MAP and MAR are Column Total, Macro Average Precision,
Macro Average Recall, True Positive, False Negative, and False Positive,
respectively.
CART test validation confirms that the operational applicability of
integrating the new Sentinel-1A data with optical data for rice mapping
in tropical and highly heterogeneous landscapes.
Ethical statement
1. This manuscript has not been published in whole or in part else­
where and it is not currently being considered for publication in another
journal;
2. All authors have been personally and actively involved in sub­
stantive work leading to the manuscript, and will hold themselves
jointly and individually responsible for its content.
References
Ahmed, K.R., Akter, S., 2017. Analysis of landcover change in southwest Bengal delta
due to floods by NDVI, NDWI and K-means cluster with landsat multi-spectral
surface reflectance satellite data. Remote Sens Appl Soc Environ 8, 168–181.
Bartsch, A., Trofaier, A.M., Hayman, G., Sabel, D., Schlaffer, S., Clark, D.B., Blyth, E.,
2012. Detection of open water dynamics with ENVISAT ASAR in support of land
surface modeling at high latitudes. Bio-geosci. 9, 703–714.
Batterham, P.J., Christensen, H., Mackinnon, A.J., 2009. Modifiable risk factors
predicting major depressive disorder at four-year follow-up: a decision tree
approach. BMC Psychiatr. 9, 1–15.
Bouvet, A., Le Toan, T., 2011. Use of ENVISAT/ASAR wide-swath data for timely rice
fields mapping in the Mekong River Delta. Remote Sens. Environ. 115, 1090–1101.
Campos-Taberner, M., García-Haro, F., Camps-Valls, G., Grau-Muedra, G., Nutini, F.,
Busetto, L., et al., 2017. Exploitation of SAR and optical Sentinel data to detect rice
crop and estimate seasonal dynamics of leaf area index. Rem. Sens. 9 (3), 248.
Choudhury, I., Chakraborty, M., 2005. SAR signature investigation of rice crop using
RADARSAT data. Int. J. Rem. Sens. 3, 519–534.
Dawe, D., Pandey, S., Nelson, A., 2010. Emerging trends and spatial patterns of rice
production. In: Rice in the Global Economy: Strategic Research and Policy Issues for
Food Security. International Rice Research Institute (IRRI), Los Ba~ nos, Philippines,
pp. 1–35.

10
T. Talema and B.T. Hailu
Elachi, C., 1988. Spaceborne Radar Remote Sensing: Applications and Techniques. IEEE
Press, 1988.
Inglada, J., Vincent, A., Arias, M., Marais-Sicre, C., 2016. Improved early crop type
identification by joint use of high temporal resolution SAR and optical image time
series. Rem. Sens. 8, 1–21.
Inoue, Y., Sakaiya, E., Wang, C., 2014. Capability of C-band backscattering coefficients
from high-resolution satellite SAR sensors to assess biophysical variables in paddy
rice. Remote Sens. Environ. 140, 257–266.
Koppe, W., Gnyp, M.L., Hutt, C., Yao, Y., Miao, Y., Chen, X., Bareth, G., 2013. Rice
monitoring with multi-temporal and dual-polarimetric TerraSAR-X data. Int. J. Appl.
Earth Obs. Geoinf. 21, 568–576.
Kuenzer, C., Knauer, K., 2013. Remote sensing of rice crop areas. Int. J. Rem. Sens. 34,
2101–2139.
Li, C., Gluer, C.C., Eastell, R., Felsenberg, D., Reid, D.M., Rox, D.M., Lu, Y., 2012. Tree-
structured subgroup analysis of receiver operating characteristic curves for
diagnostic tests. Acad. Radiol. 19, 1529–1536.
Lin, H., Chen, J., Pei, Z., Zhang, S., Hu, X., 2009. Monitoring sugarcane growth using
ENVISAT ASAR data. IEEE Trans. Geosci. Rem. Sens. 47, 2572–2580.
Mansaray, L.R., Huang, W., Zhang, D., Huang, J., Li, J., 2017. Mapping rice fields in
urban shanghai, southeast China, using sentinel-1A and landsat 8 datasets. Rem.
Sens. 9, 1–23.
Nelson, A., Setiyono, T., Rala, A., Quicho, E., Raviz, J., Abonete, P., et al., 2014. Towards
an operational SAR-based rice monitoring system in Asia: examples from 13
demonstration sites across Asia in the RIICE project. Rem. Sens. 6 (11),
10773–10812.
Nguyen, B.A., Clauss, K., Cao, S., Naeimi, V., Kuenzer, C., Wagner, W., 2015. Mapping
rice seasonality in the mekong delta with multi-year envisat ASAR WSM data. Rem.
Sens. 7, 15868–15893.
Nguyen Lam-Dao, M., 2009. Rice Crop Monitoring Using New Generation Synthetic
Aperture Radar (SAR) Imagery. Unpublished Ph.D. Thesis. University of Southern
Queensland, Queensland, Australia, pp. 1–163.
Reddy, S.R., 2004. Agronomy of Field Crops. Department of Agronomy ANGR
Agricultural University SV Agriculture College, India, pp. 1–136.
Rogan, J., Miller, J., Stow, D., Franklin, J., Levien, L., Fischer, C., 2003. Land-cover
change monitoring with classification trees using Landsat TM and ancillary data.
Photogramm. Eng. Rem. Sens. 69, 793–804.
11
Remote Sensing Applications: Society and Environment 18 (2020) 100290
Sakamoto, T., Sprague, D.S., Okamoto, K., Ishitsuka, N., 2018. Semi-automatic
classification method for mapping the rice-planted areas of Japan using multi-
temporal Landsat images. Remote Sens Appl Soc Environ 10, 7–17.
Shen, S., Yang, S., Li, B., Tan, B., Li, Z., Le Toan, T., 2009. A scheme for regional rice yield
estimation using ENVISAT ASAR data. Sci. China Earth Sci. 52, 1183–1194.
Shi, J., Dozier, J., Rott, H., 1994. Snow mapping in alpine regions with synthetic aperture
radar. IEEE Trans. Geosci. Rem. Sens. 32, 152–158.
Simonetti, D., Marelli, A., Eva, H.D., 2015. IMPACT: Portable GIS Toolbox Image
Processing and Landcover Mapping. Publication Office of the European Union,
Luxembourg, ISBN 978-92-79-50115-9. EUR 27358 EN.
Skolnik, M.I., 2008. Radar Handbook. The McGraw-Hill, USA, New York, pp. 1–1351.
Takele, A., 2010. Analysis of Rice Profitability and Marketing Chain: the Case of Fogera
Woreda, South Gonder Zone, Amhara National Regional State, Ethiopia.
Unpublished MSc Thesis, Haramaya University, Haramaya, Ethiopia, 1–118.
Torbick, N., Chowdhury, D., Salas, W., Qi, J., 2017. Monitoring rice agriculture across
Myanmar using time series Sentinel-1 assisted by Landsat 8 and PALSAR 2. Rem.
Sens. 9, 1–22.
Veloso, A., Mermo, S., Bouvet, A., Le Toan, T., Planells, M., Dejoux, J.F., Eric, C., 2017.
Understanding the temporal behavior of crops using Sentinel-1 and Sentinel-2-like
data for agricultural applications. Remote Sens. Environ. 199, 415–426.
Wang, D., Lin, H., Chen, J., Zhang, Y., Zeng, Q., 2010. Application of multi-temporal
ENVISAT ASAR data to agricultural area mapping in the Pearl River Delta. Int. J.
Rem. Sens. 31, 1555–1572.
Wu, F., Wang, C., Zhang, H., Zhang, B., Tang, Y., 2011. Rice crop monitoring in South
China with RADARSAT-2 quad-polarization SAR data. Geosci. Rem. Sens. Lett. IEEE
8, 196–200.
Yonezawa, C., Negishi, M., Azuma, K., Watanabe, M., Ishitsuka, N., Ogawa, S., Saito, G.,
2012. Growth monitoring and classification of rice fields using multitemporal
RADARSAT-2 full-polarimetric data. Int. J. Rem. Sens. 33, 5696–5711.
Zheng, H., Du, P., Chen, J., Xia, J., Li, E., Xu, Z., Li, X., Yokoya, N., 2017. Performance
evaluation of downscaling sentinel-2 imagery for land use and land cover
classification by spectral-spatial features. Rem. Sens. 9, 1274.
Zhang, Y., Wang, C., Wu, J., Qi, J., Salas, W.A., 2009. Mapping paddy rice with
multitemporal ALOS/PALSAR imagery in southeast China. Int. J. Rem. Sens. 30,
6301–6315.