A Novel Cotton Mapping Index Combining Sentinel-1 SAR and Sentinel-2

ISPRS Journal of Photogrammetry and Remote Sensing 181 (2021) 148–166
Contents lists available at ScienceDirect
ISPRS Journal of Photogrammetry and Remote Sensing

journal homepage: www.elsevier.com/locate/isprsjprs
A novel cotton mapping index combining Sentinel-1 SAR and Sentinel-2

multispectral imagery
Lan Xun a, b, Jiahua Zhang a, b, *, Dan Cao a, b, Shanshan Yang a, b, Fengmei Yao b, *
a
Key Laboratory of Digital Earth Science, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China
b
University of Chinese Academy of Sciences, Beijing 100049, China
A R T I C L E I N F O A B S T R A C T
Keywords: Cotton is an important cash crop in the world, as the main source of natural and renewable fiber for textiles.
Cotton Accurate and timely monitoring of the cotton distribution is crucial for cotton cultivation management and
Automatic mapping international trade. However, most of the previous researches on cotton identification using remotely sensed
Cotton Mapping Index
images are highly dependent on training samples, and the collection of samples is time-consuming and expensive.
Sentinel-1
Sentinel-2
To overcome this limitation, a new index, termed as Cotton Mapping Index (CMI), was developed in this study for
automatic cotton mapping using time series of Sentinel-1 synthetic aperture radar (SAR) and Sentinel-2 Multi
spectral Instrument (MSI) satellite data. Four sites in the United States (U.S.) and four sites in China were selected
to develop and assess the performance of the CMI. The spectral characteristics derived from Sentinel-2 and
backscattering coefficients derived from Sentinel-1 for cotton and non-cotton crops during the cotton growth
period were analyzed. Considering the phenology differences of crops in different regions, the features at an
adaptive window were adopted to construct the CMI. The results showed that at the peak greenness period, the
multiplication of red-edge 1 and red-edge 2 band for cotton samples were much larger than those for non-cotton
samples, whereas the spectral angle at the red band as well as the absolute values of backscattering coefficients in
vertical transmit and vertical receive (VV) polarization for cotton samples were much smaller than those for non-
cotton samples. Based on these findings, the CMI was developed to identify cotton cultivated area within the
cropland area. The overall accuracy of classification results for the sites in the U.S. was higher than 81.20%, and
the mean relative error for the sites in Xinjiang of China was 26.69%. The CMI, which incorporated optical and
radar features, had a better performance than the indices using optical features solely. The advantage of the CMI
over supervised classifiers (i.e., k-nearest neighbors, support vector machine and random forest) is that no
training samples are required. Moreover, the cotton distribution map can be obtained before the harvest using
the CMI. These results indicated the potential of the CMI for cotton mapping. The applicability of CMI in other
regions with different cropping systems and crop types needs to be further assessed in the future study.
1. Introduction information of cotton provides the basis for yield estimation, agricul
tural production management and decision making.
Cotton is an important cash crop, which plays an significant role in The traditional methods for the acquisition of crop cultivated area
economic development (Razzaq et al., 2021). As the main natural fiber, are based on field measurement and statistical sampling, which are time-
it accounts for 3 percent of the agricultural area and meets 27 percent of consuming, low efficiency and labour-intensive (Zhang et al., 2014). In
the textile demand in the world (ICAC, 2018). The cotton cultivated area the past few decades, satellite remote sensing has contributed to crop
information is of significance for a series of issues that influence human type mapping on a regional or global scale because of its advantages
well-being, ranging from textiles in the clothing manufacturing industry, such as wide coverage, high integration, fast speed and periodicity
cosmetics, medical gauze, to international trade (Akter et al., 2019; H. Li (Chamorro Martinez et al., 2021). According to the number of images
et al., 2021). The timely and accurate acquisition of spatial distribution used, remote sensing-based studies on cotton identification can be
* Corresponding authors at: Key Laboratory of Digital Earth Science, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094,
China (Jiahua Zhang). University of Chinese Academy of Sciences, Beijing 100049, China (Fengmei Yao)
E-mail addresses: [email protected] (J. Zhang), [email protected] (F. Yao).
https://doi.org/10.1016/j.isprsjprs.2021.08.021
Received 10 May 2021; Received in revised form 7 August 2021; Accepted 27 August 2021
Available online 21 September 2021
0924-2716/© 2021 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.
L. Xun et al. ISPRS Journal of Photogrammetry and Remote Sensing 181 (2021) 148–166
divided into two categories, including the identification based on single- and radar images has a better performance than using single sensor in
temporal and multi-temporal images. Li et al. (2011) extracted the the fields of land cover mapping (Ienco et al., 2019), crop classification
cotton cultivated area by combining the decision tree classifier with the (Orynbaikyzy et al., 2019; Sonobe et al., 2017) and biophysical metrics
spectral characteristics, which were derived from the Huanjing image estimation (J. Wang et al., 2019). Steinhausen et al. (2018) combined
acquired on 1st July in Dingzhuang of China. However, it is difficult to Sentinel-1 and Sentinel-2 images to classify land use and land cover
extend the classification rules and parameters on a large scale, as types in Chennai Basin, and the results showed that the classification
different crop types in different regions may exhibit similar spectral accuracy obtained by the combined images was improved compared
characteristics at the same time. To solve this problem, some researchers with that by Sentinel-2 images solely. Adrian et al. (2021) utilized the
have used multi-temporal images that can reflect crop phenology in backscattering coefficients, texture and spectral features derived from
formation for cotton mapping. For instance, Musande et al. (2012) the multi-temporal Sentinel-1 SAR and Sentinel-2 optical images for
combined multi-temporal indices (i.e., normalized difference vegetation classification, and the results showed that the fused radar and optical
index (NDVI), Transformed Normalized Difference Vegetation Index features achieved a higher overall accuracy for crop type mapping than
(TNDVI), Simple Ratio (SR), Triangular Vegetation Index (TVI) and Soil- those using radar or optical data solely.
Adjusted Vegetation Index (SAVI)) derived from Advanced Wide Field In this study, a new index, called the Cotton Mapping Index (CMI),
Sensor (AWIFS) and Linear Imaging Self Scanner Sensor (LISS-III) im was proposed to map the cotton cultivated area combining Sentinel-1
ages with fuzzy classification for cotton mapping in Aurangabad of SAR and Sentinel-2 Multispectral Instrument (MSI) images. Based on
India. The results showed that the flowering-boll opening and harvesting the preliminary analysis of the optical and radar features of cotton and
stages were the optimum temporal for cotton identification. Based on non-cotton samples during the cotton growth period, three assumptions
the annual time series of vegetation index derived from Moderate Res were proposed to select the optimal features to distinguish cotton from
olution Imaging Spectroradiometer (MODIS) images, Zhong et al. non-cotton crops. Considering the phenology difference in various re
(2011), Conrad et al. (2011) and Xun et al. (2021) combined the gions, an adaptive window was introduced to develop the CMI. The main
phenological metrics, temporal segmentation metrics and harmonic objectives of the current study are (1) to develop a novel index (CMI) for
features with the decision tree classification, classification and regres automatic cotton mapping; (2) to evaluate the capability and robustness
sion tree (CART) as well as fused representation-based algorithm for of the proposed CMI in different study sites; and (3) to assess the per
cotton identification, respectively. Based on the Sentinel-2 images ac formance of the CMI by comparing with the supervised classifiers.
quired from March to November, Huang et al. (2018) combined spectral
bands (i.e., blue, green, red, near infrared, red-edge bands), spectral 2. Materials and methods
indices (i.e., NDVI and Modified Chlorophyll Absorption Ratio Index
(MCARI)) with the random forest classifier to map the spatial distribu The overall workflow is shown in Fig. 1. Firstly, time series of radar
tion of cotton and other crops in Shawan county of China. The results and optical features for cotton and non-cotton samples were derived
showed that involving the red-edge spectral band and related index can from multi-temporal Sentinel-1 SAR and Sentinel-2 MSI satellite data.
improve the classification accuracy. Second, three assumptions were proposed based on the analysis of the
However, all the above-mentioned methods belong to the supervised spectral reflectance, spectral angle and backscattering coefficients of
classifiers, and their accuracy highly depends on the high quality and cotton and non-cotton samples during the cotton growth period. The
sufficient quantity of samples to train the classifier (Zhu et al., 2016). It’s appropriate features were then selected to develop the CMI. Finally, the
usually a time-consuming, expensive and labor-intensive task to collect proposed CMI was used to map the cotton cultivated area within the
field samples (Ashourloo et al., 2019; Zhang and Roy, 2017; Zhong et al., cropland mask, and its’ performance was evaluated and compared with
2019). The index-based mapping method has the advantage of those of the supervised classifiers.
simplicity and efficiency, which has been used in the identification of
water (L. Li et al., 2021), plastic greenhouse (Yang et al., 2017), plastic 2.1. Study sites
materials (Guo and Li, 2020), and urban biophysical compositions (Deng
and Wu, 2012). For specific crop type mapping, some features/indices The phenological period and growth status of each crop are varied in
have also been developed, such as for alfalfa (Ashourloo et al., 2018), different regions due to the influence of crop varieties, environmental
maize (Qiu et al., 2018), canola (Ashourloo et al., 2019), soybeans (Silva conditions and farmers’ decisions (Pan et al., 2012). Considering the
Junior et al., 2020), potato (Ashourloo et al., 2020), and winter wheat availability of field data, a total of eight study sites were selected to test
(Qu et al., 2021). According to the spectral characteristics of red and the potential of the proposed index, including four sites from the United
near infrared bands during the growth periods of alfalfa and potato, States (U.S.), as well as four sites from China. The distribution of these
Ashourloo et al. (2018, 2020) proposed two novel features to identify study sites is shown in Fig. 2.
the alfalfa and potato areas, respectively, and the overall accuracy was In the U.S., cotton is mainly cultivated in the southern region. Texas,
over 90%. Silva Junior et al. (2020) proposed the Perpendicular Crop Georgia, Mississippi, Oklahoma and Arkansas are the top five states with
Enhancement Index (PCEI) for soybean mapping in Mato Grosso of the largest cotton plantations in 2019. Arkansas has a humid subtropical
Brazil, and the high PCEI values indicated the high probability of soy climate, with the average maximum temperature of 34.2 ℃ in summer,
bean areas. In terms of cotton mapping, L. Wang et al. (2019) proposed average minimum temperature of − 3℃ in winter, and annual precipi
the Cotton Extraction Index (CEI) based on a single temporal of Rap tation of 1220 mm. In this study, site A is selected from the northeast of
idEye image, which is the summation of the reflectance of red-edge and Arkansas State, where the land is generally flat and soil is fertile. The
near infrared bands. The result showed that the classification accuracy main crops in this site include corn, cotton, soybeans and rice. According
achieved by the CEI was similar to those by the maximum likelihood and to the Crop Progress and Condition graphical products published by the
random forest classifiers. However, the selection of the acquisition time United States Department of Agriculture (USDA), cotton and soybeans
of satellite image is also critical to calculate the CEI for cotton mapping, are usually planted from April to June/ July, and harvested from
and the differences of crop phenology over large regions will also affect September to November. Corn and rice are usually planted from April to
the classification accuracy using single temporal image. May/ June, and harvested from August to October. In this study, site A
The features/indices proposed by the above-mentioned studies are was used to develop the index for cotton identification, and the other
mainly based on optical images. Synthetic Aperture Radar (SAR) im study sites were used as implementation sites to assess the potential and
agery has posed a key advantage over optical images in crop mapping stability of the proposed index. Site B is selected from the eastern
due to its all-weather acquisition capability (Inglada et al., 2016; Tei Arkansas and western Mississippi, with soybeans, cotton, corn and rice
mouri et al., 2019). Recent studies have shown that combining optical as the main crops. Mississippi has a humid subtropical climate with long,
149
Fig. 1. Flowchart of the identification of cotton cultivated area in this study.
hot and humid summers, as well as short and mild winters. The average Agency (ESA), which can capture images with a spatial resolution of
temperature is around 27 ◦ C in July and 6 ◦ C in January. It is entirely 10–60 m and a revisit period of 5 days. It provides 13 spectral bands
composed of lowlands, with an average elevation of 91 m. Georgia has a ranging from visible to shortwave infrared spectrum. The Level-2A
humid subtropical climate with hot and moist summers and mild win surface reflectance (SR) images were used for cotton identification at
ters. The daytime temperature in summer often exceeds 35 ◦ C, and sites A-D in this study, which have been available from the data cata
winter is characterized by mild temperature. The entire state receives logue of the Google Earth Engine (GEE) since 2019. Among all spectral
moderate to heavy precipitation, varying from 1100 mm to 1900 mm. bands, the 9 bands were used as the optical candidate features for cotton
Site C is selected from the southern region of Georgia State, where the identification in this study, as shown in Table 1. In addition, the QA60
main crops include cotton, peanuts, corn and pecans. Site D is selected band, a bitmask band with cloud mask information, was used to mask
from the southern region of Texas State. It is located in the coastal plain, opaque and cirrus clouds. Bands QA60, B5, B6, B7, B11 and B12 were
and the terrain is flat. This site has a humid subtropical climate. Summer resampled from 60 m/ 20 m to 10 m using the nearest neighbor
is hot and humid, with the average daily temperature above 32.2 ◦ C. The resampling method in the GEE platform.
temperature reaches freezing point only a few times in winter. The main Taking the sites A-D as an example, the number of all available im
crops in this site include cotton, sorghum, corn and wheat. ages with different cloud percentages is shown in Fig. 3. For each region,
Located in northwest of China, Xinjiang Uygur Autonomous Region there were over 50 percent images with the cloud cover less than 50%. In
is the largest cotton production region in China, accounting for 84.94 this study, all available Sentinel-2 observations covering the sites A-C
percent of the national total production in 2019 (National Bureau of during April-November in 2019 with cloud coverage of less than 50%
Statistics of China, 2020). Xinjiang is in the arid and semiarid inland and were considered. For site D, due to the limited number of high-quality
has a typical temperate continental climate characterized by strong images available, the images in 2020 were used. In total, 91, 168, 68
evaporation and rare rainfall, with a mean annual precipitation of and 255 Sentinel-2 images for sites A-D composed the initial image data
100–200 mm. The average temperature in January and July is − 17.1 ◦ C for cotton mapping. The Level-1C top-of-atmosphere (TOA) reflectance
and 27.1 ◦ C, respectively (Wang et al., 2012). To facilitate accuracy images for sites E-H in 2018 were used due to the unavailability of
assessment using statistical data, ten counties from Xinjiang were statistical data in 2019 when we conducted the experiment.
selected and grouped as the four study sites (E-H), including Awat, There was more than one satellite track involved to achieve full
Jiashi, Kuqa, Luntai, Makit, Jinghe, Wensu, Bole, Shule and Yopurga coverage for each study site. The number of images varies for each site,
county (Fig. 2). These sites are dominated by single cropping system, and more observations can be found along the strips where satellite
and the main crops include cotton, wheat, maize, watermelon, grape and orbits overlap each other. In some regions, there is more than one
so on. The growing season of cotton is usually from April to October. observation at different times on the same day, which may lead to dif
ferences in the observed values of spectral bands. In addition, the image
2.2. Data selection concerning cloud cover and cloud masking may also cause an
inconsistency in the time interval of the obtained image. To ensure that
2.2.1. Sentinel-2 images and preprocessing the interval of time series for each pixel is consistent, the composite
The Sentinel-2 Multi-Spectral Instrument (MSI) onboard the images for every 15 days were generated using the median value of all
Sentinel-2A and − 2B satellites is developed by the European Space observations over the 15-day period (Oliphant et al., 2019; Wilson et al.,
150
Fig. 2. Locations of study sites A-D and cotton/ non-cotton samples in the United States and sites E-H in Xinjiang of China as well as their RGB color composites of
red, near infrared and blue. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
151
Table 1 18 vegetation indices (VIs), which were commonly used in literature on

Main parameters of the spectral bands of Sentinel-2 used in this study. cotton studies, were also calculated and selected as the optical candidate
Band Band type Central Spatial Abbreviation used features to develop the cotton identification rules in this study (Table 2).
No. wavelength/ resolution/m in this study
nm 2.2.2. Sentinel-1 images and preprocessing
B2 Blue 490 10 B The Sentinel-1A and Sentinel-1B satellites carry a dual-polarization
B3 Green 560 10 G SAR instrument at C-band, which can capture images with a spatial
B4 Red 665 10 R resolution of 5–40 m and a revisit period of 12 days. Four imaging
B5 Vegetation 705 20 RE1
Red Edge 1
modes, which include Strip Map (SM), Interferometric Wide-swath (IW),
B6 Vegetation 740 20 RE2 Extra Wide-swath (EW), and Wave (WV), are provided. The IW mode
Red Edge 2 provides dual polarization imagery consisting of vertical transmit, ver
B7 Vegetation 783 20 RE3 tical receive (VV) and vertical transmit, horizontal receive (VH). The
Red Edge 3
Ground Range Detected (GRD) products acquired in IW mode were used
B8 Near infrared 842 10 NIR
B11 Shortwave 1610 20 SW1 in this study. The main parameters of the product are shown in Table 3.
infrared 1 The images covering the study area during the study period were filtered
B12 Shortwave 2190 20 SW2 on the GEE, which have been preprocessed to a calibrated, ortho-cor
infrared 2 rected product with a resolution of 10 m using the Sentinel-1 Toolbox
(Jin et al., 2019). The Refined Lee Filter with a window size of 7 × 7 was
applied to reduce noise in the Sentinel-1 images (Yommy et al., 2015).
To keep consistent with Sentinel-2 data, the 15-day composited images
were generated by calculating the median values of all observations over
the 15-day period. In addition to the VV and VH polarizations, their
combinations, such as VH/VV and VV − VH, were calculated and used as
radar candidate features to distinguish cotton from other crops due to
their potential for agricultural applications (Wang et al., 2020).
2.2.3. Reference samples

In this study, the Cropland Data Layer (CDL) from USDA was used as
a reference map to select training and validation samples for the four
study sites in the U.S. It is a raster, geo-referenced, crop-specific land
cover data layer created annually for the continental United States using
Table 3
Main parameters of Sentinel-1 IW GRD product.
Parameter Specification
Fig. 3. The number of images with different cloud percentages for sites A-D.
Wavelength C-band
Frequency 5.405 GHz
2018). Moreover, in some regions, high-quality observations may not be Spatial resolution 10 m
available because of cloud cover or other issues. The linear interpolation Repeat cycle 6 days (two satellites)
was used to fill these gaps using high-quality observation data before Number of looks (range × azimuth) 5×1
Ground range coverage 251.8 km
and after the time step.
Swath width 250 km
Apart from the 9 spectral bands as above-mentioned in Table 1, the
Table 2
The vegetation indices used in this study and their expressions using Sentinel-2 bands.
Spectral index Abbreviation Expression using Sentinel-2 bands Reference
Normalized Difference Vegetation Index NDVI (B8 − B4)/(B8 + B4) (Huete et al., 2002)
Enhanced Vegetation Index EVI 2.5 × (B8 − B4)/(B8 + 6 × B4 − 7.5 × B2 + 1) (Huete et al., 2002)
Infrared Percentage Vegetation Index IPVI B8/(B8 + B4) (Crippen, 1990)
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
Transformed Normalized Difference Vegetation Index TNDVI (B8 − B4)/(B8 + B4) + 0.5 (Tucker, 1979)
Soil Adjusted Vegetation Index SAVI 1.5 × (B8 − B4)/(B8 + B4 + 0.5) (Huete, 1988)
Optimized Soil Adjusted Vegetation Index OSAVI 1.16 × (B8 − B4)/(B8 + B4 + 0.16) (Rondeaux et al., 1996)
Green Atmospherically Resistant Vegetation Index GARI (B8 − (B3 − (B2 − B4)))/(B8 + (B3 − (B2 − B4))) (Gitelson et al., 1996)
Green Normalized Difference Vegetation Index GNDVI (B8 − B3)/(B8 + B3) (Gitelson and Merzlyak, 1998)
Sentinel-2 Red-Edge Position S2REP 705 + 35 × (((B4 + B7)/2) − B5)/(B6 − B5) (Frampton et al., 2013)
Inverted Red-Edge Chlorophyll Index IRECI (B7 − B4)/(B5/B6) (Frampton et al., 2013)
Red Edge Inflection Point REIP 700 + 40 × (((B4 + B7)/2) − B5)/(B6 − B5) (Major, 1988)
MERIS Terrestrial Chlorophyll Index MTCI (B6 − B5)/(B5 − B4) (Dash and Curran, 2007)
Pigment Specific Simple Ratio PSSRa B7/B4 (Blackburn, 1998)
Modified Chlorophyll Absorption Ratio Index MCARI ((B5 − B4) − 0.2 × (B5 − B3))(B5/B4) (Daughtry et al., 2000)
Normalized differential senescent vegetation index NDSVI (B11 − B4)/(B11 + B4) (Yin et al., 2020)
Normalized difference residue index NDRI (B4 − B12)/(B4 + B12) (Zhang et al., 2020)
Normalized difference tillage index NDTI (B11 − B12)/(B11 + B12) (Yin et al., 2020)
Land surface water index LSWI (B8 − B11)/(B8 + B11) (Yin et al., 2020)
152
moderate resolution satellite imagery and extensive agricultural ground 2.3. Developing the CMI for cotton identification
truth (https://nassgeodata.gmu.edu/CropScape/). It includes crop type
labels and is often used as a reference map for crop classification (Cai Taking site A as an example, this section mainly describes the process
et al., 2018; Sun et al., 2019; Zhong et al., 2019). The spatial resolution of developing the CMI. The time series curves of spectral reflectance,
of the CDL data during our study periods is 30 m. The CDL data for vegetation index, spectral angle, and backscattering coefficients of cot
Arkansas, Mississippi and Georgia in 2019 as well as for Texas in 2020 ton and other crops during the cotton growth period were preliminarily
have high self-reported accuracies for cotton class, with the user’s and analyzed in Section 2.3.1–2.3.3, respectively. The three assumptions
producer’s accuracy higher than 84.1% and 89.1%, respectively. were then proposed in Section 2.3.4 to select the optimal features to
Moreover, the CDL data provide a confidence layer that reflects the per- construct the CMI.
pixel prediction confidence for a given classification, where 0 represents
the least confident and 100 represents the most confident. In this study, 2.3.1. Spectral reflectance of cotton and other crops
the cotton and non-cotton samples in sites A-D were randomly selected Phenology information is a significant characteristic of the crops,
from the region with the confidence value higher than 90%. which can be exhibited in the time series of remotely sensed images and
plays a valuable role in crop classification (Zhong et al., 2016). The
2.2.4. Statistical data mean spectral reflectance and NDVI curves for cotton, corn, rice, soy
For Xinjiang, the county-level statistical cotton areas in 2018 were beans and others during April-November in site A are shown in Fig. 4,
collected from the Xinjiang Statistical Yearbook, which were used to which are calculated from the 100 samples for each class. It can be seen
compare with the detected areas of the mapping results and to assess the that the differences between cotton and the other crops were not obvious
performance of the proposed CMI. in the visible spectral bands. The peak values of the RE2, RE3 and NIR of
the cotton samples are higher than those of the other crops, which
2.2.5. Cropland mask correspond to the time when the peak of NDVI occurs.
To rule out the interference of non-croplands areas (i.e., grass, forest,
water, building bare land and others), the input satellite data were 2.3.2. Spectral angle of cotton and other crops
masked by the cropland layer. For the study sites of the U.S., a binary In addition to spectral reflectance, the relationship between spectral
cropland mask was generated using the cultivated layer derived from bands, such as spectral angle, is also an important indicator in vegeta
CDL data and resampled from the original spatial resolution (30 m) to tion remote sensing research (Khanna et al., 2007). The angle index can
10 m. For the study sites of Xinjiang, the cropland map with a spatial not only capture the magnitude of spectral reflectance as the spectral
resolution of 10 m was extracted from the Finer Resolution Observation index does, but also is related to the reflectance of adjacent bands
and Monitoring of Global Land Cover (FROM-GLC10) datasets (Gong (Palacios-Orueta et al., 2012). The spectral angle can reflect the relative
et al., 2019). relationship of the three bands and summarize the shape of the reflec
tance spectrum over the wavelength range of these bands. In the
calculation of angles, the center of each spectral band is considered to be
the vertex of a triangle, so the angle (α) at a certain vertex can be
Fig. 4. Temporal profiles of spectral reflectance (a-i) and NDVI (j) of cotton, corn, rice, soybeans and others from April to November. The shading represents the
standard deviation of each temporal profile for cotton samples.
153
growth period are shown in Fig. 5. In the visible bands, the spectral
reflectance of cotton showed a decreasing trend, while in the near
infrared band, the spectral reflectance increased first and then
decreased. The spectral angles at green, red, red-edge 1, red-edge 2, red-
edge 3, near infrared, and shortwave infrared 1 bands for cotton and
other crops during the cotton growth period are shown in Fig. 6. It can
be seen that the spectral angles at green, red, and red-edge bands for the
crops showed a trend of decreasing first and then increasing during the
cotton growth period. During the period of minimum value for cotton,
the spectral angles at red and red-edge 2 reflectance of cotton are much
smaller than those of non-cotton, which correspond to the time when the
peak of NDVI occurs.
2.3.3. Backscattering characteristics of cotton and other crops

The temporal profiles of the VV and VH polarizations as well as their
combinations (i.e., VH/VV and VV-VH) for cotton and other crops are
shown in Fig. 7, which are derived from the Sentinel-1 images. In gen
Fig. 5. Spectral curves of the cotton samples from April to November. eral, the backscattering coefficients of dual-band cross-polarization
(VH) were smaller than that of single co-polarization (VV). The back
calculated using the law of cosines as follows (Khanna et al., 2007): scattering coefficients in VV and VH for cotton were higher than those of
other crops during the middle growth periods. The absolute values of
α = arcos
a2 + b2 − c2
(1) backscattering coefficients in VH from June to August were more than
2ab twice that of VV of cotton.
where a, b and c represent Euclidian distances between three vertices.

2.3.4. Feature selection
Taking the angle at red reflectance as an example, a, b and c correspond
The maximum greenness time is one of the key indicators in vege
to the distances between the green and red, red and red-edge 1, as well
tation classification, which can be extracted from the time series of
as green and red-edge 1 band, respectively (Fig. 5).
vegetation index (Pastor-Guzman et al., 2018). Based on the above
The mean spectral curves of 100 cotton samples during the cotton
Fig. 6. Temporal profiles of spectral angles at green, red, red-edge 1, red-edge 2, red-edge 3, near infrared, and shortwave infrared 1 bands for cotton, corn, rice,
soybeans and others from April to November. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of
this article.)
154
Fig. 7. Temporal profiles of VV, VH, VH/VV and VV-VH for cotton, corn, rice, soybeans and others from April to November.
preliminary analysis, the features at the peak greenness time (maximum

|μ 1 − μ 2 |
NDVI) exhibit much difference between cotton and other crops, thus M= (3)
could be considered to be used to distinguish cotton from non-cotton
σ1 + σ2
crops. Moreover, considering the phenology differences of crops in
where μ1 and μ2 represent the mean values of two classes, σ 1 and σ2
different regions, an adaptive window was adopted to develop the CMI.
represent the standard deviations of two classes. For a certain feature, it
Specifically, the features at the peak greenness time for each pixel were
indicates a good separability between the two classes when the M value
used to develop the CMI. To further determine the appropriate features
is higher than 1, while the poorly separable when the M value is less than
to distinguish cotton from other crops, the following three assumptions
1.
were proposed.
The JM distance between a pair of class-specific functions is given by:
Firstly, we assumed that cotton has a maximum or minimum value in
a certain spectral band or vegetation index at the peak greenness time JM = 2(1 − e− B ) (4)
compared to other crops. In other words, cotton differs significantly
from other crops in a certain feature, which is helpful to the identifi 1 2 1 σ2 + σ 22
B = (μ1 − μ2 )2 2 + ln( 1 ) (5)
cation of the cotton cultivated area. Thus, we proposed the following 8 2
σ1 + σ2 2 2σ 1 σ 2
criteria to select feature 1 from the 27 optical candidate features
It indicates that the two classes are poorly separated when the JM
mentioned in Section 2.2.1.
value is less than 1 (Deng and Wu, 2012).
feature 1 = argmaxi (Di ) (2) Similarly, the second and third assumptions were proposed that the
spectral angle or radar feature was larger or smaller for cotton than that
where i represents the serial number of optical candidate features, for other crops at the peak greenness time. Therefore, the feature 2 was
ranging from 1 to 27. Di represents the separability between cotton and selected from the 7 spectral angles including the angle at green, red, red-
non-cotton samples at the ith feature. In this study, two indicators edge 1, red-edge 2, red-edge 3, near infrared and shortwave infrared 1.
including the M index (Kaufman and Remer, 1994) and Jeffries-Matusita The feature 3 was selected from the 4 radar candidate features including
(JM) distance (Swain and Davis, 1978) were used, which have been VV, VH, VH/VV and VV − VH.
widely used in separability analysis (Guo and Li, 2020; Yang et al., All the above proposed assumptions were used to select the optimal
2017). The M index is defined as the absolute value of the difference of features to develop the CMI for cotton identification. Specifically,
two mean values divided by the sum of the standard deviations of two adequate number of samples of cotton (1000) and other crops (1000) in
classes, which can be calculated as follows (Guo and Li, 2020): site A were randomly selected with reference to the CDL data. The peak
greenness time for each sample was determined by the time when the
155
Fig. 8. Histograms of spectral reflectance (a-i) and vegetation index (j-zz) values of cotton and non-cotton samples at the peak greenness time.
156
maximum NDVI occurs. Then, the candidate feature values corre margin between them with a small number of training cases. It usually
sponding to the peak greenness time for each sample were obtained and performs well in handling a limited number of training samples and
analyzed to search for the optimal features that can discriminate cotton processing high-dimensional data (Zhang et al., 2020). In this study, the
from other crops. The independent t-test was used to investigate the radial basis function (RBF) kernel was used due to its’ robust perfor
statistical significance of the differences between the mean values of mance in remote sensing image classification (Hu et al., 2019). The
optimal features for cotton and other crops. Finally, the CMI was con kernel parameter γ and penalty factor C need to be set to train the SVM
structed using these features. classifier. The former determines how the data is distributed in the new
feature space and the latter controls the penalty associated with mis
classified training samples. The parameters were set to the default values
2.4. Accuracy assessment and comparison of the fitcsvm function in Matlab software.
As an ensemble classifier, the RF adopts multiple decision trees and
2.4.1. Accuracy assessment then assigns classification labels by majority vote. It has been widely
In this study, for sites A-D, the classification accuracy of the mapping used for land cover mapping and crop mapping based on remotely
results achieved by the proposed CMI was assessed by comparing with sensed images (Breiman, 2001; Pareeth et al., 2019; Zhang and Yang,
the CDL map. The overall accuracy (OA), user’s accuracy (UA) and 2020). Two parameters need to be set for the RF classifier, including the
producer’s accuracy (PA) were calculated from the confusion matrix. number of trees and the number of prediction variables used in each
Moreover, the F1-score was calculated using the UA and PA as follows node to grow the tree. In this study, they were set to 500 and square root
(Guo and Li, 2020): of the number of input variables, respectively (Belgiu and Drăgu, 2016).
UA × PA The time series NDVI during the cotton growth period was used as input
F1 − score = 2 × (6)
UA + PA into the KNN, SVM and RF classifiers, respectively. The training samples
were randomly selected from the region with the confidence value of the
For sites E-H, the cotton areas detected by the CMI were compared
CDL map larger than 90%.
with the statistical cotton areas at the county level. The mean relative
error (MRE) and determination coefficient (R2) between the detected
3. Results
and statistical cotton areas were calculated.
3.1. Development of the CMI

2.4.2. Accuracy comparison
To further assess the performance of the CMI, the accuracy of the
3.1.1. Analysis of the first assumption
cotton map obtained by the CMI was compared with those by the CEI (L.
Wang et al., 2019) and three supervised classifiers (i.e., k-nearest To develop an efficient index to detect the cotton cultivated area, the
objective of the first assumption was to select the optimal reflectance or
neighbors (KNN), support vector machine (SVM) and random forest
(RF)), respectively. The red-edge 1 and near infrared bands of the vegetation index from the optical candidate features to discriminate
cotton from other crops. The distributions of the 27 optical candidate
Sentinal-2 images were used to calculate the CEI.
As one of the simple machine learning algorithms, the principle of features at the peak greenness time are shown in Fig. 8, which is based
the KNN is that each sample can be represented by its nearest K neigh
boring values. The label of a sample is defined as the category corre Table 4
sponding to the majority of the K most similar samples in the feature Independent t-test between cotton and other crops in terms of the mean values of
feature 1, feature 2, CMI and CMI-S2 using 1000 samples for each class.
space (Wilson et al., 2018). During the classification process, the cate
gory of a sample is only determined by the category of the nearest one or T-test between cotton and ~ Corn Rice Soybeans Others
several samples (Chen et al., 2020). In this study, the K value was Number of samples 1000 1000 1000 1000
determined by comparing the classification accuracy under different K Sig (2-tailed) feature 1 0.000 0.000 0.000 0.000
values. The results showed that the performance of the KNN in cotton feature 2 0.000 0.000 0.000 0.000
CMI 0.000 0.000 0.000 0.000
mapping was optimal when the K was set as 4.
CMI-S2 0.000 0.000 0.000 0.000
SVM is a non-parametric supervised classifier derived from statistical
learning theory, which separates different classes by maximizing the
Fig. 9. The M index and JM distance between cotton and non-cotton samples using the multiplication of every two candidate features. The squares on the diagonal
line represent the M and JM values using the single feature. The horizontal and vertical axes represent the 27 optical candidate features, i.e., B, G, R, RE1, RE2, RE3,
NIR, SW1, SW2, NDVI, EVI, IPVI, TNDVI, SAVI, OSAVI, GARI, GNDVI, IRECI, MCARI, PSSRA, MTCI, S2REP, REIP, NDSVI, NDRI, NDTI, and LSWI.
157
Fig. 10. Histograms of spectral angles of cotton and non-cotton samples at the peak greenness time.
Fig. 11. The diagram of selection of feature 3 using the sliding window. TPG represents the composite period of image when the maximum NDVI occurs.
158
on the 1000 cotton samples and 1000 non-cotton samples in site A. It can
be seen that the differences between cotton and non-cotton samples for
the RE1, RE2 and MCARI values are significant, with the M (JM) values
of 0.87 (0.65), 0.86 (0.62) and 1.04 (0.85), respectively. The values of
these features for most of the cotton samples were much larger than
those for non-cotton samples. For the other optical candidate features,
the overlap between the cotton and non-cotton samples was large,
indicating the less informative for cotton identification.
Although the RE1, RE2 and MCARI performed better than other
features in distinguishing cotton from non-cotton samples, the M and JM
values were still much lower. Therefore, the addition and multiplication
combinations of every two optical candidate features were tested to
further amplify the differences between cotton and non-cotton samples.
Due to relatively better results, the multiplication of two candidate
features was adopted in this study to distinguish cotton from non-cotton.
The M and JM values of the multiplication of every two features are
shown in Fig. 9. It can be seen that the results incorporated the RE1 or
MCARI were much better than those of others. Among all the combi Fig. 12. The distribution of CMI values of cotton and non-cotton samples in site
nations, the multiplication of RE1 and RE2 reflectance performed best, A. For each box, the central mark and diamond indicate the median and mean
with the M and JM values of 1.36 and 1.20, respectively. Therefore, the value, respectively. The left and right edges of the box represent the 25th and
75th percentiles, respectively.
multiplication of the red-edge 1 and red-edge 2 reflectance at the peak
greenness time was selected as feature 1 to constitute the CMI for cotton
identification. It is expected that this value will be maximized for cotton for separation of cotton from other crops as follows:
and minimal for other crops. ρPG PG
RE1 × ρRE2 × θ
The independent t-test was employed to examine the statistical sig CMI = PG
( ∑PG t
) (7)
α × abs t=PG− 2 βVV
nificance of the differences between mean values of feature 1 for cotton
R
and other crops. As shown in Table 4, the sig (2-tailed) values are equal where ρ and α represent the spectral reflectance and spectral angle,
to zero, indicating that the differences between mean values of cotton respectively. β represents the backscattering coefficients. θ is a constant
and other crops are significant. variable, and is set to 1000 to avoid too many decimals of the CMI in this
study. t represents the composite period of the image, and abs represents
3.1.2. Analysis of the second assumption absolute value.
The distributions of seven spectral angles of cotton and non-cotton Moreover, to test the performance of the combination of radar fea
samples at the peak greenness time were analyzed (Fig. 10). It can be tures and optical features in cotton mapping, we also constructed an
seen that the differences of spectral angles at the red and green reflec index using spectral features alone for comparison, as follows:
tance during the peak greenness time between cotton and non-cotton
samples were significant, with the M (JM) values of 1.08 (0.88) and CMI-S2 = (ρPG PG PG
RE1 × ρRE2 )/αR (8)
0.88 (0.64), respectively. The values of these two features of cotton
It is expected that the CMI and CMI-S2 values will be maximized for
samples were much smaller than those of non-cotton samples. The
cotton and minimal for other crops.
spectral angle at the red reflectance was selected as feature 2 to
To investigate the differences in CMI values between cotton and non-
construct the CMI due to its better separability. Moreover, the sig (2-
cotton crops, the 1000 samples for cotton, corn, rice, soybean and others
tailed) values obtained by t-test were equal to zero, demonstrating that
were randomly selected from the CDL map with the confidence value
the differences between mean values of selected feature 2 for cotton and
higher than 90% in site A. The distribution of the CMI values for these
other crops were significant (Table 4).
samples is shown in Fig. 12. It can be seen that the mean CMI values of
cotton are obviously larger than those of other crops. The mean values of
3.1.3. Analysis of the third assumption
CMI for cotton, corn, rice and soybeans were 0.74, 0.24, 0.21 and 0.32,
Due to the relatively low separability when using radar features
respectively. The M and JM values between cotton and non-cotton
alone, we investigated the performance of combining radar features
samples were 1.45 and 1.30, respectively, indicating the relatively
with feature 1 and feature 2 for cotton mapping (Section 3.1.4). A sliding
good separability between the cotton and non-cotton samples. More
window was adopted to search for the optimal feature 3 that could
over, the sig (2-tailed) values obtained by t-test were equal to zero,
perform best when combined with feature 1 and feature 2. Specifically,
demonstrating that the differences between mean values of CMI for
the features within different temporal windows derived from Sentinel-1
cotton and other crops were significant (Table 4).
were tested. As shown in Fig. 11, the window size was tested from 1 to 3,
with the window sliding around the maximum NDVI period. According
to the experiments, the summation of backscattering coefficients in VV 3.2. Determination of the optimum threshold value
polarization during the peak greenness and previous two composite
periods was selected as the feature 3. An empirical method was adopted to select the appropriate threshold
value of the CMI, which has been successfully used to determine the
3.1.4. Construction of CMI threshold in previous studies (Ashourloo et al., 2019). A total number of
According to the above results, for cotton, the multiplication of the 2000 samples were randomly selected from the CDL map with the
reflectivity of red-edge 1 and red-edge 2 was higher than that of the confidence value higher than 90% in site A, including 1000 cotton
other crops at the peak greenness time, while the spectral angle at red samples and 1000 non-cotton samples. The classification accuracy cor
reflectance was lower than that of the other crops. The absolute value of responding to different CMI thresholds with step size of 0.02 were
summation of the backscattering coefficients in VV polarization was calculated. The experiment was repeated 20 times. As shown in Fig. 13
lower for cotton than that for other crops. To further amplify the dif (a), with the increase of the threshold value of CMI, the user’s accuracy,
ferences, the larger value was used as the numerator and the smaller overall accuracy and F1-score showed a trend of increasing first and then
value as the denominator to constitute the Cotton Mapping Index (CMI) decreasing. When the threshold value was small, the omission error for
159
Fig. 13. Variations of classification accuracy with different CMI thresholds for sites A-D.
intervals was also analyzed. As shown in Fig. 14, the consistency in
Table 5
creases with the increase of the confidence values of the CDL map,
Confusion matrix for classification results with the CMI threshold value of 0.50
indicating that the mapping results derived by the CMI have good
for study sites A-D.
accuracy.
Study CDL CMI map PA UA OA F1- The CMI-derived cotton maps as well as the consistency maps be
sites map (%) (%) (%) score
Cotton Non-
(%)
tween the CDL and CMI-derived maps for sites A-D are shown in Fig. 15.
cotton In general, the distributions of cotton cultivated areas on CDL and CMI
A Cotton 935 65 93.50 93.03 93.25 93.27 map were similar for four sites. As shown in Fig. 15 (e-h), the red and
Non- 70 930 green areas represent the consistent cotton and non-cotton areas be
cotton tween the CDL and CMI map, respectively, representing the regions
B Cotton 886 114 88.60 90.69 89.75 89.63
Non- 91 909
where CMI has a relatively high accuracy. Compared with the CDL map,
cotton the cotton area detected by the CMI was overestimated, and most of
C Cotton 873 127 87.30 77.81 81.20 82.28 these cotton areas corresponded to the uncertain regions with confi
Non- 249 751 dence value less than 90% on the CDL map.
cotton
D Cotton 878 122 87.80 86.33 86.95 87.06
Non- 139 861 3.3.2. Application of CMI in Xinjiang sites
cotton The four sites in Xinjiang were used to further verify the applicability
of CMI for cotton mapping. The cotton areas were identified using the
optimal threshold value of the CMI obtained in the site A. The MRE and
cotton was small due to the relatively large CMI values of cotton, thus the R2 between the statistical and CMI-detected cotton areas were
the producer’s accuracy was relatively high. With the increase of 43.44% and 0.74 at the county level, respectively. As shown in Fig. 16,
threshold value, the cotton samples with low CMI values could not be most of the points are above the 1:1 line, indicating that the CMI slightly
identified, thus the producer’s accuracy showed the decreasing ten overestimates the cotton area when the threshold value is 0.50. There
dency. In this study, the overall accuracy reached the peak in site A fore, the threshold value of the CMI was fine-tuning using an empirical
when the CMI threshold was 0.5. Thus, it was considered to be an method. The R2 and MRE achieved by different threshold values were
appropriate threshold and used for the identification of cotton cultivated calculated, with the step of 0.02 from 0.5 to 0.7. The results showed that
areas in other sites. The same method was applied to CMI-S2, and the the optimal threshold value for these regions was 0.56, with the R2 and
optimal threshold value for CMI-S2 was 0.013 in site A. MRE of 0.60 and 26.69%, respectively. Thus, this threshold value was
adopted for cotton mapping in these regions in 2018 (Fig. 17).
3.3. Accuracy assessment of the proposed CMI
3.3.1. Accuracy assessment for sites A-D 3.4. Comparison with other methods
As shown in Table 5, when the threshold value of CMI was set as
0.50, the overall accuracy and F1-score of the CMI-derived cotton map The performance of the CMI was further assessed by comparing the
for site A were 93.25% and 93.27%, respectively. Then it was applied to classification accuracy of the CMI-derived map with those achieved by
other sites to further verify the validity of the CMI and the threshold. The the other two indices (CMI-S2, CEI) and three supervised classifiers
results showed that the overall accuracies for sites B, C and D were (KNN, SVM, RF) using the validation samples at sites A-D. The 20 groups
89.75%, 81.20% and 86.95%, respectively. The producer’s accuracy of validation samples were randomly generated, and each group con
was higher than the user’s accuracy for sites A, C and D, indicating that sisted of 1000 cotton and 1000 non-cotton samples. Considering the
the omission error was lower than the commission error in these sites. differences of phenology in different regions, the optimal CEI threshold
The producer’s accuracy was lower than the user’s accuracy for sites B, and optimal temporal for each region were determined by empirical
indicating that the commission error was lower than the omission error method.
in this site. As shown in Fig. 18, the CMI performed better than the RF, SVM and
In this study, the samples with confidence values higher than 90% KNN classifiers in terms of the mean values of overall accuracy and F1-
were randomly selected from the CDL map to develop the CMI and assess score in site A and site B, indicating that the performance of CMI without
the accuracy for sites A-D (Table 5). In addition, the consistency between using training samples was similar to those of supervised classifiers. The
the CMI-derived cotton map and the CDL map at different confidence overall accuracy and F1-score achieved by the CMI and CMI-S2 was
160
Fig. 14. Consistency between the CMI-derived and CDL-derived cotton map at different confidence intervals of the CDL map.
similar, both of which were superior to that of the CEI. Also, both the angle at the red band as well as backscattering coefficients in VV po
optimal CMI and CMI-S2 thresholds were determined in site A and then larization were used to develop the CMI. The results demonstrated that
applied for four sites, while the optimal CEI threshold with the optimal the CMI can be effectively used to map the cotton distribution in the
temporal of image was searched first and then used for each site. These eight sites of this study without the use of training samples. Specifically,
results indicated the advantages of adaptive window of the proposed the CMI has the following advantages over existing methods.
CMI and CMI-S2. In site C and site D, the overall accuracy and F1-score Firstly, in this study, the time series of Sentinel-1 and Sentinel-2
achieved by three indices were lower than those of the KNN, SVM and images during the cotton growth periods were used as the initial data
RF classifiers, indicating that the supervised classifiers were robust when sets. The features at/around the peak greenness time for each pixel were
the training samples were available. In the absence of samples, among used to develop the CMI for cotton identification, which was based on an
the three indices, the performance of CMI was significantly higher than adaptive window. The phenology differences in different regions can be
that of CMI-S2 and CEI, indicating the advantages of the combination of considered in this study when compared with that using the single
optical and radar features in cotton mapping. temporal image solely. The comparison results between the CMI-S2 and
CEI in Fig. 18 demonstrated this advantage.
4. Discussion Secondly, the features derived from both Sentinel-2 MSI and
Sentinel-1 SAR images were jointly adopted to develop the CMI and
4.1. Advantages of the proposed index contribute to the ability to identify the cotton cultivated area. Previous
studies have shown the potential of the red-edge band in crop classifi
As an important cash crop and strategic material in the world, the cation using supervised classifiers (Zhang et al., 2020). The analysis in
spatial distribution information of cotton is of great significance to the this study showed that the multiplication of red-edge 1 and red-edge 2
monitoring of cotton growth and yield estimation. In recent years, ma reflectance for cotton samples were much larger than those for non-
chine learning algorithms have been proved to be an effective method to cotton samples at the peak greenness time. Therefore, these two bands
map the spatial distribution of various crops (He et al., 2019). However, were chosen to develop the CMI based on the first assumption of this
most of these methods belong to supervised classifiers, which strongly study. In addition to the spectral reflectance, the spectral angle is also an
depend on the samples to train the classifier (Dong et al., 2020). It still important indicator in vegetation studies (Khanna et al., 2007; Palacios-
remains a challenge in cotton mapping when the samples are unavai Orueta et al., 2012). Based on the second assumption of this study, the
lable. There is a great demand for developing a method that can be used analysis results showed that the spectral angle at the red band of
for automated mapping the cotton cultivated area. Sentinel-2 data for cotton samples were much smaller than those for
To this end, a novel index termed CMI, was proposed using the time non-cotton samples at the peak greenness time. With the above two
series of Sentinel-1 SAR and Sentinel-2 MSI images, and its ability for groups of optical features, the optimal radar features were explored. The
cotton mapping was examined in the eight sites selected from the U.S. results showed that the incorporation of VV performed better than
and China. The reflectance of red-edge 1 and red-edge 2 bands, spectral others for cotton identification in this study. The combination of optical
161
Fig. 15. False color composites of red, near infrared and blue of Sentinel-2 images overlaid with cotton maps detected by the CMI (in red color) for sites A-D (a-d).
(e)-(h) show the comparison of the CMI and CDL map. ‘Others’ represents the regions with confidence value less than 90% of the CDL map. (For interpretation of the
references to color in this figure legend, the reader is referred to the web version of this article.)
increasing trend and gradually stabilized. The variation range of overall

accuracy reduces with the increase of the number of training samples,
indicating that there is still a high uncertainty in the case of a small
number of training samples for supervised classifiers.
Moreover, timeline is an important factor when considering a crop
type map, as the early map before the end of the season can help monitor
growth, determine crop drought conditions, manage irrigation and
predict crop yields, thus benefiting farmers and policymakers (Inglada
et al., 2016; Vaudour et al., 2015). Taking sites A-D as an example, the
satellite images with different time series lengths were used to test the
potential of CMI in early season cotton mapping. The input data started
in April and increased by half a month at a time. The classification ac
curacy achieved by input data with different lengths were compared
(Fig. 20). With the length of time series increases, the overall accuracy
showed a trend of increase. For sites A-C, the accuracy reached the
maximum and gradually stabilized after August, indicating that using
satellite data during day of year (DOY) 90–240 could achieve an
acceptable accuracy in these sites. For site D, the accuracy tends to be
Fig. 16. Comparison of CMI-derived and statistical cotton area at the county stable using the time series data during DOY 90–180, which was earlier
level with the CMI thresholds of 0.50 and 0.56, respectively.
than for sites A-C. The reason is that the peak greenness time of cotton in
site D (the southern Texas State) is much earlier than those in the other
and radar features performed better than using optical features solely in three sites. This indicates that the cotton map can be obtained before
cotton identification. The comparison results between the CMI and CMI- harvest using the CMI.
S2 in Fig. 18 demonstrated this advantage.
Third, compared to the supervised classifiers, the proposed CMI is
convenient and performs well even if without training samples in 4.2. Uncertainty and potential refinements
identifying the cotton cultivated area, which is a significant advantage
given the lack of survey samples available. Taking site A and site B as an Although with the above-mentioned advantages of the proposed
example, the classification accuracies of the KNN, SVM and RF classifiers method, there are still some limitations and uncertainties in cotton
under different number of training samples were also compared. The 20 mapping. Firstly, in this study, the threshold value of the CMI was
groups of training samples were randomly generated to train the clas determined as 0.50 at site A, and its performance at sites B-D was
sifier, and the classification accuracy was evaluated with the same acceptable. This indicated that this threshold can be adopted in the re
validation samples. As shown in Fig. 19, with the increase of the number gion where the samples are not available to determine an appropriate
of training samples, the mean values of overall accuracy show an threshold value. However, as shown in Table 5, the accuracy of the
mapping results in sites B-D using this value was slightly lower than that
162
Fig. 17. The cotton distribution map detected by the CMI overlaid with the Google images for sites E-H.
Fig. 18. Classification accuracy of cotton mapping results achieved by the CMI, CMI-S2, CEI, KNN, SVM and RF methods in sites A-D, respectively. RE represents the
absolute value of relative error of cotton pixels between the mapping results and CDL map, which corresponds to the right vertical axis. The error bar represents the
upper and lower bound of the accuracy assessed by 20 groups of validation samples.
163
Fig. 19. Overall accuracy achieved by the KNN, SVM and RF classifiers with different numbers of training samples.
Fig. 20. Variations in the classification accuracy of the CMI-derived results obtained by using satellite images with different time series lengths in sites A-D.
in site A. Due to the availability of the adequate samples for cotton and slightly affected by different crop types at different sites. The results
other crops, the optimal CMI thresholds for sites B-D were also deter indicated that better performance may be achieved when a small num
mined using the same method in Section 3.2 (Fig. 13 (b)). The results ber of reference data (i.e., training samples or statistics) can be used to
showed that the optimal threshold value for site B was the same as that fine-tune the CMI threshold. The applicability of the CMI and the
for site A. When the threshold value was 0.58 and 0.54, the overall ac determination of threshold in other regions need to be further assessed
curacy reached the peak for site C and site D, respectively, both of which in the future study. More features could be explored and selected to
were much higher than that using the threshold value of 0.50. For sites further improve the performance of the CMI as well as the stability of the
E-H, the evaluation results based on the county-level statistics showed optimal CMI threshold for automatic cotton mapping.
that the R2 was high whereas the cotton area was a little overestimated Secondly, in the current study, the development and application of
by the CMI with the threshold value of 0.50. It achieved a better per the CMI were only considered in the single cropping systems. According
formance using the threshold value of 0.56, which was fine-tuned using to the CDL map, there is a small percentage of double cropping system
statistical data. The reason may be that the performance of the CMI was area in the sites A-D. Taking site A as an example, the areas of winter
164
wheat-soybeans, winter wheat-corn and winter wheat-cotton account Ashourloo, D., Shahrabi, H.S., Azadbakht, M., Aghighi, H., Matkan, A.A, Radiom, S.,
2018. A novel automatic method for alfalfa mapping using time series of Landsat-8
for 0.488%, 0.019% and 0.007% of the whole study area, respectively.
OLI data. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 11, 4478–4487. https://
Among the pixels of these three classes, 96.80% were correctly classified doi.org/10.1109/JSTARS.2018.2874726.
using the CMI. Considering the relatively small percentage of double Ashourloo, D., Shahrabi, H.S., Azadbakht, M., Aghighi, H., Nematollahi, H.,
cropping system area in the current study sites, the applicability of the Alimohammadi, A., Matkan, A.A., 2019. Automatic canola mapping using time series
of sentinel 2 images. ISPRS J. Photogramm. Remote Sens. 156, 63–76. https://doi.
CMI in double/multiple cropping systems needs to be further explored in org/10.1016/j.isprsjprs.2019.08.007.
the future study. It is important for the application of the CMI in the Ashourloo, D., Shahrabi, H.S., Azadbakht, M., Rad, A.M., Aghighi, H., Radiom, S., 2020.
cotton producing countries such as China (Yellow River basin and A novel method for automatic potato mapping using time series of Sentinel-2 images.
Comput. Electron. Agric. 175, 105583. https://doi.org/10.1016/j.
Yangtze River basin), India, Pakistan and Brazil with much double/ compag.2020.105583.
multiple cropping systems areas. Moreover, the red-edge 1 and red-edge Belgiu, M., Drăgu, L., 2016. Random forest in remote sensing: A review of applications
2 bands of Sentinel-2 were used to calculate the CMI, both of which were and future directions. ISPRS J. Photogramm. Remote Sens. 114, 24–31. https://doi.
org/10.1016/j.isprsjprs.2016.01.011.
resampled from 20 m to 10 m. The uncertainty of the resampling was not Blackburn, G.A., 1998. Quantifying chlorophylls and carotenoids at leaf and canopy
considered in this study. In the future research, the influence of scales: An evaluation of some hyperspectral approaches. Remote Sens. Environ. 66,
resampling on classification accuracy can be further analyzed in cotton 273–285. https://doi.org/10.1016/S0034-4257(98)00059-5.
Breiman, L.E.O., 2001. Random Forests. Mach. Learn. 45, 5–32. https://doi.org/
identification. 10.1023/A:1010933404324.
Cai, Y., Guan, K., Peng, J., Wang, S., Seifert, C., Wardlow, B., Li, Z., 2018. A high-
performance and in-season classification system of field-level crop types using time-
5. Conclusion
series Landsat data and a machine learning approach. Remote Sens. Environ. 210,
35–47. https://doi.org/10.1016/j.rse.2018.02.045.
In this study, a novel index termed Cotton Mapping Index (CMI) was Chamorro Martinez, J.A., Cué La Rosa, L.E., Feitosa, R.Q., Sanches, I.D.A., Happ, P.N.,
developed using Sentinel-1 SAR and Sentinel-2 MSI images, which was a 2021. Fully convolutional recurrent networks for multidate crop recognition from
multitemporal image sequences. ISPRS J. Photogramm. Remote Sens. 171, 188–201.
convenient method for automatic identification of cotton cultivated area https://doi.org/10.1016/j.isprsjprs.2020.11.007.
without training samples. The characteristics of temporal profiles of Chen, Z., Jia, K., Xiao, C., Wei, D., Zhao, X., Lan, J., Wei, X., Yao, Y., Wang, B., Sun, Y.,
cotton samples during its growth period were analyzed to develop the Wang, L., 2020. Leaf area index estimation algorithm for GF-5 hyperspectral data
based on different feature selection and machine learning methods. Remote Sens. 12
CMI. Based on the analysis results, the reflectance at the red-edge 1 and (13), 2110. https://doi.org/10.3390/rs12132110.
red-edge 2 bands, the spectral angle at the red band, as well as back Conrad, C., Colditz, R.R., Dech, S., Klein, D., Vlek, P.L.G., 2011. Temporal segmentation
scattering coefficients in VV polarization, were combined to constitute of MODIS time series for improving crop classification in Central Asian irrigation
systems. Int. J. Remote Sens. 32, 8763–8778. https://doi.org/10.1080/01431161.20
the CMI. Four sites from the U.S. and four sites from China were used to 10.550647.
assess the performance of the CMI. The results indicated that the cotton Crippen, R.E., 1990. Calculating the vegetation index faster. Remote Sens. Environ. 34
cultivated area can be identified using the CMI with the overall accuracy (1), 71–73. https://doi.org/10.1016/0034-4257(90)90085-Z.
Dash, J., Curran, P.J., 2007. Evaluation of the MERIS terrestrial chlorophyll index
higher than 81.20% for the sites from the U.S. as well as the mean
(MTCI). Adv. Sp. Res. 39 (1), 100–104. https://doi.org/10.1016/j.asr.2006.02.034.
relative error of 26.69% for the sites from Xinjiang of China. Compared Daughtry, C.S.T., Walthall, C.L., Kim, M.S., Colstoun, E.B. de, III, J.E.M., 2000.
with the supervised classifiers such as KNN, SVM and RF, the proposed Estimating corn leaf chlorophyll concentration from leaf and canopy reflectance.
Remote Sens. Environ. 74, 229–239. https://doi.org/10.1016/s0034-4257(00)
CMI achieved a similar performance without using the training dataset.
00113-9.
Compared with the CMI-S2 and CEI using optical features only, the CMI, Deng, C., Wu, C., 2012. BCI: A biophysical composition index for remote sensing of urban
which incorporated the optical and radar features, achieved a better environments. Remote Sens. Environ. 127, 247–259. https://doi.org/10.1016/j.
performance in cotton mapping. The comparison results of the CMI, rse.2012.09.009.
Dong, J., Fu, Y., Wang, J., Tian, H., Fu, S., Niu, Z., Han, W., Zheng, Y., Huang, J.,
CMI-S2 and CEI also demonstrated the advantage of using the features Yuan, W., 2020. Early-season mapping of winter wheat in China based on Landsat
within an adaptive window for cotton mapping. Moreover, the cotton and Sentinel images. Earth Syst. Sci. Data 12, 3081–3095. https://doi.org/10.5194/
cultivated area can be identified before harvest using the CMI. With the essd-12-3081-2020.
Frampton, W.J., Dash, J., Watmough, G., Milton, E.J., 2013. Evaluating the capabilities
promising results of this study, the applicability of the CMI in other of Sentinel-2 for quantitative estimation of biophysical variables in vegetation.
regions with different cropping systems and crop types needs to be ISPRS J. Photogramm. Remote Sens. 82, 83–92. https://doi.org/10.1016/j.
further assessed in the future study. isprsjprs.2013.04.007.
Gitelson, A.A., Kaufman, Y.J., Merzlyak, M.N., 1996. Use of a green channel in remote
sensing of global vegetation from EOS- MODIS. Remote Sens. Environ. 58 (3),
289–298. https://doi.org/10.1016/S0034-4257(96)00072-7.
Declaration of Competing Interest Gitelson, A.A., Merzlyak, M.N., 1998. Remote sensing of chlorophyll concentration in
higher plant leaves. Adv. Sp. Res. 22 (5), 689–692. https://doi.org/10.1016/S0273-
1177(97)01133-2.
The authors declare that they have no known competing financial Gong, P., Liu, H., Zhang, M., Li, C., Wang, J., Huang, H., Clinton, N., Ji, L., Li, W., Bai, Y.,
interests or personal relationships that could have appeared to influence Chen, B., Xu, B., Zhu, Z., Yuan, C., Ping Suen, H., Guo, J., Xu, N., Li, W., Zhao, Y.,
the work reported in this paper. Yang, J., Yu, C., Wang, X., Fu, H., Yu, L., Dronova, I., Hui, F., Cheng, X., Shi, X.,
Xiao, F., Liu, Q., Song, L., 2019. Stable classification with limited sample:
transferring a 30-m resolution sample set collected in 2015 to mapping 10-m
Acknowledgments resolution global land cover in 2017. Sci. Bull. 64 (6), 370–373. https://doi.org/
10.1016/j.scib.2019.03.002.
Guo, X., Li, P., 2020. Mapping plastic materials in an urban area: Development of the
This work was jointly supported by the CAS Strategic Priority normalized difference plastic index using WorldView-3 superspectral data. ISPRS J.
Research Program (No. XDA19030402), the Shandong Key Research and Photogramm. Remote Sens. 169, 214–226. https://doi.org/10.1016/j.
Development Project (No. 2018GNC110025), the Natural Science isprsjprs.2020.09.009.
He, T., Xie, C., Liu, Q., Guan, S., Liu, G., 2019. Evaluation and comparison of random
Foundation of China (Nos. 42071425, 41871253). The authors would forest and A-LSTM networks for large-scale winter wheat identification. Remote
like to express great appreciations to the anonymous reviewers and the Sens. 11 (14), 1665. https://doi.org/10.3390/rs11141665.
editor for their valuable comments and suggestions. Hu, Q., Sulla-Menashe, D., Xu, B., Yin, H., Tang, H., Yang, P., Wu, W., 2019.
A phenology-based spectral and temporal feature selection method for crop mapping
from satellite time series. Int. J. Appl. Earth Obs. Geoinf. 80, 218–229. https://doi.
References org/10.1016/j.jag.2019.04.014.
Huang, S., Yang, L., Chen, X., Yao, Y., 2018. Study of typical arid crops classification
based on machine learning. Spectrosc. Spectr. Anal. 38, 3169–3176. https://doi.org/
Adrian, J., Sagan, V., Maimaitijiang, M., 2021. Sentinel SAR-optical fusion for crop type
10.3964/j.issn.1000-0593(2018)10-3169-08.
mapping using deep learning and Google Earth Engine. ISPRS J. Photogramm.
Huete, A, Didan, K, Miura, T, Rodriguez, E.P, Gao, X, Ferreira, L.G, 2002. Overview of
Remote Sens. 175, 215–235. https://doi.org/10.1016/j.isprsjprs.2021.02.018.
the radiometric and biophysical performance of the MODIS vegetation indices.
Akter, T., Islam, A.K.M.A., Rasul, M.G., Kundu, S., Khalequzzaman, Ahmed, J.U., 2019.
Remote Sens. Environ. 83 (1-2), 195–213. https://doi.org/10.1016/S0034-4257(02)
Evaluation of genetic diversity in short duration cotton (Gossypium hirsutum L.).
00096-2.
J. Cott. Res. 2, 2–7. https://doi.org/10.1186/s42397-018-0018-6.
165
Huete, A.R, 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25 (3), GIScience Remote Sens 54, 918–938. https://doi.org/10.1080/15481603.2017.1
295–309. 351149.
ICAC, 2018. International Cotton Advisory Committee. USA. Steinhausen, M.J., Wagner, P.D., Narasimhan, B., Waske, B., 2018. Combining Sentinel-1
Ienco, D., Interdonato, R., Gaetano, R., Ho Tong Minh, D., 2019. Combining Sentinel-1 and Sentinel-2 data for improved land use and land cover mapping of monsoon
and Sentinel-2 satellite image time series for land cover mapping via a multi-source regions. Int. J. Appl. Earth Obs. Geoinf. 73, 595–604. https://doi.org/10.1016/j.
deep learning architecture. ISPRS J. Photogramm. Remote Sens. 158, 11–22. https jag.2018.08.011.
://doi.org/10.1016/j.isprsjprs.2019.09.016. Sun, P., Congalton, R.G., Pan, Y., 2018. Using a simulation analysis to evaluate the
Inglada, J., Vincent, A., Arias, M., Marais-Sicre, C., 2016. Improved early crop type impact of crop mapping error on crop area estimation from stratified sampling. Int.
identification by joint use of high temporal resolution sar and optical image time J. Digit. Earth 12 (9), 1046–1066. https://doi.org/10.1080/
series. Remote Sens. 8 (5), 362. https://doi.org/10.3390/rs8050362. 17538947.2018.1499827.
Jin, Z., Azzari, G., You, C., Di Tommaso, S., Aston, S., Burke, M., Lobell, D.B., 2019. Swain, P.H., Davis, S.M., 1978. Remote sensing: the quantitative approach. Remote Sens.
Smallholder maize area and yield mapping at national scales with Google Earth Quant. approach. PAMI-3 (6), 713–714. https://doi.org/10.1109/
Engine. Remote Sens. Environ. 228, 115–128. https://doi.org/10.1016/j. TPAMI.1981.4767177.
rse.2019.04.016. Teimouri, N., Dyrmann, M., Jørgensen, R.N., 2019. A novel spatio-temporal FCN-LSTM
Kaufman, Y.J., Remer, L.A., 1994. Detection of forests using mid-IR reflectance: An network for recognizing various crop types using multi-temporal radar images.
application for aerosol studies. IEEE Trans. Geosci. Remote Sens. 32, 672–683. Remote Sens. 11, 1–18. https://doi.org/10.3390/rs11080893.
https://doi.org/10.1109/36.297984. Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring
Khanna, S., Palacios-Orueta, A., Whiting, M.L., Ustin, S.L., Riaño, D., Litago, J., 2007. vegetation. Remote Sens. Environ. 8 (2), 127–150. https://doi.org/10.1016/0034-
Development of angle indexes for soil moisture estimation, dry matter detection and 4257(79)90013-0.
land-cover discrimination. Remote Sens. Environ. 109 (2), 154–165. https://doi.org/ Vaudour, E., Noirot-Cosson, P.E., Membrive, O., 2015. Early-season mapping of crops
10.1016/j.rse.2006.12.018. and cultural operations using very high spatial resolution Pléiades images. Int. J.
Li, H., Pan, Z., He, S., Jia, Y., Geng, X., Chen, B., Wang, L., Pang, B., Du, X., 2021. QTL Appl. Earth Obs. Geoinf. 42, 128–141. https://doi.org/10.1016/j.jag.2015.06.003.
mapping of agronomic and economic traits for four F2 populations of upland cotton. Wang, J., Xiao, X., Bajgain, R., Starks, P., Steiner, J., Doughty, R.B., Chang, Q., 2019.
J. Cott. Res. 4 (1) https://doi.org/10.1186/s42397-020-00076-y. Estimating leaf area index and aboveground biomass of grazing pastures using
Li, L., Su, H., Du, Q., Wu, T., 2021. A novel surface water index using local background Sentinel-1, Sentinel-2 and Landsat images. ISPRS J. Photogramm. Remote Sens. 154,
information for long term and large-scale Landsat images. ISPRS J. Photogramm. 189–201. https://doi.org/10.1016/j.isprsjprs.2019.06.007.
Remote Sens. 172, 59–78. https://doi.org/10.1016/j.isprsjprs.2020.12.003. Wang, L., Liu, J., Yao, B., Gao, J., Ji, F., 2019. Construction and application of cotton
Li, M., Zhao, G., Qin, Y., 2011. Extraction and monitoring of cotton area and growth characteristic spectral index based on Rapideye image. China Agric. Informatics 31,
information using remote sensing at small scale: A case study in Dingzhuang town of 25–37. https://doi.org/10.12105/j.issn.1672-0423.20190503.
Guangrao county, China. In: International Conference on Computer Distributed Wang, S., Di Tommaso, S., Faulkner, J., Friedel, T., Kennepohl, A., Strey, R., Lobell, D.B.,
Control and Intelligent Environmental Monitoring (CDCIEM). 816–823. https://doi. 2020. Mapping crop types in southeast India with smartphone crowdsourcing and
org/10.1109/CDCIEM.2011.569. deep learning. Remote Sens. 12, 1–42. https://doi.org/10.3390/RS12182957.
Major, G.G.B.J., 1988. High spectral resolution: determination of spectral shifts between Wang, T., Yan, C.Z., Song, X., Xie, J.L., 2012. Monitoring recent trends in the area of
the red and the near infrared, in: Int. Arch. Photogramm. Remote Sens. aeolian desertified land using Landsat images in China’s Xinjiang region. ISPRS J.
Musande, V., Kumar, A., Kale, K., 2012. Cotton crop discrimination using fuzzy Photogramm. Remote Sens. 68, 184–190. https://doi.org/10.1016/j.
classification approach. J. Indian Soc. Remote Sens. 40 (4), 589–597. https://doi. isprsjprs.2012.01.001.
org/10.1007/s12524-012-0201-z. Wilson, B.T., Knight, J.F., McRoberts, R.E., 2018. Harmonic regression of Landsat time
National Bureau of Statistics of China, 2020. China Statistical Yearbook. China Statistics series for modeling attributes from national forest inventory data. ISPRS J.
Press, Beijing, China. Photogramm. Remote Sens. 137, 29–46. https://doi.org/10.1016/j.
Oliphant, A.J., Thenkabail, P.S., Teluguntla, P., Xiong, J., Gumma, M.K., Congalton, R.G., isprsjprs.2018.01.006.
Yadav, K., 2019. Mapping cropland extent of Southeast and Northeast Asia using Xun, L., Zhang, J., Cao, D., Wang, J., Zhang, S., Yao, F., 2021. Mapping cotton cultivated
multi-year time-series Landsat 30-m data using a random forest classifier on the area combining remote sensing with a fused representation-based classification
Google Earth Engine Cloud. Int. J. Appl. Earth Obs. Geoinf. 81, 110–124. https://doi. algorithm. Comput. Electron. Agric. 181, 105940. https://doi.org/10.1016/j.
org/10.1016/j.jag.2018.11.014. compag.2020.105940.
Orynbaikyzy, A., Gessner, U., Conrad, C., 2019. Crop type classification using a Yang, D., Chen, J., Zhou, Y., Chen, Xiang, Chen, Xuehong, Cao, X., 2017. Mapping plastic
combination of optical and radar remote sensing data: a review. Int. J. Remote Sens. greenhouse with medium spatial resolution satellite data: Development of a new
40 (17), 6553–6595. https://doi.org/10.1080/01431161.2019.1569791. spectral index. ISPRS J. Photogramm. Remote Sens. 128, 47–60. https://doi.org/
Palacios-Orueta, A., Huesca, M., Whiting, M.L., Litago, J., Khanna, S., Garcia, M., 10.1016/j.isprsjprs.2017.03.002.
Ustin, S.L., 2012. Derivation of phenological metrics by function fitting to time-series Yin, L., You, N., Zhang, G., Huang, J., Dong, J., 2020. Optimizing feature selection of
of Spectral Shape Indexes AS1 and AS2: Mapping cotton phenological stages using individual crop types for improved crop mapping. Remote Sens. 12 (1), 162. https://
MODIS time series. Remote Sens. Environ. 126, 148–159. https://doi.org/10.1016/j. doi.org/10.3390/rs12010162.
rse.2012.08.002. Yommy, A.S., Liu, R., Wu, A.S., 2015. SAR image despeckling using refined lee filter. In:
Pan, Y., Li, L., Zhang, J., Liang, S., Zhu, X., Sulla-Menashe, D., 2012. Winter wheat area 2015 7th International Conference on Intelligent Human-Machine Systems and
estimation from MODIS-EVI time series data using the Crop Proportion Phenology Cybernetics (IHMSC 2015), pp. 260–265. https://doi.org/10.1109/IHMSC.2015.
Index. Remote Sens. Environ. 119, 232–242. https://doi.org/10.1016/j. 236.
rse.2011.10.011. Zhang, F., Yang, X., 2020. Improving land cover classification in an urbanized coastal
Pareeth, S., Karimi, P., Shafiei, M., De Fraiture, C., 2019. Mapping agricultural landuse area by random forests: The role of variable selection. Remote Sens. Environ. 251,
patterns from time series of Landsat 8 using random forest based hierarchial 112105. https://doi.org/10.1016/j.rse.2020.112105.
approach. Remote Sens. 11 (5), 601. https://doi.org/10.3390/rs11050601. Zhang, H., Kang, J., Xu, X., Zhang, L., 2020. Accessing the temporal and spectral features
Pastor-Guzman, J., Dash, J., Atkinson, P.M., 2018. Remote sensing of mangrove forest in crop type mapping using multi-temporal Sentinel-2 imagery: A case study of Yi’an
phenology and its environmental drivers. Remote Sens. Environ. 205, 71–84. County, Heilongjiang province. China. Comput. Electron. Agric. 176, 105618.
https://doi.org/10.1016/j.rse.2017.11.009. https://doi.org/10.1016/j.compag.2020.105618.
Qiu, B., Huang, Y., Chen, C., Tang, Z., Zou, F., 2018. Mapping spatiotemporal dynamics Zhang, H.K., Roy, D.P., 2017. Using the 500 m MODIS land cover product to derive a
of maize in China from 2005 to 2017 through designing leaf moisture based consistent continental scale 30 m Landsat land cover classification. Remote Sens.
indicator from Normalized Multi-band Drought Index. Comput. Electron. Agric. 153, Environ. 197, 15–34. https://doi.org/10.1016/j.rse.2017.05.024.
82–93. https://doi.org/10.1016/j.compag.2018.07.039. Zhang, J., Feng, L., Yao, F., 2014. Improved maize cultivated area estimation over a large
Qu, C., Li, P., Zhang, C., 2021. A spectral index for winter wheat mapping using multi- scale combining MODIS-EVI time series data and crop phenological information.
temporal Landsat NDVI data of key growth stages. ISPRS J. Photogramm. Remote ISPRS J. Photogramm. Remote Sens. 94, 102–113. https://doi.org/10.1016/j.
Sens. 175, 431–447. https://doi.org/10.1016/j.isprsjprs.2021.03.015. isprsjprs.2014.04.023.
Razzaq, A., Zafar, M.M, Ali, A., Hafeez, A., Batool, W., Shi, Y., Gong, W., Yuan, Y., 2021. Zhong, L., Hawkins, T., Biging, G., Gong, P., 2011. A phenology-based approach to map
Cotton germplasm improvement and progress in Pakistan. J. Cott. Res. 4 (1) https:// crop types in the San Joaquin Valley. California. Int. J. Remote Sens. 32 (22),
doi.org/10.1186/s42397-020-00077-x. 7777–7804. https://doi.org/10.1080/01431161.2010.527397.
Rondeaux, G., Steven, M., Baret, F., 1996. Optimization of soil-adjusted vegetation Zhong, L., Hu, L., Yu, L., Gong, P., Biging, G.S., 2016. Automated mapping of soybean
indices. Remote Sens. Environ. 55 (2), 95–107. https://doi.org/10.1016/0034-4257 and corn using phenology. ISPRS J. Photogramm. Remote Sens. 119, 151–164.
(95)00186-7. https://doi.org/10.1016/j.isprsjprs.2016.05.014.
Silva Junior, C.A. da, Leonel-Junior, A.H.S., Rossi, F.S., Correia Filho, W.L.F., Zhong, L., Hu, L., Zhou, H., Tao, X., 2019. Deep learning based winter wheat mapping
Santiago, D. de B., Oliveira-Júnior, J.F. de, Teodoro, P.E., Lima, M., Capristo- using statistical data as ground references in Kansas and northern Texas. US. Remote
Silva, G.F., 2020. Mapping soybean planting area in midwest Brazil with remotely Sens. Environ. 233, 111411. https://doi.org/10.1016/j.rse.2019.111411.
sensed images and phenology-based algorithm using the Google Earth Engine Zhu, Z., Gallant, A.L., Woodcock, C.E., Pengra, B., Olofsson, P., Loveland, T.R., Jin, S.,
platform. Comput. Electron. Agric. 169, 105194. https://doi.org/10.1016/j. Dahal, D., Yang, L., Auch, R.F., 2016. Optimizing selection of training and auxiliary
compag.2019.105194. data for operational land cover classification for the LCMAP initiative. ISPRS J.
Sonobe, R., Yamaya, Y., Tani, H., Wang, X., Kobayashi, N., Mochizuki, K., 2017. Photogramm. Remote Sens. 122, 206–221. https://doi.org/10.1016/j.
Assessing the suitability of data from Sentinel-1A and 2A for crop classification. isprsjprs.2016.11.004.
166

A Novel Cotton Mapping Index Combining Sentinel-1 SAR and Sentinel-2

Uploaded by

Copyright:

Available Formats

A Novel Cotton Mapping Index Combining Sentinel-1 SAR and Sentinel-2

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Novel Cotton Mapping Index Combining Sentinel-1 SAR and Sentinel-2

Uploaded by

Copyright:

Available Formats

ISPRS Journal of Photogrammetry and Remote Sensing 181 (2021) 148–166

Contents lists available at ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

A novel cotton mapping index combining Sentinel-1 SAR and Sentinel-2

Fig. 1. Flowchart of the identification of cotton cultivated area in this study.

Table 1 18 vegetation indices (VIs), which were commonly used in literature on

2.2.3. Reference samples

2.3.3. Backscattering characteristics of cotton and other crops

where a, b and c represent Euclidian distances between three vertices.

preliminary analysis, the features at the peak greenness time (maximum

3.1. Development of the CMI

increasing trend and gradually stabilized. The variation range of overall

You might also like