Crop-specific phenomapping by fusing Landsat and Sentinel data with MODIS time series

Abstract

Agricultural production and food security highly depend on crop growth and condition throughout the growing season. Timely and spatially explicit information on crop phenology can assist in informed decision making and agricultural land management. Remote sensing can be a powerful tool for agricultural assessment. Remotely sensed data is ideally suited for both large-scale and field-level analyses due to the wide variability of datasets with diverse spatiotemporal resolution. To derive crop-specific phenometrics, we fused time series from Landsat 8 and Sentinel 2 with Moderate-resolution Imaging Spectroradiometer (MODIS) data. Using a linear regression approach, synthetic Landsat 8 and Sentinel 2 data were created based on MODIS imagery. This fusion-process resulted in synthetic imagery with radiometric characteristics of original Landsat 8 and Sentinel 2 data. We created four different time series using synthetic data as well as a mix of original and synthetic data. The extracted time series of phenometrics consisting of both synthetic and original data showed high detail in the final phenomaps which allowed intra-field level assessment of crops. In-situ field reports were used for validation. Our phenometrics showed only a few days of deviation for most crops and datasets. The proposed data integration method can be applied in areas where data from a single high-resolution source is scarce.

Full text

Full Terms & Conditions of access and use can be found at
https://www.tandfonline.com/action/journalInformation?journalCode=tejr20
European Journal of Remote Sensing
ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tejr20
Crop-specific phenomapping by fusing Landsat
and Sentinel data with MODIS time series
Jonas Schreier , Gohar Ghazaryan & Olena Dubovyk
To cite this article: Jonas Schreier , Gohar Ghazaryan & Olena Dubovyk (2020): Crop-specific
phenomapping by fusing Landsat and Sentinel data with MODIS time series, European Journal of
Remote Sensing, DOI: 10.1080/22797254.2020.1831969
To link to this article: https://doi.org/10.1080/22797254.2020.1831969
© 2020 The Author(s). Published by Informa
UK Limited, trading as Taylor & Francis
Group.
Published online: 12 Oct 2020.
Submit your article to this journal
Article views: 5
View related articles
View Crossmark data

Crop-specic phenomapping by fusing Landsat and Sentinel data with MODIS
time series
Jonas Schreier
a,b
, Gohar Ghazaryan
a,b
and Olena Dubovyk
a,b
a
Center for Remote Sensing of Land Surfaces (ZFL), University of Bonn, Bonn, Germany;
b
Remote Sensing Research Group (RSRG),
Department of Geography, University of Bonn, Bonn, Germany
ABSTRACT
Agricultural production and food security highly depend on crop growth and condition
throughout the growing season. Timely and spatially explicit information on crop phenology
can assist in informed decision making and agricultural land management. Remote sensing can
be a powerful tool for agricultural assessment. Remotely sensed data is ideally suited for both
large-scale and eld-level analyses due to the wide variability of datasets with diverse spatio-
temporal resolution. To derive crop-specic phenometrics, we fused time series from Landsat 8
and Sentinel 2 with Moderate-resolution Imaging Spectroradiometer (MODIS) data. Using
a linear regression approach, synthetic Landsat 8 and Sentinel 2 data were created based on
MODIS imagery. This fusion-process resulted in synthetic imagery with radiometric character-
istics of original Landsat 8 and Sentinel 2 data. We created four dierent time series using
synthetic data as well as a mix of original and synthetic data. The extracted time series of
phenometrics consisting of both synthetic and original data showed high detail in the nal
phenomaps which allowed intra-eld level assessment of crops. In-situ eld reports were used
for validation. Our phenometrics showed only a few days of deviation for most crops and
datasets. The proposed data integration method can be applied in areas where data from
a single high-resolution source is scarce.
ARTICLE HISTORY
Received 27 November 2019
Revised 25 September 2020
Accepted 30 September 2020
KEYWORDS
Data-fusion; phenometrics;
high-resolution; crops
Introduction
Agricultural crop production and food security highly
depend on crop growth and condition. Timely and spa-
tially explicit information on growth stages is important
for agricultural land management. Remote sensing ana-
lysis opens up possibilities to conduct crop-assessment
across large areas. Remotely sensed data has long been
used for vegetation-analyses (Kogan, 1987; R. Lee et al.,
2002; Tian et al., 2016). the use of vegetation index (VI)
time series is usually used as a base to derive phenological
metrics (R. Lee et al., 2002; Reed & Brown, 2005). The
use of VI time series to derive phenology development
patterns from space is well documented. However, most
studies rely on the use of coarse spatial resolution
data from the National Oceanic and Atmospheric
Administration Advanced Very High-Resolution
Radiometer (NOAA AVHRR) (R. Lee et al., 2002; Tian
et al., 2016) or moderate resolution sensors, such as
Moderate-resolution Imaging Spectroradiometer
(MODIS) (Karlsen et al., 2008; R. Lee et al., 2002).
SPOT data has also successfully been applied for pheno-
metric analyses using TIMESAT, for example, for winter
wheat in China (Lu et al., 2014). High spatial resolution
data, such as from Landsat 8 or Sentinel 2, has been
applied less frequently for phenology analyses (Eklundh
et al., 2012; Skakun et al., 2019).
Due to cloud-cover, high spatial resolution data
might be scarce for certain study areas and time-
frames. Data fusion has been applied by several studies
to overcome a lack of good-quality imagery. Feng Gao
et al. (2017) fused Landsat 5, 7, 8, and MODIS data to
map phenology at field scale for Iowa, the United
States of America (USA). Images were fused using
the spatial and temporal adaptive reflectance fusion
model (STARFM) approach (Gao et al., 2015). The
results of this study were reasonable; however, the
amount of cloud-contaminated Landsat imagery
impacted the analyses significantly. Q. Li et al. (2015)
fused MODIS with Landsat data using the enhanced
algorithm ESTARFM for crop classification. The
authors achieved a good degree of accuracy and
underline that a fusion of MODIS and Landsat is
practical (Q. Li et al., 2015). Complex approaches
like (E)STARFM are one option to achieve data fusion.
On the other hand, simpler methods, such as linear
regression approaches, satisfying result for fusion with
significantly reduced computing intensity, even with
high resolution data (M. H. Lee et al., 2017; Sankey &
Glenn, 2011). Siachalou et al. (2015) fused Landsat 7
data with very high-resolution Rapideye imagery for
vegetation monitoring and classification, which
resulted in satisfactory accuracy. The authors pro-
posed an additional usage of Sentinel 2 data for “a
CONTACT Jonas Schreier s6jsschr@uni-bonn.de Center for Remote Sensing of Land Surfaces (ZFL), University of Bonn, Bonn 53113, Germany
EUROPEAN JOURNAL OF REMOTE SENSING
https://doi.org/10.1080/22797254.2020.1831969
© 2020 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits
unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

more dense set of observations” (Siachalou et al.,
2015). Skakun et al. (2019) assessed winter wheat
yield on central Ukraine using a 30 m spatial resolu-
tion harmonized Landsat 8 and Sentinel 2 product as
input for yield models (Claverie et al., 2018; Skakun
et al., 2019). VIs as well as surface reflectance values
were used for this task. Results of this study under-
lined the potential of fused high-resolution data for
crop and phenology-related analyses.
High-resolution optical sensors potentially allow
for better differentiation of fields, visibility of patterns
on an intra-field level and are less prone to mixed
pixels which may impact data quality (ESA, 2015).
This offers new perspectives in terms of remotely
sensed phenology.
In this paper, we focus on the integration of high-
resolution Sentinel 2, Landsat 8, and MODIS imagery
to analyze crop-specific phenology for a case study in
Ukraine. The objectives of this study were: (i) genera-
tion and analysis of detailed and high-resolution phe-
nometrics based on multi-sensor satellite imagery, as
well as (ii) bridging data gaps using a data-fusion
approach involving Sentinel, Landsat, and MODIS
data.
Materials and methods
Study area
Bila Tserkva in Kiev region, central Ukraine, was cho-
sen as the test site (Figure 1). (Kussul et al., 2012, 2015;
Skakun et al., 2016).
Ukraine is one of Europe’s top agricultural produ-
cers. The state-wide major crops include wheat, sun-
flower, rapeseed, soybeans, sugar beet, and maize
(Food and Agriculture Organization, 2008; World
Data Center, n.d.). Bila Tserkva district, the focus-
area of this study, is covered by more than 60% crop-
land. In this study, we focused on the analysis of the
main crops in this region, namely sunflower, maize,
and soy.
Data and preprocessing
The 250 m resolution MODIS, 30 m resolution
Landsat 8 and high-resolution 10 m Sentinel 2A data
were used for analysis (Barsi et al., 2014; ESA, n.d.;
MODIS Web, n.d.). We acquired data for the year
2016. Due to cloud-contamination, we could not
include data starting from February or earlier.
Therefore, data from March to October was selected
for all sensors to preserve a homogenous period across
datasets. We acquired 8-day MODIS Red and near-
infrared (NIR) bands surface reflectance product. The
acquired MOD09Q1 product is provided corrected for
atmospheric and aerosol effects (NASA LP DAAC,
2014). These data were later used to calculate VIs at
high temporal intervals (MODIS Web, n.d.; NASA LP
DAAC, 2014).
We acquired atmospherically corrected Red and
NIR bands of Landsat 8 surface-reflectance products.
We used integrated fmask data for cloud-masking (Z.
Zhu et al., 2015; Z. Zhu & Woodcock, 2014). Three
overlapping scenes from two paths with different
Figure 1. The location of the study area in the Kyiv region, Ukraine.
2J. SCHREIER ET AL.

overpassing times covered the study area, resulting in
a dense time series with a revisit interval of about
a week (Figure 2).
Next, we acquired high-resolution Sentinel 2A ima-
gery. Level 1 C whole scenes were downloaded for the
desired time frame (ESA, 2015; Müller-Wilm, 2016).
Only tile T35UQR was needed to cover our study area.
Revisiting times differ from region to region for
Sentinel 2A (Li & Roy, 2017). For the Bila Tserkva
district, acquisition intervals varied between three and
seven days. Sentinel 2B had not yet launched at the
time of this study.
The availability of the cloud-free observations var-
ied highly amongst sensors. MODIS provided high-
quality eight-day composites, mitigating issues caused
by cloud coverage (Parkinson et al., 2000). In contrast,
Landsat 8 and Sentinel 2A were heavily affected by
cloudy conditions during the crop growing season in
2016 (Figure 3). There were also scenes with low scene
cloud-cover but with most clouds located inside the
study area, as well as there were images with many
clouds mostly located outside of the study area.
Moreover, some images were taken during dense
haze conditions. Thus, we selected images for further
analyses not only based on the scenes’ cloud-cover
percentages but also based on manual assessment of
each available scene.
We conducted atmospheric correction using ESA’s
Sentinel Application Platform (SNAP) as well as the
Sen2Cor plug-in (Mather & Koch, 2011; Müller-Wilm,
2016). Sen2Cor (v2.2.1) was also used for cloud-masking.
The final selection of the 12 Landsat 8 and 14
Sentinel 2A images used for analysis is shown in
Figure 3. For both Landsat 8 and Sentinel 2A, images
for the early season (March) were mostly taken during
obstructive weather conditions and had to be omitted.
Cloudy conditions resulted in large data gaps for the
early growing stages from April to June for both sensors.
We used all preprocessed images to calculate the
Normalized Difference Vegetation Index (NDVI)
(Rouse et al., 1974). NDVI time series were used as
the baseline of the phenology analysis.
For each crop selected for this study (maize, soy,
sunflower), we conducted a selection of fields for phe-
nology analyses considering the availability of field
data for validation. For crop fields’ identification, we
used the crop map of 2016 provided by the Space
Research Institute (Kussul et al., 2012, 2015; Skakun
et al., 2016). Only fields inside the region covered by
all Sentinel 2A and Landsat 8 scenes were considered
for the analysis. In total, we chose 30 maize fields, 20
soy fields, and 18 sunflower fields.
In-situ phenology data per crop was provided by
the State Hydrometeorological Services of Ukraine.
The dataset included different growing-phases and
their dates on the district level for five different
crops for the year 2016. We used this dataset for
validation of the remotely sensed phenometrics. In
this dataset, there were some gaps. For maize, the
harvest data was missing for Bila Tserkva in 2016.
Therefore, the latest stage recorded was full matur-
ity, which was subsequently taken as an estimate for
the end of the season. Two phenology stages could
be linked to the start of the season for maize: emer-
gence and fifth leaf. We picked the fifth leaf stage as
a proxy for the start of the season as emerging
plants most likely were too small to be detected by
the sensors. Sunflower statistics contained harvest
Figure 2. Three Landsat 8 scenes (in blue) cover the study area (in red).
EUROPEAN JOURNAL OF REMOTE SENSING 3

dates as well as the second leaf stage. However, these
were only recorded for a district about 60 km east
from the study area. Consequently, these dates only
could be used as an estimate. Soybeans contained
dates for the fifth leaf as well as harvest dates for
Bila Tserkva which we used as validation data.
Methodology
Figure 4 summarizes the workflow of the analysis.
Data fusion
To increase the temporal density of Landsat 8 and
Sentinel 2 images, we applied a data-fusion approach.
To create synthetic Sentinel 2 and Landsat 8 data, we
used a linear regression model based on temporally
dense MODIS imagery. This approach was chosen as it
is straightforward in implementation and it reportedly
yields good results (Sankey & Glenn, 2011). Due to
ground truth data only being available for the year
2016 as well as large data-gaps caused by clouds espe-
cially during the beginning of the growing season,
a fusion based on Sentinel-2 and Landsat 8 data
alone was not feasible. Additionally, only Sentinel-2A
data was available as Sentinel-2B was not in orbit by
2016, further reducing the amount of imagery (Li &
Roy, 2017).
Images from the three sensors used in this study
were characterized by slightly different radiometric
and spatial resolution (Barsi et al., 2014; ESA, n.d.;
MODIS Web, n.d.). All three analyzed types of imagery
possess sufficient spatial resolution for the analyses of
agricultural fields in the study area. For the image
fusion, an initial random point sampling was conducted
on a sample-pair of Landsat and MODIS images
acquired on the same day (21.08.2016). This first assess-
ment showed the viability of the NDVI as a regression-
based with an R2-value of 0.69. As a normalized index,
NDVI is less vulnerable to remaining traces of atmo-
spheric effects on band reflectances and radiometric
discrepancies (Zhu & Lei, 2018).
The implementation of the fusion was conducted as
follows. First, we paired each acquired MODIS scene
(8-day temporal interval) with its temporally closest
match of Sentinel 2 (or Landsat 8). Second, each
MODIS image was resampled to match Sentinel 2 (or
Landsat 8) resolution. Third, we used a 3500-point
random sampling of NDVI values to conduct a linear
regression between each image-pair of one MODIS
image and one Sentinel-2 (or Landsat 8) image.
Fourth, resulting regression models were then applied
Figure 3. All available Landsat 8 and Sentinel 2 Scenes from March to October 2016 for the study site (a). Final image selection
after filtering unusable imagery for Landsat 8 (b) and Sentinel 2 (c). Dark bars represent mostly cloud-free conditions specifically
within the study area. Brighter bars signify images with some cloud-cover over the study area. Because of the size of the overall
scenes, even at small scene cloud cover values, the study area might be affected. Scene cloud cover [%] is indicated above the
bars.
4J. SCHREIER ET AL.

to the spatially resampled MODIS scenes for every
MODIS 8-day time step. This fusion routine resulted
in two datasets: a synthetic Sentinel 2-time series and
a synthetic Landsat 8 time series. Both are based on
MODIS and as such have an interval of 8 days.
Phenology derivation
For satellite-based phenometrics calculation, the changes
of the NDVI during the growing season were observed.
The bigger a crop grows during its different stages, the
more ground is covered by it causing a higher proportion
of green biomass per pixel. This, in turn, is reflected by
higher NDVI values (Todd & Hoffer, 1998) (Figure 5)
We analyzed four different time series (Figure 4),
namely: (i) the synthetic Landsat 8 time series, (ii) the
synthetic Sentinel 2-time series, (iii) the time series
consisting of original Landsat 8 data which was tempo-
rally gap-filled using synthetic Landsat 8 data. (hereafter
Figure 4. The workflow of the analysis.
Figure 5. Example of TIMESAT parameters for one season. The green curve shows the original NDVI time series, the red curve the
TIMESAT smoothed NDVI. Parameters: (a) start of the season, (b) peak time of the season, (c) end of the season. The growing
season is marked in grey (modified after L. Eklundh & Jönsson, 2012).
EUROPEAN JOURNAL OF REMOTE SENSING 5

denoted as “mixed Landsat time series”); and (iv) the
time series consisting of original Sentinel 2A data which
was gap-filled using synthetic Sentinel 2 data (there-
after, denoted as “mixed Sentinel time series”).
TIMESAT software was used for the derivation of
phenological parameters for each of the four-time
series (Lars Eklundh & Jönsson, 2015). We calculated
the following phenometrics: start of the season (SOS),
peak time (peak), and end of the season (EOS, Figure
5). Derivation of SOS and EOS requires a threshold
value that can differ by surface, study area, vegetation
type, and available data. Based on empirical tests,
a threshold of 0.2 was used for this study. Gaussian
fitting was selected to create a smooth NDVI curve.
Optimal input-data for TIMESAT should be pro-
vided in equal temporal intervals (L. Eklundh &
Jönsson, 2012). However, we could not achieve per-
fectly homogenous temporal time steps for either
mixed time series (Figure 4) due to the different scenes
for Landsat imagery and variable revisiting times of
Sentinel acquisitions. We selected images manually,
keeping the temporal intervals as consistent as possi-
ble in addition to maintaining an adequate share of
original Sentinel (or Landsat) scenes. Overall, the
mixed Sentinel time series (iv) could be constructed
more densely due to a larger amount of Sentinel
images available. Both mixed time series (iii, iv) start
and end on the 29
th
of March and 31
st
of October.
The values for all four different NDVI and phe-
nology datasets (i, ii, iii, iv) were plotted for the
selected fields and summarized for each field.
NDVI boxplots were created for each crop using
monthly means. We created raster images of the
phenological parameters and extracted values for
each field. The resulting phenology was validated
with the in-situ field-reports. Results for different
crops and time series (i, ii, iii, iv) were compared to
one another.
Results and discussion
Phenology results for all four time series (Figure 6, i to
iv) were mostly precise and revealed similar patterns.
Figure 6 shows the seasonal development of NDVI
values for each crop in the selected fields. Mixed
time series (iii, iv) generally showed higher values
during the peak months of growth and a slight shift
of peak time. The first and last two months were
mostly consistent across datasets. Only mixed
Sentinel time series (iv) showed slightly lowered
NDVI in April. Soy and maize NDVI series were
similar across the datasets. August showed the highest
variation of values within each dataset, possibly caused
by harvest in some of the fields.
Figure 7 shows the phenometrics (SOS, peak, EOS)
derived from each time series (i to iv). For the mixed
time series (iii, iv), the peak time was calculated later
compared to the purely synthetic time series (i, ii).
SOS was estimated slightly later for the mixed time
series (iii, iv) compared to the other datasets. Peak
Figure 6. Monthly NDVI values for the selected fields across the synthetic Landsat 8 (i), synthetic Sentinel 2 (ii), mixed Landsat (iii),
and mixed Sentinel (iv) time series. Dashed lines represent mean values.
6J. SCHREIER ET AL.

time showed the highest spread: values stretched from
the end of June to the end of July (Figure 7). EOS was
distributed mostly uniformly for most fields with
a spread of approximately two weeks (Figure 8).
Figure 8 shows phenometrics in field-level for both
mixed time series (iii, iv) and one synthetic set (i).
Phenomapping results based on the mixed time series
(iii, iv) retained their original resolution and detail.
Figure 7. Phenometrics derived from synthetic Landsat 8 (i), synthetic Sentinel 2 (ii), mixed Landsat (iii), and mixed Sentinel (iv)
time series.
Figure 8. The field-level phenomaps revealing the high spatial level of detail of the mixed time series (iii, iv) results. EOS is shown.
EUROPEAN JOURNAL OF REMOTE SENSING 7

Variations were observable on the field-level for the
mixed Sentinel dataset (iv). Visually, mixed Sentinel
time series showed the highest level of detail at the
intra-field level, revealing phenology patterns inside
a field. Mixed Landsat results (iii) retain clear field
boundaries and a good degree of intra-field detail.
The level of detail of the mixed time series (iii, iv)
phenomaps shows the potential of the proposed
method.
Figure 9 compares the calculated phenometrics
across all datasets and crops with the field data. The
best performance was achieved for maize followed by
sunflower. EOS values showed a smaller variability than
SOS values for most crops and datasets. This could be
related to a sharp NDVI decline at harvest-time in
contrast to a mostly smooth, steady rise of NDVI values
throughout the beginning stages of growth. Both mixed
time series (iii, iv) achieved good overall performance.
For soy, the mixed Sentinel time series (iv) per-
formed best with a deviation of 5 days for EOS and
2 days for SOS compared to the ground truth data. All
other datasets (i, ii, iii) showed higher deviations as
well as high spreads for the beginning of the season.
End of the season was calculated within an error-
margin of fewer than two weeks among all dataset
medians; Synthetic Sentinel (ii) showing 13, synthetic
Landsat (i) 10 and mixed Landsat time series (iii)
5 days of the error.
The phenometrics for maize were calculated more
accurately than soy overall. For the start of the season,
the best performing time series was mixed Sentinel (iv)
with a deviation of 10 days from the in-situ recordings.
However, all other datasets (i, ii, iii) performed simi-
larly well with errors of 13 and 14 days. For the end of
the season, the synthetic Landsat (i) matched the exact
date. Synthetic Sentinel (ii) missed the date by 3 days,
both mixed time series (iii, iv) by 12 days. The start of
the season for sunflower was determined very accu-
rately by synthetic Sentinel (ii), synthetic Landsat (i) as
well as mixed Landsat (iii) with 1 to 2 days of deviation
as well as marginal spread among fields. Mixed
Sentinel (iv) was an outlier, showing about two
weeks of error. The end of the season was best esti-
mated by both mixed time series (iii, iv): mixed
Sentinel (iv) missed the in-situ date by 8 days, mixed
Landsat (iii) by 5 days. Both synthetic time series (i, ii)
deviated by about two weeks. The temporal accuracy
in general is comparable with other studies using
TIMESAT, especially for the start of the season (Lu
et al., 2014). The results for the start of the season were
more reliable than analyses which applied a modified
TIMESAT approach on a different vegetation cover
(grassland) as well as different types of sensors such as
Sentinel 1 (Stendardi et al., 2019).
It should be noted that we connected phenological
stages provided by the in-situ data to SOS and EOS as
optimally as possible. However, a small margin of
error might be attributed to those linkages.
Moreover, the slightly uneven intervals of the mixed
time series (iii, iv) might have affected the results to
a minor degree. Some of the variations can be
explained by the inter-field management differences
(e.g., timing of the harvest).
TIMESAT interprets every loaded dataset as evenly
distributed, causing limitations in terms of its usability
for mixed time series (L. Eklundh & Jönsson, 2012;
Eklundh & Jönsson, 2015). This circumstance also
resulted in good quality images of Landsat and
Sentinel having to be omitted to keep time-intervals
as homogenous as possible for the mixed datasets (iii,
iv). Leaving out these images might have resulted in
a loss of overall quality of achieved results. Using
different vegetational indices, such as EVI, might also
slightly affect phenology output (L. Li et al., 2014; Tan
et al., 2011). Further applications over larger areas may
need an adaptive threshold selection for different
crops (Huang et al., 2019) or the use of methods that
do not use thresholds for Phenometrics estimation,
such as the use of extreme values of the derivative of
the seasonal cycle (Forkel et al., 2015).
Figure 9. Calculated values for SOS and EOS for each time series across all selected fields. Values on the day of the year (DOY).
8J. SCHREIER ET AL.

As for the acquired datasets and their preprocessing,
the performance of cloud-masks was satisfactory. Some
fragments of cloud-shadows remained, which resulted
in slight changes in the selected fields to avoid shadow-
contaminated pixels. Approaches using object-based
cloud detection might achieve good results for masking
shadows, as the clouds themselves have mostly been
masked well (Zhang et al., 2014). Overall, Landsat
fmask (Zhu & Woodcock, 2012) provided more reliable
results than Sentinel 2 Sen2Cor-flags.
Although the overlaying patterns are closely
matched across the different time series, some degree
of differentiation can be observed. It is possible that
data gaps that are temporally close to critical stages,
such as harvesting, could cause these slight differences.
Gaps were especially prevalent during the early season
which might explain the minimally higher diversity of
SOS values. It can be observed that NDVI values were
overall more saturated for Sentinel 2 data and the
related time series. Another vegetation index, such as
EVI, might reveal more comparable levels without
oversaturating. Furthermore, the spectral response
function for Sentinel 2A was changed at one point.
This could also have an effect on the data and the
results (ESA, 2018; Hagolle, 2018).
Despite being simple, the synthetic image creation
approach showed overall good results. Linear regression
model R2-values mostly reached from 0.2 to 0.7. Lowest
correlations were observed in early spring and for image
pairs with high temporal differences in the acquisition.
Landsat pairs with around a week or less of temporal
difference to their MODIS counterpart resulted in good
regression results with R2-values of 0.5 to 0.7 (Figure 10).
Larger time-gaps resulted in R2-values of 0.25 and
below. Other outliers could be a result of unforeseen
changes in the field. MODIS data can be affected by
mixed pixels. For Bila Tserkva district, this was not
a significant issue, as most fields are large enough to be
covered by multiple MODIS pixels: the average field size
is around 20 ha for the Kyiv region which Bila Tserkva
district is part of (State Statistics Service of Ukraine, n.d.).
It was out of the scope of this study to evaluate different
fusion approaches (Bannari et al., 1995).
Both the synthetic Landsat and Sentinel time
series (i, ii) show an overestimation of NDVI-
values in April despite the fusions’ correlation R2-
values being high for both datasets. Seasonal pat-
terns are evident in general. Some areas showed an
increased amount of outlier-values. However, these
were determined to be forests and built-up areas
and as such not relevant for the study. Crop-
specific seasonal signatures were visibly identifiable
in NDVI for both datasets (see Figure 6). With the
emergence of a new, Harmonized Sentinel-Landsat
Figure 10. Selection of scatterplots showing the relation between Landsat 8 and MODIS NDVI for image pairs. Regression lines are
shown in red.
EUROPEAN JOURNAL OF REMOTE SENSING 9

product (HLS), these signatures may be observable
even clearer and the derivation of phenometrics can
be tested over larger areas (Claverie et al., 2018). For
future years with available field-data, this product or
another type of fusion between Sentinel 2 and
Landsat 8 data could be utilized to follow up on
the study.
Conclusion
We performed a derivation of crop-specific using
multi-source remote sensing data as well as a data-
fusion. Three of the study area’s most important crops
were analyzed: maize, sunflower, and soy. Synthetic
Landsat 8 and Sentinel 2 VI data were created based on
MODIS to receive dense time series.
The proposed method shows great potential and
good accuracy for field-level analysis of crop-
phenology. The high spatial resolution phenomaps
achieved in this study provided spatially-explicit infor-
mation at a field level on crop phenology variation that
could further support crop management activities and
has not been previously available for the study site.
Mixed time series containing synthetic, as well as
original data, resulted in phenomaps with great spa-
tial detail: Phenology differences on an intra-field
level could be identified. In-situ field reports were
used as validation data for the phenometrics. All
datasets showed good accuracy, only deviating few
days from on-field reported growing-stages for most
crops and datasets. The linear regression-based
fusion method achieved good results for the creation
of synthetic Landsat 8 and Sentinel 2 based on
MODIS imagery. Integrating radar data or testing
harmonized Landsat 8/Sentinel 2 product time series
from NASA could be a next step to overcome the
limitations of optical data caused by cloud-over and
achieve even denser time series. Further research
could also utilize and compare different phenology-
derivation approaches.
Acknowledgments
The crop-maps and on-field crop-reports were kindly pro-
vided by the Ukrainian Space Research Institute, the
National Academy of Sciences, SSA Ukraine, and the
Ukrainian State Hydrometeorological Services.
Landsat 8 and MODIS data is available from and courtesy
of the U.S. Geological Survey. Copernicus Sentinel 2 data:
courtesy of ESA.
Data Availability Statement
The raw data used for this study (Landsat-8 surface reflec-
tance, MODIS surface reflectance, Sentinel-2A imagery) can
be acquired via USGS (Landsat-8, MOD09Q1.006: https://
earthexplorer.usgs.gov/) and ESA respectively (Sentinel-2A:
https://scihub.copernicus.eu/dhus/#/home). The crop maps
and ground truth data on phenology are subject to permis-
sion of third parties (Ukrainian Space Research Institute,
National Academy of Sciences, SSA Ukraine, and Ukrainian
State Hydrometeorological Services).
Disclosure statement
No potential conflict of interest was reported by the authors.
Funding
Research support was provided by the German Federal
Ministry of Education and Research through its Global
Resource Water (GRoW) funding initiative (Project:
GlobeDrought, grant no. 02WGR1457A-F)
ORCID
Olena Dubovyk http://orcid.org/0000-0002-7338-3167
References
Bannari, A., Morin, D., Bonn, F., & Huete, A. R. (1995).
A review of vegetation indices. Remote Sensing Reviews,
13(1–2), 95–120. https://doi.org/10.1080/
02757259509532298
Barsi, J., Lee, K., Kvaran, G., Markham, B., & Pedelty, J.
(2014). The spectral response of the Landsat-8 opera-
tional land imager. Remote Sensing, 6(10), 10232–10251.
https://doi.org/10.3390/rs61010232
Claverie, M., Ju, J., Masek, J. G., Dungan, J. L.,
Vermote, E. F., Roger, J.-C., Skakun, S. V., & Justice, C.
(2018). The harmonized Landsat and Sentinel-2 surface
reflectance data set [Special issue]. Remote Sensing of
Environment, 219, 145–161. https://doi.org/10.1016/j.rse.
2018.09.002
Eklundh, L., & Jönsson, P. (2012). TIMESAT 3.2 with par-
allel processing-Software Manual. Lund University. http://
web.nateko.lu.se/TIMESAT/docs/TIMESAT32_soft
ware_manual.pdf
Eklundh, L., Ardö, J., Jönsson, P., & Sjöström, M. (2012).
High resolution mapping of vegetation dynamics from
Sentinel-2. Malmö Universtity. http://muep.mau.se/bit
stream/handle/2043/14624/Sentinel_paper_Eklundhetal.
pdf?sequence=2&isAllowed=y
Eklundh, L., & Jönsson, P. (2015). TIMESAT: A software
package for time-series processing and assessment of
vegetation dynamics. In: Kuenzer C., Dech S., Wagner
W. (Eds.) Remote Sensing Time Series. Remote Sensing
and Digital Image Processing, 22, 141–158. Springer
Cham. https://doi.org/10.1007/978-3-319-15967-6_7
ESA. (2015). Sentinel −2 user Handbook (2nd ed.). European
Space Agency Standard Document. https://sentinel.esa.
int/documents/247904/685211/Sentinel-2_User_
Handbook
ESA. (2018). Sentinel-2 mission status report 119. European
Commission. https://sentinel.esa.int/documents/247904/
3347201/Sentinel-2-Mission-Status-Report-119-16-Dec
-2017-05-Jan-2018
ESA. (n.d.). Radiometric - Resolutions - Sentinel-2 MSI - User
guides - Sentinel Online. European Commission. https://
sentinel.esa.int/web/sentinel/user-guides/sentinel-2-msi/
resolutions/radiometric
10 J. SCHREIER ET AL.

Food and Agriculture Organization. (2008). Country report
on the state of plant genetic resources for food and
agriculture. National Centre for Plant Genetic Resources
of Ukraine, Kharkiv. http://www.fao.org/docrep/013/
i1500e/Ukraine.pdf
Forkel, M., Migliavacca, M., Thonicke, K., Reichstein, M.,
Schaphoff, S., Weber, U., & Carvalhais, N. (2015).
Codominant water control on global interannual varia-
bility and trends in land surface phenology and
greenness. Global Change Biology, 21(9), 3414–3435.
https://doi.org/10.1111/gcb.12950
Gao, F., Hilker, T., Zhu, X., Anderson, M., Masek, J.,
Wang, P., & Yang, Y. (2015). Fusing Landsat and
MODIS data for vegetation monitoring. IEEE
Geoscience and Remote Sensing Magazine, 3(3), 47–60.
https://doi.org/10.1109/MGRS.2015.2434351
Gao, F., Anderson, M. C., Zhang, X., Yang, Z., Alfieri, J. G.,
Kustas, W. P., Mueller, R., Johnson, D. M., &
Prueger, J. H. (2017). Toward mapping crop progress at
field scales through fusion of Landsat and MODIS
imagery. Remote Sensing of Environment, 188, 9–25.
https://doi.org/10.1016/j.rse.2016.11.004
Hagolle, O. (2018). Revised spectral bands for Sentinel-2A.
Séries Temporelles. http://www.cesbio.ups-tlse.fr/multi
temp/?p=12618
Huang, X., Liu, J., Zhu, W., Atzberger, C., & Liu, Q. (2019).
The optimal threshold and vegetation index time series
for retrieving crop phenology based on a modified
dynamic threshold method. Remote Sensing, 11(23),
2725. https://doi.org/10.3390/rs11232725
Karlsen, S. R., Tolvanen, A., Kubin, E., Poikolainen, J.,
Høgda, K. A., Johansen, B., Danks, F. S., Aspholm, P.,
Wielgolaski, F. E., & Makarova, O. (2008). MODIS-NDVI
-based mapping of the length of the growing season in
northern Fennoscandia. International Journal of Applied
Earth Observation and Geoinformation, 10(3), 253–266.
https://doi.org/10.1016/j.jag.2007.10.005
Kogan, F. N. (1987). Vegetation index for areal analysis of
crop conditions. Proceedings of 18th Conference on
Agricultural and Forest Meteorology (pp. 103–106).
Kussul, N., Skakun, S., Shelestov, A., Kravchenko, O., &
Kussul, O. (2012). Crop classification in Ukraine using satel-
lite optical and SAR images. International Journal“
Information Models and Analyses”, 2(2), 118–122. ITHEA.
http://www.foibg.com/ijima/vol02/ijima02-02-p03.pdf
Kussul, N., Skakun, S., Shelestov, A., Lavreniuk, M.,
Yailymov, B., & Kussul, O. (2015). Regional scale crop
mapping using multi-temporal satellite imagery. The
International Archives of Photogrammetry, Remote
Sensing and Spatial Information Sciences, 40(7), 45.
https://doi.org/10.5194/isprsarchives-XL-7-W3-45-2015
Lee, M. H., Lee, S. B., Eo, Y. D., Kim, S. W., Woo, J.-H., &
Han, S. H. (2017). A comparative study on generating
simulated Landsat NDVI images using data fusion and
regression method—the case of the Korean Peninsula.
Environmental Monitoring and Assessment, 189(7), 333.
https://doi.org/10.1007/s10661-017-6034-z
Lee, R., Yu, F., Price, K. P., Ellis, J., & Shi, P. (2002).
Evaluating vegetation phenological patterns in Inner
Mongolia using NDVI time-series analysis.
International Journal of Remote Sensing, 23(12),
2505–2512. https://doi.org/10.1080/01431160110106087
Li, J., & Roy, D. P. (2017). A global analysis of Sentinel-2A,
Sentinel-2B and Landsat-8 data revisit intervals and
implications for terrestrial monitoring. Remote Sensing,
9(9), 902. https://doi.org/10.3390/rs9090902
Li, L., Friedl, M. A., Xin, Q., Gray, J., Pan, Y., & Frolking, S.
(2014). Mapping crop cycles in China using MODIS-EVI
time series. Remote Sensing, 6(3), 2473–2493. https://doi.
org/10.3390/rs6032473
Li, Q., Wang, C., Zhang, B., & Lu, L. (2015). Object-based
crop classification with Landsat-MODIS enhanced
time-series data. Remote Sensing, 7(12), 16091–16107.
https://doi.org/10.3390/rs71215820
Lu, L., Wang, C., Guo, H., & Li, Q. (2014). Detecting winter
wheat phenology with SPOT-VEGETATION data in the
North China Plain. Geocarto International, 29(3),
244–255. https://doi.org/10.1080/10106049.2012.760004
Mather, P. M., & Koch, M. (2011). Computer processing of
remotely-sensed images: An introduction. John Wiley &
Sons.
MODIS Web. (n.d.). Specifications. NASA. Retrieved
January 24, 2018, from https://modis.gsfc.nasa.gov/
about/specifications.php
Müller-Wilm, U. (2016). Sentinel-2 MSI—Level-2A proto-
type processor installation and user manual (Darmstadt,
Germany). Telespazio VEGA. http://step.esa.int/thirdpar
ties/sen2cor/2.2.1/S2PAD-VEGA-SUM-0001-2.2.pdf
NASA LP DAAC. (2014). Surface reectance 8-Day L3 global
250m. United States Geological Survey. https://lpdaac.
usgs.gov/dataset_discovery/modis/modis_products_
table/mod09q1
Parkinson, C. L., Greenstone, R., & Closs, J. (2000). EOS
data products Handbook (Vol. 2). NASA. https://ntrs.
nasa.gov/archive/nasa/casi.ntrs.nasa.gov/20010069261.
pdf
Reed, B. C., & Brown, J. F. (2005). Trend analysis of
time-series phenology derived from satellite data. 3rd
International Workshop on the Analysis of Multi-
Temporal Remote Sensing Images 2005 (pp. 166–168).
Rouse, J., Jr, Haas, R. H., Schell, J. A., & Deering, D. W.
(1974). Monitoring vegetation systems in the Great Plains
with ERTS. NASA Special Publication, 351, 309.
Sankey, T., & Glenn, N. (2011). Landsat-5 TM and lidar
fusion for sub-pixel juniper tree cover estimates in
a western rangeland. Photogrammetric Engineering and
Remote Sensing, 77(12), 1241–1248. https://doi.org/10.
14358/PERS.77.12.1241
Siachalou, S., Mallinis, G., & Tsakiri-Strati, M. (2015).
A Hidden Markov models approach for crop classifica-
tion: Linking crop phenology to time series of
multi-sensor remote sensing data. Remote Sensing, 7(4),
3633–3650. https://doi.org/10.3390/rs70403633
Skakun, S., Kussul, N., Shelestov, A. Y., Lavreniuk, M., &
Kussul, O. (2016). Efficiency assessment of multitemporal
C-band Radarsat-2 intensity and Landsat-8 surface reflec-
tance satellite imagery for crop classification in Ukraine.
IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing, 9(8), 3712–3719.
https://doi.org/10.1109/JSTARS.2015.2454297
Skakun, S., Vermote, E., Franch, B., Roger, J.-C., Kussul, N.,
Ju, J., & Masek, J. (2019). Winter wheat yield assessment
from Landsat 8 and Sentinel-2 data: Incorporating sur-
face reflectance, through phenological fitting, into regres-
sion yield models. Remote Sensing, 11(15), 1768. https://
doi.org/10.3390/rs11151768
State Statistics Service of Ukraine. (n.d.). Statistical
Information. Government of Ukraine. Retrieved January
24, 2018, from http://www.ukrstat.gov.ua/
Stendardi, L., Karlsen, S. R., Niedrist, G., Gerdol, R.,
Zebisch, M., Rossi, M., & Notarnicola, C. (2019).
Exploiting time series of Sentinel-1 and Sentinel-2
EUROPEAN JOURNAL OF REMOTE SENSING 11

imagery to detect Meadow phenology in mountain
regions. Remote Sensing, 11(5), 542. https://doi.org/10.
3390/rs11050542
Tan, B., Morisette, J. T., Wolfe, R. E., Gao, F., Ederer, G. A.,
Nightingale, J., & Pedelty, J. A. (2011). An enhanced
TIMESAT algorithm for estimating vegetation phenology
metrics from MODIS data. IEEE Journal of Selected Topics
in Applied Earth Observations and Remote Sensing, 4(2),
361–371. https://doi.org/10.1109/JSTARS.2010.2075916
Tian, F., Brandt, M., Liu, Y. Y., Verger, A., Tagesson, T.,
Diouf, A. A., Rasmussen, K., Mbow, C., Wang, Y., &
Fensholt, R. (2016). Remote sensing of vegetation dynamics
in drylands: Evaluating vegetation optical depth (VOD) using
AVHRR NDVI and in situ green biomass data over West
African Sahel. Remote Sensing of Environment, 177, 265–276.
https://doi.org/10.1016/j.rse.2016.02.056
Todd, S. W., & Hoffer, R. M. (1998). Responses of spectral
indices to variations in vegetation cover and soil
background. Photogrammetric Engineering and Remote
Sensing, 64(9), 915–922. https://www.asprs.org/wp-content/
uploads/pers/1998journal/sep/1998_sep_915-921.pdf
World Data Center. (n.d.). Ukraine: Agricultural overview.
World Data Center Ukraine. Retrieved January 28, 2018,
from http://wdc.org.ua/en/node/29
Zhang, Y., Guindon, B., & Li, X. (2014). A robust approach
for object-based detection and radiometric characteriza-
tion of cloud shadow using haze optimized
transformation. IEEE Transactions on Geoscience and
Remote Sensing, 52(9), 5540–5547. https://doi.org/10.
1109/TGRS.2013.2290237
Zhu, W., & Lei, H. (2018). Urban vegetation coverage mon-
itoring technology based on NDVI. 2018 7th
International Conference on Energy, Environment and
Sustainable Development (ICEESD 2018).
Zhu, Z., Wang, S., & Woodcock, C. E. (2015). Improvement
and expansion of the Fmask algorithm: Cloud, cloud
shadow, and snow detection for Landsats 4–7, 8, and
Sentinel 2 images. Remote Sensing of Environment, 159,
269–277. https://doi.org/10.1016/j.rse.2014.12.014
Zhu, Z., & Woodcock, C. E. (2012). Object-based cloud and
cloud shadow detection in Landsat imagery. Remote
Sensing of Environment, 118, 83–94. https://doi.org/10.
1016/j.rse.2011.10.028
Zhu, Z., & Woodcock, C. E. (2014). Automated cloud, cloud
shadow, and snow detection in multitemporal Landsat
data: An algorithm designed specifically for monitoring
land cover change. Remote Sensing of Environment, 152,
217–234. https://doi.org/10.1016/j.rse.2014.06.012
12 J. SCHREIER ET AL.