remote sensing

Article
Impact of STARFM on Crop Yield Predictions: Fusing MODIS
with Landsat 5, 7, and 8 NDVIs in Bavaria Germany
Maninder Singh Dhillon 1, * , Thorsten Dahms 1,2 , Carina Kübert-Flock 3 , Adomas Liepa 1 ,
Thomas Rummler 4 , Joel Arnault 5 , Ingolf Steffan-Dewenter 6 and Tobias Ullmann 1

1 Department of Remote Sensing, Institute of Geography and Geology, University of Würzburg,

97074 Wurzburg, Germany; [email protected] (T.D.); [email protected] (A.L.);
[email protected] (T.U.)
2 Gauss Center for Geodesy and Geoinformation, Federal Agency for Cartography and Geodesy,
60598 Frankfurt am Main, Germany
3 Department of Remote Sensing, Hessian State Agency for Nature Conservation, Environment and
Geology (HLNUG), 65203 Wiesbaden, Germany; [email protected]
4 Department of Applied Computer Science, Institute of Geography, University of Augsburg,
86159 Augsburg, Germany; [email protected]
5 Institute of Meteorology and Climate Research, Karlsruhe Institute of Technology, Campus Alpin,
82467 Garmish-Partenkirchen, Germany; [email protected]
6 Department of Animal Ecology and Tropical Biology, University of Würzburg, 97074 Wurzburg, Germany;
[email protected]
* Correspondence: [email protected]; Tel.: +49-151-684-27047

Abstract: Rapid and accurate yield estimates at both field and regional levels remain the goal
of sustainable agriculture and food security. Hereby, the identification of consistent and reliable
methodologies providing accurate yield predictions is one of the hot topics in agricultural research.
This study investigated the relationship of spatiotemporal fusion modelling using STRAFM on crop
yield prediction for winter wheat (WW) and oil-seed rape (OSR) using a semi-empirical light use
efficiency (LUE) model for the Free State of Bavaria (70,550 km2 ), Germany, from 2001 to 2019. A
synthetic normalised difference vegetation index (NDVI) time series was generated and validated by
Citation: Dhillon, M.S.; Dahms, T.; fusing the high spatial resolution (30 m, 16 days) Landsat 5 Thematic Mapper (TM) (2001 to 2012),
Kübert-Flock, C.; Liepa, A.; Rummler,
Landsat 7 Enhanced Thematic Mapper Plus (ETM+) (2012), and Landsat 8 Operational Land Imager
T.; Arnault, J.; Steffan-Dewenter, I.;
(OLI) (2013 to 2019) with the coarse resolution of MOD13Q1 (250 m, 16 days) from 2001 to 2019.
Ullmann, T. Impact of STARFM on
Except for some temporal periods (i.e., 2001, 2002, and 2012), the study obtained an R2 of more than
Crop Yield Predictions: Fusing
MODIS with Landsat 5, 7, and
0.65 and a RMSE of less than 0.11, which proves that the Landsat 8 OLI fused products are of higher
8 NDVIs in Bavaria Germany. Remote accuracy than the Landsat 5 TM products. Moreover, the accuracies of the NDVI fusion data have
Sens. 2023, 15, 1651. https:// been found to correlate with the total number of available Landsat scenes every year (N), with a
doi.org/10.3390/rs15061651 correlation coefficient (R) of +0.83 (between R2 of yearly synthetic NDVIs and N) and −0.84 (between
RMSEs and N). For crop yield prediction, the synthetic NDVI time series and climate elements (such
Academic Editor: Guido D’Urso
as minimum temperature, maximum temperature, relative humidity, evaporation, transpiration, and
Received: 10 February 2023 solar radiation) are inputted to the LUE model, resulting in an average R2 of 0.75 (WW) and 0.73
Revised: 13 March 2023 (OSR), and RMSEs of 4.33 dt/ha and 2.19 dt/ha. The yield prediction results prove the consistency
Accepted: 15 March 2023 and stability of the LUE model for yield estimation. Using the LUE model, accurate crop yield
Published: 18 March 2023
predictions were obtained for WW (R2 = 0.88) and OSR (R2 = 0.74). Lastly, the study observed a
high positive correlation of R = 0.81 and R = 0.77 between the yearly R2 of synthetic accuracy and
modelled yield accuracy for WW and OSR, respectively.
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland. Keywords: MOD13Q1; precision agriculture; fusion; sustainable agriculture; decision making; winter
This article is an open access article wheat; oil-seed rape; crop models
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).

Remote Sens. 2023, 15, 1651. https://doi.org/10.3390/rs15061651 https://www.mdpi.com/journal/remotesensing

Remote Sens. 2023, 15, 1651 2 of 35

1. Introduction
Accurate predictions of grain yield at both field and regional scales remain a goal for
sustainable agriculture and food security [1,2]. The delivery of timely crop monitoring
and accurate crop yield estimates is of great value for the formulation of food policies, the
regulation of food prices, and agricultural management and is urgently needed for the
development of sustainable agriculture [3,4]. Among different crop types, oil-seed rape
(OSR) (Brassica napus) and winter wheat (WW) (Triticum aestivum) are major crops with high
economic value for animal feed, biodiesel production, pollination, biodiversity, and human
consumption in the European Union [5,6]. In Germany, WW (total production in 2016 was
24.6 million tons) and OSR (4.9 million tons) are crops of significant importance, generally
cultivated as high input and conventionally managed monocultures [6–9]. The future
climatic changes and increasing climatic variability have diverted the increasing grain
yield trend of these crops towards maintaining yield stability [7]. Therefore, the accurate
yield estimates of WW and OSR could contribute positively to agricultural management
practises and optimise resource use to stabilise yields in the future.
Remote sensing (RS) technology can be used to determine and monitor the features of
the earth’s surface by providing synoptic, timely, and cost-effective information about the
earth’s surface [10,11]. Many studies have implemented RS-based methodologies to esti-
mate the crop production of different crop types at different geographical locations [12–17].
Landsat (L), SPOT, World View, and Sentinal-2 (S) satellite data with a medium spatial
resolution of 10–100 m were utilised to assess and estimate agricultural production at
regional and local scales [1,10]. The availability of historical RS data since 1972 has also
increased the potential of science to invest, design, and implement accurate and reliable
methodologies by validating the methods with old yield data sets [18–20]. Until now,
various studies have implemented different methodologies (such as interpolation [21,22],
extrapolation [23,24], vegetation indices [25,26], linear regression models [27], crop growth
models (CGMs) [17,28], machine learning (ML) [29–31], and deep learning (DL) [32,33]
using the RS data and accurately predicted crop yields in almost every corner of the world.
However, to adequately justify their methods’ reliability, stability, and preciseness, very
few studies have consistently tested their methodologies for yield prediction for more than
five years.
CGMs using the RS data as input parameters successfully attempted to estimate crop
yields by covering vast spatial scales and updating the information temporally [17,34–37].
Many CGMs have been used in crop monitoring for different design purposes, regional
environments, and crop types [35]. Some very famous models driven by various factors
such as radiation, water, or soil are named as AquaCrop [38], soil–water–atmosphere–plant
(SWAP) [39], agricultural production systems simulator (APSIM) [40], simple and universal
crop growth simulator (SUROS) [41], semi-empiric light use efficiency (LUE) model [42],
world food study model (WOFOST) [43], Carnegie–Ames–Stanford Approach (CASA) [44],
and the simple algorithm for yield estimate (SAFY) model [45]. However, most CGMs
are complicated and time-consuming and require many input parameters that could be
difficult to obtain or substitute through RS data. LUE and AquaCrop are proven to be more
precise, accurate, and reliable by the previous literature [17]. However, their performance
stability is not determined, as no study has analysed their performance for more than two
years at the same study site.
Crop yield prediction at regional, national, and global scales has been conducted based
on both climate data and RS data [46]. Temperature, solar radiation, and precipitation,
as well as the normalised difference vegetation index (NDVI) and leaf area index (LAI),
are generally considered the primary climatic and satellite-based input variables used in
CGMs [47,48]. Therefore, the quality of RS input to CGMs might impact the accuracy of
the predicted yield. Even though the RS has broadened the spatial and temporal range
of CGMs, the cloud and shadow gaps in the optical satellite data can hinder or limit
CGMs from producing accurate yield results [49,50]. Many studies have successfully used
multitemporal data fusion, combining the data obtained from two different sensors with
Remote Sens. 2023, 15, 1651 3 of 35

different spatial and temporal scales, to fill the data gaps [17,51–53]. Due to its public
availability of code and simplicity of usage, the spatial and temporal adaptive reflectance
fusion model (STARFM) [54] is widely used to combine L/S with the moderate resolution
imaging spectroradiometer (MODIS) for the application of crop monitoring [55–58]. In a
previous study, we tested blending different high (L (30 m, 16 days) and S (10 m, 5–6 days))
and coarse (MODIS: MCD43A4, MOD13Q1, MOD09GQ, and MOD09Q1) spatial resolution
products for different land use classes using the STARFM. The study found that both
L-MOD13Q1 (30 m, 16 days) (R2 = 0.62 and RMSE = 0.11) and S-MOD13Q1 (10 m, 16 days)
(R2 = 0.68 and RMSE = 0.13) are suitable for the application of agricultural monitoring, with
the former having the upper hand due to its fast and easy processing with lesser storage
requirements [59].
Thus, the present study uses the L-MOD13Q1 NDVI product and high-resolution
climate parameters (2 km, eight days) as inputs to the LUE model (considered the most
accurate, precise, and reliable [17]) for predicting crop yields of WW and OSR at a regional
scale for Bavaria from 2001 to 2019. This long-term yield prediction of both crop types
would investigate the stability and preciseness of the LUE model by validating the modelled
yield with district level Bayerisches Landesamt für Statistik (LfStat) data of Bavaria with a
95% confidence interval. The specific research objectives include: (i) finding the potential
of STARFM for blending the long-term NDVI time series; (ii) investigating the preciseness
and stability of the LUE model by validating the modelled yield at district level in Bavaria
from 2001 to 2019; and (iii) exploring the impact of the fused NDVI input time series on the
accuracy of the modelled yields.

2. Materials and Methods

The general workflow of the study is shown in Figure 1. The flow diagram is divided
into three parts: (1) data fusion; (2) generation and validation of L-MOD13Q1 NDVI time
series from 2001 to 2019; and (3) comparative analysis between fused (L-MOD13Q1) and
non-fused (L-MOD13Q1) products in crop yield modelling 2019; and then, modelling crop
yields using L-MOD13Q1 NDVI for WW and OSR from 2001 to 2019. The first part was a
testing phase that investigated the suitable synthetic NDVI product (which is L-MOD13Q1)
for the agricultural class of Bavaria for the year 2019 (completed in the preceding work [59]).
The second section is an extension of the first section, and it generates and validates the
NDVI time series of L-MOD13Q1 for eighteen more years (i.e., from 2001 to 2018) using
the same methodology as the previous section (as used for 2019). In the third section,
the output NDVI time series of part 2 and the climate elements are used as inputs to
the LUE model, which estimates the crop yields of WW and OSR from 2001 to 2019 in
Bavaria. The satellite NDVI and the climate data are selected for the respective starts
and ends of the seasons for WW and OSR from 2001 to 2019. Both inputs are masked
for WW and OSR using the InVeKos data that was available from 2005 to 2019 (source:
www.ec.europa.eu/info/index_en, accessed on 21 June 2021).
As crop field information was unavailable from 2001 to 2004, InVeKos field data from
2005 to 2009 was used to classify the WW and OSR fields in their respective years. Finally,
the obtained crop yield is validated using the LfStat data at the regional level in Bavaria (the
regional map is shown in Figure 2). Because the validation data is available at a regional
scale, the field outputs of every region were converted to a single regional value by totalling
the pixel values of every field. The satellite data sets were downloaded and preprocessed
in Google Earth Engine (GEE), and the fusion analysis is performed in R (version 4.0.3)
using R-Studio.
Remote
RemoteSens. 2023,15,
Sens.2023, xx,1651
x FOR PEER REVIEW 4 of 35

Figure 1. The conceptual framework of the study is divided into three parts: Part 1 states the
Figure 1. The conceptual framework of the study is divided into three parts: Part 1 states the data
data fusion for 2019 to investigate the best synthetic NDVI time series product (this section was
fusion for 2019 to investigate the best synthetic NDVI time series product (this section was already
already completed in our previous study [59]); Part 2 generates and validates the synthetic NDVI
completed in our previous study [59]); Part 2 generates and validates the synthetic NDVI time se-
time series
ries from 2001from 2001 for
to 2019 to the
2019 for theL-MOD13Q1;
product product L-MOD13Q1; and Part the
and Part 3 performs 3 performs
comparativethe analysis
compar-
ative analysis
to compare thetoperformance
compare the of performance of fused
fused (L-MOD13Q1) (L-MOD13Q1)
and and non-fused
non-fused (MOD13Q1) NDVI(MOD13Q1)
time series
NDVI
in croptime
yieldseries in crop
prediction for yield
2019 andprediction for 2019
then estimates andand then estimates
validates and validates
the crop yield for Bavariathebycrop
in-
putting
yield forthe L-MOD13Q1
Bavaria time the
by inputting series and climate
L-MOD13Q1 elements
time to a climate
series and semi-empiric Light
elements to Use Efficiency
a semi-empiric
(LUE)Use
Light model; STARFM
Efficiency = Spatial
(LUE) model; and Temporal
STARFM Adaptive
= Spatial andReflectance Fusion Model;
Temporal Adaptive NDVI =Fusion
Reflectance Nor-
malised Difference Vegetation Index; L-MOD09GQ = Landsat-MOD09GQ;
Model; NDVI = Normalised Difference Vegetation Index; L-MOD09GQ = Landsat-MOD09GQ; L-MOD09Q1 = Landsat-
MOD09Q1; L-MCD43A4 = Landsat-MCD43A4; L-MOD13Q1 = Landsat-MOD13Q1; S-MOD09GQ =
L-MOD09Q1 = Landsat-MOD09Q1; L-MCD43A4 = Landsat-MCD43A4; L-MOD13Q1 = Landsat-
Sentinel-2-MOD09GQ; S-MOD09Q1 = Sentinel-2-MOD09Q1; S-MCD43A4 = Sentinel-2-MCD43A4;
MOD13Q1; S-MOD09GQ = Sentinel-2-MOD09GQ; S-MOD09Q1 = Sentinel-2-MOD09Q1;
S-MOD13Q1 = Sentinel-2-MOD13Q1; PAR is photosynthetically active radiation, and FPAR is the
S-MCD43A4
fraction of PAR= Sentinel-2-MCD43A4;
absorbed by the canopy. S-MOD13Q1 = Sentinel-2-MOD13Q1;
APAR = Absorbed PhotosyntheticallyPAR is photosyntheti-
Active Radiation.
cally active radiation, and FPAR is the fraction of PAR absorbed by the canopy. APAR = Absorbed
Photosynthetically Active Radiation.
Remote Sens. 2023, xx, x FOR PEER REVIEW
Remote Sens. 2023, 15, 1651 5 of 35

Figure 2. Overview
2. Overview
Figure of of
thethe study
study region.
region. The
The LCLC mapofofBavaria
map Bavariaisisobtained
obtainedby bycombining
combiningmultiple
multiple
inputs of of
inputs landcover
landcover maps,
maps,suchsuchas
asthe
theAmtliche Topographisch-KartographischeInformations
Amtliche Topographisch-Kartographische Informations Sys-
System,
tem, Integrated
Integrated Administration
Administration Control
Control System
System (which
(which provides
provides thethe crop
crop field
field information),
information), andand
the
theCorine
CorineLC,LC, into
into oneone
map.map. Agriculture
Agriculture (peach(peach
green) green) dominates
dominates mainly inmainly in the northwest
the northwest andof
and southeast
southeast of Bavaria, while forest and grassland classes (dark green and yellow, respectively) dom-
Remote Sens. 2023, 15, 1651 6 of 35

Bavaria, while forest and grassland classes (dark green and yellow, respectively) dominate in the
northeast and south. The LC map is overlayed by the district map of Bavaria. The enlargement
(displayed with a dark red box on the top right map) shows the urban area of the city of Würzburg,
with the oil-seed rape (OSR) fields (dark orange) and the winter wheat (WW) fields (dark green) in
2019. A brief description of the regions of Bavaria is shown in Figure A1.

2.1. Study Area

The study area is Bavaria, which is one of the federal states of Germany located
between 47◦ N and 50.5◦ N and between 9◦ E and 14◦ E (Figure 2). As the largest state
of Germany, Bavaria covers an area of approx. 70,550 km2 , covering almost one-fifth
of Germany. The diverse topography of the region, with higher elevations in the south
(Bavarian Alps) and east (Bavarian Forest and Fichtel Mountains), impacts the climate of
the state. The mean annual temperature ranges from −3.3 ◦ C to 11 ◦ C, and the mean annual
precipitation sums range from approx. 500 to above 3100 mm. In 2019, about 36.91% of the
area of the state was covered by forest, and 31.67% by agriculture [59]. More than half of the
arable land is used to grow cereals, where WW predominates with 37%, followed by winter
barley (25%), summer barley (12%), and grain maize (8%) [60]. While OSR predominates
in the oil-producing crops in the state. The federal state is divided into 71 Landkreise
(rural districts) and 26 Kreisfreie Städte (city districts). A brief description of the regions of
Bavaria is shown in Figure A1.

2.2. Data
The study collected satellite data (with different spatial and temporal resolutions),
climate data, and vector data for the period of 2001 to 2019. A brief description of the data
used in the present study, with their spatial and temporal resolutions and references, is
shown in Table 1.

Table 1. A summary of the collected datasets for fusion modelling and winter wheat’s (WW) and oil-
seed rape’s (OSR) crop modeling. The satellite data used for fusion and crop modelling are Landsat
5, 7, and 8 and Moderate Resolution Imaging Spectroradiometer (MODIS) MOD13Q1; the climate
parameters are minimum temperature (◦ C) (Tmin), maximum temperature (◦ C) (Tmax), dewpoint
temperature (◦ C) (Tdew), relative humidity (%) (RH), evaporation (mm) (Ep), transpiration (mm)
(Tp), and solar radiation (MJm−2 day−1 ) (Rs); Shuttle Radar Topography Mission (SRTM) elevation
data of Bavaria; InVeKos data provides the fields of WW and OSR for Bavaria from 2005 to 2019; the
Bayerisches Landesamt für Statistik (LfStat) data provides the crop yield information (dt/ha) of WW
and OSR at district level in Bavaria from 2001 to 2019.

Data Product Name Resolution Spatial-Temporal References

Tmin, Tmax, Tdew, 2000 m, 1 day https://www.uni-augsburg.de/de/fakultaet/fai/geo/
Climate data
RH, Ep, Tp, Rs 2001–2019 (accessed on 21 June 2021)
30 m, 16 days
Landsat www.usgs.gov (accessed on 21 June 2021)
Satellite data 2001–2019
250 m, 16 days
MODIS (MOD13Q1) www.lpdaac.usgs.gov (accessed on 21 June 2021)
2001–2019
https://www.usgs.gov/centers/eros
Elevation data SRTM 30 m
(accessed on 15 December 2022)
www.ec.europa.eu/info/index_en
InVeKos 2005–2019
(accessed on 21 June 2021)
Vector data
https://www.statistikdaten.bayern.de/genesis/online/
LfStat 2001–2019
(accessed on 21 June 2021)

2.2.1. Satellite Data

The present study used L-MOD13Q1 (30 m, 16 days) NDVI time series generated
by [59] as an input to the LUE model for nearly two decades (2001 to 2019). The L-
MOD13Q1 time series needed a pair of high (Landsat: high pair) and coarse (MODIS:
low pair) spatial resolution data for fusing together to generate a cloud and shadow-free
Remote Sens. 2023, 15, 1651 7 of 35

synthetic time series using the STARFM algorithm. With the aim of generating a continuous
cloud-free and shadow-free time series (that covers the time frame of 2001 to 2019), high-
pair data sets such as Landsat 5 Thematic Mapper (TM) (1984 (launched)-2013 (ended)),
Landsat 7 Enhanced Thematic Mapper Plus (ETM+) (1999–2003 (stripes in the data after
this date due to scan line corrector failure)), and Landsat 8 Operational Land Imager
(OLI) (2013-present) were used. The Landsat data arrived with different spectral bands,
i.e., coastal/aerosol, blue, green, red, near-infrared (NIR), shortwave infrared (SWIR) 1,
and SWIR 2. The snow, shadow, and cloud cover were removed from the Landsat data
using the “pixel_qa” quality assessment band generated using the C function of the mask
(CFMask) algorithm. The number of cloud-free scenes (0% cloud cover based on CFMask)
available every year (N) is shown in Table 2. Due to the difference in surface reflectance
and atmospheric conditions, there is a considerable variation between the spectral values
of Landsat sensors, which may have significant influences depending on the Landsat
data application [61]. Therefore, the study performed the inter-sensing harmonisation of
the NDVI bands (calculated using NIR and red bands) of Landsat sensors, applying the
coefficients proposed by [61] and derived using ordinary least squares (OLS) regression.
The pre-processing steps were performed using the platform Google Earth Engine.

Table 2. A summary of the collected cloud-and shadow-free Landsat 5, Landsat 7, and Landsat
8 datasets available every year with their day of the years (DOYs) between the start and end of the
seasons of WW and OSR from 2001 to 2019. N is the total number of Landsat scenes available per
year for WW and OSR.

Year N DOYs Year N DOYs

2001 2 81, 161 2011 7 65, 81, 113, 129, 145, 177, 225
2002 3 33, 145, 161 2012 5 49, 65, 81, 129, 145
2003 4 65, 129, 177, 193 2013 5 65, 129, 161, 193, 209
2004 4 33, 65, 97, 161 2014 6 65, 81, 113, 161, 177, 209
2005 6 17, 65, 81, 97, 177, 241 2015 4 65, 97, 145, 209
2006 6 33, 129, 145, 161, 177, 193 2016 8 17, 65, 81, 113, 129, 161, 177, 193
2007 6 49, 81, 113, 145, 161, 193 2017 4 97, 129, 145, 225
2008 6 65, 81, 129, 145, 177, 193 2018 7 49, 81, 113, 129, 145, 177, 193
2009 6 33, 97, 113, 145, 161, 209 2019 5 49, 81, 145, 177, 193
2010 5 33, 113, 129, 145, 193

The Landsat products were generated using the Landsat Ecosystem Disturbance
Adaptive Processing System (LEDAPS), which applies atmospheric correction, geometric
correction, and calibration procedures to the raw data. During the atmospheric correction
step, the impact of atmospheric scattering and absorption is removed from the raw data,
and a surface reflectance product is generated independent of atmospheric effects. The
geometric correction corrected the viewing angles to remove the effects of the satellite’s
position and attitude at the time of image acquisition. This correction ensures that the pixels
are accurately georeferenced and aligned with each other. Lastly, the calibration procedures
applied during the LEDAPS processing correct for spectral band configuration, ensuring
that the reflectance values across different spectral bands are consistent and accurate.
In addition, for the low pair, the study selected the MOD13Q1 V6 product, which
provides an NDVI value per pixel with 250 m spatial and 16-day temporal resolution.
Based on the quality information, pixels with noise (NDVI values <−1 and >+1) were
masked out. Both the day of acquisition and quality information were considered while
generating the NDVI values from the product. For crop modelling, this study input the
eight-day satellite datasets from the stem elongation phases till the flowering stages of
both WW and OSR. The parameters for the growth season of OSR were taken from a
literature review that accurately monitored the growth timing and condition of the crop
based on NDVI and the normalised difference yellowness index (NDYI) (calculated using
the green and blue bands [62]) using the unmanned aerial vehicles (UAVs) in Germany [63].
Remote Sens. 2023, 15, 1651 8 of 35

The phenological stages for WW were referenced from the literature that detected the
phonological development of the crop using the time series of Sentinel-1 and Sentinel-2 in
Germany [64]. The study compared the phenology results with the BBCH scale (Biologische
Bundesanstalt, Bundessortenamt, and CHemische Industrie), which is a system used
worldwide by research and administration to standardise phenologically similar growth
stages of multiple plant species [64,65]. Therefore, the start (the stem elongation phase)
and end (the flowering stage) of the seasons of OSR and WW were taken as 15 February to
20 April [63] and 15 April to 30 June from 2001 to 2019 [64], respectively.

2.2.2. Climate Data

For this study, the climate data from 2001 to 2019 with one-day temporal resolution
were obtained by dynamically downscaling the ECWMF reanalysis 5th generation (ERA5)
dataset to a horizontal grid resolution of 2000 m using the hydrologically enhanced weather
research and forecasting model [66–68]. The ERA5 data were provided by the European
Centre for medium-range weather forecasts. A detailed analysis of the downscaling ap-
proach is provided by [69,70]. The climate data were used as one of the inputs to the
LUE model, which requires temperature, solar radiation, evapotranspiration, and relative
humidity (Figure 1). Prior to input to the model, all climate elements were synchronised
with the LUE model by aggregating them into eight days of temporal periods. Similar to
the satellite data, the present study considered the eight-day climate data for the same start
and end of the seasons for WW and OSR as described in the Section 2.2.1.

2.2.3. Elevation Data

The study made use of the shuttle radar topography mission (SRTM) digital elevation
data for Bavaria [71]. The data had a spatial resolution of 30 m. For this study, the SRTM was
used to correlate modelled crop yields with the elevation above sea level. The visualisation
of the data is shown in Figure A2.

2.2.4. InVeKos Data

The present study made use of the InVeKos data to obtain the field base information
of WW and OSR from 2005 to 2019 for Bavaria. The InVeKos data were collected through
the integrated administration control system (www.ec.europa.eu/info/index, accessed
on 21 June 2021), which was available for all agricultural plots in European Union (EU)
countries by allowing farmers to graphically indicate their agricultural area.

2.2.5. LfStat Data

The Bayerisches Landesamt für Statistik (LfStat) provided the crop yield information
for 29 crop categories, including WW and OSR, in Bavaria on a district level from 2001 to
2019 (source: www.statistikdaten.bayern.de/genesis/online/, accessed on 21 June 2021,
Statistics Code: 41241). The LfStat data were used to validate the modelled yield informa-
tion of the LUE model. The validation results were used to check the model’s accuracy,
consistency, and stability in generating the yield results in the region. The validation was
limited to the rural regions, and the city districts were excluded (Figure A1).

2.3. Method
2.3.1. STARFM
The STARFM method [54] was used to fuse Landsat and MOD13Q1 to generate the
synthetic NDVI time series with high spatial and temporal resolution from 2001 to 2019. As
this paper is an extension of our previous paper, the detailed methodology of STARFM’s
generation of L-MOD13Q1 time series was explained in [17,59].

2.3.2. LUE Model

The LUE model was based on the light use efficiency principle [72,73], and it was
coupled with the RS data by using a similar methodology as [17,42]. The model was based
Remote Sens. 2023, 15, 1651 9 of 35

on a semi-empirical approach and calculated the FPAR [74] and aboveground biomass (at
an eight-day temporal resolution) as follows:

EOS
Biomass = ∑ (PAR ∗ FPAR)∗ ∈ (1)
SOS

FPAR = 1.222 ∗ NDVI + 0.1914 (2)

EOS
∑ Tminmin 0 ∗ VPD0 ∗ Ks ∗ ∈o


∈= (3)
SOS

where PAR is photosynthetically active radiation (MJ m−2 d−1 ), FPAR is the fraction of PAR
absorbed by the canopy, SOS and EOS are the start and end of seasons of WW and OSR, and
є is the actual light-use efficiency (g C M J−1 ). The total aboveground biomass calculated by
the LUE model is equivalent to the net primary productivity (NPP) (kg ha−1 yr−1 ). A brief
explanation of the model with a flow diagram was described in our previous study [17].
The specific model was not only selected for its performance but also for its high processing
speed and low requirement of input parameters as compared to the other CGMs. The
model was calibrated by using values from the previous literature, as follows: The study
used a minimum lethal temperature value of −2 ◦ C for both WW and OSR [75–77]. In the
other studies, the optimal minimum values of temperature for WW and OSR at growth
stages were 10 ◦ C and 12 ◦ C, respectively [75–77]. For the vapour pressure deficit (VPD),
the present study followed [78], which had analysed the environmental impact on leaf gas
exchange in WW with minimum and maximum values of 1.5 and 4.0 kPa, respectively. The
value for optimal light use efficiency was used as 3 gC/MJ [79].

2.3.3. Sensitivity Analysis

The study performed a sensitivity analysis of the LUE model for both WW and OSR
in Bavaria from 2001 to 2019. During the analysis, the impact of climate stress factors was
nullified, and the biomass was calculated by replacing the actual light use efficiency (ε)
values with the optimal (εo ) values (Equation (4)).

EOS
Biomass = ∑ (PAR ∗ FPAR)∗ ∈o (4)
SOS

2.3.4. Statistical Analysis

Both the STARFM NDVI and the LUE-modelled crop yield of WW and OSR were
validated using the observed NDVI and LfStat crop yield (with 95% confidence intervals)
from 2001 to 2019, respectively. The quality (R2 ) and the precision (root mean square
error (RMSE)) of the obtained results were calculated using the linear regression model
(LRM), which aimed to establish a linear relationship between the referenced NDVI/or
measured yield (an independent variable) and the synthetic NDVI/or modelled yield (a
dependent variable). The correlation plots between the number of Landsat scenes and the
synthetic NDVI accuracy from 2001 to 2019 were generated by calculating the correlation
coefficient (R) (Equation (5)). R values lie between −1 (strong negative correlation between
two variables) and 1 (strong positive correlation between two variables). The statistical
parameters used to validate the accuracy of modelled yield and synthetic NDVI are R2
(Equation (6)), mean error (ME) (Equation (7)), and RMSE (Equation (8)). The Equation (9)
calculates the yield percent difference (%), which was calculated for every region of Bavaria.
The yield percent difference was analysed in six categories: less than −4, −4 to −2, −2 to 0,
0 to 2, 2 to 4, and more than 4.

n(∑ Oi ∗Pi ) − (∑ Oi )(∑ Pi )

R= r    , (5)
2 2
( n ∑ Oi ) − (∑ Oi ) ( n ∑ Pi ) − (∑ Pi )
2 2
Remote Sens. 2023, 15, 1651 10 of 35

2
2 ((∑ Pi − P0 ) Oi − O0 )
R =
2
2 , (6)
∑ Pi − P0 ) ∑ Oi − O0 )
1 n
n i∑
ME = (Oi − Pi )2 (7)
=1
√
RMSE = ME, (8)
! !
referenced yieldy − modelled yieldy
Yield Percent Difference = Mean ∗ 100 (9)
referenced yieldy

where Pi is the predicted value, Oi is the observed value, P’ is the predicted mean, O0 is
the observed mean value, and n is the total number of observations, referenced yieldy
is the LfStat yield of every district from 2001 to 2019, and modelled yieldy is the LUE
generated yield of every district from 2001 to 2019. The significance of the obtained results
was obtained by observing the probability value (p-value), which was calculated using
the LRM with a H0 that there is no correlation between the referenced and the modelled
or synthetic values and an H1 that the correlation exists. The test was performed at a
significance level (or alpha (α)) of 0.05. A p-value lower than 0.05 indicated that the model
is significant, and it rejected the H0 that there was no correlation. The correlation was
calculated between the accuracy of synthetic NDVI and crop yield on a yearly basis using
Equation (3). This calculated the relationship of data fusion with crop yield prediction
results by the LUE model.

3. Results
3.1. Validation of Synthetic Remote Sensing Time Series from 2001 to 2019
For nineteen years (i.e., from 2001 to 2019), the STARFM performed significantly
for yearly synthetic output (having a p-value < 0.05); this rejected the H0 of the linear
regression model that there was no correlation between the synthetic and referenced NDVI
(Figure 3a–s). After generating the yearly scatter plots, the synthetic products’ R2 and
RMSE values were analysed. Among all years, the highest accuracy and precision were
obtained for 2016 and 2018, with an average R2 of 0.75 and RMSE of 0.09. For 2005, 2006,
2007, 2009, 2011, 2013, 2014, 2017, and 2019, the R2 values were higher than 0.60 and the
RMSE values were lower than 0.12. In other years, such as 2003, 2004, 2008, and 2010, the
R2 and RMSE values lied within 0.60 to 0.62 and 0.10 to 0.14, respectively. However, the
rest of the temporal period (i.e., 2001, 2002, 2012, and 2015) resulted in lower R2 (<0.60) and
RMSE (>0.13) values.
The results proved that the yearly accuracy assessment of the synthetic products is
impacted by the total number of Landsat scenes (N) available every year (Figure 4a,b). A
high positive (R = +0.83) and negative (R = −0.84) correlation was seen between the yearly
quality (R2 ) and preciseness (RMSE) of the synthetic NDVI products with N, respectively.
For example, 2011, 2016, and 2018 were the most accurate years (R2 > 0.68 and RMSE = 0.09)
with a total N of more than 7. Similarly, 2001 and 2002 had the least R2 (< 0.50) and highest
RMSE (> 0.15) with the fewest available Landsat scenes (N = 2/3) in both years. The overall
accuracy of L-MOD13Q1 for nineteen years was R2 of 0.62 and RMSE of 0.12, with an
average of 5 N every year.
 Remote
Remote Sens. Sens.
2023, 15,2023,
1651xx, x FOR PEER REVIEW 12 of 3711 of 35

(a) (b) (c) (d) (e)

(f) (g) (h) (i) (j)

(k) (l) (m) (n) (o)

Figure 3. Cont.
Remote Sens. 2023, 15, 1651 Remote Sens. 2023, xx, x FOR PEER REVIEW 12 of 35 13

(p) (q) (r) (s)

Figure 3. The scatter plots Figure 3. The scatter

(a–s) compare plots (a–s)
the accuracies compare(referenced
of Landsat the accuracies
NDVI)of Landsat (referenced(synthetic
with L-MOD13Q1 NDVI) with L-MOD13Q1
NDVI) (synthetic
for 2001 to 2019. TheNDVI)
valuesfor
of2001
the to 2019.
values
2 of the statistical parameters such as R 2 and RMSE and the total number of Landsat scenes available every year (N) are displayed at the to
statistical parameters such as R and RMSE and the total number of Landsat scenes available every year (N) are displayed at the top of each plot. Every plot contains
each plot. Every plot contains a solid line (1:1 line) that is used to visualise the correlation of pixels between the referenced and synthetic N
a solid line (1:1 line) that is used to visualise the correlation of pixels between the referenced and synthetic NDVI values. The dashed line represents the regression
values. The dashed line represents the regression line. The colour of scatter plots depicts the density of points (yellow: low, blue: high).
line. The colour of scatter plots depicts the density of points (yellow: low, blue: high).
Remote Sens. 2023, xx, x FOR PEER REVIEW
Remote Sens. 2023, 15, 1651 13 of 35

Remote Sens. 2023, xx, x FOR PEER REVIEW

(a) (b)
Figure 4. The correlation plots between the total number of Landsat scenes per year (N) and (a) R22
Figure 4. The correlation plots between the total number of Landsat scenes per year (N) and (a) R
values and (b) RMSE values obtained during the accuracy assessment of referenced and synthetic
values and (b) RMSE values obtained during the accuracy assessment of referenced and synthetic
NDVI products from 2001 to 2019. The correlation coefficient refers to R (see Equation (5)).
NDVI products from 2001 to 2019. The correlation coefficient refers to R (see Equation (5)).

(a) On comparing the yearly fusion results on a DOY basis, (b) the DOYs 113, 129, and
193 had 4.
theThe
highest average accuracy 2
Figure correlation plots betweenwith an Rnumber
the total of more than 0.65
of Landsat and per
scenes a RMSE lesser
year (N) andthan
(a) R2
0.10 (Figure 5a,b). The DOYs of 33 to 97 and 145 to 177, with low R 2 (<0.60) and high RMSE
values and (b) RMSE values obtained during the accuracy assessment of referenced and synthetic
(>0.11), were obtained.
NDVI products from 2001 to 2019. The correlation coefficient refers to R (see Equation (5)).

(a) (b)

Figure 5. The day of the year (DOY)-based comparison of correlation coefficients between (a) R2
values and (b) RMSE values obtained during the accuracy assessment of referenced and synthetic
NDVI products from 2001 to 2019. The correlation coefficient refers to R (see Equation (5)).

(a) (b)
3.2. Comparative Analysis between Crop Yield Accuracy of MOD13Q1 and L-MOD13Q1 Using
the Light Use Efficiency Model in 2019 2
Figure5. 5.The
Figure Thedaydayofofthe year
the year(DOY)-based
(DOY)-basedcomparison
comparisonofofcorrelation
correlationcoefficients
coefficientsbetween
between(a)(a)
R R2
Figure
valuesand
values 6a–c
and(b) displayed
(b)RMSE
RMSEvalues the crop
valuesobtained yield
obtainedduringaccuracies
duringthe
theaccuracybetween
accuracyassessmentthe modelled
assessmentofofreferencedand
referencedandreferenced
and synthetic
synthetic
crop
NDVI yields
NDVI productsof WW
products from and
from2001 OSR
2001
toto obtained
2019.
2019. The
The with different
correlation
correlation satellite
coefficient
coefficient products
refers
refers toto
RR using
(see
(see the(5)).
Equation
Equation LUE
(5)). mod-
el in 2019. The figures show that the fused product (L-MOD13Q1) obtained a higher R2
3.2.
3.2.Comparative
(0.81) and a lowerAnalysis
Comparative RMSE
Analysisbetween
(3.91
between Crop
dt/ha) Yield
than
Crop theAccuracy
Yield ofofMOD13Q1
non-fused
Accuracy product
MOD13Q1 and
andL-MOD13Q1
(MOD13Q1: Using
R2 = 0.70
L-MOD13Q1 and
Using
the
RMSE Light
the LightUse
= 4.77 Efficiency
dt/ha)
Use Model
for both
Efficiency Model in
WW 2019
in and
2019OSR. Analysing the ME of both products with LUE,
Figure
Figure6a–c
the L-MOD13Q1 6a–cdisplayed
resulted inthe
displayed athecrop
cropyield
lower ME accuracies
(3.04
yield dt/ha) between
accuraciesthan the the
between themodelled
MOD13Q1 and
(3.50
modelled referenced
dt/ha)
and (Fig-
referenced
crop6c).
ure yields of WW and OSR obtained with different satellite products using the LUE model
crop yields of WW and OSR obtained with different satellite products using the LUE mod-
in 2019. The figures show that the fused product (L-MOD13Q1) obtained a higher R2 (0.81) 2
el in 2019. The figures show that the fused product (L-MOD13Q1) obtained a higher R
and a lower RMSE (3.91 dt/ha) than the non-fused product (MOD13Q1: R2 =2 0.70 and
(0.81) and a lower RMSE (3.91 dt/ha) than the non-fused product (MOD13Q1: R = 0.70 and
RMSE = 4.77 dt/ha) for both WW and OSR. Analysing the ME of both products with LUE,
RMSE = 4.77 dt/ha) for both WW and OSR. Analysing the ME of both products with LUE,
the L-MOD13Q1 resulted in a lower ME (3.04 dt/ha) than the MOD13Q1 (3.50 dt/ha)
the L-MOD13Q1 resulted in a lower ME (3.04 dt/ha) than the MOD13Q1 (3.50 dt/ha) (Fig-
(Figure 6c).
ure 6c).
 Remote
Remote Sens.
Sens. 2023,
2023, 15,xx, x FOR PEER REVIEW
1651 14 of 35

(a) (b)

(c)

Figure6.6.The
Figure Thedot
dotplots
plotscompare
comparethe
theaccuracies (a)R2R,2(b)
accuracies(a) , (b)RMSE,
RMSE,and
and(c)
(c)ME
MEofofreferenced
referenceddata
data(at
(at
95% confidence intervals) and modelled yields
95% confidence intervals) and modelled yields obtained from multi-source
multi-source data:
data:MOD13Q1
MOD13Q1and andL-
MOD13Q1 in
L-MOD13Q1 in2019.
2019.

3.3.
3.3.Statistical
StatisticalAnalysis
Analysisbetween
betweenReference
Referenceand
andModelled
ModelledCropCropYields
YieldsofofWW
WWandandOSR
OSRfrom
2001 to 2019 Using the Light Use Efficiency Model
from 2001 to 2019 using the Light Use Efficiency Model
For
Forboth
both WW
WW and OSR, the
and OSR, theLUE
LUEmodel
model performed
performed significantly
significantly for for
everyevery
year year
(hav-
(having a p-value < 0.05); this rejected the H 0 of the linear regression model
ing a p-value < 0.05); this rejected the H0 of the linear regression model that there was that thereno
was no correlation
correlation between between the referenced
the referenced and modelled
and modelled cropfrom
crop yield yield2001
fromto2001
2019to(Figure
2019
(Figures A3a–s and A4a–s). After generating the scatter plots, all crop yield products’ R2 ,2
A3a–s and Figure A4a–s). After generating the scatter plots, all crop yield products’ R ,
RMSE, and ME values were analysed. For both WW and OSR, the years 2007 through
RMSE, and ME values were analysed. For both WW and OSR, the years 2007 through 2018
2018 and 2019 were the most accurate years where the estimated crop yield resulted in 2
and 2019 were the most accurate years where the estimated crop yield resulted in high R
high R2 values (>0.79). However, both 2018 and 2019 for WW resulted in higher RMSE
values (>0.79). However, both 2018 and 2019 for WW resulted in higher RMSE (4.74 and
(4.74 and 4.98 dt/ha) and ME (3.46 and 3.71 dt/ha) values, respectively (Figure A3). The
4.98 dt/ha) and ME (3.46 and 3.71 dt/ha) values, respectively (Figure A3). The remaining
remaining years for WW showed a similar trend in R2 (>0.65), RMSE (<4.50 dt/ha), and
years for WW showed a similar trend in R2 (>0.65), RMSE (<4.50 dt/ha), and ME (<3.60
ME (<3.60 dt/ha) values, with the exceptions of 2001 and 2013, which had RMSE values
dt/ha) values, with the exceptions of 2001 and 2013, which had RMSE values more than
more than 5.40 dt/ha and ME values more than 4.30 dt/ha. Similarly, for OSR, the RMSE
5.40 dt/ha and ME values more than 4.30 dt/ha. Similarly, for OSR, the RMSE values for
values for 2001, 2005, and 2012 resulted in higher RMSE (>3.22 dt/ha) and ME (>2.47 dt/ha)
2001, 2005, and 2012 resulted in higher RMSE (>3.22 dt/ha) and ME (>2.47 dt/ha) (Figure
(Figure A4). A mostly, similar trend in R2 values was observed in the OSR, with values
A4). A mostly, similar trend in R2 values was observed in the OSR, with values ranging
ranging from 0.63 to 0.80. The overall accuracies of both WW and OSR for 19 years were
from 0.63astoR0.80.
recorded 2 of The
0.79 overall
and 0.86 accuracies
and RMSE of both WWdt/ha
of 4.51 and OSR
andfor2.4719dt/ha,
years were recorded
respectively
as R 2 of 0.79 and 0.86 and RMSE of 4.51 dt/ha and 2.47 dt/ha, respectively (Figure 7a,b).
(Figure 7a,b). Negative correlations were seen between the regional mean elevations and
Negative
the modelled correlations were(−
yields of WW seen
0.30)between
and OSR the(−regional mean elevations
0.38), respectively (Figureand the modelled
8a,b).
yields of WW (−0.30) and OSR (−0.38), respectively (Figure 8a,b).
 Remote
Remote Sens.
Sens. 2023,
2023, 15,xx, x FOR PEER REVIEW
1651 15 of 35
Remote Sens. 2023, xx, x FOR PEER REVIEW

(a) (b)
(a) (b)
Figure
Figure The
7. 7. Thescatter plots
scatter plotscompare
compare the accuracies
the accuracies ofofthe modelled
the modelled andand referenced
referencedyieldsyields(at(ata 95%
a 95%
confidence
confidence interval)
interval) of of
(a)(a)
WW WW and and
(b) (b)
OSROSRfor for
19 19 years
years together
together (i.e.,
(i.e., from from
2001
Figure 7. The scatter plots compare the accuracies of the modelled and referenced yields (at a 95% 2001
to to 2019).
2019). The The
valuesval-
ues
the of the statistical
ofconfidence interval)
statistical parameters
of
parameters (a)such
WWasand such asOSR
R2 ,(b)
RMSER2,(dt/ha),
RMSE (dt/ha),
for 19 years ME (dt/ha),
together
ME (dt/ha), (i.e.,
and and
from
total total
2001 number
number to 2019).of
of points (n)points
The val-
are
(n) are
ues displayed
of the
displayed the top at
at statistical the top
of parameters
each of Every
plot. each
suchplot. Every
as Rcontains
plot 2, RMSE plot contains
a(dt/ha),
solid a(dt/ha),
ME(1:1
line solid
line) lineand(1:1
that line)
istotal
used tothat
number isofused
visualise theto
points
visualise the correlation of pixels between the modelled and referenced yield values. The dashed
correlation of pixels between the modelled and referenced yield values. The dashed line representsto
(n) are displayed at the top of each plot. Every plot contains a solid line (1:1 line) that is used
line represents
visualise the regression
the line.
correlation line.between
of colours
pixels Differentthe colours of theand points display different years.
the regression Different of the pointsmodelled
display differentreferenced
years. yield values. The dashed
line represents the regression line. Different colours of the points display different years.

(a) (b)
(a) Figure 8. The scatter plots correlating the modelled yield and (b)
regional mean elevation for (a) WW
and (b)8.OSR.
Figure The The dashed
scatter plotsline represents
correlating thethe regression
modelled yieldline.
andDifferent
regionalcolours of the points
mean elevation display
for (a) WW
Figure 8. The
different cropscatter
typesplots
(greencorrelating the orange
for represents
WW and modelledforyield
OSR). and regional
TheDifferent mean
correlation elevationrefers
coefficient for (a) WW
and (b) OSR. The dashed line the regression line. colours of the pointstodisplay
R (see
and (b) OSR.
Equation
different The
(5)).
crop dashed
types linefor
(green represents
WW andthe regression
orange line.The
for OSR). Different colours
correlation of the points
coefficient refers display
to R (see
different crop
Equation (5)). types (green for WW and orange for OSR). The correlation coefficient refers to R (see
3.4. Sensitivity
Equation (5)). Analysis
3.4. Sensitivity Analysis
The sensitivity
3.4. Sensitivity Analysisanalysis compared the model’s performance by excluding the effect
of climate stress factors
The sensitivity fromcompared
analysis 2001 to 2019
the for both performance
model’s WW and OSR byinexcluding
Bavaria. The LUE-
the effect
The sensitivity analysis compared the model’s performance by excluding the effect
modelled
of climate yield
stressshowed a higher
factors from 2001correlation
to 2019 forwith
boththe
WW referenced
and OSRyield when the
in Bavaria. Theclimate
LUE-
of climate stress factors from 2001 to 2019 for both WW and OSR in Bavaria. 2 The LUE-
stress factors
modelled yieldwere included,
showed andcorrelation
a higher vice versa.with
Thethemodel showed
referenced higher
yield whenR the
andclimate
lower
modelled yield showed a higher correlation with the referenced yield when the climate
RMSE values
stressfactors when
factorswere compared
wereincluded,
included,and with
andvice the yield
viceversa.
versa.The values
Themodel obtained
modelshowed during
showedhigher the
higherR2R and
2 sensitivity
andlower
lower
stress
analysisvalues
RMSE (Figure 9a,b).compared
when The overall accuracies
with the obtained
yield values during theduring
obtained sensitivity
the analysis of
sensitivity
RMSE values when compared with the yield values obtained during the sensitivity analysis
both WW
analysis and OSR
(Figure forThe
9a,b). 19 years were
overall recorded
accuracies as R2 ofduring
obtained 0.68 and
the 0.78 and RMSE
sensitivity of 5.88
analysis of
(Figure 9a,b). The overall accuracies obtained during the sensitivity analysis of both WW
dt/ha
both
and OSR
and
WW 3.41
forand
dt/ha,
OSR for
19 years
respectively
were19recorded (Figure
years were 9c,d). and
as recorded
R2 of 0.68as R2 0.78
of 0.68
andand
RMSE0.78ofand
5.88RMSE
dt/haofand5.88
dt/ha and 3.41 dt/ha, respectively
3.41 dt/ha, respectively (Figure 9c,d). (Figure 9c,d).
Remote
Remote Sens.
Sens. 2023,
2023, xx,
15, x1651
FOR PEER REVIEW 16 of 35

(a) (b)

(c) (d)
Figure 9. The bar plots show the yearly comparison of accuracies (a) R2 values and (b) RMSE val-
Figure 9. The bar plots show the yearly comparison of accuracies (a) R2 values and (b) RMSE
ues obtained from the referenced yields (at a 95% confidence interval), with LUE-modelled yields
values obtained
including climatefrom thefactors
stress referenced
(darkyields
blue)(at a 95%
and confidence interval),
LUE-modelled with LUE-modelled
yields excluding yields
the climate stress
including climate stress factors (dark blue) and LUE-modelled yields excluding
factors (sensitivity analysis) (light blue). The scatter plots compare the accuracies of the modelled the climate stress
factors (sensitivity analysis) (light blue). The scatter plots compare the accuracies
and referenced yields (at a 95% confidence interval) of (c) WW and (d) OSR for 19 years together of the modelled and
referenced
(i.e., from 2001yieldsto(at2019).
a 95%Theconfidence
values interval) of (c) WW
of the statistical and (d) OSR
parameters for 19
such as years together
R , RMSE
2 (i.e., from
(dt/ha), ME
(dt/ha), and total
2001 to 2019). Thenumber
values ofofthe
points (n) are
statistical displayedsuch
parameters at the
as R 2 , RMSE
top of each plot. Every
(dt/ha), plot contains
ME (dt/ha), and totala
solid
number lineof(1:1 line)(n)
points that is displayed
are used to visualise theofcorrelation
at the top each plot. of pixels
Every between
plot containsthea modelled and line)
solid line (1:1 ref-
erenced yieldtovalues.
that is used visualiseThethedashed line represents
correlation the regression
of pixels between line. Different
the modelled colours of
and referenced the points
yield values.
display different years.
The dashed line represents the regression line. Different colours of the points display different years.

3.5. Statistical
3.5. StatisticalAnalysis
Analysis between
between Reference
Reference and
and Modelled
Modelled Crop
Crop Yields
Yields of
of WW
WW and
and OSR
OSR from
2001 to 2019 Using the Light Use Efficiency Model at Regional Level
from 2001 to 2019 using the Light Use Efficiency Model at Regional Level
On comparing
On comparing the
the long-term
long-term crop
crop yield
yield at
at the
the regional
regional level,
level, the
the yearly
yearly spatial
spatial change
change
from the mean referenced and modelled yield was displayed for
from the mean referenced and modelled yield was displayed for both WW and OSR (Fig-both WW and OSR
(Figures 10 and 11). For WW, most of the regional yield lied between 65
ures 10 and 11). For WW, most of the regional yield lied between 65 and 80 dt/ha (Figure and 80 dt/ha
(Figure
10). 10). Districts
Districts suchRegen,
such as, as, Regen, Freyung-Grafenau,
Freyung-Grafenau, BadTölz-Wolfratshausen,
Bad Tölz-Wolfratshausen, and and
Garmisch-Partenkirchen, the average percent difference was calculated
Garmisch-Partenkirchen, the average percent difference was calculated as −25.10% (LUE:as − 25.10% (LUE:
~75 dt/ha), −18.68% (~60 dt/ha), −8.08% (~62 dt/ha), and −5.58% (~65 dt/ha), which
~75 dt/ha), −18.68% (~60 dt/ha), −8.08% (~62 dt/ha), and −5.58% (~65 dt/ha), which
showed that the model highly overestimated the crop yield values as compared to the
showed that the model highly overestimated the crop yield values as compared to the ref-
referenced yield (Figures 12a and A5a). The positive yield percent difference (where
erenced yield (Figures 12a and A5a). The positive yield percent difference (where the
the model underestimated the crop yield) between 0 and 4% had an accuracy greater
model underestimated the crop yield) between 0 and 4% had an accuracy greater than
than 0.80 as compared to the negative yield percent difference between −4 and 0% with
0.80 as compared to the negative yield percent difference between −4 and 0% with an ac-
an accuracy less than 0.70 (Figure 13). Similarly, the model underestimated the crop
curacy less than 0.70 (Figure 13). Similarly, the model underestimated the crop yield of
yield of Oberallgäu, Miltenberg, Deggendorf, and Dachau with 4.65% (~78 dt/ha), 3.91%
Oberallgäu, Miltenberg, Deggendorf, and Dachau with 4.65% (~78 dt/ha), 3.91% (~68
Remote Sens. 2023, xx,
15, x1651
FOR PEER REVIEW 17 of 35

dt/ha), 3.30% 3.30%

(~68 dt/ha), (~75 dt/ha)
(~75 and 3.15%
dt/ha) and(~78 dt/ha),
3.15% (~78respectively. Similarly for
dt/ha), respectively. OSR, the
Similarly formodel
OSR,
the model overestimated
overestimated the yield forthe yield for Aichach-Friedberg,
Aichach-Friedberg, Deggendorf,Deggendorf, Dingolfing-Lindau,
Dingolfing-Lindau, Traunstein,
Traunstein, Unterallgäu,
Unterallgäu, Dachau,Miltenberg
Dachau, Rottal-Inn, Rottal-Inn,and
Miltenberg
Günzberg andwith
Günzberg with 7.13%
7.13% (LUE: (LUE:
~38 dt/ha),
~38
5.12%dt/ha), 5.12%4.91%
(~39 dt/ha), (~39 dt/ha), 4.91%
(~37 dt/ha), (~37(~35
4.80% dt/ha), 4.80%
dt/ha), 4.53%(~35 dt/ha),
(~36 dt/ha),4.53%
4.36%(~36
(~38 dt/ha),
dt/ha),
4.36% (~38
4.25% (~35 dt/ha), 4.25% (~35
dt/ha), 4.24% (~37 dt/ha), 4.24%
dt/ha) and (~37 (~34
4.06% dt/ha) and 4.06%
dt/ha), (~34 dt/ha),
respectively respectively
(Figures 12b and
(Figures
A5b). 12b andunlike
However, A5b).WW,
However, unlike
both the over-WW,
and both the over- and
underestimation ofunderestimation
OSR yield valuesof OSR
result-
yield
ed in avalues
similarresulted
increaseinand
a similar increase
decrease and decrease
in accuracy in accuracy (Figure 13).
(Figure 13).

Figure
Figure 10.
10. Spatial
Spatial distribution
distribution of
of mean
mean referenced
referenced yield (2001–2019) and
yield (2001–2019) the year-wise
and the year-wise predicted
predicted
yield for WW from 2001 to 2019 using the LUE model for the state of Bavaria. The white colour
yield for WW from 2001 to 2019 using the LUE model for the state of Bavaria. The white colour rep-
resents no available data. A detailed map of the administrative regions of Bavaria is shown in Figure
represents no available data. A detailed map of the administrative regions of Bavaria is shown in
A1.
Figure A1.

3.6. Correlation Analysis between the Accuracy Assessments of the Input Synthetic Products and
the Crop Yield Modelling
The bar and scatter plots compared and linked the yearly accuracies of the input
synthetic time series with the crop yield modelling for WW and OSR from 2001 to 2019,
respectively (Figures 14 and 15). For WW, the correlation coefficient showed a higher
positive correlation of 0.81 between the R2 of synthetic accuracy and the modelled yield
accuracy (Figure 15a). Except 2015 (yield R2 : 0.77, synthetic R2 : 0.53) and 2013 (yield
R2 : 0.71, synthetic R2 : 0.65), where the fusion accuracies were negatively correlated with
crop yield accuracy (Figure 14a). Similarly, for OSR, the correlation coefficient was found to
be 0.77 (Figure 15b). For 2001 and 2002, the fusion accuracy was lower (R2 < 0.50); however,
the crop yield accuracy for the same years resulted in an R2 of more than 0.65 (Figure 14b).
Remote Sens.
Remote 2023,
Sens. 15,xx,
2023, 1651
x FOR PEER REVIEW 18 of 35

Figure11.11.Spatial
Figure Spatialdistribution
distributionofofmean
meanreferenced
referencedyield
yield(2001–2019)
(2001–2019)and
andthe
the year-wisepredicted
year-wise predicted
yield for OSR from 2001 to 2019 using the LUE model for the state of Bavaria. The white
yield for OSR from 2001 to 2019 using the LUE model for the state of Bavaria. The white colour colour rep-
resents no available data. A detailed map of the administrative regions of Bavaria is shown
represents no available data. A detailed map of the administrative regions of Bavaria is shown in
in Figure
A1.
Figure A1.

3.7. Visualisation of the Modelled Crop Biomass and the NDVI of Different Years at a Field Level
The side-by-side spatial visualisation of the input synthetic NDVI product (DOY 169:
18 June) and the WW-modelled biomass for selected years (2005, 2013, and 2019) is shown
in Figure 16, respectively. For every year, the spatial trend of crop biomass and NDVI
in every field was seen differently. Likewise, NDVI values were rising from 2005 to
2019 from 0.4 to 0.8, and crop biomass had been observed rising from less than
550 g/m2 in 2005 to more than 850 g/m2 in 2019. In most of the fields, the crop biomass
was dependent on the higher NDVI values. The NDVI values higher than 0.8 impacted
higher crop biomass of more than 850 g/m2 in almost every year. In 2005, the average field
crop biomass resulted in less than 650 g/m2 ; however, in 2019, the crop biomass resulted in
more than 650 g/m2 .

(a)
 Figure 11. Spatial distribution of mean referenced yield (2001–2019) and the year-wise predicted
Remote Sens. 2023, 15, 1651
yield for OSR from 2001 to 2019 using the LUE model for the state of Bavaria. The white colour rep-
19 of 35
resents no available data. A detailed map of the administrative regions of Bavaria is shown in Figure
A1.

Remote Sens. 2023, xx, x FOR PEER REVIEW

(a)

(b)
Figure 12. The dot plots show the district-wise distribution of modelled yield for (a) WW and (b)
Figure 12. The dot plots show the district-wise distribution of modelled yield for (a) WW and (b) OSR,
OSR, from 2001 to 2019. The green colour depicts the modelled yield of WW, the orange colour
from 2001 to 2019. The green colour depicts the modelled yield of WW, the orange colour depicts the
depicts the modelled yield of OSR, and the grey colour depicts both referenced yields of WW and
modelled
OSR. yield of OSR, and the grey colour depicts both referenced yields of WW and OSR.
 (b)
Figure 12. The dot plots show the district-wise distribution of modelled yield for (a) WW and (b)
Remote Sens. 2023, 15, 1651 OSR, from 2001 to 2019. The green colour depicts the modelled yield of WW, the orange colour 20 of 35
depicts the modelled yield of OSR, and the grey colour depicts both referenced yields of WW and
OSR.

Remote Sens. 2023, xx, x FOR PEER REVIEW

Remote Sens. 2023, xx, x FOR PEER REVIEW

Figure13.
Figure Theline
13.The lineplots
plotscompare
compare thethe accuracies
accuracies with with
the themeanmean
yieldyield percent
percent difference
difference (as calculated
(as calcu-
latedaccuracy
in Equation (Figure 15a). Except
OSR2015 (yield
years R
2: 0.77, synthetic R2: 0.53) and 2013 (yield R2:
in Equation (9)) (9))
for for
WW WW and
and OSR for19
for 19 years (i.e.,
(i.e., from
from 2001
2001to to
2019).
2019). TheThe
accuracies
accuraciesof WW of WW and
and accuracy (Figure :15a). Except
where2015 (yield R2−4
: 0.77,
to −2,synthetic 0Rto :2,0.53) andand2013 (yield R2:
2
0.71,are
OSR synthetic
analysed R2in 0.65),
six categories the fusion
(less than −4,accuracies were negatively
−2 to 0, 2 tocorrelated
4, morewiththancrop
OSR0.71,
are analysed
synthetic in six categories (less than − 4, − 4 to − 2, − 2 to 0, 0 to 2, 2 to 4, and more than 4%)
4%) yield
of yield percentR(Figure
accuracy 2: 0.65), where the fusion accuracies were negatively correlated with crop
difference.14a).
TheSimilarly, for OSR,
negative range shows thethe
correlation coefficient
overestimation, and the was found to
positive
of yield
yield
range percent
beshowsaccuracy
0.77 (Figure difference.
(Figure
15b). For
the underestimation The
14a).
2001negative
ofSimilarly,
and
the 2002,
modelledrange
forthe
OSR,
yieldshows
fusion thethe
values byoverestimation,
correlation
accuracy
the LUE was lower and
coefficient
compared (Rto <the
2was positive
found
0.50);
the to range
how-
refer-
shows
enced
be the
yield
0.77
ever, underestimation
values.
the cropThe
(Figure green
15b).
yield of
For the modelled
colour
2001
accuracy depicts
andthe
for WW,
2002, yield
same and
the values
the orange
fusion
years by the
accuracy
resulted LUE
colour
in was
an compared
depicts of OSR.
R2 lowermore to2than
(R the referenced
< 0.50); (Fig- yield
0.65how-
values.
ure The
ever, thegreen
14b). colouraccuracy
crop yield depicts WW, and
for the the orange
same colour depicts
years resulted in an R2OSR.
of more than 0.65 (Fig-
3.6. ure
Correlation
14b). Analysis between the Accuracy Assessments of the Input Synthetic Products and
the Crop Yield Modelling
The bar and scatter plots compared and linked the yearly accuracies of the input
synthetic time series with the crop yield modelling for WW and OSR from 2001 to 2019,
respectively (Figures 14 and 15). For WW, the correlation coefficient showed a higher
positive correlation of 0.81 between the R2 of synthetic accuracy and the modelled yield

(a) (b)
(a) (b)
Figure 14. The bar plots compare the yearly (a) R2 2and (b) RMSE values of estimated OSR yield
Figure 14. The
(orange),
baryield
plots compare the yearly (a) R and (b)from
RMSE values of The
estimated OSR yield
Figure 14.WW
The bar (green),
plots and the
compare synthetic
yearly NDVI
(a) R2 (purple)
and (b) RMSE 2001 toof2019.
values estimatedunits
OSRof the
yield
(orange), WW
RMSE values
(orange), yield (green),
WWofyield and
both (green),
WW andand synthetic
OSRsyntheticNDVI
yields areNDVI (purple) from 2001 to 2019. The units of
dt/ha.(purple) from 2001 to 2019. The units of thethe RMSE
values
RMSEof both
valuesWW andWW
of both OSRandyields
OSRare dt/ha.
yields are dt/ha.

(a) (b)
(a) (b)
Figure 15. The correlation plots between R of synthetic NDVI time series and R2 of modelled yield
2

time series
Figure for correlation
15. The (a) WW (green)
plotsand (b) OSR
between (orange),
R22of syntheticfrom 2001
NDVI to 2019.
time seriesThe
andcorrelation coefficient
R2 of modelled yield
Figure
time 15.
refers toThe
series correlation
R (see
for Equation
(a) plots between
(5)). and
WW (green) R (orange),
(b) OSR of synthetic
fromNDVI
2001 totime
2019.series and R2 of coefficient
The correlation modelled yield
timerefers
series
to for (a)Equation
R (see WW (green)
(5)). and (b) OSR (orange), from 2001 to 2019. The correlation coefficient
3.7.
refers toVisualisation of the(5)).
R (see Equation Modelled Crop Biomass and the NDVI of Different Years at a Field Level
3.7. Visualisation of the Modelled
The side-by-side Crop Biomass
spatial visualisation ofand
the the NDVI
input of Different
synthetic NDVI Years at a Field
product (DOY Level
169:
The and
18 June) side-by-side spatial visualisation
the WW-modelled biomass forof selected
the inputyears
synthetic
(2005,NDVI
2013, product
and 2019) (DOY 169:
is shown
18
in June)
Figureand16,the WW-modelled
respectively. biomass
For every forthe
year, selected
spatialyears
trend(2005,
of crop2013, and 2019)
biomass and isNDVI
shown in
in Figure
every field16, respectively.
was For every
seen differently. year, NDVI
Likewise, the spatial
valuestrend
wereofrising
crop from
biomass
2005and NDVI
to 2019 in
from
every
0.4 to field was crop
0.8, and seen biomass
differently.
had Likewise, NDVI values
been observed rising were
from rising from550
less than 2005
g/mto2 2019 from
in 2005 to
2 2
Remote Sens. 2023, xx, x FOR PEER REVIEW

Remote Sens. 2023, 15, 1651 21 of 35

Figure
Figure16.
16.The
Theside-by-side
side-by-sidevisualisation
visualisationofofsynthetic
syntheticNDVI
NDVIproducts
productsobtained
obtainedonon1818June
June2005,
2005,
2013
2013and
and2019
2019(left)
(left)with
withthe
theWW
WWbiomass
biomassobtained
obtainedfrom
fromthe
theLUE
LUEmodelled
modelledforforthe
theyears
yearsofof2005,
2005,
2013
2013and
and2019
2019(right).
(right).

4.4.Discussion
Discussion
4.1.Quality
4.1. QualityAssessment
AssessmentofofSynthetic
SyntheticRemote
RemoteSensing
SensingTime
TimeSeries
Seriesfrom
from2001
2001toto2019
2019
Thepresent
The presentstudy
studyinvestigates
investigatesthe
thepotential
potentialofofthe
theSTARFM
STARFMover
overthe
theBavarian
Bavarianstate
state
of Germany to generate the synthetic NDVI time series from 2001 to 2019 by selecting
of Germany to generate the synthetic NDVI time series from 2001 to 2019 by selecting the
best-performing high (Landsat) and low (MODIS) pair obtained for the agricultural
the best-performing high (Landsat) and low (MODIS) pair obtained for the agricultural class
from the previous literature. Many studies prefer using ESTARFM (Enhanced STARFM)
class from the previous literature. Many studies prefer using ESTARFM (Enhanced
for better fusion accuracy [80,81]; however, some studies found STARFM performing
STARFM) for better fusion accuracy [80,81]; however, some studies found STARFM per-
significantly better than ESTARFM [82,83]. Simple in its design, faster to implement, and
forming significantly better than ESTARFM [82,83]. Simple in its design, faster to im-
capable of fusing the entire state of Bavaria (which covers almost one-fifth of the area
plement, and capable of fusing the entire state of Bavaria (which covers almost one-fifth
of Germany) for two decades, the study finds STARFM to be more advantageous over
of the area of Germany) for two decades, the study finds STARFM to be more advanta-
ESTARFM. ESTARM was complex, time-consuming, and computationally expensive for
geous over ESTARFM. ESTARM was complex, time-consuming, and computationally
covering extensive data for extended periods [84,85]. One of the strengths of ESTARFM
Remote Sens. 2023, 15, 1651 22 of 35

is that it incorporates additional information, such as a land-cover map, to improve the

accuracy of the fusion [80]. The study incorporates Bavaria’s accurate and updated land
cover map into the STARFM to balance its input requirements with the ESTARFM. It
provided homogeneity to the STARFM and increased its fusion accuracy (as discussed
briefly in our previous study [59]).
In our previous study, we found that L-MOD13Q1 (30 m, 16 days) (R2 = 0.62 and
RMSE = 0.11) was suitable for the application of agricultural monitoring due to its fast
and easy processing with lesser storage requirements [59]. Moreover, the present study
focuses on two decades (2001 to 2019); therefore, the paper generates and validates a
Landsat-based synthetic NDVI time series (L-MOD13Q1) due to its continuous availability
since 1982 with a maximum resolution of 30 m. As NDVI is among the most effective and
widely used vegetation indices, many spatiotemporal fusion-based studies have used it
as their primary input [5,30,36,37]. However, many spatiotemporal fusion algorithms are
based on reflectance fusion, which requires more processing time and storage than NDVI
(or one-band blending) fusion [86,87]. Having high computation power with fewer storage
problems for the long-term time series of 2001 to 2019 for complete Bavaria (70,550 km2 ),
the research uses the strategy “index-then-blend” (IB), which generates the NDVI from
Landsat and MOD13Q1 before they are blended for fusion [88]. The IB strategy is used in
multiple works of the literature with highly accurate and precise fusion outputs [17,59,88].
The analysis found that the accuracies of the fusion products are dependent on the
available number of Landsat scenes per year (N) [59], such that the higher N, the higher
the fusion accuracy of the synthetic NDVI product in a respective year. For instance, the
positive R (+0.75) shows the positive correlation between R2 of yearly synthetic NDVIs
and N (representing the higher quality of the fused product), and the negative R (−0.73)
shows the negative correlation between RMSEs and N (representing the higher precision).
However, as the research made use of Landsat 8 Operational Land Imager (OLI) (from
2013 to 2019) and Landsat 5 Thematic Mapper (TM) (from 2001 to 2013), it was found that
Landsat OLI-based fusion with MOD13Q1 resulted in higher accuracy as compared to
Landsat TM [89]. For example, the years 2001, 2002, 2004, 2005, and 2012 (using Landsat
5 and 7) have a lower R2 (<0.60) and a higher RMSE (>0.12) than the remaining years (using
Landsat 8). The reason could be that Landsat 8 has improved upon the quality of Landsat
5 and 7, offering improved data accuracy. Moreover, the accuracy of the year 2012 is affected
due to the gaps generated by the scan line corrector (SLC) failure in Landsat 7.
On comparing the fusing results on a DOY basis, the study finds that the few cloud-
free DOYs could create large gaps between the available Landsat scenes that might affect
the accuracy of the fusion product [17,59]. For example, the DOYs 33 to 97 (N = ~6) result
in a low R2 (0.54) and a high RMSE (0.16) as compared to the DOYs 113 to 193 (N = ~8),
which have a high R2 (0.64) and a low RMSE (0.10).

4.2. Impact of Synthetic Time Series on Crop Yield Modelling

The objective of the present study is to generate and validate the long-term crop yield
time series using the semi-empiric LUE model, which has proven to be more reliable,
precise, and simple in the previous literature [17,42]. The present study validates the
crop yield results of WW and OSR obtained by inputting the synthetic NDVI and climate
elements to the LUE model at a regional scale in Bavaria from 2001 to 2019. However, before
generating the long-term time series using the synthetic NDVI product, the study finds the
potential of fused (L-MOD13Q1) in crop yield prediction by comparing its accuracies with
the non-fused (MOD13Q1) product in 2019. The study obtains higher crop yield accuracy
with the L-MOD13Q1 (R2 = 0.81 and RMSE = 3.91 dt/ha) than the MOD13Q1 (R2 = 0.70 and
RMSE = 4.77 dt/ha) irrespective of the crop type (Figure 6a,b). It proves the importance of
high-resolution synthetic data for accurate modelling of crop yields.
After generating the long-term crop yield time series, the research finds the significant
yearly performance of the model for both WW and OSR; however, some years obtained
higher accuracy than the others. For example, 2007, 2018, and 2019 are the most accurate
Remote Sens. 2023, 15, 1651 23 of 35

years, with R2 values of more than 0.79 for both crop types. However, the RMSEs of both
2018 and 2019 are relatively higher (>4.74 dt/ha) than in the other years. Similarly, 2011 and
2016, with a higher number of N (~6), result in lower crop yield accuracy than 2007, 2008,
and 2011 (N = ~8). This might be due to the impact of climate variables inputted into the
LUE model (discussed briefly in Section 4.3).
The study discusses briefly how the quality of the input data fusion product impacts
the accuracy of the CGM. For example, due to the low quality of synthetic NDVI products
in 2001, 2002, and 2012, which might impact the accuracy of the input FPAR products
generated, the yield prediction accuracy of both WW and OSR is low. The analysis tries
to prove that even though synthetic time series would be the preferable solution to input
a CGM for yield prediction when the quality of the combined fusion product is low, it
could negatively affect the crop yield estimation. In relevance to the above point, high
positive correlations have been seen when the accuracies of the synthetic NDVI time series
are plotted with the accuracies of modelled crop yield from 2001 to 2019 for WW (R = 0.81)
and OSR (0.77). For example, the quality of the NDVI time series for the years 2016 and
2018 is higher with R2 (>0.73), and the crop yield accuracies are also higher with R2 of
0.83 (WW)/0.81 (OSR), 0.85/0.83, respectively (Figure 3p,r). Similarly, the striped data
collected from Landsat 7 in 2012 has deteriorated the quality of the synthetic NDVI product
(R2 = 0.51; RMSE = 0.13), which further negatively affected the crop yield estimations for
WW (R2 = 0.62; RMSE = 5.40 dt/ha) and OSR (R2 = 0.49; RMSE = 4.13 dt/ha) (Figure 3l).
Moreover, the Landsat images were available at different times of the year. This has an
impact on the prediction accuracy of both crops. For example, the WW yield results are
more accurate than the OSR because the synthetic data in late spring and early summer
(DOYs 113 to 193) is usually more precise.
The study compares the long-term crop yield time series by calculating the average per-
cent change from the referenced and modelled yields for both crop types. Previous studies
found that the elevation plays a significant role in impacting the regional crop yield [90,91].
Most of the studies found lower crop productivity at higher elevations due to complex
topography, different climates, and management practices [92,93]. Moreover, the cropping
intensity at lower elevations is higher as compared to the higher elevations. The survey
finds negative correlations between the mean regional elevations and the crop yields of WW
(−0.30) and OSR (−0.38). The model is precarious in specific regions, especially the districts
at higher elevations in the south (Bavarian Alps) and east (Bavarian Forest and Fichtel
Mountains) of Bavaria for both WW and OSR. In regions such as Regen, Freyung-Grafenau,
Bad Tölz-Wolfratshausen, and Garmisch-Partenkirchen, the model highly overestimates
the crop yield, and for regions such as Oberallgäu, Miltenberg, Deggendorf, and Dachau, it
underestimates the yield as compared to the referenced yield for WW. This overestimation
of WW yield values has resulted in a decrease in accuracy. The model shows yearly stability
in predicting crop yields of WW between 65 and 80 dt/ha for most of the regions. The
positive yield percent change (where the model underestimated the crop yield) between
0 and 4% had higher accuracy (R2 > 0.80) as compared to the percent change between
−4 and 0% (R2 < 0.70). For 48 of the 71 total districts, the model performs relatively well,
with a percent change between –2% and +2%. However, unlike WW, both the over- and
underestimation-yield values have resulted in a similar increase and decrease in the ac-
curacy of OSR. The positive and negative yield percent change (where the model under-
and over-estimates the crop yield, respectively) between 0 and +/−4% had an accuracy
of more than 0.80. For OSR, the model overestimates the yield for Aichach-Friedberg,
Deggendorf, Dingolfing-Lindau, Traunstein, Unterallgäu, Dachau, Rottal-Inn, Miltenberg,
and Günzberg and underestimates the yield for Roth, Regen, Kronach, Kitzingen, and Bad
Tölz-Wolfratshausen. However, for the 27 districts with OSR, the model performs steadily.
Interestingly, the regions where the model’s performance went unstable were primarily
located in the southern alps, except for Regen, Freyung-Grafenau, Kitzingen, Roth, and
Miltenberg. The reason could be the instability of the model at higher elevations or the bad
quality of the synthetic NDVI products in specific districts. The quality of the synthetic
Remote Sens. 2023, 15, 1651 24 of 35

NDVI product might vary for these regions as the districts have no horizontal or vertical
overlay of Landsat scenes within the path row, limiting their coverage frequency.

4.3. Sensitivity Analysis

Besides the impact of data fusion, climate variables play an essential role in affecting
the accuracy of crop yield predictions [31,94,95]. To analyse the impact of climate elements,
the study performs sensitivity analysis, where the LUE model calculates the crop yields
of WW and OSR without including the climate stress factors from 2001 to 2019. As the
referenced yield is already influenced by the climate, the results of the study show that the
accuracy of crop yield predictions worsens with the exclusion of climate variables, with
a lower R2 (0.68 (WW)/0.74 (OSR)) and a higher RMSE (5.88/3.41 dt/ha). However, an
increase in R2 (0.79/0.86) and decrease in RMSE (4.51/2.57 dt/ha) have been seen when
the climate effect is included in the model. As the relationship between climate and crop
yield undergoes significant shifts, it might be the reason that some years (2011 and 2016)
with higher N (8) obtained lower crop yield accuracy than years (2007, 2018 and 2019) with
comparably lower N (6). Moreover, our previous study, which made use of the machine
learning approach with the LUE model, identified the impact of every individual climate
element used in crop yield predictions [31].
Furthermore, many studies stated that the availability of coarse climate data nega-
tively affected yield prediction accuracy. In a previous study, the coarse spatial resolution
of climate data (ECMWF: ~80 km) used to estimate the biomass resulted in low R2 and
high RMSE using CGMs by inputting coarse synthetic NDVI products [17,42]. However,
while inputting high spatial resolution NDVI products, the low impact of the high spatial
resolution of climate elements is observed. The present study inputs high spatial resolution
climate data time series (2 km, daily) to the LUE model, resulting in stable yearly accuracies
from 2001 to 2019. Notably, selecting climate thresholds according to the geographical
location and crop types is essential in achieving high crop yield accuracy [96–98]. Differ-
ent climate thresholds are used for WW and OSR, resulting in accurate and stable yield
predictions in Bavaria during the study period.

4.4. Validation at the District Level

The crop yield validation for the more extended time series of 2001 to 2019 is performed
using the LfStat crop yield data (used for validation at a 95% confidence interval) for WW
and OSR provided by the Bavarian State Office of Statistics. As the validation data set
is provided at a regional scale, pixel-based yield information is converted for both crop
types to the regional level. However, transferring the field-based information to the district
level could result in some uncertainties in the validation process. For example, in some
regions of southern Bavaria (Bad Tölz-Wolfratshausen, Garmisch-Partenkirchen, Traunstein,
Unterallgäu, and Oberallgäu), where the model’s performance is volatile, this might be due
to the uncertainty occurring while transferring the pixel-level information to the district
level. The availability of fewer fields of WW and OSR in those regions might be the reason
for the model’s instability, as the validation data recorded high yield values for the same
districts. Therefore, future work should aim to validate crop yield results at the field level,
which could help achieve more precise results. Additionally, the availability of field data
for FPAR, an important input to the LUE model, would help to validate the FPAR product
generated using the NDVI time series.

5. Conclusions
The present study investigates the relationship of spatiotemporal fusion modelling
using STRAFM on crop yield prediction for winter wheat (WW) and oil-seed rape (OSR)
using a semi-empirical light use efficiency (LUE) model for Bavaria, Germany, from 2001 to
2019. The research paper concludes the findings as follows:
(i) To find the potential of STARFM for long-term time series, the paper generates
and validates a synthetic normalised difference vegetation index (NDVI) time series
Remote Sens. 2023, 15, 1651 25 of 35

blending the high spatial resolution (30 m, 16 days) of Landsat 5 Thematic Mapper
(TM) (2001 to 2012), Landsat 7 Enhanced Thematic Mapper Plus (ETM+) (2012), and
Landsat 8 Operational Land Imager (OLI) (2013 to 2019) with the coarse resolution of
MOD13Q1 (250 m, 16 days) from 2001 to 2019. Overall, the average accuracy of data
fusion for nineteen years has an R2 of 0.66 and an RMSE of 0.11. The accuracy of data
fusion is found to be dependent on the number of Landsat scenes available per year
(N). The higher the N, the more accurate is the synthetic NDVI time series per year.
(ii) To investigate the stability and precision of the LUE model in crop yield prediction,
the paper inputs the synthetic NDVI time series and climate elements to the crop
model to estimate and validate yearly crop yields for WW and OSR from 2001 to
2019. The validation of crop yield at regional scale results in an average R2 of 0.79
(WW)/0.86 (OSR) and an RMSE of 4.51 dt/ha/2.46 dt/ha, respectively.
(iii) Identifying the impact of the input data fusion product on the accuracy assessment
of the LUE model, high positive correlations are seen when the accuracies of the
synthetic NDVI time series are plotted with the accuracies of modelled crop yield
from 2001 to 2019 for WW (R = 0.81) and OSR (0.77).
The present study recommends validating crop yields at the field scale, as transferring
the pixel-based information to the district level could cause uncertainties in the validation
process. The accurate crop yield predictions from the analysis for WW and OSR could
be further used for the application of biodiversity, where the impact of land use diversity
on crop yields could be estimated. The ease of using spatiotemporal modelling with crop
growth models would not be limited to Bavaria. The study’s methodology could also be
tested by coupling machine/deep learning (ML/DL) approaches with CGMs, which might
help to include more climate elements to achieve more precise results. Lastly, the study’s
two decades of accurate yield estimations could strengthen trust in the decision(/policy)
making to achieve sustainability in agriculture.

Author Contributions: Conceptualization, M.S.D. and T.U.; Methodology, M.S.D.; Software, M.S.D.;
Validation, M.S.D.; Formal analysis, M.S.D. and T.U.; Investigation, M.S.D.; Resources, M.S.D., C.K.-F.,
A.L., T.R., J.A., I.S.-D. and T.U.; Data curation, M.S.D., C.K.-F., T.R. and J.A.; Writing—original draft,
M.S.D.; Writing—review & editing, M.S.D. and T.U.; Visualization, M.S.D. and T.U.; Supervision, T.U.;
Project administration, M.S.D., T.D. and C.K.-F.; Funding acquisition, I.S.-D. and T.U. All authors
have read and agreed to the published version of the manuscript.
Funding: This research received no external funding. This publication is supported by the Open
Access Publication Fund of the University of Wuerzburg. The research is a part of the LandKlif project
funded by the Bavarian Ministry of Science and the Arts via the Bavarian Climate Research Network
(bayklif: https://www.bayklif.de/, accessed on 1 March 2019).
Data Availability Statement: Publicly available satellite datasets were analysed in this study. This
data can be found here: (MODIS SR) LP DAAC. 2019. MOD13Q1 v006; https://lpdaac.usgs.
gov/products/mod13q1v006/ (accessed on 10 September 2019); (MODIS SR); Landsat 5; https:
//www.usgs.gov/ (accessed on 21 June 2021); Landsat 7; https://www.usgs.gov/ (accessed on
21 June 2021); (Landsat 8 SR) Landsat 8; https://www.usgs.gov/ (accessed on 21 April 2021);
shuttle radar topography mission (SRTM) 30 m; https://www2.jpl.nasa.gov/srtm/ (accessed on
20 December 2022).
Acknowledgments: The authors thankfully express gratitude to the National Aeronautics and Space
Administration (NASA) Land Processed Distributed Active Archive Center (LP DAAC) for MODIS;
U.S. Geological Survey (USGS) Earth Resources Observation and Science (EROS) Center for Landsat;
The Copernicus Sentinel-2 mission for the Sentinel-2 data; Professorship of Ecological Services,
University of Bayreuth, for the updated Land Cover Map of Bavaria 2019; Jie Zhang and Sarah Redlich,
both Department of Animal Ecology and Tropical Biology, University of Wuerzburg, developed the
initial concept and methodology for the Land Cover Map of Bavaria; together with Melissa Versluis,
Rebekka Riebl, Maria Hänsel, Bhumika Uniyal, Thomas Koellner (all Professorship of Ecological
Services, University of Bayreuth) this concept and methodology were further developed; Melissa
Versluis (Professorship of Ecological Services, University of Bayreuth) has created the Land Cover Map
of 2019 based on the concluding methodology for whole Bavaria; we acknowledge the data providers
Remote Sens. 2023, 15, 1651 26 of 35

for the base layers ATKIS (Bayerische Vermessungsverwaltung), IACS (Bayerisches Staatsministerium
für Ernährung, Landwirtschaft und Forsten) and Corine LC (European Environment Agency).;
University of Augsburg for providing the climate data from 2001 to 2019; Simon Sebold for providing
all technical and computational help during the complete fusion processing of Landsat/Sentinel-
2 and MODIS; Department of Remote Sensing, University of Wuerzburg, for providing all necessary
resources and support; Google Earth Engine platform for faster downloading and preprocessing the
large data sets of Bavaria; The Comprehensive R Archive Network (CRAN) for running the STARFM,
and validating and visualizing the data outputs using R packages such as raster, rgdal, ggplot2,
ggtools, and xlsx; ArcGIS 10.8 and ArcGIS pro 2.2.0 maintained by Environmental Systems Research
Institute (Esri) for generating maps and Bavarian Land Cover Map of 2019.
Conflicts of Interest: The authors declare no conflict of interest.
Remote Sens. 2023, xx, x FOR PEER REVIEW

Appendix A

Figure A1.A1.Detailed
Figure Detailed map
map of ofadministrative
administrative regions
regions of Bavaria
of Bavaria (Landkreise
(Landkreise und Städte
und kreisfreie kreisfreie
in Städte
Bayern). The
in Bayern). Thenames
namesofofthe
thedistricts are are
districts translated from from
translated German to English
German as: Unterfranken
to English as
as: Unterfranken as
Lower Franconia, Mittelfranken as Middle Franconia, Oberfranken as Upper Franconia, Oberpfalz
Lower Franconia, Mittelfranken as Middle Franconia, Oberfranken as Upper Franconia, Oberpfalz
as Upper Palatinate, Oberbayern as Upper Bavaria, and Niederbayern as Lower Bavaria. (Source:
as Upper Palatinate, Oberbayern
https://www.gifex.com/, accessed onas12Upper
January Bavaria,
2023). and Niederbayern as Lower Bavaria. (Source:
https://www.gifex.com/, accessed on 12 January 2023).
Remote Sens.
Remote2023,
Sens.xx, x FOR
2023, PEER REVIEW
15, 1651 27 of 35

Figure
Figure A2.A2.
TheThe digitalelevation
digital elevation map
mapofofBavaria. TheThe
Bavaria. mapmap
is generated from shuttle
is generated from radar topography
shuttle radar topog-
mission
raphy (SRTM)
mission digital
(SRTM) elevation
digital data. The
elevation elevation
data. ranges from
The elevation 93 m
ranges to 2943
from 93 mm.to 2943 m.
 Remote Sens. 2023, xx, x FOR PEER REVIEW 30 of 37
Remote Sens. 2023, 15, 1651 28 of 35

(a) (b) (c) (d) (e)

(f) (g) (h) (i) (j)

(k) (l) (m) (n) (o)

Figure A3. Cont.

Remote Sens. 2023, 15, 1651 Remote Sens. 2023, xx, x FOR PEER REVIEW 29 of 35 31 of
Remote Sens. 2023, xx, x FOR PEER REVIEW 31 of 37

(p) (q) (r) (s)

(p) (q) (r) (s)
Figure
Figure A3. The scatter plots (a–s)A3. The scatter
compare plots (a–s)
the accuracies ofcompare
modelledtheandaccuracies
referencedofyields
modelled
of WW andfor
referenced yields
2001 to 2019. of values
The WW for of 2001 to 2019. The
the statistical values of
parameters the as
such statistical p
2
Figure A3. The scatter
rametersplotssuch
(a–s)ascompare
R2, RMSE the accuracies
(dt/ha), and MEof modelled and
(dt/ha) are referenced
displayed yields
at the topofofWW
eachfor 2001
plot. to 2019.
Every plot The values
contains of theline
a solid statistical pa-that is used to v
(1:1 line)
R , RMSE (dt/ha),
rameters andasME
such R 2, (dt/ha)
RMSE are displayed at(dt/ha)
the topare
of each plot. Every plotof
contains a solid line (1:1contains
line) that solid
is used to(1:1
visualise the is
correlation of pixels
ualise the (dt/ha), andof
correlation MEpixels between displayed at theand
the modelled top each plot.
referenced Every
yield plot
values. The greena colour line line)
of scatter that
plots used toWW.
represents vis-
between the modelled and referenced yield values. The green colour of scatter plots represents WW.
ualise the correlation of pixels between the modelled and referenced yield values. The green colour of scatter plots represents WW.

(a) (b) (c) (d) (e)

(a) (b) (c) (d) (e)

(f) (g) (h) (i) (j)

(f) (g) (h) (i) (j)

Figure A4. Cont.

Remote Sens. 2023,
Remote 15,2023,
Sens. 1651 xx, x FOR PEER REVIEW 32 of 3730 of 35

(k) (l) (m) (n) (o)

(p) (q) (r) (s)

Figure
Figure A4. TheA4. The plots
scatter scatter plots
(a–s) (a–s) compare
compare the accuracies
the accuracies of modelled
of modelled and referenced
and referenced yields of yields
OSR forof 2001
OSR tofor2019.
2001The
to 2019.
valuesThe
of values of the statistical
the statistical parameterspa-such as
rameters
R2 , RMSE such
(dt/ha), andasME
R2, (dt/ha)
RMSE (dt/ha), and MEat(dt/ha)
are displayed the topare
of displayed at the top
each plot. Every plotofcontains
each plot. Every
a solid plot
line (1:1contains a solid
line) that line
is used to(1:1 line) that
visualise the is used to vis-
correlation of pixels
betweenualise the correlation
the modelled of pixels between
and referenced the modelled
yield values. The orangeand referenced
colour yield
of scatter values.
plots The orange
represents OSR.colour of scatter plots represents OSR.
 Remote Sens. 2023, xx, x FOR PEER REVIEW
Remote Sens. 2023, 15, 1651 31 of 35

(a) (b)
Figure A5. The regional scale average yield percent difference between the referenced and the modelled
Figure A5. The regional scale average yield percent difference between the referenced and the
yield from 2001 to 2019 (a) WW, (b) OSR. The yield percent difference is calculated in Equation (9).
modelled yield from 2001 to 2019 (a) WW, (b) OSR. The yield percent difference is calculated in
Equation (9).
References
1. Mueller, N.D.; Gerber, J.S.; Johnston, M.; Ray, D.K.; Ramankutty, N.; Foley, J.A. Closing yield gaps through nutrient and
References
water management. Nature 2012, 490, 254–257.
1. Mueller, N.D.; Gerber, J.S.; Johnston, M.; Ray, D.K.; Ramankutty, N.; Foley, J.A. Closing yield gaps through nutrient and water
2. Bian, C.; Shi, H.; Wu, S.; Zhang, K.; Wei, M.; Zhao, Y.; Sun, Y.; Zhuang, H.; Zhang, X.; Chen, S. Prediction of Field-Scale Wheat
management. Nature 2012, 490, 254–257. [CrossRef] [PubMed]
Yield Using Machine Learning Method and Multi-Spectral UAV Data. Remote Sens. 2022, 14, 1474.
2. Bian, C.; Shi, H.; Wu, S.; Zhang, K.; Wei, M.; Zhao, Y.; Sun, Y.; Zhuang, H.; Zhang, X.; Chen, S. Prediction of Field-Scale Wheat
3. Fritz, S.; See, L.; Bayas, J.C.L.; Waldner, F.; Jacques, D.; Becker-Reshef, I.; Whitcraft, A.; Baruth, B.; Bonifacio, R.; Crutchfield, J.
Yield Using Machine Learning Method and Multi-Spectral UAV Data. Remote Sens. 2022, 14, 1474. [CrossRef]
A comparison of global agricultural monitoring systems and current gaps. Agric. Syst. 2019, 168, 258–272.
3. Fritz, S.; See, L.; Bayas, J.C.L.; Waldner, F.; Jacques, D.; Becker-Reshef, I.; Whitcraft, A.; Baruth, B.; Bonifacio, R.; Crutchfield, J. A
4. Ziliani, M.G.; Altaf, M.U.; Aragon, B.; Houborg, R.; Franz, T.E.; Lu, Y.; Sheffield, J.; Hoteit, I.; McCabe, M.F. Early season
comparison of global agricultural monitoring systems and current gaps. Agric. Syst. 2019, 168, 258–272. [CrossRef]
prediction of within-field crop yield variability by assimilating CubeSat data into a crop model. Agric. For. Meteorol. 2022, 313,
4. Ziliani, M.G.; Altaf, M.U.; Aragon, B.; Houborg, R.; Franz, T.E.; Lu, Y.; Sheffield, J.; Hoteit, I.; McCabe, M.F. Early season prediction
108736.
of within-field crop yield variability by assimilating CubeSat data into a crop model. Agric. For. Meteorol. 2022, 313, 108736.
5. Eurostat, Waste Statistics- Electrical and Electronic Equipment. 2019. Available online: https://ec.europa.eu/eurostat/statistics-
[CrossRef]
explained/index.php (accessed on 7 August 2020).
5. Eurostat, Waste Statistics- Electrical and Electronic Equipment. 2019. Available online: https://ec.europa.eu/eurostat/statistics-
6. Alarcón-Segura, V.; Grass, I.; Breustedt, G.; Rohlfs, M.; Tscharntke, T. Strip intercropping of wheat and oilseed rape enhances
explained/index.php (accessed on 7 August 2020).
biodiversity and biological pest control in a conventionally managed farm scenario. J. Appl. Ecol. 2022, 59, 1513–1523.
6. Alarcón-Segura, V.; Grass, I.; Breustedt, G.; Rohlfs, M.; Tscharntke, T. Strip intercropping of wheat and oilseed rape enhances
7. Macholdt, J.; Honermeier, B. Yield stability in winter wheat production: A survey on German farmers’ and advisors’ views.
biodiversity and biological pest control in a conventionally managed farm scenario. J. Appl. Ecol. 2022, 59, 1513–1523. [CrossRef]
Agronomy 2017, 7, 45.
7. Macholdt, J.; Honermeier, B. Yield stability in winter wheat production: A survey on German farmers’ and advisors’ views.
8. Lutter, S.; Giljum, S.; Gözet, B.; Wieland, H.; Manstein, C. The Use of Natural Resources: Report for Germany 2018; Bundesamt, U.,
Agronomy 2017, 7, 45. [CrossRef]
Ed.; Federal Environment Agency: Dessau, German, 2018.
8. Lutter, S.; Giljum, S.; Gözet, B.; Wieland, H.; Manstein, C. The Use of Natural Resources: Report for Germany 2018; Bundesamt, U.,
9. UFOP. Union zur Förderung von oel-und Proteinpflanzen E.V. Availabe online: https://www.ufop.de/ (accessed on 13 July
Ed.; Federal Environment Agency: Dessau, Germany, 2018.
2021).
9. UFOP. Union zur Förderung von oel-und Proteinpflanzen E.V. Available online: https://www.ufop.de/ (accessed on
10. 13Ali,
JulyA.M.;
2021). Abouelghar, M.A.; Belal, A.-A.; Saleh, N.; Younes, M.; Selim, A.; Emam, M.E.; Elwesemy, A.; Kucher, D.E.;
Magignan, S. Crop Yield
10. Ali, A.M.; Abouelghar, M.A.; Prediction Using
Belal, A.-A.; Multi
Saleh, N.; Sensors
Younes, Remote
M.; Selim,Sensing. Egypt.
A.; Emam, J. Remote
M.E.; Sens. A.;
Elwesemy, Space Sci. 2022,
Kucher, D.E.;25, 711–716. S.
Magignan,
11. Crop
Justice, C.; Townshend, J.; Vermote, E.; Masuoka, E.; Wolfe, R.; Saleous, N.; Roy, D.; Morisette,
Yield Prediction Using Multi Sensors Remote Sensing. Egypt. J. Remote Sens. Space Sci. 2022, 25, 711–716. J. An overview of MODIS Land
data processing and product status. Remote Sens. Environ. 2002, 83, 3–15.
11. Justice, C.; Townshend, J.; Vermote, E.; Masuoka, E.; Wolfe, R.; Saleous, N.; Roy, D.; Morisette, J. An overview of MODIS Land
12. data
Ahmad, I.; Ghafoor,
processing A.; Bhatti,
and product M.I.;
status. Akhtar,
Remote I.-u.H.;
Sens. Ibrahim,
Environ. 2002, M.
83, Satellite remote sensing and GIS-based crops forecasting &
3–15. [CrossRef]
estimation system in Pakistan. Crop Monit. Improv. Food Secur. 2014, 28, 95-109.
12. Ahmad, I.; Ghafoor, A.; Bhatti, M.I.; Akhtar, I.-u.H.; Ibrahim, M. Satellite remote sensing and GIS-based crops forecasting &
13. estimation
Karila, K.;system
Nevalainen, O.; Krooks,
in Pakistan. A.; Karjalainen,
Crop Monit. Improv. FoodM.; Kaasalainen,
Secur. S. Monitoring changes in rice cultivated area from SAR
2014, 28, 95–109.
and optical satellite images in Ben Tre and Tra Vinh Provinces
13. Karila, K.; Nevalainen, O.; Krooks, A.; Karjalainen, M.; Kaasalainen, S. Monitoring in Mekong Delta,changes
Vietnam.inRemote Sens. 2014,
rice cultivated area6,from
4090–4108.
SAR and
14. optical
Friedl,satellite images in Ben Tre and Tra Vinh Provinces in Mekong Delta, Vietnam. Remote Sens. 2014, 6, 4090–4108.5[CrossRef]
M.A.; Sulla-Menashe, D.; Tan, B.; Schneider, A.; Ramankutty, N.; Sibley, A.; Huang, X. MODIS Collection global land
cover:M.A.;
14. Friedl, Algorithm refinements
Sulla-Menashe, D.;and
Tan,characterization
B.; Schneider, A.; of Ramankutty,
new datasets.N.; Remote Sens.
Sibley, A.;Environ.
Huang, 2010, 114, 168–182.
X. MODIS Collection 5 global land
15. cover:
Lobell, D.B. The use of satellite data for crop yield gap analysis. Field Crops Res. 2013, 143, 56–64.
Algorithm refinements and characterization of new datasets. Remote Sens. Environ. 2010, 114, 168–182. [CrossRef]
Remote Sens. 2023, 15, 1651 32 of 35

15. Lobell, D.B. The use of satellite data for crop yield gap analysis. Field Crops Res. 2013, 143, 56–64. [CrossRef]
16. Ogutu, B.O.; Dash, J. An algorithm to derive the fraction of photosynthetically active radiation absorbed by photosynthetic
elements of the canopy (FAPAR ps) from eddy covariance flux tower data. New Phytol. 2013, 197, 511–523. [CrossRef] [PubMed]
17. Dhillon, M.S.; Dahms, T.; Kuebert-Flock, C.; Borg, E.; Conrad, C.; Ullmann, T. Modelling Crop Biomass from Synthetic Remote
Sensing Time Series: Example for the DEMMIN Test Site, Germany. Remote Sens. 2020, 12, 1819. [CrossRef]
18. Wulder, M.A.; Loveland, T.R.; Roy, D.P.; Crawford, C.J.; Masek, J.G.; Woodcock, C.E.; Allen, R.G.; Anderson, M.C.; Belward, A.S.;
Cohen, W.B. Current status of Landsat program, science, and applications. Remote Sens. Environ. 2019, 225, 127–147. [CrossRef]
19. Wulder, M.A.; White, J.C.; Loveland, T.R.; Woodcock, C.E.; Belward, A.S.; Cohen, W.B.; Fosnight, E.A.; Shaw, J.; Masek, J.G.; Roy,
D.P. The global Landsat archive: Status, consolidation, and direction. Remote Sens. Environ. 2016, 185, 271–283. [CrossRef]
20. Wulder, M.A.; Masek, J.G.; Cohen, W.B.; Loveland, T.R.; Woodcock, C.E. Opening the archive: How free data has enabled the
science and monitoring promise of Landsat. Remote Sens. Environ. 2012, 122, 2–10. [CrossRef]
21. Souza, E.; Bazzi, C.; Khosla, R.; Uribe-Opazo, M.; Reich, R.M. Interpolation type and data computation of crop yield maps is
important for precision crop production. J. Plant Nutr. 2016, 39, 531–538. [CrossRef]
22. Mariano, C.; Monica, B. A random forest-based algorithm for data-intensive spatial interpolation in crop yield mapping. Comput.
Electron. Agric. 2021, 184, 106094. [CrossRef]
23. Nemecek, T.; Weiler, K.; Plassmann, K.; Schnetzer, J.; Gaillard, G.; Jefferies, D.; García–Suárez, T.; King, H.; i Canals, L.M.
Estimation of the variability in global warming potential of worldwide crop production using a modular extrapolation approach.
J. Clean. Prod. 2012, 31, 106–117. [CrossRef]
24. Atamanyuk, I.; Kondratenko, Y.; Shebanin, V.; Sirenko, N.; Poltorak, A.; Baryshevska, I.; Atamaniuk, V. Forecasting of cereal crop
harvest on the basis of an extrapolation canonical model of a vector random sequence. CEUR Workshop Proc. 2019, II, 302–315.
25. Bolton, D.K.; Friedl, M.A. Forecasting crop yield using remotely sensed vegetation indices and crop phenology metrics. Agric. For.
Meteorol. 2013, 173, 74–84. [CrossRef]
26. Johnson, M.D.; Hsieh, W.W.; Cannon, A.J.; Davidson, A.; Bédard, F. Crop yield forecasting on the Canadian Prairies by remotely
sensed vegetation indices and machine learning methods. Agric. For. Meteorol. 2016, 218, 74–84. [CrossRef]
27. Ramesh, D.; Vardhan, B.V. Analysis of crop yield prediction using data mining techniques. Int. J. Res. Eng. Technol. 2015, 4, 47–473.
28. Mo, X.; Liu, S.; Lin, Z.; Xu, Y.; Xiang, Y.; McVicar, T. Prediction of crop yield, water consumption and water use efficiency with a
SVAT-crop growth model using remotely sensed data on the North China Plain. Ecol. Model. 2005, 183, 301–322. [CrossRef]
29. Ghadge, R.; Kulkarni, J.; More, P.; Nene, S.; Priya, R. Prediction of crop yield using machine learning. Int. Res. J. Eng. Technol.
(IRJET) 2018, 5, 2237–2239.
30. Van Klompenburg, T.; Kassahun, A.; Catal, C. Crop yield prediction using machine learning: A systematic literature review.
Comput. Electron. Agric. 2020, 177, 105709. [CrossRef]
31. Dhillon, M.S.; Dahms, T.; Kuebert-Flock, C.; Rummler, T.; Arnault, J.; Stefan-Dewenter, I.; Ullmann, T. Integrating random forest
and crop modeling improves the crop yield prediction of winter wheat and oil seed rape. Front. Remote Sens. 2023, 3, 109.
[CrossRef]
32. Elavarasan, D.; Vincent, P.D. Crop yield prediction using deep reinforcement learning model for sustainable agrarian applications.
IEEE Access 2020, 8, 86886–86901. [CrossRef]
33. Kuwata, K.; Shibasaki, R. Estimating crop yields with deep learning and remotely sensed data. In Proceedings of the 2015 IEEE
International Geoscience and Remote Sensing Symposium (IGARSS), Milan, Italy, 26–31 July 2015; pp. 858–861.
34. Zhuo, W.; Fang, S.; Gao, X.; Wang, L.; Wu, D.; Fu, S.; Wu, Q.; Huang, J. Crop yield prediction using MODIS LAI, TIGGE weather
forecasts and WOFOST model: A case study for winter wheat in Hebei, China during 2009–2013. Int. J. Appl. Earth Obs. Geoinf.
2022, 106, 102668. [CrossRef]
35. Kasampalis, D.A.; Alexandridis, T.K.; Deva, C.; Challinor, A.; Moshou, D.; Zalidis, G. Contribution of remote sensing on crop
models: A review. J. Imaging 2018, 4, 52. [CrossRef]
36. Mirschel, W.; Schultz, A.; Wenkel, K.O.; Wieland, R.; Poluektov, R.A. Crop growth modelling on different spatial scales—A wide
spectrum of approaches. Arch. Agron. Soil Sci. 2004, 50, 329–343. [CrossRef]
37. Murthy, V.R.K. Crop growth modeling and its applications in agricultural meteorology. Satell. Remote Sens. GIS Appl. Agric.
Meteorol. 2004, 235, 235–261.
38. Iqbal, M.A.; Shen, Y.; Stricevic, R.; Pei, H.; Sun, H.; Amiri, E.; Penas, A.; del Rio, S. Evaluation of the FAO AquaCrop model for
winter wheat on the North China Plain under deficit irrigation from field experiment to regional yield simulation. Agric. Water
Manag. 2014, 135, 61–72. [CrossRef]
39. Van Dam, J.C.; Huygen, J.; Wesseling, J.; Feddes, R.; Kabat, P.; Van Walsum, P.; Groenendijk, P.; Van Diepen, C. Theory of SWAP
Version 2.0. In Simulation of Water Flow, Solute Transport and Plant Growth in the Soil-Water-Atmosphere-Plant Environment; DLO
Winand Staring Centre: Wageningen, The Netherlands, 1997.
40. Wang, E.; Robertson, M.; Hammer, G.; Carberry, P.S.; Holzworth, D.; Meinke, H.; Chapman, S.; Hargreaves, J.; Huth, N.; McLean,
G. Development of a generic crop model template in the cropping system model APSIM. Eur. J. Agron. 2002, 18, 121–140.
[CrossRef]
41. Spitters, C.; Van Keulen, H.; Van Kraalingen, D. A simple and universal crop growth simulator: SUCROS87. In Simulation and
Systems Management in Crop Protection; Pudoc: Wageningen, The Netherlands, 1989; pp. 147–181.
Remote Sens. 2023, 15, 1651 33 of 35

42. Shi, Z.; Ruecker, G.R.; Mueller, M.; Conrad, C.; Ibragimov, N.; Lamers, J.; Martius, C.; Strunz, G.; Dech, S.; Vlek, P.L.G. Modeling
of cotton yields in the amu darya river floodplains of Uzbekistan integrating multitemporal remote sensing and minimum field
data. Agron. J. 2007, 99, 1317–1326. [CrossRef]
43. Van Diepen, C.A.v.; Wolf, J.; Van Keulen, H.; Rappoldt, C. WOFOST: A simulation model of crop production. Soil Use Manag.
1989, 5, 16–24. [CrossRef]
44. Potter, C.S.; Randerson, J.T.; Field, C.B.; Matson, P.A.; Vitousek, P.M.; Mooney, H.A.; Klooster, S.A. Terrestrial ecosystem
production: A process model based on global satellite and surface data. Glob. Biogeochem. Cycles 1993, 7, 811–841. [CrossRef]
45. Duchemin, B.; Maisongrande, P.; Boulet, G.; Benhadj, I. A simple algorithm for yield estimates: Evaluation for semi-arid irrigated
winter wheat monitored with green leaf area index. Environ. Model. Softw. 2008, 23, 876–892. [CrossRef]
46. Schwalbert, R.A.; Amado, T.; Corassa, G.; Pott, L.P.; Prasad, P.V.; Ciampitti, I.A. Satellite-based soybean yield forecast: Integrating
machine learning and weather data for improving crop yield prediction in southern Brazil. Agric. For. Meteorol. 2020, 284, 107886.
[CrossRef]
47. Kern, A.; Barcza, Z.; Marjanović, H.; Árendás, T.; Fodor, N.; Bónis, P.; Bognár, P.; Lichtenberger, J. Statistical modelling of crop
yield in Central Europe using climate data and remote sensing vegetation indices. Agric. For. Meteorol. 2018, 260, 300–320.
[CrossRef]
48. Shammi, S.A.; Meng, Q. Use time series NDVI and EVI to develop dynamic crop growth metrics for yield modeling. Ecol. Indic.
2021, 121, 107124. [CrossRef]
49. Gevaert, C.M.; García-Haro, F.J. A comparison of STARFM and an unmixing-based algorithm for Landsat and MODIS data fusion.
Remote Sens. Environ. 2015, 156, 34–44. [CrossRef]
50. Roy, D.P.; Ju, J.; Lewis, P.; Schaaf, C.; Gao, F.; Hansen, M.; Lindquist, E. Multi-temporal MODIS–Landsat data fusion for relative
radiometric normalization, gap filling, and prediction of Landsat data. Remote Sens. Environ. 2008, 112, 3112–3130. [CrossRef]
51. Benabdelouahab, T.; Lebrini, Y.; Boudhar, A.; Hadria, R.; Htitiou, A.; Lionboui, H. Monitoring spatial variability and trends of
wheat grain yield over the main cereal regions in Morocco: A remote-based tool for planning and adjusting policies. Geocarto Int.
2019, 36, 2303–2322. [CrossRef]
52. Htitiou, A.; Boudhar, A.; Lebrini, Y.; Hadria, R.; Lionboui, H.; Elmansouri, L.; Tychon, B.; Benabdelouahab, T. The performance
of random forest classification based on phenological metrics derived from Sentinel-2 and Landsat 8 to map crop cover in an
irrigated semi-arid region. Remote Sens. Earth Syst. Sci. 2019, 2, 208–224. [CrossRef]
53. Lebrini, Y.; Boudhar, A.; Htitiou, A.; Hadria, R.; Lionboui, H.; Bounoua, L.; Benabdelouahab, T. Remote monitoring of agricultural
systems using NDVI time series and machine learning methods: A tool for an adaptive agricultural policy. Arab. J. Geosci. 2020,
13, 796. [CrossRef]
54. Gao, F.; Masek, J.; Schwaller, M.; Hall, F. On the blending of the Landsat and MODIS surface reflectance: Predicting daily Landsat
surface reflectance. IEEE Trans. Geosci. Remote Sens. 2006, 44, 2207–2218.
55. Cui, J.; Zhang, X.; Luo, M. Combining Linear pixel unmixing and STARFM for spatiotemporal fusion of Gaofen-1 wide field of
view imagery and MODIS imagery. Remote Sens. 2018, 10, 1047. [CrossRef]
56. Lee, M.H.; Cheon, E.J.; Eo, Y.D. Cloud Detection and Restoration of Landsat-8 using STARFM. Korean J. Remote Sens. 2019,
35, 861–871.
57. Xie, D.; Zhang, J.; Zhu, X.; Pan, Y.; Liu, H.; Yuan, Z.; Yun, Y. An improved STARFM with help of an unmixing-based method
to generate high spatial and temporal resolution remote sensing data in complex heterogeneous regions. Sensors 2016, 16, 207.
[CrossRef] [PubMed]
58. Zhu, L.; Radeloff, V.C.; Ives, A.R. Improving the mapping of crop types in the Midwestern US by fusing Landsat and MODIS
satellite data. Int. J. Appl. Earth Obs. Geoinf. 2017, 58, 1–11.
59. Dhillon, M.S.; Dahms, T.; Kübert-Flock, C.; Steffan-Dewenter, I.; Zhang, J.; Ullmann, T. Spatiotemporal Fusion Modelling Using
STARFM: Examples of Landsat 8 and Sentinel-2 NDVI in Bavaria. Remote Sens. 2022, 14, 677. [CrossRef]
60. Miller, J. Agriculture and Forestry in Bavaria: Facts and Figures 2002; Bayerisches Staatsministerium für Landwirtschaft und Forsten:
München, Germany, 2002.
61. Roy, D.P.; Kovalskyy, V.; Zhang, H.; Vermote, E.F.; Yan, L.; Kumar, S.; Egorov, A. CFEDharacterization of Landsat-7 to Landsat-
8 reflective wavelength and normalized difference vegetation index continuity. Remote Sens. Environ. 2016, 185, 57–70. [CrossRef]
62. Sulik, J.J.; Long, D.S. Spectral indices for yellow canola flowers. Int. J. Remote Sens. 2015, 36, 2751–2765. [CrossRef]
63. Zamani-Noor, N.; Feistkorn, D. Monitoring Growth Status of Winter Oilseed Rape by NDVI and NDYI Derived from UAV-Based
Red–Green–Blue Imagery. Agronomy 2022, 12, 2212. [CrossRef]
64. Harfenmeister, K.; Itzerott, S.; Weltzien, C.; Spengler, D. Detecting phenological development of winter wheat and winter barley
using time series of sentinel-1 and sentinel-2. Remote Sens. 2021, 13, 5036. [CrossRef]
65. Meier, U.; Bleiholder, H.; Buhr, L.; Feller, C.; Hack, H.; Heß, M.; Lancashire, P.D.; Schnock, U.; Stauß, R.; Van Den Boom, T. The
BBCH system to coding the phenological growth stages of plants–history and publications. J. Kult. 2009, 61, 41–52.
66. Hersbach, H.; Bell, B.; Berrisford, P.; Hirahara, S.; Horányi, A.; Muñoz-Sabater, J.; Simmons, A. The ERA5 global reanalysis. Q. J.
R. Meteorol. Soc. 2020, 146, 1999–2049. [CrossRef]
67. Gochis, D.; Barlage, M.; Dugger, A.; FitzGerald, K.; Karsten, L.; McAllister, M.; McCreight, J.; Mills, J.; RafieeiNasab, A.; Read, L.
The WRF-Hydro modeling system technical description, (Version 5.0). NCAR Tech. Note 2018, 107. [CrossRef]
Remote Sens. 2023, 15, 1651 34 of 35

68. Skamarock, W.C.; Klemp, J.B.; Dudhia, J.; Gill, D.O.; Liu, Z.; Berner, J.; Wang, W.; Powers, J.G.; Duda, M.G.; Barker, D.M. A
Description of the Advanced Research WRF Model Version 4; National Center for Atmospheric Research: Boulder, CO, USA, 2019;
Volume 145, p. 145.
69. Arnault, J.; Rummler, T.; Baur, F.; Lerch, S.; Wagner, S.; Fersch, B.; Zhang, Z.; Kerandi, N.; Keil, C.; Kunstmann, H. Precipitation
sensitivity to the uncertainty of terrestrial water flow in WRF-Hydro: An ensemble analysis for central Europe. J. Hydrometeorol.
2018, 19, 1007–1025. [CrossRef]
70. Rummler, T.; Arnault, J.; Gochis, D.; Kunstmann, H. Role of lateral terrestrial water flow on the regional water cycle in a complex
terrain region: Investigation with a fully coupled model system. J. Geophys. Res. Atmos. 2019, 124, 507–529. [CrossRef]
71. Farr, T.G.; Rosen, P.A.; Caro, E.; Crippen, R.; Duren, R.; Hensley, S.; Kobrick, M.; Paller, M.; Rodriguez, E.; Roth, L. The shuttle
radar topography mission. Rev. Geophys. 2007, 45. [CrossRef]
72. Monteith, J.L. Solar radiation and productivity in tropical ecosystems. J. Appl. Ecol. 1972, 9, 747–766. [CrossRef]
73. Monteith, J.L. Climate and the efficiency of crop production in Britain. Philos. Trans. R. Soc. Lond. B Biol. Sci. 1977, 281, 277–294.
74. Asrar, G.; Myneni, R.; Choudhury, B. Spatial heterogeneity in vegetation canopies and remote sensing of absorbed photosyntheti-
cally active radiation: A modeling study. Remote Sens. Environ. 1992, 41, 85–103. [CrossRef]
75. Single, W.V. Frost injury and the physiology of the wheat plant. J. Aust. Inst. Agric. Sci. 1985, 51, 128–134.
76. Habekotté, B. A model of the phenological development of winter oilseed rape (Brassica napus L.). Field Crops Res. 1997, 54, 127–136.
[CrossRef]
77. Hodgson, A. Repeseed adaptation in Northern New South Wales. II.* Predicting plant development of Brassica campestris L.
and Brassica napus L. and its implications for planting time, designed to avoid water deficit and frost. Aust. J. Agric. Res. 1978,
29, 711–726. [CrossRef]
78. Russell, G.; Wilson, G.W. An Agro-Pedo-Climatological Knowledge-Base of Wheat in Europe; Brussels (Belgium) EC/JRC: Brussels
Belgium, 1994.
79. Djumaniyazova, Y.; Sommer, R.; Ibragimov, N.; Ruzimov, J.; Lamers, J.; Vlek, P. Simulating water use and N response of winter
wheat in the irrigated floodplains of Northwest Uzbekistan. Field Crops Res. 2010, 116, 239–251. [CrossRef]
80. Zhu, X.; Chen, J.; Gao, F.; Chen, X.; Masek, J.G. An enhanced spatial and temporal adaptive reflectance fusion model for complex
heterogeneous regions. Remote Sens. Environ. 2010, 114, 2610–2623. [CrossRef]
81. Tewes, A.; Thonfeld, F.; Schmidt, M.; Oomen, R.J.; Zhu, X.; Dubovyk, O.; Menz, G.; Schellberg, J. Using RapidEye and MODIS
data fusion to monitor vegetation dynamics in semi-arid rangelands in South Africa. Remote Sens. 2015, 7, 6510–6534. [CrossRef]
82. Ghosh, R.; Gupta, P.K.; Tolpekin, V.; Srivastav, S. An enhanced spatiotemporal fusion method–Implications for coal fire monitoring
using satellite imagery. Int. J. Appl. Earth Obs. Geoinf. 2020, 88, 102056. [CrossRef]
83. Xue, J.; Leung, Y.; Fung, T. A Bayesian data fusion approach to spatio-temporal fusion of remotely sensed images. Remote Sens.
2017, 9, 1310. [CrossRef]
84. Chen, B.; Huang, B.; Xu, B. Comparison of spatiotemporal fusion models: A review. Remote Sens. 2015, 7, 1798–1835. [CrossRef]
85. Guo, Y.; Wang, C.; Lei, S.; Yang, J.; Zhao, Y. A framework of spatio-temporal fusion algorithm selection for landsat NDVI time
series construction. ISPRS Int. J. Geo-Inf. 2020, 9, 665. [CrossRef]
86. Dong, T.; Liu, J.; Qian, B.; Zhao, T.; Jing, Q.; Geng, X.; Wang, J.; Huffman, T.; Shang, J. Estimating winter wheat biomass by
assimilating leaf area index derived from fusion of Landsat-8 and MODIS data. Int. J. Appl. Earth Obs. Geoinf. 2016, 49, 63–74.
[CrossRef]
87. Walker, J.J.; De Beurs, K.M.; Wynne, R.H.; Gao, F. Evaluation of Landsat and MODIS data fusion products for analysis of dryland
forest phenology. Remote Sens. Environ. 2012, 117, 381–393. [CrossRef]
88. Chen, X.; Liu, M.; Zhu, X.; Chen, J.; Zhong, Y.; Cao, X. “Blend-then-Index” or “Index-then-Blend”: A theoretical analysis for
generating high-resolution NDVI time series by STARFM. Photogramm. Eng. Remote Sens. 2018, 84, 65–73. [CrossRef]
89. Poursanidis, D.; Chrysoulakis, N.; Mitraka, Z. Landsat 8 vs. Landsat 5: A comparison based on urban and peri-urban land cover
mapping. Int. J. Appl. Earth Obs. Geoinf. 2015, 35, 259–269. [CrossRef]
90. Thomson, A.M.; Brown, R.A.; Ghan, S.J.; Izaurralde, R.C.; Rosenberg, N.J.; Leung, L.R. Elevation dependence of winter wheat
production in eastern Washington State with climate change: A methodological study. Clim. Chang. 2002, 54, 141–164. [CrossRef]
91. Bhatt, D.; Maskey, S.; Babel, M.S.; Uhlenbrook, S.; Prasad, K.C. Climate trends and impacts on crop production in the Koshi River
basin of Nepal. Reg. Environ. Chang. 2014, 14, 1291–1301. [CrossRef]
92. Semwal, R.; Maikhuri, R. Structure and functioning of traditional hill agroecosystems of Garhwal Himalaya. Biol. Agric. Hortic.
1996, 13, 267–289. [CrossRef]
93. Anderson, M.C.; Hain, C.R.; Jurecka, F.; Trnka, M.; Hlavinka, P.; Dulaney, W.; Otkin, J.A.; Johnson, D.; Gao, F. Relationships
between the evaporative stress index and winter wheat and spring barley yield anomalies in the Czech Republic. Clim. Res. 2016,
70, 215–230. [CrossRef]
94. Cabas, J.; Weersink, A.; Olale, E. Crop yield response to economic, site and climatic variables. Clim. Chang. 2010, 101, 599–616.
[CrossRef]
95. Sidhu, B.S.; Mehrabi, Z.; Ramankutty, N.; Kandlikar, M. How can machine learning help in understanding the impact of climate
change on crop yields? Environ. Res. Lett. 2023, 18, 024008. [CrossRef]
96. Grace, J. Temperature as a determinant of plant productivity. Symp. Soc. Exp. Biol. 1988, 42, 91–107.
Remote Sens. 2023, 15, 1651 35 of 35

97. Porter, J.R.; Gawith, M. Temperatures and the growth and development of wheat: A review. Eur. J. Agron. 1999, 10, 23–36.
[CrossRef]
98. Porter, J.R.; Moot, D.J. Research beyond the means: Climatic variability and plant growth. In International Symposium on Applied
Agrometeorology and Agroclimatology; Office for Official Publication of the European Commission: Luxembourg, 1998; pp. 13–23.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.