You are on page 1of 15

Computers and Electronics in Agriculture 186 (2021) 106173

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture


journal homepage: www.elsevier.com/locate/compag

Original papers

Large-scale winter catch crop monitoring with Sentinel-2 time series and
machine learning–An alternative to on-site controls?
Christian Schulz a, *, Ann-Kathrin Holtgrave a, b, Birgit Kleinschmit a
a
Technische Universität Berlin, Institute for Landscape Architecture and Environmental Planning, Geoinformation in Environmental Planning Lab, Straße des 17. Juni
145, 10623 Berlin, Germany
b
Johann Heinrich von Thünen-Institute, Federal Research Institute for Rural Areas, Forestry and Fisheries, Bundesallee 64, 38116 Brunswick, Germany

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

Keywords: The European legislative guidelines on the Common Agricultural Policy (CAP) lead to an obligation of area-wide
Random forest information about cross-compliant Greening measures in EU countries. But by on-site controlling, agricultural
Cover crop authorities can monitor only a small portion of all registered parcels. With the Copernicus Programme, the freely
Normalized difference vegetation index (NDVI)
available Sentinel-1 and Sentinel-2 imagery increasingly supports remote sensing-based agricultural monitoring
Temporal metrics
on a large-scale. However, most prototypes lack the topic of winter catch cropping. Therefore, we developed a
Phenological profile
new machine learning method for catch crop monitoring and detection at the parcel-level. To gain training and
test data for supervised machine learning, we collected winter catch crop parcel data from different years
(2016–2019) and four federal states in Germany. Normalized Difference Vegetation Index (NDVI) time series
were calculated for each parcel from Sentinel-2 data within the typical winter catch crop cultivation season
(July-April). We revealed distinctive temporal patterns of catch cropping and developed nineteen descriptive
features for automatized prediction. Then, we trained fifteen Random Forest classifiers comprising different
regions and years and conducted a multi-level validation to identify the model with the highest robustness on
new data. The Random Forest classifier trained with the input data from all federal states and years outperformed
the other models. It reached a mean prediction accuracy of 84% for both classes (catch crop and non-catch crop)
across eleven different spatio-temporal domains. Under optimal annual weather conditions it reached accuracies
close to 90%. Anomalies caused by heat waves and early frost events were found to have a high influence on the
phenology of catch crops and thus lead to reduced prediction accuracies. From the set of predictors, those
features with the highest importance measured the correlation between the observed time series and a simulated
NDVI phenological profile. We concluded that catch cropping parcels are automatically separable from parcels
with other winter cultivations (e.g., winter cereals, grasslands, fallow) with Sentinel-2 NDVI time series data.
Different catch crop subgroups (i.e., seed mixes) could not be differentiated by our approach due to very similar
phenological profiles. Nonetheless, the used approach allows for large-scale winter catch crop monitoring and
supports authorities in the selection of parcels with high demand for on-site controlling. By merging the training
datasets from different federal states and years, we could overcome the typical spatial and temporal overfitting
problem in machine learning. Therefore, the study’s final classifier can be reliably transferred to new datasets in
Germany and other regions with similar bio-geographical conditions.

1. Introduction In this study, we address the example of Germany, where catch crops are
mostly cultivated as winter green cover between late summer until at
Catch crops—often referred to as cover crops—are crop mix cultiva­ least the 15th of February when the cultivation is plowed and the main
tions usually grown in summer or winter between the main cropping summer crop is directly sown (German Federal Ministry of Food and
seasons and are subsidized as Greening measures for farmland diversi­ Agriculture, 2015). The used seed mixes have a large variety of species
fication and soil protection (Magdoff and Van Es, 2009; OSCAR, 2016). such as oilseed radish (Raphanus sativus var. oleiformis), white mustard

* Corresponding author.
E-mail addresses: christian.schulz.1@tu-berlin.de (C. Schulz), ann-kathrin.holtgrave@thuenen.de (A.-K. Holtgrave), birgit.kleinschmit@tu-berlin.de
(B. Kleinschmit).

https://doi.org/10.1016/j.compag.2021.106173
Received 18 December 2020; Received in revised form 9 April 2021; Accepted 12 April 2021
Available online 6 May 2021
0168-1699/© 2021 Elsevier B.V. All rights reserved.
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

(Sinapsis alba), red clover (Trifolium pratense), phacelia (Phacelia tana­ However, few remote sensing projects in Europe and North America
cetifolia), buckwheat (Fagopyrum esculentum) and black oat (Avena stri­ have dealt with the monitoring of Greening measures for farmland
gosa). More information on commonly used seed mixes and their links to diversification in agriculture. Parcel data like the Integrated Adminis­
management practices can be found in the Cover Crop and Living Mulch tration and Control System (IACS) and agricultural control data by the
Toolbox (OSCAR, 2016). federal ministries abide by strict data protection regulations which
Winter catch crops protect agricultural soils from erosion and pre­ limits the set of possible computing platforms and data archives. The
vent nitrogen contamination of the groundwater caused by over­ agricultural practices monitoring product of the sen4CAP project integrated
fertilization during winter time (Dabney et al., 2001; Hargrove, 1991; a prototype for catch crop assessment using the CREODIAS cloud
Wilke and Snapp, 2008; Meisinger et al., 1991). As green fertilizers, they computing platform (European Space Agency, 2018). They calculated
reduce the need for artificial fertilizers and therefore promote sustain­ temporal markers from S1 and S2 data for the estimation of harvest dates
able land use and climate protection (Thorup-Kristensen, 1994; Delgado and growth on catch crop parcels in the pilot countries Czech Republic,
et al., 2007). Furthermore, they are important nutrition sources and Lithuania, Romania and the Netherlands (European Space Agency,
provide habitats for many species, enriching the biodiversity in rural 2018). Thieme et al. (2020) and Peredo et al. (2019) introduced the
landscapes (Delgado et al., 2007). These and other ecosystem services application CCROP for winter cover crop performance and biomass
provided by catch cropping were summarized by Blanco-Canqui et al. estimation with the Google Earth Engine cloud computing platform
(2015) and Żuk-Gołaszewska et al. (2019). using time series from Landsat and S2 data. Previous studies on the
Winter catch cropping became an important subsidized measure for control for cover crop subsidy payments in the U.S. focused on mono-
agricultural parcels with ecological priority in the European Union (EU), temporal SPOT and Landsat images (Hively et al., 2009, 2015). All
which accounts for approximately 5% of farmlands in Germany (German these prototypes can generally benefit the estimation of the success of
Federal Ministry of Food and Agriculture, 2015). It supports the aim of catch/cover cropping. However, the aforementioned approaches require
biodiversity-rich landscape features on agricultural land by the EU biodi­ a parcel-wise interpretation and—more importantly—specialized
versity strategy for 2030 (European Commission, 2020). For the past remote sensing knowledge for decision-making from the user. A fully-
twenty years, the EU Water Framework Directive (Directive 2000/60/ automatic monitoring and detection approach, based on a replicable
EC) has promoted catch cropping to protect water bodies from over­ set of catch crop-specific metrics and temporal features, would be more
fertilization. The main legal background for the on-site monitoring and efficient and practical for the controlling authorities.
the obligation of area-wide information on catch cropping areas origi­ By using Copernicus earth observation data and machine learning
nated from the European legislative guidelines on the Common Agri­ classifiers, we aim to support the controlling agencies in their legal
cultural Policy (CAP) (EU Regulation No 1307/2013). With the revision obligation for CAP CC monitoring. Thus, the objective of this study is the
of the Commission Implementing Regulation (EU Regulation 2018/ development of a new, fully-automated method for large-scale winter
746), new techniques from remote sensing became evident source of catch crop detection based on the example of Germany. Considering the
data for agricultural monitoring. challenge of area-wide monitoring and reduction of on-site controls, we
German federal agriculture authorities are legally bound to monitor defined two main research questions: 1) Can winter catch cropping be
the parcels registered for catch cropping and decide on subsidization monitored from Sentinel-2 image time series data? 2) Which spatio-temporal
under the CAP Greening guidelines. So far, monitoring was done by on- limitations arise from fully-automated machine learning methods?
site controls where conditions such as the sowing dates, the ground
coverage rates, the tillage dates and the used seed mixes are surveyed. 2. Materials and methods
These controls require large numbers of inspectors to survey several
hundred thousand parcels within a short time frame of each catch crop 2.1. Study area
season. As a consequence, a very low percentage of approximately 5% of
all registered parcels are able to be controlled each year. The study area comprises the federal states of North Rhine-
To fulfill the legal requirements of area-wide agricultural monitoring Westphalia (NW), Lower Saxony (LS), Brandenburg (BB), and
and the reduction of on-site controls, alternative methods from the field Rhineland-Palatinate (RP). These federal states belong to the geospatial
of remote sensing are strongly encouraged by European authorities. units Northern Lowlands and Central Uplands of Germany (Ssymank,
With the Copernicus Sentinel missions managed by the European Space 1994). With approximately 6,000,000 ha, they comprise 36.6% of the
Agency (ESA), a large archive of free satellite imagery has been available agricultural area in Germany (Destatis, 2019). Due to differing historical
since 2014. As a consequence, many studies cover the topics of crop and conditions, the study area represents a gradient of agricultural land­
meadow identification and the monitoring of managing practices such scapes ranging from small, less-intensively used parcels (RP) to very large,
as harvesting, mowing and crop rotations based on satellite data of the high-intensively used parcels (BB). According to the Köppen-Geiger clas­
sensors Sentinel-1 (S1) (e.g., Tamm et al., 2016; Bargiel, 2017, Taravat sification scheme, all four states have warm summers and no dry season
et al., 2019; Garioud et al., 2020) and Sentinel-2 (S2) (e.g., Immitzer (Beck et al., 2018). Geospatial samples about catch crop cultivation in
et al., 2016; Kolecka et al., 2018; Sitokonstantinou et al., 2018; Griffiths these federal states were made available by agricultural authorities and
et al., 2019a, 2019b; Ottosen et al., 2019; Preidl et al., 2020; Solano- farmers during the timeStamp project (Thünen Institute, 2021). The
Correa et al., 2020) or both sensors (e.g., Veloso et al., 2017; Kussul datasets and their respective years are outlined within the study area
et al., 2018; Van Tricht et al., 2018; Pipia et al., 2019; D’Andrimont map (Fig. 1).
et al., 2020; Holtgrave et al., 2020). The Dynamic Cropland Mask (Valero
et al., 2016) of the Sen2-Agri project by ESA and the planned Copernicus 2.2. Data
Land Monitoring Service High Resolution Layer Crop Types by the European
Environmental Agency (Copernicus Programme, 2020) are two exam­ 2.2.1. Agricultural parcel data
ples of the vast application of S1 and S2 data for automatized agricul­ For training data acquisition, we collected approximately 150,000
tural monitoring on the large scale. Misra et al. (2020) outlined the parcel geometries with winter catch crop cultivation and 3,600,000
major benefits of S2 time series data for phenology research in crop parcel geometries without catch crop cultivation from different federal
mapping and yield estimation. Moreover, studies from the European agricultural agencies. The eleven raw datasets were from four different
Commission’s Joint Research Centre (Devos et al., 2017; D’Andrimont years (2016–2019) and four federal countries (Table 1). For LS and NW,
et al., 2018; Devos et al., 2018a; Devos et al., 2018b) show that the we received on-site control information of the catch crop parcels. For BB,
extraction of temporal markers from S1 and S2 time series data is crucial we received catch crop registered parcels without on-site control in­
to quantify traits of phenology and detect management events. formation. For RP, one survey dataset with voluntary information from

2
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

Fig. 1. Study area and agricultural parcel datasets used in this study. Overview map: Administrative borders and available samples within the federal states. Detail
maps: Parcel examples of the main class datasets with a 10 km diameter circle from different years and the federal states of Brandenburg (BB16/17/18), Lower
Saxony (LS16/17/18), North Rhine-Westphalia (NW16/17/18/19) and Rhineland-Palatinate (RP18). Bar charts: Number of catch crop subclass samples in North
Rhine-Westphalia 2019 (NW19_s) and Rhineland-Palatinate (RP18_s). Administrative borders: GADM, 2020. Projection: ETRS89 / UTM zone 32 N (EPSG: 25832).

farmers was available. For the datasets of NW and RP, partial informa­ Vegetation Index (NDVI), the Enhanced Vegetation Index, the
tion on the cultivated seed mixes and winter crops was available. Chlorophyll-Red-Edge Index, and the Soil Adjusted Vegetation Index
From the raw data, we created eleven main class datasets for model (Schulz et al., 2019). Within all indices, we found similar temporal
training and testing. For a better comparison of the machine learning trends and patterns for each class. Since the class based characteristics
classifiers trained from samples in different spatio-temporal domains, and differences were found to be most pronounced in the NDVI time
we decided on the reduction of each parcel dataset to an equal and series, we selected this index for further analysis. Within the same step,
randomly selected number of 1000 parcels. To prevent possible over­ we decided on the timeframe from July to April as most suitable for
weighting problems of a specific class in the machine learning models, winter catch crop monitoring as it comprises the harvest event of the
both classes (catch crop and non-catch crop) were equally proportioned to previous summer crop and the sowing of the following crop.
500 samples each. The final main class datasets are available in the The NDVI derived from multi-spectral, optical satellite data is
supplements of this article. commonly used for vegetation monitoring in remote sensing (e.g.,
We also created two subclass datasets with different crop mix sub­ Kuenzer et al., 2015; Ban and Yousif, 2016). As a ratio of the reflected
groups from the records of RP in 2018 and NW in 2019 (Fig. 1). The red and near-infrared spectra of light, the NDVI allows for the estimation
samples of each subclass dataset were aggregated into groups by the of the chlorophyll content of vegetation surfaces. It ranges between − 1
species or seed information given in the raw data. For the NW19_s and +1 with high values close to + 1 indicating dense and phenologi­
dataset, we chose 1,035 catch crop parcels for which information was cally active vegetation surfaces and values near to 0 and lower indi­
available from on-site controls and classified them into eight catch crop cating non-vegetated or phenologically non-active surfaces. NDVI
subgroups. For all other 1,372 samples, no or unclear species data were observations collected as time series facilitate the measurement of
available. For the RP18_s dataset, 276 samples with seed information on spectral-temporal profiles (e.g., Förster et al., 2012; Verger et al., 2016),
both main classes catch crop and non-catch crop were available. All other phenological chronologies (e.g., Gerstmann et al., 2016) or trends and
samples were tagged as ‘unknown’ or ‘other’. We aggregated them to 5 abrupt changes (e.g., Verbesselt et al., 2010), which is especially useful
catch crop subclasses and 8 non-catch crop subclasses, with 55 and 221 for crop monitoring on the pixel- or parcel-levels.
samples respectively (Fig. A.2). The subclass datasets are available as For the NDVI time series extraction, the GFZ Time Series System for
supplements of this article. Sentinel-2 (GTS2) (German Research Centre for Geosciences, 2020) was
used. This archive provides S2 Level 2A products with a very high spatial
2.2.2. NDVI time series data accuracy and cloud masks improved for Germany. This allowed us to
In the initial stage of this study, we decided on the best spectral index attain spatially precise data even for very small parcels and to reduce the
for time series-based agricultural monitoring. Thus, we compared time number of observations negatively affected by clouds. As input bands for
series of different multispectral indices from S2 data for the catch crop the NDVI calculation, we used bands 4 (red) and 8 (near-infrared), each
and non-catch crop parcels of RP18 including the Normalized Difference with 10 m spatial resolution.

3
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

Table 1 and observation gaps. The tables are available in the supplements of this
Raw datasets for training data extraction from the federal states of Brandenburg, article.
Lower Saxony, North Rhine-Westphalia, and Rhineland-Palatinate. Fig. 2 gives an overview of the temporal quality of the unfilled NDVI
Federal state Record Parcel number Origin Source* time series datasets. Over the different input years, the temporal
year
catch non-
coverage and resolution increased, with median observation numbers
crop catch ranging between 10 and 13 in 2016 and 24 to 40 in 2018. The higher
crop number of observations in recent years can mainly be explained by the
Brandenburg 2016 2,895 159,123 ecological MLUK BB launch of the S2B satellite in March 2017. Across all datasets, the
priority observation numbers have a large range, with a minimum of 3 and a
parcels maximum of 54. Observation numbers are strongly influenced by parcel
Brandenburg 2017 2,562 161,530 ecological MLUK BB sizes, leading to a reduced number of valid (i.e., “non-cloudy”) at smaller
priority
parcels
parcels. This also explains why in RP, a federal state with very small
Brandenburg 2018 2,757 210,201 ecological MLUK BB parcels, the median observation number of 24 is much lower compared
priority to other datasets from 2018, with 39 to 40.
parcels
Lower Saxony 2016 62,000 on-site MELV NI
2.3. Methods

control data
– 909,676 InVeKoS MELV NI
parcels 2.3.1. Exploratory data analysis
Lower Saxony 2017 32,700 – on-site MELV NI The Exploratory data analysis (EDA) (Tukey, 1977) provides us with
control data
statistical tools for developing hypotheses on the used data related to our
– 913,743 ecological SLA NI
priority research questions. Our first goal was to reveal phenological cycles and
parcels time frames for winter catch cropping across the NDVI time series of the
Lower Saxony 2018 32,000 – on-site MELV NI main class datasets. Our second goal was to find pattern variations
control data within the main classes by analyzing the NDVI time series of the subclass
912,988 ecological SLA NI
datasets.

priority
parcels
North Rhine- 2016 2,910 – on-site LWK NRW 2.3.2. Feature engineering
Westphalia control data Based on the outcomes of the EDA, we developed a set of descriptive
85,849 ecological MULNV

features with high information rate for catch crop monitoring and likely
priority NRW
parcels
to be highly predictive for automated catch crop detection. These fea­
North Rhine- 2017 3,336 – on-site LWK NRW tures were then used as input variables (i.e., predictors) for the training
Westphalia control data and testing of machine learning models.
– 86,060 ecological MULNV
priority NRW
2.3.3. Prediction modelling
parcels
North Rhine- 2018 4,269 272 on-site LWK NRW For the prediction of the target classes catch crop and other with the
Westphalia control data predictors created in the feature engineering process, we chose Random
– 137,704 InVeKoS MULNV Forest (RF) (Breiman, 2001) as it is known as a robust and transparent
parcels NRW
supervised machine learning approach. Fig. 3 outlines the basic work­
North Rhine- 2019 2,407 – on-site LWK NRW
Westphalia control data
flow for developing a single RF model in our study. First, we calculated
– 137,704 InVeKoS MULNV the filled NDVI time series with daily resolution based on the full parcel
parcels NRW geometries and satellite image data. Then, for all samples (i.e., parcels), a
Rhineland- 2018 110 241 voluntary AgroScience set of predictors was extracted from the time series and labeled ac­
Palatinate survey/
cording to the input classes catch crop and non-catch crop. Then, the
interviews
predictors dataset was split into training (70%) and test (30%) samples.
*
Ministerium für Landwirtschaft, Umwelt und Klimaschutz Brandenburg With the training samples, the model was carried out using 10-fold
(MLUK BB); Niedersächsisches Ministerium für Ernährung, Landwirtschaft und cross-validation repeated three times. Finally, the predictive model
Verbraucherschutz (MELV); Servicezentrum Landentwicklung und Agrarför­
was validated with the test data in a model-internal accuracy
derung Niedersachsen (SLA); Landwirtschaftskammer Nordrhein-Westfalen
assessment.
(LWK NRW); Ministerium für Umwelt, Landwirtschaft, Natur- und Ver­
braucherschutz des Landes Nordrhein-Westfalen (MULNV NRW); Institute for
Based on several combinations of the main class datasets, fifteen RF
Agroecology, RLP AgroScience GmbH (AgroScience). classifiers were created. The classifiers were named after their input
datasets comprising data of a single region and year (rf_BB16/17/18
rf_LS16/17/18, rf_NW16/17/18/19, rf_RP18), of multiple regions with
To extract NDVI time series vectors from the GTS2 archive, the
the same year (rf_ALL16/17/18), and data of all regions and years
following steps were conducted for each parcel: a) download of the S2
(rf_ALL). a) For the single-region, single-year RF models, each initial
image time series, b) calculation of the NDVI for each image, c)
dataset consisted of 1000 samples except for RP18, where only 351
computation of the mean NDVI pixel value, d) calculation of an equi­
samples were available. In order to have both classes equally propor­
distant time series with daily time steps, and e) linear temporal inter­
tioned in RP18, we raised the number of catch crop samples by dupli­
polation of the missing values within the time series. By steps d) and e),
cation and reduced the non-catch crop samples, resulting in a dataset
two univariate NDVI time series tables with daily resolution coded by a
with 440 samples (220 samples per class). b) For the multi-regional,
unique parcel-ID were created. One table comprises the real observa­
single-year RF models, all training and test datasets with the same
tions (i.e, the unfilled NDVI time series). This table was used for quality
input year were aggregated. c) For the multi-regional, multi-temporal
assessment regarding the number of available observations and the
RF model, the training and test datasets of all models were aggregated.
length of data gaps within the time series. The other table comprises the
Supervised machine learning models such as RF often perform well
real observations with interpolated values (i.e., the filled NDVI time se­
within their own input data, while performing worse for the prediction
ries). This step is necessary for the uniform calculation and comparison
of other data. To find the most robust model with a high temporal,
of temporal metrics throughout datasets with different observation dates
spatial, and class-level transferability, we conducted a multi-level

4
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

Fig. 2. Observation dates from July to April of the unfilled Sentinel-2 NDVI time series data from Brandenburg 2016–2018 (BB16/17/18), Lower Saxony 2016–2018
(LS16/17/18), North Rhine-Westphalia 2016–2019 (NW16/17/18/19), and Rhineland-Palatinate 2018 (RP18). Median and range refers to the number of available
observations across all parcels of each dataset. For better visualization, the number of time series for each plot was reduced to 50 random samples.

target class catch crop. The SPEC is the proportion of non-catch crop
samples that were correctly predicted as the target class other. Secondly,
we cross-validated all models with their respective test datasets according to
the OA, SENS and SPEC to determine the strengths and limitations of
each model on the spatial and temporal level. Thirdly, we assessed the
variable importances of all predictors. The variable importance is a sta­
tistical metric for the weight of each predictor in the RF model which
was used to define the most useful features for automated catch crop
detection. Finally, we compared the Out-Of-Bag probability (PROB) rates
of the best classifier for the prediction of the target class catch crop on the
main class and subclass datasets to define prediction strengths and
limitations on the class-level. The PROB is an estimate for each sample to
be correctly labeled by the proportion of the trees that voted for the
target class. It shows the likelihood for the test samples being predicted
as catch crop, with values ranging from 0 to 1. The higher the PROB value
of a sample is, the higher the model weights the likelihood that it is a
parcel with catch crop cultivation. Samples with a PROB higher than 0.5
were labeled by the RF model as catch crop. Samples with lower values
Fig. 3. Flowchart for data pre-processing, training and testing using the were labeled as other.
example of a single Random Forest model based on time series from Sentinel-2
Level 2A (S2 L2A) products from the GTS2 archive (German Research Centre for
2.3.4. Software environment
Geosciences, 2020).
The algorithms and RF models developed for this study were
implemented via virtual Docker containers into the timeStamp software
validation of the RF prediction models. First, we compared the overall prototype which allows for large-scale automatized catch crop analysis
accuracies (OAs), sensitivities (SENS) and specificities (SPEC) from the on the parcel-level (Thünen Institute, 2021). This software saves the
model-internal accuracy assessments. The OA is the proportion of raster data from the GTS2 archive as parcel-wise clipped image time
correctly classified samples according to their input labels. The SENS is series into a PostGIS database. All further processing steps in this study
the proportion of catch crop samples that were correctly predicted as the were performed with the statistical computing language R (RStudio

5
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

Team, 2020). For raster data manipulation within the PostGIS database features to be more robust and comparable over different spatio-
and downloading NDVI time series, we used the packages rpostgis temporal domains.
(Bucklin and Basille, 2019) and RPostgreSQL (Conway et al., 2017). For We then developed nineteen descriptive features and statistical
time series filling and calculation of the predictors, we used the packages metrics. All parameters are described in detail in Table 2. In summary,
zoo (Zeileis et al., 2020), hydroGOF (Zambrano-Bigiarini, 2020), tsou­ they can be divided into univariate and bivariate features. The univar­
tliers (de Lacalle, 2019), and changepoint (Killick et al., 2016). For RF iate features measure the strength and progression of the NDVI during
modelling, we used the packages caret (Kuhn et al., 2020) and ran­ autumn and spring (e.g., minima, maxima, trends) or points in time
domForest (Liaw and Wiener, 2018). For better reproducibility, the input where specific conditions and thresholds are reached (e.g., level shifts,
data, scripts and models are available in the supplements of this article. maxima). The bivariate features measure the correlation of the observed
time series to a standardized catch crop NDVI time series profile. The
3. Results results for the main class and subclass datasets are available in the
supplements and were summarized in Fig. A.4.
3.1. Exploratory data analysis
3.3. Prediction modelling
3.1.1. Main class datasets
For nine of the eleven main class datasets, we found generalizable 3.3.1. Model-internal validation
trends and unique temporal patterns for catch crop cultivation. As Table 3 summarizes the model-internal statistics of all RF models.
shown in Fig. 4, the NDVI time series of the class catch crop commonly There is a large variety across all models with OAs ranging from 0.76 to
shows a strong positive trend between October and December and a 0.91, SENS from 0.78 to 0.91, and SPEC from 0.65 to 0.94. According to
negative trend between January and April. NDVI time series of the class the OA, rf_BB16 performed the worst and rf_RP18 performed the best of
non-catch crop always have a negative or no trend in autumn and a all models. The multi-region, single-year models showed more similar
positive trend in spring. For the dataset RP18, the typical pattern of the OA, SENS and SPEC values than the single-region, single-year models.
class catch crop was much weaker and more similar to the class non-catch The model rf_ALL reached OA, SENS and SPEC with values around 0.83.
crop. For BB16, no visual difference between the two classes was found.
In the NDVI time series heatmaps, additional temporal patterns were 3.3.2. Cross-validation
found (Fig. A.1). First, in the catch crop time series of some datasets (e. Table 4 shows the cross-validation results of all RF models against
g., LS17), abrupt NDVI drops in July and mid-February could be their respective test datasets. Each column of the confusion matrix
observed. Secondly, the time series within the class non-catch crop represents a model while each row represents a test dataset. A detailed
showed larger temporal variance in each dataset compared to the time version of this table with information on the SENS, SPEC and the mean
series of the class catch crop. We also see that these patterns are not values of each metric can be found in Table A.3.
clearly visible in the datasets of BB16 and RP18.
(a) The single-region, single-year models reached OAs of 0.35 to 0.88
3.1.2. Subclass datasets across the test datasets of the other models. The models rf_BB16
For nine of the ten catch crop subclasses within the dataset NW19_s, and rf_RP18 showed the lowest OAs with 0.35 to 0.69 across all
we found positive NDVI trends in autumn and negative trends in spring datasets except their own. And conversely, other models per­
(Fig. 5). This observation fits the general temporal pattern of catch crop formed worse on BB16 and RP18. The mean SENS values of all
cultivation found in the main class datasets. The subclass grass-clover mix models (0.31 to 0.76) are lower than the mean SPEC values of all
differs from the other subclasses with a general negative NDVI trend models (0.66 to 0.90). Some models showed very low SENS
until October and a slow positive growth until March. values compared to other datasets. For instance, rf_NW19 reaches
For most of the catch crop subclasses of the dataset RP18_s (Fig. A.2), a SENS of 0.01 for BB16. A higher detection rate was found across
we found less typical results. However, the NDVI curves of most catch models with the same input year. For instance, the model
crop subclasses show a slight positive peak between October and rf_NW18 performed better on the datasets BB18, LS18 and RP18
January which matches the main class pattern for catch crops. The mean than models from other years. In contrast, no higher detection
NDVI curves of nearly all non-catch crop subclasses fit the mean NDVI rate was found across models with the same input region. For
time series of the main class datasets for non-catch crops. instance, the model rf_NW16 did not perform better on the
datasets NW17, NW18 and NW19 than other models.
3.2. Feature engineering (b) The multi-region, single-year models reached similarly high ac­
curacies as the single-region models. They showed high OAs on
Transferring administrative requirements from on-site controlling to its respective test data and low accuracies on the test data of other
the feature engineering process was a major challenge of this study. models. Additionally, they had high performances for datasets
Thus, the set of descriptive features should cover the following aspects: within the same input year and low performances for datasets
a) catch cropping-specific metrics for a reliable and reproducible from other input years. For instance, the model rf_ALL18 reached
monitoring of agricultural management practices (e.g., sowing, har­ OAs above 0.80 in BB18, LS18, NW18, and RP18. For datasets
vesting, plowing) and phenotypical stages (>40% ground coverage); from other years, it reached OAs below 0.80.
and b) metrics for machine learning-based detection and differentiation (c) The multi-region, multi-year model rf_ALL, trained with samples
of parcels with the target classes catch crop and other. from all models, reached OAs between 0.67 and 0.91 across all
Based on the outcomes of the EDA, we pre-defined temporal win­ test datasets. It had the highest mean OA of all models (0.83) and
dows by the day of year (DOY) in which agricultural management events the lowest standard deviation (0.05). In addition, the mean SENS
occur and differences between the classes catch crop and non-catch-crop and mean SPEC, with 0.82 and 0.84 respectively, were the most
most likely arise. These temporal windows include the time frame from similar compared to all other models.
catch crop sowing in late summer until winter dormancy (DOY 182 -
DOY 356) and the time frame from winter dormancy until tillage in 3.3.3. Variable importance
spring (DOY 1 - DOY 120). To consider the spatio-temporal variability of The variable importance is shown in Fig. 6. From the set of nineteen
phenophases caused by different weather conditions and climate and predictors, the bivariate features rpearson_182_365 and rpear­
due to potentially long data gaps in the cloudy winter period, the win­ son_001_120 showed a variable importance of more than 50% across all
dows were intentionally chosen to be wide. This attribute allows for the models. Those predictors that quantify trends and sums in spring and

6
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

Fig. 4. Original NDVI time series observations (black dots) and its mean average LOESS smoothed trend (orange line) of the eleven main class datasets split into the
input classes catch crop and non-catch crop. For better visualization, the number of time series for each plot was reduced to 50 random samples. LOESS (Locally
Estimated Scatterplot Smoothing) is a local regression filter and was set up with a 10% span for smoothing the time series and showing the weighted average profile.
(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

7
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

Fig. 5. NDVI time series (black dots) for different catch crop subclasses and its mean average LOESS smoothed trend (orange curve) based on the NW19_s dataset.
The number of time series for each plot was randomly reduced to a maximum of 20 for better visualization. The LOESS filter was set up with a 10% span for
smoothing the time series and showing the weighted average profile. The original sample number (n) for each class is stated in the upper right corner of each plot.
(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

late autumn had an importance of less than 25% on average across all biomass estimation (CCROP, Thieme et al., 2020; Peredo et al., 2019)
models. The level shift predictors that measure the DOY and number of and temporal markers for agricultural practices monitoring (sen4CAP,
abrupt changes in the NDVI time series had the lowest importance of less European Space Agency, 2018) have been developed, methods for
than 10%. automated winter catch crop detection are still lacking. Based on the
example of CAP Greening parcels in Germany and S2 data, we developed
3.3.4. Probability rates novel algorithms for feature extraction from NDVI time series to mea­
Fig. 7 summarizes the PROB rates of the model rf_ALL for the samples sure the temporal stages of winter catch cropping and for machine
of the class catch crop in the main class datasets. We found mean PROB learning-based classification.
values between 0.6 and 0.9 across the eleven datasets. The lowest values According to an EDA of eleven training datasets, specific temporal
and high ranges were found for the datasets RP18 and BB16. For samples patterns for catch crop and non-catch crop parcels were identified for
of the class non-catch crop (not shown in Fig. 7), we found mean values different years. The strong positive NDVI trend in autumn relates to the
between 0.1 and 0.2 across all datasets. major growing period of catch crops which starts earlier and is stronger
In Fig. 8, the PROB rates of the model rf_ALL for the subclass input than in other winter cultures. The negative trend in spring relates to the
datasets are summarized. Within the dataset NW19_s, samples of the plowing of the catch crop while most other crops are allowed to continue
catch crop subclass grass-clover mix had the lowest mean PROB rates of growing. We expect the unique temporal pattern for winter catch crop
approximately 0.6 and a range between 0.4 and 0.8. The samples of all cultivation stages generally applies across upcoming years at the
other groups reached mean values above 0.8 and a range between 0.7 research area. The different observation numbers and data gaps within
and 0.9. Within the dataset RP18_s, we found PROB values ranging be­ the NDVI time series datasets shown in Fig. 2 did not exert visually a
tween 0.3 and 0.9 with mean values between 0.7 and 0.8 for samples of major influence on the overall patterns found in the exploratory data
the catch crop subclasses. analysis. For all years, we can observe similar class-specific increases
and decreases in spring and autumn (Fig. 4). However, two of the eleven
4. Discussion input datasets showed variations from the unique catch crop phenology
which led to lower prediction accuracies of the developed models. For
4.1. Detectability of catch crops example, the catch crop cultivation season 2016/2017 in Germany was
highly influenced by hot and dry weather conditions until late
Even though precise metrics for parcel-based performance and September and frosts in mid-November. The resulting slow and patchy

8
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

Table 2
Descriptive features for catch crop monitoring and prediction derived from NDVI time series.
Predictor Description Example for a real catch crop parcel

ndvi_min_182_365 The parameter corresponds to the minimum of photosynthetically active


NDVI minimum between 1 July and 31 Dec. vegetation cover of the parcel. The value is expected to be low with a preliminary
harvest event before catch crop cultivation.

ndvi_max_182_365 The parameter corresponds to the maximum of photosynthetically active


NDVI maximum between 1 July and 31 Dec. vegetation cover of the parcel. The value is expected to be high with a successful
catch crop cultivation.

ndvi_mean_182_365 The parameter corresponds to an active vegetation cover of the parcel. The value is
NDVI mean between 1 July and 31 Dec. expected to be high with a successful catch crop cultivation.

ndvi_mean_var_182_365 The parameter corresponds to the temporal steadiness of the vegetation activity of
NDVI mean variance between 1 July and 31 Dec. the parcel. The value is expected to be high with abrupt changes caused by
harvesting, plowing or frost.

ndvi_trend_196_273 The parameter corresponds to the strength of the vegetation activity growth on the
NDVI trend slope between 15 July and 30 Sept. parcel. The value is expected to be high with a successful catch crop cultivation.
Negative values correspond to a harvest or plowing event within the timeframe.

ndvi_trend_213_258 The parameter corresponds to the strength of the vegetation activity growth on the
NDVI trend slope between 1 August and 15 Sept. parcel. The value is expected to be high with a successful catch crop cultivation.
Negative values correspond to a harvest or plowing event within the timeframe.

ndvi_trend_274_332 The parameter corresponds to the strength of the vegetation activity growth on the
NDVI trend slope between 1 Oct. and 28 Nov. parcel. The value is expected to be high with a successful catch crop cultivation.
Negative values correspond to a harvest or plowing event within the timeframe.

ndvi_sum_232_272 The parameter corresponds to the strength of the vegetation activity on the parcel.
NDVI sum between 22 August and 29 Sept. The value is expected to be high with a successful catch crop cultivation.

ndvi_sum_278_348 The parameter corresponds to the strength of the vegetation activity on the parcel.
NDVI sum between 5 Oct. and 14 Dec. The value is expected to be high with a successful catch crop cultivation.

ndvi_sum_001_046 The parameter corresponds to the strength of the vegetation activity on the parcel.
NDVI sum between 1 Jan. and 15 Feb. The value is expected to be high with a successful catch crop cultivation.

doy_ndvi_max_182_365 The parameter corresponds to the date of the maximum. The value is expected in
Day of year of the NDVI maximum between 1 July late autumn between the sowing date of the catch crop and the start of winter
and 31 Dec. dormancy.

doy_ndvi_min_182_365 The parameter corresponds to the date of the minimum. The value is expected in
Day of year of the NDVI minimum between 1 July late summer between the harvest date of the preliminary main crop and the sowing
and 31 Dec. date of the catch crop.

rpearson_ts_sim_182_365 The parameter corresponds to the similarity of the observed time series with an
Pearson’s r of the simulated and the observed NDVI averaged (i.e., simulated) time series for catch crop cultivation (Verger et al., 2016;
time series between 1 July and 31 Dec. Heupel et al., 2019). The value is expected to be high with a successful catch crop
cultivation.

rpearson_ts_sim_1_120 The parameter corresponds to the similarity of the observed time series with an
Pearson’s r of the simulated and the observed NDVI averaged (i.e., simulated) time series for catch crop cultivation (Verger et al., 2016;
time series between 1 Jan. and 30 April
(continued on next page)

9
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

Table 2 (continued )
Predictor Description Example for a real catch crop parcel

Heupel et al., 2019). The value is expected to be high with a successful catch crop
cultivation.

number_level_shifts_182_365 Outliers of the type ’level shift’ (de Lacalle, 2019) indicate abrupt changes in the
Number of level shifts between 1 July and 31 Dec. time series. The value is expected to be higher with the number of harvesting,
plowing, frost and dryness events.

doy_first_level_shift_182_365 The parameter corresponds to the date of the first outlier of the type ’level shift’ (de
Day of year of the first NDVI level shift between 1 Lacalle, 2019). The value is expected to correlate with the plowing/harvest date of
July and 31 Dec. the preliminary main crop in late summer.

doy_first_level_shift_001_120 The parameter corresponds to the date of the first outlier of the type ’level shift’ (de
Day of year of the first NDVI level shift between 1 Lacalle, 2019). The value is expected to correlate with the plowing/harvest date of
Jan. and 30 April the catch crop, which is allowed after the 16th of February (DOY 47).

doy_changepoint_182_365 The parameter corresponds to the date of the changepoint of the type ’cpt.mean’ (
Day of year of the NDVI changepoint between 1 July Killick et al., 2016). The value is expected between the sowing date of the catch
and 31 Dec. crop in late summer and the highest growth rates in late autumn.

doy_changepoint_001_120 The parameter corresponds to the date of the changepoint of the type ’cpt.mean’ (
Day of year of the NDVI changepoint between 1 Jan. Killick et al., 2016). The value is expected between the winter dormancy of the
and 30 April catch crop and the highest growth rates in early spring.

If both predictors have high values, the likelihood of a catch crop sample
Table 3
to be correctly labeled increases substantially. Despite the other metrics
Model-internal statistics for all Random Forest models including the number of
having a lower importance for the RF classifiers, they are useful for
training samples (n train), number of test samples (n test), overall accuracy
(OA), sensitivity (SENS) and specificity (SPEC).
reproducible monitoring of phenological stages and agricultural man­
agement practices at catch crop parcels.
Model n train n test OA SENS* SPEC**
The poor performance of the single-region, single-year models for
rf_BB16 700 300 0.76 0.87 0.65 catch crop detection across other spatio-temporal domains indicates an
rf_BB17 700 300 0.86 0.78 0.94 overfitting problem. This issue negatively affects the application of RF
rf_BB18 700 300 0.86 0.81 0.91
rf_LS16 700 300 0.84 0.85 0.83
classifiers across large regions. However, with the balanced OA, SENS
rf_LS17 700 300 0.88 0.91 0.85 and SPEC values of the final classifier rf_ALL shown in the cross-
rf_LS18 700 300 0.87 0.89 0.85 validation, we achieved a high robustness regarding the prediction of
rf_NW16 700 300 0.83 0.89 0.78 both classes across all spatio-temporal domains.
rf_NW17 700 300 0.80 0.85 0.75
rf_NW18 700 300 0.89 0.88 0.90
rf_NW19 700 300 0.82 0.82 0.82 4.2. Detectability of catch crop subclasses
rf_RP18 308 132 0.91 0.91 0.91
rf_ALL16 2100 900 0.78 0.82 0.74
rf_ALL17 2100 900 0.84 0.83 0.85 In the EDA of the datasets NW19_s and RP18_s, we found similar
rf_ALL18 2408 1032 0.88 0.88 0.88 temporal patterns for most subclasses of the main class catch crop. On the
rf_ALL 6608 2832 0.83 0.83 0.84 one hand, this means most subclass samples are equally predictable as
*
Proportion of catch crop test samples that were correctly labeled as catch the target class catch crop. On the other hand, this result indicates that
crop. individual seed mixes cannot be differentiated with our data and
**
Proportion of non-catch crop test samples that were correctly labeled as methods. Only seed mixes with a grass component (e.g., Lolium, Phleum,
other. Festuca, Avena) appeared to sprout later and develop more slowly than
other seed mixes. According to the PROB rates, the temporal pattern of
development and early die-off of catch crops (Deutscher Wetterdienst, the subclass grass-clover mix was likely a major source for misclassifi­
2016) especially influenced the quality of the BB16 training data set. cation within the NW19_s dataset. However, more data with subclass
Thus, under specific annual weather conditions such as heat waves or information is needed for a comprehensive analysis. In applying remote
early frost, phenological stages and growth rates become less pro­ sensing methods to differentiate catch crop subgroups, texture-based
nounced in the NDVI signal and the chosen method may work more indices on very high resolution aerial imagery could potentially pro­
weakly. vide new insights.
The developed set of descriptive features, which measured cultiva­
tion stages and phenological patterns, facilitated automated catch crop 4.3. Spatio-temporal limitations
detection. In particular, the variables rpearson_ts_sim_182_365 and
rpearson_ts_sim_1_120, which measure the correlation of the observed Our method was developed for agricultural monitoring in Germany.
time series to an averaged catch crop profile, assisted in class separation. Thus, the classifiers are limited to regions with temperate climate and

10
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

non-irrigated agriculture leading to a relatively distinct seasonality of

Confusion matrix for the overall accuracies of the Random Forest models (columns) and their respective test datasets (rows). For additional information on the sensitivity and specificity and the mean values for each

rf_ALL
different crop types. Similar applications to other climate zones or other

0.67
0.84
0.84
0.78
0.83
0.87
0.76
0.82
0.85
0.88
0.90
0.91
0.88
0.74
0.83
land surfaces are not addressed in this study, but could be further
investigated in future studies.
On the temporal level, we found clear differences (i.e., phenological

rf_NW19
shifts) within the input datasets, most likely caused by year-specific

0.83
0.52
0.77
0.82
0.72
0.50
0.72
0.68
0.64
0.74
0.82
0.71
0.76
0.75
0.85
weather conditions (Fig. 5). To avoid temporal limitations, we
selected long time frames of at least two months for feature extraction.
Additionally, we incorporated input data from different years into the
rf_ALL18

training process to overcome overfitting problems. Another option to


address this issue would be phenological corrections of the time series
0.50
0.68
0.77
0.65
0.75
0.78
0.63
0.72
0.84
0.88
0.90
0.89
0.88
0.71
0.76
data based on weather data before feature extraction.
On the spatial-level, we did not find differences within the input
datasets. This led to the assumption that there are no specific regional
rf_RP18

variances of the farmers’ practices in catch cropping. Therefore, special


0.58
0.91
0.65
0.58
0.60
0.58
0.58
0.54
0.57
0.56
0.60
0.52
0.56
0.69
0.57
strategies on training data selection or phenological corrections were
not necessary in this study. Nonetheless, phenological variations caused
by different climate conditions potentially reduce the applicability on
rf_NW18

large scales. Thus, for the application for the full scale of Germany, we
0.89
0.61
0.82
0.73
0.73
0.50
0.69
0.74
0.64
0.70
0.78
0.66
0.71
0.80
0.87

would propose to integrate training data from all bio-geographical re­


gions, including the alpine mountain regions, into the model training
process.
rf_LS18

0.50
0.65
0.72
0.62
0.70
0.75
0.65
0.70
0.80
0.87
0.88
0.65
0.82
0.72
0.72

4.4. Variable importance

Although all nineteen features are useful for parcel-based moni­


rf_BB18

toring, few were highly important for predictive modelling. From the
0.50
0.79
0.80
0.70
0.78
0.83
0.70
0.77
0.86
0.82
0.88
0.64
0.83
0.72
0.77

variable importance and PROB rates, we drew three conclusions. First,


the level shift variables, which should measure the summer crop har­
vesting dates and the catch crop plowing date, had unexpectedly low
rf_ALL17

importance for the predictive models. We believe that this is caused by


0.77
0.72
0.76
0.49
0.77
0.78
0.68
0.85
0.87
0.79
0.84
0.76
0.83
0.82
0.61

the chosen time frame (beginning on the 1st of July), which does not
fully comprise the summer crop harvest time frame. In addition, we
assume that many time series have too few observations for a reliable
calculation of these metrics. Therefore, level shifts were often unable to
rf_NW17

be calculated or wrongly calculated. Secondly, the implementation of


0.47
0.71
0.75
0.64
0.69
0.77
0.80
0.75
0.63
0.74
0.70
0.55
0.67
0.66
0.69

the two correlation variables, which measure Pearson’s r to an average


NDVI profile, were unexpectedly helpful. They served as a good proxy
for the separation of the two classes in many input datasets, leading to a
rf_LS17

high mean variable importance. However, a reduction of the model


0.58
0.78
0.68
0.75
0.50
0.75
0.77
0.67
0.81
0.88
0.69
0.80
0.75
0.84
0.84

input features to these two parameters may be misleading. All features


with mean variable importances higher than 25% in the training data
(see Fig. 6) are needed for consistent performance across different
rf_BB17

spatio-temporal domains. For instance, in the dataset BB16, where the


0.52
0.83
0.76
0.70
0.86
0.86
0.72
0.81
0.74
0.80
0.79
0.60
0.75
0.71
0.76

average temporal pattern of the catch crop parcels was quite unspecific,
the r Pearson metrics performed comparably low for the prediction (see
Fig. A.4) but other features counterbalanced this. Thirdly, we found that
rf_ALL16

features that cover the spring time are important for accurate catch crop
0.63
0.75
0.69
0.76
0.69
0.83
0.83
0.78
0.78
0.83
0.66
0.76
0.76
0.76
0.77

prediction. However, administrative control timeframes for subsidy


payments currently end in December and thus all the information on
decision-making must be provided at an early stage. Alternative detec­
rf_NW16

tion models solely trained for the first half of catch crop cultivation
0.82
0.79
0.80
0.67
0.79
0.70
0.76
0.48
0.83
0.83
0.72
0.80
0.80
0.69
0.76

could potentially support these administrative constraints. However, the


administrative deadlines for CAP monitoring could be reversely adapted
to the remote sensing-based monitoring time frame.
rf_LS16

0.48
0.84
0.81
0.71
0.78
0.82
0.70
0.76
0.82
0.80
0.81
0.68
0.79
0.70
0.76

4.5. Predictive model improvements


model, we refer to Table A.3.

For further application of the proposed method, five aspects of the


training and validation process of the models could be improved.
rf_BB16
Model

0.58
0.48
0.41
0.49
0.76
0.49
0.40
0.55
0.49
0.49
0.35
0.44
0.55
0.45
0.38

(a) Using S2 data, we developed an easy and reproducible approach


with a short processing time. However, we are also limited to
optical, mono-sensor information with a temporal resolution of
Data-set
Table 4

ALL16

ALL17

ALL18
NW16

NW17

NW18

NW19
RP18
BB16

BB17

BB18
LS16

LS17

LS18

approximately 5 days and potentially large data gaps in winter­


ALL

time due to cloudy weather. To expand the number of

11
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

Fig. 6. Variable importance across all models (left) and the model rf_ALL (right).

Fig. 7. Out-Of-Bag probability rates for the prediction of the target class catch crop in the main class datasets calculated with the model rf_ALL. Left: Values for catch
crop samples in Brandenburg (BB16/17/18), Lower Saxony (LS16/17/18), North Rhine-Westphalia (NW16/17/18/19), and Rhineland-Palatinate (RP18). Right:
Example of values calculated for catch crop and non-catch crop parcels in the dataset RP18. High values > 0.5 indicate catch cropping.

observations and ensure a consistent temporal coverage, we elevation and climate information to the set of predictors could
could add NDVI time series data from the Landsat 8 sensor or increase the RF classifiers’ accuracy significantly.
Radar Vegetation Index (RVI) (Kim et al., 2012; Pipia et al., 2019) (c) We expect that the classifier performs more robust to new data if
time series from the Sentinel-1 sensor to the set of input data. As equal training samples from all years would be integrated into
can be seen in Fig. 10, all three sources show distinctive temporal training. In the cross-validation, the model RF_ALL performs
patterns for catch cropping, but there are also sensor-specific better for test samples in the year 2018 and worse for the years
biases in pattern strengths and temporal progress. Thus, the set 2016, 2017 and 2019. This maybe is caused by the imbalanced
of predictors would need to be adapted for each sensor or the number of available training datasets for each year. For the pre­
multi-sensor data would need to be harmonized. The Harmonized diction of new data this could lead to a problem of overfitting to
Landsat Sentinel-2 surface reflectance dataset (Claverie et al., specific weather (i.e., phenological) conditions alike the year
2018) used in Thieme et al. (2020) would be a good alternative to 2018.
the used data if available in 10 m spatial resolution. (d) We believe that weather data could largely explain the spatio-
(b) Our study area comprises samples with similar climatic condi­ temporal variations found in this study. For instance, an
tions from the Northern Lowlands and Central Uplands of Ger­ extremely hot September and first frosts in mid-November in
many. However, on the country-scale, climatic variations may 2016 led to a weak growth and early die-off of the catch crops
reduce the models’ applicability. Thus, the integration of (Deutscher Wetterdienst, 2016). In comparison, the mild weather

12
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

Fig. 8. Out-Of-Bag probability rates


for the prediction of the target class
catch crop in the two subclass datasets
calculated with the model rf_ALL.
Left: Values for all samples from the
NW19_s dataset subdivided into ten
catch crop subclasses. Right: Values
for all samples from the RP18_s data­
set subdivided into five catch crop
subclasses (red) and eight non-catch
crop subclasses (blue). (For interpre­
tation of the references to colour in
this figure legend, the reader is
referred to the web version of this
article.)

Fig. 10. Comparison of the Normalized Difference Vegetation Index (NDVI) from Sentinel-2 and Landsat 8 and the Radar Vegetation Index (RVI) from Sentinel-1
based on the LS18 main class dataset. Each plot is subdivided into the classes catch crop and non-catch crop with 100 samples for each class. Sources: Google
Earth Engine Landsat 8 Collection 1 Tier 1 8-Day NDVI Composite dataset (Gorelick et al., 2017). JKI Sentinel-1 Gamma0 backscatter time series data archive (Julius
Kühn Institute, 2020).

conditions with high rain sums in the years 2017 and 2018 had a 5. Conclusions
positive influence on the catch crop development. However, the
integration of up-to-date weather or drought data into the models We developed an effective and easily reproducible machine learning-
is challenging and further analysis on pattern-process relation­ based method for parcel-level winter catch crop monitoring and detec­
ships needs to be conducted. tion. The implementation of the algorithms into the timeStamp software
(e) The quality of the developed monitoring metrics often remains prototype brings a major contribution to the control of subsidized par­
unclear. Thus, a great benefit for the alternating feature engi­ cels in Germany under CAP CC guidelines. With the extraction of nine­
neering and validation process would be the provision of precise teen cultivation-specific features from S2 NDVI time series, our
documentation on agricultural management events (e.g., har­ predictive models reached overall accuracies of up to 90% for test
vesting, sowing, and plowing) and phenology stages (e.g., first datasets under optimal (i.e., mild and sunny) weather conditions. This
sprout shooting, maximum soil coverage, and start of winter high detection rate was possible because of a generalizable and unique
dormancy). Such on-site measurements are urgently needed for temporal pattern for catch cropping which was found across the time
further improving machine learning-based methods with remote series datasets of four years and four federal states in Germany.
sensing time series. Despite its full-scale, automated applicability, the presented moni­
toring method also has disadvantages compared to the conventional on-

13
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

site controlling approach. First, different catch crop seed mixes cannot Ban, Y., Yousif, O., 2016. Change Detection Techniques: A Review. In: Ban, Y. (Ed.),
Multitemporal Remote Sensing – Methods and Applications. Springer, Cham,
be distinguished with the used data. Secondly, if the phenological con­
pp. 19–43. https://doi.org/10.1007/978-3-319-47037-5_2.
ditions strongly vary from the average catch crop pattern, due to region- Bargiel, D., 2017. A new method for crop classification combining time series of radar
specific or year-specific weather conditions, the predictive models images and crop phenology information. Remote Sens. Environ. 198, 369–383.
become less precise. The assimilation of weather, elevation or multi- https://doi.org/10.1016/j.rse.2017.06.022.
Beck, H., Zimmermann, N., McVicar, T., Vergopolan, N., Berg, A., Wood, E., 2018.
sensor satellite data could overcome limitations on the spatial, tempo­ Present and future Köppen-Geiger climate classification maps at 1-km resolution.
ral and class-level. More precise temporal in-situ documentation of dates Sci. Data 5, 180214. https://doi.org/10.1038/sdata.2018.214.
with agricultural events (e.g., harvesting, tillage, and plowing) and Blanco-Canqui, H., Shaver, T., Lindquist, J., Shapiro, C., Elmore, R., Francis, C.,
Herger, G., 2015. Cover Crops and Ecosystem Services: Insights from Studies in
specific phenology stages (e.g., sprouting, growth, winter dormancy) Temperate Soils. Agron. J. 107 (6), 2449–2474. https://doi.org/10.2134/
could greatly improve the feature engineering process and model agronj15.0086.
validation. Breiman, L., 2001. Random Forests. Mach. Learn. 45 (1), 5–32 https://doi.org/10.1023.
Bucklin, D., Basille, M., 2019. rpostgis: Linking R with a PostGIS Spatial Database. The R
Nonetheless, within the framework of a cost-efficient and area-wide Journal 10 (1), 251–268 https://doi.org/10.32614/RJ-2018-025.
monitoring of CAP Greening measures, our approach can greatly support Claverie, M., Ju, J., Masek, J., Dungan, J., Vermote, E., Roger, J.-C., Skakun, S.,
the authorities in their controlling obligation. During the early stage of Justice, C., 2018. The harmonized Landsat and Sentinel-2 surface reflectance data
set. Remote Sens. Environ. 219, 145–161. https://doi.org/10.1016/j.
monitoring, the authorities can select critical parcels which most require rse.2018.09.002.
on-site controls by interpreting the set of temporal features. Further­ Conway, J., Eddelbuettel, D., Nishiyama, T., Prayaga, S., Tiffin, N., 2017. RPostgreSQL: R
more, the classifier automatically detects catch cropping parcels and Interface to the ’PostgreSQL’ Database System. R Package Version 0.6-2. https://CR
AN.R-project.org/package=RPostgreSQL.
evaluates the likelihood for each parcel that the cultivation was suc­
Copernicus Programme, 2020. Copernicus Land Monitoring Service High Resolution
cessful within the fixed timeframe. Thus, our winter catch cropping Layer Crop Types (HRL-Crop Types). https://land.copernicus.eu/news/consultation-
monitoring approach is a reliable alternative to on-site controls. process-now-open-for-hrl-crop-type.
Dabney, S., Delgado, J., Reeves, D., 2001. Using winter cover crops to improve soil and
water quality. Commun. Soil Sci. Plant Anal. 32, 1221–1250. https://doi.org/
CRediT authorship contribution statement 10.1081/CSS-100104110.
de Lacalle, J., 2019. tsoutliers: Detection of Outliers in Time Series. R Package Version
0.6-8. https://CRAN.R-project.org/package=tsoutliers.
Christian Schulz: Conceptualization, Methodology, Software,
Delgado, J., Dillon, M., Sparks, R., Essah, S., 2007. A decade of advances in cover crops.
Writing - original draft, Formal analysis, Validation. Ann-Kathrin J. Soil Water Conserv. 62 (5), 110A–117A.
Holtgrave: Writing - review & editing. Birgit Kleinschmit: Conceptu­ Destatis, 2019. Statistisches Bundesamt: Landwirtschaftlich genutzte Fläche in
alization, Methodology, Supervision, Funding acquisition. Deutschland nach Bundesländern in den Jahren 2017 bis 2019 (in 1.000 Hektar).
206265.
Deutscher Wetterdienst, 2016. Deutscher Wetterdienst zum Agrarwetter im Herbst 2016.
Trockener Herbst brachte Zyklus von Winterkulturen und Bäumen durcheinander.
Declaration of Competing Interest (“The weather in Germany in autumn 2016. Warm, sunny and too dry”). Press
release from 2016-12-20. https://www.dwd.de/DE/presse/pressemitteilungen/DE/
The authors declare that they have no known competing financial 2016/20161220_agrarwetter_herbst.pdf.
Devos, W., Fasbender, D., Lemoine, G., Loudjani, P., Milenov, P., Wirnhardt, C., 2017.
interests or personal relationships that could have appeared to influence
Discussion document on the introduction of monitoring to substitute OTSC. JRC
the work reported in this paper. Technical Reports. https://doi.org/10.2760/258531.
Devos, W., Lemoine, G., Milenov, P., Fasbender, D., 2018a. Technical guidance on the
decision to go for substitution of OTSC by monitoring. JRC Technical Reports.
Acknowledgments
https://doi.org/10.2760/693101.
Devos, W., Lemoine, G., Milenov, P., Fasbender, D., Loudjani, P., Wirnhardt, C., Sima, A.,
This article is a contribution of timeStamp (Entwicklung eines Griffiths, P., 2018b. Second discussion document on the introduction of monitoring
to substitute OTSC: rules for processing applications in 2018–2019. JRC Technical
automatisierten Hinweissystems für die Kontrolle von Agrarförderflä­
Reports. https://doi.org/10.2760/344612.
chen) funded by the German Ministry for Economics and Technology European Commission, 2020. Communication from the Commission to the European
(BMWi) and managed by the German Aerospace Centre (DLR) (grant Parliament, the Council, the European Economic and Social Committee and the
number: 50EW1704C). The authors are grateful for the support and Committee of the Regions. EU Biodiversity Strategy for 2030. Bringing nature back
into our lives. COM/2020/380 final.
feedback by all partners and authorities involved in this project. We European Space Agency, 2018. Sen4CAP - Sentinels for Agricultural Monitoring. http://e
thank Dr. Daniel Spengler (German Research Centre for Geosciences) sa-sen4cap.org.
and Dr. Holger Lilienthal (†) (Julius Kühn Institute) for providing us pre- Förster, S., Kaden, K., Förster, M., Itzerott, S., 2012. Crop type mapping using spectral-
temporal profiles and phenological information. Comput. Electr. Agric. 89, 30–40.
processed Sentinel-2 and -1 data products. Dr. Charlotte Tönshof https://doi.org/10.1016/j.compag.2012.07.015.
(Thuenen Institute), Dr. Andrea Ackermann (Thuenen Institute), M.Sc. GADM, 2020. GADM country data (version 3.6). https://gadm.org.
Niklas Keck (RLP AgroScience GmbH) and Dipl.-Geogr. Gregor Sintrup Garioud, A., Valero, S., Giordano, S., Mallet, C., 2020. On the joint exploitation of optical
and SAR satellite imagery for grassland monitoring. Int. Arch. Photogramm. Remote
(RLP AgroScience GmbH) kindly provided the parcel data collected in Sens. Spatial Inf. Sci. https://doi.org/10.5194/isprs-archives-XLIII-B3-2020-591-20
this project. M.Sc. Stenka Vulova kindly gave helpful feedback for the 20.
manuscript and revised our English. The authors would also like to German Research Centre for Geosciences, 2020. GFZ Time Series System for Sentinel-2.
https://gitext.gfz-potsdam.de/gts2.
thank the two anonymous reviewers for their valuable comments and
German Federal Ministry of Food and Agriculture, 2015. Umsetzung der EU-Agrarreform
suggestions during the revision process of this article. in Deutschland. https://www.bmel.de.
Gerstmann, H., Doktor, D., Gläßer, C., Möller, M., 2016. PHASE: A geostatistical model
for the Kriging-based spatial prediction of crop phenology using public phenological
Appendix A. Supplementary material and climatological observations. Comput. Electr. Agric. 127, 726–738. https://doi.
org/10.1016/j.compag.2016.07.032.
Supplementary data to this article can be found online at https://doi. Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., Moore, R., 2017.
Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sens.
org/10.1016/j.compag.2021.106173.
Environ. 202, 18–27. https://doi.org/10.1016/j.rse.2017.06.031.
Griffiths, P., Nendel, C., Pickert, J., Hostert, P., 2019a. Towards national-scale
References characterization of grassland use intensity from integrated Sentinel-2 and Landsat
time series. Remote Sens. Environ. 238 (1), 111124 https://doi.org/10.1016/j.
rse.2019.03.017.
D’Andrimont, R., Taymans, M., Lemoine, G., Ceglar, A., Yordanov, M., van der Velde, M.,
Griffiths, P., Nendel, C., Hostert, P., 2019b. Intra-annual reflectance composites from
2020. Detecting flowering phenology in oil seed rape parcels with Sentinel-1 and -2
Sentinel-2 and Landsat for national-scale crop and land cover mapping. Remote Sens.
time series. Remote Sens. Environ. 239, 111660 https://doi.org/10.1016/j.
Environ. 220, 111124 https://doi.org/10.1016/j.rse.2019.03.017.
rse.2020.111660.
Hargrove, W. (Ed.), 1991. Cover Crops for Clean Water. Soil and Water Conservation
D’Andrimont, R., Lemoine, G., Van der Velde, M., 2018. Targeted Grassland Monitoring
Society, Ankeny, Iowa.
at Parcel Level Using Sentinels, Street-Level Images and Field Observations. Remote
Sens. 10 (8), 1300. https://doi.org/10.3390/rs10081300.

14
C. Schulz et al. Computers and Electronics in Agriculture 186 (2021) 106173

Heupel, K., Spengler, D., Itzerott, S., 2019. A Progressive Crop-Type Classification Using Preidl, S., Lange, M., Doktor, D., 2020. Introducing APiC for regionalised land cover
Multitemporal Remote Sensing Data and Phenological Information. PFG 86, 53–69. mapping on the national scale using Sentinel-2A imagery. Remote Sens. Environ.
https://doi.org/10.1007/s41064-018-0050-7. 240, 111673 https://doi.org/10.1016/j.rse.2020.111673.
Hively, W., Lang, M., McCarty, G., Keppler, J., Sadeghi, A., McConnell, L., 2009. Using RStudio Team, 2020. RStudio: Integrated Development for R. Software Version 3.6. htt
satellite remote sensing to estimate winter cover crop nutrient uptake efficiency. p://www.rstudio.com.
J. Soil Water Conserv. 64 (5), 303–313. https://doi.org/10.2489/jswc.64.5.303. Schulz, C., Keck, N., Kleinschmit, B., 2019. Reduction of On-site Controls of Catch Crop
Hively, W., Duiker, S., McCarty, G., Prabhakara, K., 2015. Remote sensing to monitor Fields with Sentinel-2 and Sentinel-1 Phenological Reference Profiles. In: Conference
cover crop adoption in southeastern Pennsylvania. Soil Water Conserv. 70 (6), Proceedings, MultiTemp Shanghai 2019. https://doi.org/10.1109/Multi-
340–352. https://doi.org/10.2489/jswc.70.6.340. Temp.2019.8866901.
Holtgrave, A.-K., Röder, N., Ackermann, A., Erasmi, S., Kleinschmit, B., 2020. Comparing Sitokonstantinou, V., Papoutsis, I., Kontoes, C., Arnal, A., Andrés, A., Zurbano, J., 2018.
Sentinel-1 and -2 Data and Indices for Agricultural Land Use Monitoring. Remote Scalable Parcel-Based Crop Identification Scheme Using Sentinel-2 Data Time-Series
Sens. 12 (18), 2919. https://doi.org/10.3390/rs12182919. for the Monitoring of the Common Agricultural Policy. Remote Sens. 10 (6), 911.
Immitzer, M., Vuolo, F., Atzberger, C., 2016. First Experience with Sentinel-2 Data for https://doi.org/10.3390/rs10060911.
Crop and Tree Species Classifications in Central Europe. Remote Sens. 8 (3), 166. Solano-Correa, Y., Bovolo, F., Bruzzone, L., Fernandez-Prieto, D., 2020. A Method for the
https://doi.org/10.3390/rs8030166. Analysis of Small Crop Fields in Sentinel-2 Dense Time Series. IEEE Trans. Geosci.
Julius Kühn Institute, 2020. Sentinel-1 time series archive. Level-1 GRD Gamma0 Remote Sens. 58, 2150–2164. https://doi.org/10.1109/TGRS.2019.2953652.
calibrated backscatter products. WCS web map service. https://flf.julius-kuehn. Ssymank, A., 1994. Neue Anforderungen im europäischen Naturschutz: Das
de/webdienste/webdienste-des-flf.html. Schutzgebietssystem Natura 2000 und die FFH-Richtlinie der EU. Natur und
Killick, R., Haynes, K., Eckley, I., Fearnhead, P., Lee, J., 2016. changepoint: Methods for Landschaft 69 (9), 395–406.
Changepoint Detection. R Package Version 2.2.2. https://CRAN.R-project.org/p Tamm, T., Zalite, K., Voormansik, K., Talgre, L., 2016. Relating Sentinel-1
ackage=changepoint. Interferometric Coherence to Mowing Events on Grasslands. Remote Sens. 8 (10),
Kim, Y., Jackson, T., Bindlish, R., Lee, H., Hong, S., 2012. Radar vegetation index for 802. https://doi.org/10.3390/rs8100802.
estimating the vegetation water content of rice and soybean. IEEE Geosci. Remote Taravat, A., Wagner, M., Oppelt, N., 2019. Automatic Grassland Cutting Status Detection
Sens. Lett. 9 (4), 564–568. https://doi.org/10.1109/LGRS.2011.2174772. in the Context of Spatiotemporal Sentinel-1 Imagery Analysis and Artificial Neural
Kolecka, N., Ginzler, C., Pazur, R., Price, B., Verburg, P., 2018. Regional Scale Mapping Networks. Remote Sens. 11 (6), 71. https://doi.org/10.3390/rs11060711.
of Grassland Mowing Frequency with Sentinel-2 Time Series. Remote Sens. 10 (8), Thieme, A., Yadav, S., Oddo, P., Fitz, J., McCartney, S., King, L., Keppler, J., McCarty, G.,
1221. https://doi.org/10.3390/rs10081221. Hively, W., 2020. Using NASA Earth observations and Google Earth Engine to map
Kuenzer, C., Dech, S., Wagner, W., 2015. Remote Sensing Time Series Revealing Land winter cover crop conservation performance in the Chesapeake Bay watershed.
Surface Dynamics: Status Quo and the Pathway Ahead. In: Kuenzer, C., Dech, S., Remote Sens. Environ. 248 https://doi.org/10.1016/j.rse.2020.111943.
Wagner, W. (Eds.), Remote Sensing Time Series – Revealing Land Surface Dynamics. Thorup-Kristensen, K., 1994. The effect of nitrogen catch crop species on the nitrogen
Springer, Cham, pp. 1–24. https://doi.org/10.1007/978-3-319-15967-6_1. nutrition of succeeding crops. Fertilizer Research 37, 227–234. https://doi.org/
Kuhn, M., Wing, J., Weston, S., Williams, A., Keefer, C., Engelhardt, A., Cooper, T., 10.1007/BF00748941.
Mayer, Z., Kenkel, B., R Core Team, Benesty, M., Lescarbeau, R., Ziem, A., Scrucca, Thünen Institute, 2021. Development of an automated support system to monitor
L., Tang, Y., Candan, C., Hunt, T., 2020. caret: Classification and Regression agricultural funding sites (timeStamp). Research project funded by the German
Training. R Package Version 6.0-84. https://CRAN.R-project.org/package=caret. Ministry for Economics and Technology (BMWi). Lead: Institute of Rural Studies,
Kussul, N., Mykola, L., Shelestov, A., Skakun, S., 2018. Crop inventory at regional scale in Thünen Institute. FKZ: 50EW1704. http://timestamp.lup-umwelt.de/.
Ukraine: Developing in season and end of season crop maps with multi-temporal Tukey, J., 1977. Exploratory Data Analysis. Addison-Wesley, Reading, MA. https://doi.
optical and SAR satellite imagery. Eur. J. Remote Sens. 51, 627–636. https://doi. org/10.1007/978-0-387-32833-1_136.
org/10.1080/22797254.2018.1454265. Van Tricht, K., Gobin, A., Gilliams, S., Piccard, I., 2018. Synergistic Use of Radar
Liaw, A., Wiener, M., 2018. randomForest: Breiman and Cutler’s Random Forests for Sentinel-1 and Optical Sentinel-2 Imagery for Crop Mapping: A Case Study for
Classification and Regression. R Package Version 4.6-14. https://CRAN.R-project.or Belgium. Remote Sens. 10 (10), 1642. https://doi.org/10.3390/rs10101642.
g/package=randomForest. Valero, S., Morin, D., Inglada, J., Sepulcre, G., Arias, M., Hagolle, O., Dedieu, G.,
Magdoff, F., Van Es, H., 2009. Building soils for better crops, sustainable soil Bontemps, S., Defourny, P., Koetz, B., 2016. Production of a Dynamic Cropland Mask
management. Sustainable Agriculture Publications. ISBN 978-1-888626-13-1. by Processing Remote Sensing Image Series at High Temporal and Spatial
Meisinger, J., Hargrove, W., Mikkelsen, R., Williams, J., Benson, V., 1991. Effects of Resolutions. Remote Sens. 8 (1), 55. https://doi.org/10.3390/rs8010055.
cover crops on groundwater quality. In: Cover crops for clean water. Proc. Veloso, A., Mermoz, S., Bouvet, A., Le Toan, T., Planells, M., Dejoux, J., Ceschia, E.,
conference, Jackson, 1991, 57–68. 2017. Understanding the temporal behavior of crops using Sentinel-1 and Sentinel-2-
Misra, G., Cawkwell, F., Wingler, A., 2020. Status of Phenological Research Using like data for agricultural applications. Remote Sens. Environ. 199, 415–426. https://
Sentinel-2 Data: A Review. Remote Sens. 12 (17), 2760. https://doi.org/10.3390/ doi.org/10.1016/j.rse.2017.07.015.
rs12172760. Verbesselt, J., Hyndman, R., Newnham, G., Culvenor, D., 2010. Detecting trend and
OSCAR, 2016. Optimising Subsidiary Crop Applications in Rotations (OSCAR): A seasonal changes in satellite image time series. Remote Sens. Environ. 114 (1),
collaborative European research project to develop more sustainable systems of 106–114. https://doi.org/10.1016/j.rse.2009.08.014.
conservation agriculture and increase the diversity of cover crops and living Verger, A., Kandasamy, S., Baret, F., 2016. Temporal Techniques in Remote Sensing of
mulches. Research Project. 2012–2016 https://web5.wzw.tum.de/oscar/. Global Vegetation. In: Ban, Y. (Ed.), Multitemporal Remote Sensing – Methods and
Ottosen, T., Lommen, S., Skjøth, C., 2019. Remote sensing of cropping practice in Applications. Springer, Cham, pp. 217–232. https://doi.org/10.1007/978-3-319-
Northern Italy using time-series from Sentinel-2. Comput. Electron. Agric. 157, 47037-5_11.
232–238. https://doi.org/10.1016/j.compag.2018.12.031. Wilke, B., Snapp, S., 2008. Winter cover crops for local ecosystems: linking plant traits
Peredo, J., Wayman, C., Whong, B., Thieme, A., Kline, L., Yadav, S., Eder, B., Lenske, V., and ecosystem function. J. Sci. Food Agric. 88, 551–557. https://doi.org/10.1002/
Portillo, D., McCartney, S., Fitz, J., Oddo, P., Keppler, J., Hively, D., Bolten, J., jsfa.3149.
McCarty, G., Lyon, A., 2019. Utilizing Landsat and Sentinel-2 to remotely monitor Zambrano-Bigiarini, M., 2020. hydroGOF: Goodness-of-Fit Functions for Comparison of
and evaluate the performance of winter cover crops throughout Maryland. The Simulated and Observed Hydrological Time Series. R Package Version 0.4-0. https
International Archives of the Photogrammetry, Remote Sensing and Spatial ://CRAN.R-project.org/package=hydroGOF.
Information Sciences, Volume XLII-3/W11, 2020. PECORA 21/ISRSE 38 Joint Zeileis, A., Grothendieck, G., Ryan, J., Ulrich, J., Andrews, F., 2020. zoo: S3
Meeting, 6–11 October 2019, Baltimore, Maryland, USA. https://doi.org/10.5194/is Infrastructure for Regular and Irregular Time Series (Z’s Ordered Observations). R
prs-archives-XLII-3-W11-125-2020. Package Version 1.8-8. https://CRAN.R-project.org/package=zoo.
Pipia, L., Muñoz-Marí, J., Amin, E., Belda, S., Camps-Valls, G., Verrelst, J., 2019. Fusing Żuk-Gołaszewska, K., Wanic, M., Orzech, K., 2019. The role of catch crops in field plant
optical and SAR time series for LAI gap filling with multioutput Gaussian processes. production – A review. J. Elem. 24 (2), 575–587. https://doi.org/10.5601/
Remote Sens. Environ. 235, 111452 https://doi.org/10.1016/j.rse.2019.111452. jelem.2018.23.3.1662.

15

You might also like