Computers and Electronics in Agriculture 176 (2020) 105640

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

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

Discovering weather periods and crop properties favorable for coffee rust T
incidence from feature selection approaches
Emmanuel Lassoa, , David Camilo Corralesa,b, Jacques Avelinoc,d,e,f, Elias de Melo Virginio Filhoe,
⁎

Juan Carlos Corralesa

a
University of Cauca, Telematics Engineering Group, Popayán, Colombia
b
INRAe, Université Fédérale de Toulouse, UMR 1248 AGIR, F-31326 Castanet-Tolosan, France
c
Bioagresseurs, Université de Montpellier, CIRAD, Montpellier, France
d
CIRAD, UPR Bioagresseurs, 30501 Turrialba, Costa Rica
e
CATIE, 7170 Turrialba, Costa Rica
f
IICA AP. 55, 2200 Coronado, San José, Costa Rica

ARTICLE INFO ABSTRACT

Keywords: Coffee Leaf Rust (CLR) is a disease that leads to considerable losses in the worldwide coffee industry; as those
Coffea arabica that have been reported recently in Colombia and Central America. The early detection of favorable conditions
Hemileia vastatrix for epidemics could be used to improve decision making for the coffee grower and thus reduce the losses due to
Crop disease the disease. Researchers tried to predict the occurrence of the disease earlier through statistical and machine
Dimensionality reduction
learning models from crop properties, disease indicators and weather conditions. These studies considered the
Machine learning
Model explanation
impact of weather variables in a common period for all. Assuming that the dynamics of weather that most impact
the development of the disease occur in the same time periods is simplistic. We propose an approach to discover
the time period (window) for each weather variables and crop related features that most explain a future ob­
served CLR incidence, in order to obtain a prediction model through machine learning. The selection of the
variables more related with coffee rust incidence and rejection of the features with no significant contribution of
information in machine learning tasks were approached from Feature Selection methods (Filter, Wrapper,
Embedded). In this way, a CLR incidence prediction model based on the features with the greatest impact on the
development of the disease was obtained. Moreover, the use of SHapley Additive exPlanations allowed us to
identify the impact of features in the model prediction. The monitoring of coffee rust incidence is the most
important predictor, since it provides information about current inoculum and this determines how much can
the incidence grow or decrease. Temperature is a determining driver for germination and penetration phases in
days 9 to 6 and 4 to 1 before the date of prediction. Additionally, the amount of rain determines whether
uredospore dispersal or washing conditions occurred. The mean absolute error expected in the model is 6.94% of
incidence, trained with XGBoost algorithm and the dataset reduced by Embedded method. The estimation of the
disease incidence 28 days later can be used to improve decision making in control and nutrition practices.

1. Introduction
Coffee Leaf Rust (CLR) is one of the diseases of coffee plants that
cause more injuries in trees and crop losses (Avelino et al., 2008). The
causal agent is the fungus Hemileia vastatrix Berk. & Broome (1869).
The disease cycle is composed by propagule germination, penetration
through stomata into the leaf, colonization of leaf tissue, sporulation
through stomata and dispersal which comprises propagule release, its
transport and deposition on coffee leaves. The uredospore is the only
known propagule. The latent period, i.e. the time between germination
and sporulation, is a key parameter of the epidemic: the shorter it is, the
more intense the epidemic (Avelino and Rivas, 2013). The first symp­
toms are yellowish spots that appear on the underside of leaves. These
spots then grow and produce uredospores displaying a typical orange
colour. Chlorotic spots can be observed on the upper surface of the
leaves. During the last stage, lesions become necrotic (Avelino et al.,
2008). The disease affects coffee leaves causing defoliation and, in the
worst-case scenario, death of branches and heavy crop losses. For

Corresponding author.
⁎

E-mail addresses: eglasso@unicauca.edu.co (E. Lasso), dcorrales@unicauca.edu.co, davidcamilo.corralesmunoz@inra.fr (D.C. Corrales), jacques.avelino@cirad.fr,
javelino@catie.ac.cr (J. Avelino), eliasdem@catie.ac.cr (E. de Melo Virginio Filho), jcorral@unicauca.edu.co (J.C. Corrales).

https://doi.org/10.1016/j.compag.2020.105640
Received 8 April 2020; Received in revised form 4 June 2020; Accepted 16 July 2020
Available online 05 August 2020
0168-1699/ © 2020 Elsevier B.V. All rights reserved.
E. Lasso, et al.
example, in Colombia after the 2008 epidemic, production decreased by
30% from 2008 to 2011, compared with 2007; while in Central America
the production decreased by 16% after the epidemic of 2012–2013
(Avelino et al., 2008). Reductions in production generate a negative
impact on the livelihood of coffee growers and agricultural workers, as
harvesters. Disease controls imply greater investment, which makes the
farmer’s situation even more precarious. Additionally, the majority of
coffee varieties planted in Latin America are still susceptible to coffee
rust, covering 80% of the area in Central America in 2012 (Avelino
et al., 2008). Among the cultivated species, Coffea arabica is the most
severely attacked (Rivillas et al., 2011). Despite the development of
CLR-resistant coffee varieties, such as in Colombia where more than
60% of its coffee crops are planted with resistant varieties (Avelino
et al., 2008), new rust races have appeared capable of breaking this
resistance (Talhinhas et al., 2017).
One of the most used ways for the CLR assessment is the calculation
of its incidence. A plant unit (normally, a leaf) is categorized according
to whether it presents the symptoms of the disease or not (Madden
et al., 2007). After classifying a representative sample of plant units, the
CLRI (Coffee Leaf Rust Incidence) corresponds to the average propor­
tion of leaves infected over the total analyzed. CLRI is a continuous
variable ranging 0–100. The main limitation of this measurement for
CLRI monitoring is the possible error in the categorization of infected or
healthy leaves. Also, the incidence is a descriptor of the dynamics of
both CLR and coffee plant (Merle et al., 2019). Although the definition
of incidence is uniformly accepted, there are many different ways of
choosing the set of plants and leaves to be examined, and even of de­
termining if a leaf is diseased or not. CLRI is not therefore necessarily
comparable between different trials if the sampling method was dif­
ferent.
Weather, shade level, fruit load and crop management are four of
the principal drivers for the development of CLR development (Avelino
and Rivas, 2013; Rayner, 1961; Avelino et al., 2006). Each phase of CLR
has its own weather requirements and specific durations for these re­
quirements.
Temperature affects propagule germination, penetration, coloniza­
tion and sporulation phases. For germination, the optimum is around 22
° C (Nutman et al., 1963), while daily average temperatures around 28
° C favors sporulation (Merle et al., 2020) and temperatures of 25 ° C
shorten the latent period (Leguizamon, 1985). Temperatures of 22–28
° C that favor germination and lower temperatures (13–16 °C) that favor
the formation of appressoria over the stomata, structures that facilitates
the penetration phase, allow the infection to occur in less than 6 h in
presence of free water (De Jong et al., 1987).
The fungus requires the presence of a layer of water on the under­
side of the leaves to germinate (Nutman et al., 1963; Waller et al.,
2007). Water is also important for dispersal, particularly via splashing,
i.e. the dispersal in raindrops after impacting lesions with uredospores.
However, if the rains are very abundant and intense, the uredospores
can be eliminated by washing (Kushalappa and Eskes, 1989). As CLR is
an obligate parasite, needing living leaves for its survival, any released
uredospore that cannot reach a coffee leaf will not contribute anymore
to the epidemic growth.
Relative humidity is an indirect measurement of leaf wetness. This
condition can be derived from the number of hours with relative hu­
midity of the air above a specific limit, usually 90% or 95% (Sutton
et al., 1984).
The physiological characteristics (particularly in relation with fruit
load) of the coffee tree has an influence on the latent period of the
disease. For susceptible crops with high fruit load and favorable
weather conditions, the latent period can last less than 2 weeks. It is
longer (up to several months) on the oldest leaves of low yielding coffee
plants in cold and dry conditions (Rayner, 1961).
The presence of shade on coffee crops has an effect on the disease
(Boudrot et al., 2016), since it maintains very narrow thermal ampli­
tude values and favors a constant high air relative humidity (Arcila
2
Computers and Electronics in Agriculture 176 (2020) 105640
et al., 2007). It also affects other drivers involved in the CLR cycle, such
as rain, wind, fruiting load and soil moisture (López-Bravo et al., 2012).
The balance of these effects is still controversial.
Crop management (fertilization, diseases controls) drives CLR epi­
demic. However, crop management is limited by the economic capacity
of the coffee grower. The continuous monitoring of the disease allows
the application of fungicides with no excess, at an appropriate time as
soon as CLRI reaches a certain level, usually 5% (Zambolim, 2016),
reducing further CLR intensity and impacts. On the other hand, ferti­
lizer applications contribute to the recovery of coffee tree due to the
action of nutrients on vegetative growth (Villarreyna et al., 2019).
The observed CLRI is also a result of coffee plant growth (Avelino
et al., 2004; Kushalappa et al., 1983; Merle et al., 2019) and previous
CLRI values, as proxies for the estimation of the inoculum stock that
will potentially cause new infections if the right conditions are met
(Kushalappa et al., 1983). The importance of host growth lies in the fact
that, in a growing season, an apparent dilution of CLRI occurs when
new healthy leaves appear, decreasing the proportion of infected leaves
(Ferrandino, 2008; Merle et al., 2019; López-Bravo et al., 2012). This
decrease does not imply that the conditions for the pathogen are not
favorable. Similarly, fall of non-rusted leaves, for diverse reasons, will
increase CLRI (Kushalappa, 1981).
The severe impacts due to CLR epidemics have pointed out the need
for an early warning system that could be used to support decision
making for disease control (Avelino et al., 2008; Lasso et al., 2018). In
order to estimate epidemic growth, some approaches to modeling CLR
have been proposed in the last decades. These models have been both
descriptive and predictive. The first approach was proposed in 1961 by
Rayner (1961) in which an attempt was made to explain the latent
period of CLR using a multiple regression. The regression was applied to
data of mean maximum and minimum temperatures that occured
during this latent period. In later years, a series of approaches by
Kushalappa et al. (1980), Kushalappa (1981), Kushalappa and Ludwig
(1982), Kushalappa et al. (1983), Kushalappa et al. (1984), Kushalappa
and Eskes (1989) through regressions were used to explain CLR latent
period, the CLR infection rate, proportion of leaf area infected and the
CLRI growth rate. The CLR infection rate corresponds to the apparent
growth rate that takes into account the proportion of leaves and leaf
area rusted (Kushalappa and Ludwig, 1982) and depends on the amount
of healthy plant material available for infection. One of the most im­
portant contributions of Kushalappa was the development of a quasi-
mechanistic model that integrated knowledge about the disease pro­
cesses and the drivers that affect it, by the generation of the NSRMP
(Net Survival Ratio for Monocyclic Process) variable. NSRMP quantifies
the favorable events for the development of the epidemic, which oc­
curred in a period of 28 days before the date of prediction (DP)
(Kushalappa et al., 1983) to explain infection rate 14 to 28 days after
DP. The study carried out by Kushalappa (1981) tried to forecast the
CLR infection rate making use of stepwise regression and different
prediction intervals. The variables that most explained the CLR infec­
tion rate were: proportion of leaf area occupied by visible uredospores,
minimum and maximum temperatures, rainfall and new leaves formed;
in a prediction interval of 56 days after DP. Santacreo et al. (1984)
studied favorable conditions for CLRI. They related the same weather
variables as those of the study of Kushalappa (1981), plus the total
leaves present in the coffee tree and previous CLRI.
Recent research works (Meira et al., 2008; Lasso et al., 2015;
Corrales et al., 2016; Lasso et al., 2017; Corrales et al., 2018) have
proposed the use of supervised learning models for CLRI or CLRI growth
rate as a function of weather conditions, crop management and crop
properties (shade, coffee planting density, altitude). The weather vari­
ables ocurring in the 28 days (a single time period or window) prior to
CLRI assessment were considered, following Kushalappa et al. (1983)
proposal. For the prediction of the numerical value of the CLRI, re­
gression-based algorithms were used, while for the CLR infection rates,
expressed as nominal values, the algorithms correspond to classification
E. Lasso, et al.
task. In Corrales et al. (2016) an unsupervised learning algorithm is
used in order to group similar weather conditions and crop properties
into clusters, which are integrated into the training set of a supervised
learning algorithm for CLR modeling. In these approaches, the pre­
paration of the training set for learning tasks follows the considerations
of expert knowledge on CLR.
The use of CLR expert knowledge was used in Meira et al. (2008),
Lasso et al. (2015) to generate composite variables representing fa­
vorable conditions for the disease, i.e. leaf wetness was addressed in a
variable that quantifies the average temperature in hours when the
relative humidity were above 90% or 95%.
Fig. 1 shows the evolution of the methods used for CLR modeling,
grouped by those using simple or multiple linear regression (Regres­
sion), regression based on Machine Learning algorithms (ML Regres­
sion), classification algorithms and other techniques. Since the 2000s,
the rise of the use of machine learning techniques to address disease
modeling can be seen.
However, most of these studies have not explored different weather
windows within the 28 days proposed by Kushalappa. Assuming that all
weather variables impacting the development of the disease occur in
the same time periods is therefore simplistic (Coakley et al., 1988), as
demonstrated by Merle et al. (2020). Statistical methods (and the
modeling process itself) allow to identify the most influential periods of
each weather variable (Bugaud et al., 2016; Leandro-Muñoz et al.,
2017). This approach generates an increase in the number of variables
(called features in data modeling and analysis tasks) included in the
modeling processes.
The objective of this work was to model the CLRI in order to study
particularly the impact of weather variables, characterized in different
weather time periods, called windows. The window approach generated
large dimensional datasets. The most important variables were selected
through Feature Selection (FS) methods. FS allowed to avoid redundant
and noisy data, improving the CLRI modeling. In addition to weather
variables, we incorporated host growth, shade level, and management
type. The monitoring of CLRI was incorporated as a response variable.
Machine learning tasks allowed us to produce a model for CLRI pre­
diction that is best sensitive to these characteristics. Moreover, the
analysis of the contribution that each of the most important variables
had in the model predictions allowed us a more detailed understanding
of the favorable conditions for the disease development.
3
Computers and Electronics in Agriculture 176 (2020) 105640
Fig. 1. Evolution of the methods used for CLR
modeling. The size of each bubble represents
the number of studies for each method.
Regression: Simple linear, Multiple linear.
Classification: Support Vector Machine,
Decision Tree, K-nearest Neighbors, Random
Forest, Bayesian Networks. ML - Regression:
Artificial Neural Networks, Regression Tree.
Other: Probability theory, Math and Statistics.
2. Materials and methods
We proposed an approach to discover the weather windows and
variables that most explain a future observed CLRI. A dataset with the
resulting features was used to obtain a prediction model through ma­
chine learning. The different stages of our approach are presented in
Fig. 2.
Weather monitoring information is broken down into windows of
different duration, and associated with crop property information. A
feature selection process is applied to obtain the best features for the
modeling of the CLRI and discard the irrelevant ones. Next, the re­
sulting datasets are used to train different machine learning algorithms
and obtain their respective model. To establish the best combination
between sets of selected features and machine learning algorithms, the
model with the lowest prediction error is selected. Finally, the highly
correlated variables are cleaned and the impact of the values of the final
features on the CLRI prediction generated by the model is analyzed.
Each subprocess and element is reported below.
2.1. Data source
The data used corresponds to a long term experiment of coffee-
based agroforestry systems established in Costa Rica in 2000, described
by Haggar et al. (2011), studying ecological processes that promote
sustainability and higher coffee productivity, under different crops
conditions. CLRI is one of the studied variables. This trial was carried
out in the Tropical Agricultural Research and Higher Education Center
(CATIE) (Rossi et al., 2011) at coordinates 9° 53′ 44″ North and 83° 38′
07″ West. Detailed information has been continuously collected, which
makes it a unique experiment in the area. The experiment has a total
area of 9.2 ha., located at 685 meter above sea level, in soils with a
franc-clay texture. The variety of coffee is Caturra of the species Coffea
arabica, susceptible to most CLR races. The crop management (fertili­
zation, diseases and pests control) has two strategies: organic and
conventional. Organic management uses chicken manure and organic
matter (coffee pulp) at two intensity levels. Conventional management
has also two levels. The high conventional level uses the complete
technical package for maximizing productivity including pesticides and
herbicides application (copper based fungicide (50% Cu) in 1 kg ha 1
doses combined with a systematic product (cyproconazole 10% WG) in
0.4 liter ha−1 doses), and fertilization (300 kg N ha−1, 20 kg P ha−1,
150 kg K ha−1). The medium conventional level is a less intense level,
using a half-dose of inputs compared to high conventional level (Haggar
E. Lasso, et al.
Fig. 2. Modules to discover the weather windows and features that most explain a future observed CLRI.
et al., 2011). There are 20 treatments configured with different com­
binations of six types of shaded and full sun exposed crops, and the two
management strategies mentioned above. The treatments are replicated
in three blocks. For our purpose, we only considered the two levels of
conventional management and the following shade conditions: dense
shaded crops those that contained the combination Chloroleucon
(Cashá - Ab.i) (C) + Erythrina (Poró) (E), and Full Sun (FS). CLRI as­
sesment was done monthly through the selection of 10 branches in 10
coffee trees, counting the total number of leaves and those with CLR
symptoms.
Daily weather data were obtained from the CATIE Meteorological
Station located in its campus, in Turrialba, Costa Rica, at an altitude of
602 meters above sea level, at coordinates 9° 53′ North Latitude and 83°
38′ West Longitude. The average meteorology in the experiment loca­
tion between 2002 and 2014 is: precipitation 3037 mm/year, air tem­
perature 22 °C, relative humidity 89.6%.
We used the information of weather and CLRI monitoring from April
2002 to December 2014. To avoid redundancy issues, we only used the
data from one of the blocks according to the process carried out: the
data from block 1 for the FS and modeling process, while the data of
block 2 for results validation and model explanations.
2.2. Datasets preprocessing
2.2.1. Weather
From the meteorological station data, we took the following
weather variables: maximum (tMax) and minimum (tMin) air tem­
perature, average (tAvg) air temperature calculated over the day,
thermal amplitude (tAmp) which represents the difference between
maximum and minimum temperatures, average (hAvg) and minimum
(hMin) relative humidity, number of days with precipitation greater or
equal to 1 mm (rDay) and daily precipitation (pre). The quality of the
weather station data was evaluated from the search for outliers, out-of-
range values and verification of null values (lack of value). The initial
scale of weather data was daily. Since the scale of disease monitoring
was monthly, a time series resample was necessary to correspond to the
same time scale as that of disease monitoring.
We only considered meteorological data in a period of 14 days be­
fore the date on which CLRI was measured in each month, which would
be the date of prediction (DP). This period was called the main time
4
Computers and Electronics in Agriculture 176 (2020) 105640
frame (MTF). We did not use a larger period, since CLRI at DP is already
the result of the meteorological conditions that mainly occurred in the
previous month, considering that the latent period is one month long on
average (Waller, 1982). CLRI at DP provides measurement of the in­
oculum stock available for new infections.
In order to generate features in shorter times and identify which of
them are most related with the CLRI to be predicted, i.e. at
DP + 28 days, we analyzed sub-frames within the MTF sequentially,
called windows (Coakley et al., 1988). Each new window begins one day
after the start time of the previous one. If s is the size of the set and i is
the size of the window, the MTF can be divided into s - i + 1 windows.
Fig. 3 shows the windows before DP that we obtained. The date of
predicted incidence (DPI) is 28 days after DP. The CLRI measured in DP
was called current CLRI (cCLRI) while the one in DPI predicted incidence
(pCLRI). The feature index represents the corresponding range of days
(before DP). For example, tMax7-4 represents the maximum tempera­
ture between days 7 and 4 before DP.
The weather data were divided into 4 types of windows, according
to their size (Fig. 3), in the following way:
• 14D: Single window of 14 consecutive days (i = 14); one feature for
each weather variable.
• 7D: 8 windows of 7 consecutive days (i = 7); 8 features for each
weather variable.
• 4D: 11 windows of 4 consecutive days (i = 4); 11 features for each
weather variable.
• 3D: 12 windows of 3 consecutive days (i = 3); 12 features for each
weather variable.
2.2.2. Other variables
Shade condition (shade) and management type (mgmt) were coded
as dummy variables that takes a binary value to indicate shaded crop
(1) or full sun (0), for shade; and conventional high (1) or conventional
medium (0) for mgmt. A variable related to the growth of the host
(coffee tree) (hGrowth), characterized by the difference between the
number of leaves in coffee trees in day 14 and day 1 before DP, was also
codified as a dummy variable; its value is 1 if the number of leaves
increases and 0 otherwise. Additionally, the two numerical variables
related to the disease, cCLRI and pCLRI, were included in the datasets
for each window. The data quality verification was similar to that done
E. Lasso, et al. Computers and Electronics in Agriculture 176 (2020) 105640

Fig. 3. Set of windows for weather variables according to window size. The feature index represents the corresponding range of days before the date of prediction (DP
– green point). For example, for a 4 consecutive days window, tMax7-4 represents the maximum temperature between days 7 and 4 before DP. The date of predicted
incidence (DPI - orange cross) is 28 days after DP. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this
article.)

for weather data. Incidence values above 100% were found and re­ and Anuradha, 2019): search direction, search strategy, stop criteria
moved. In our approach, it is necessary to have the measurement of two and validation. The search direction defines the way the set of features
consecutive months of the disease, thus, months with no data were is reduced to find the best reduced subset. In the case of backward
discarded. searching, it starts from an empty set of features that will be filled re­
cursively in each iteration. Conversely, in the case of backward
searching, each iteration recursively removes a feature from the total
2.3. Feature selection
set of features. The search strategy refers to how the features that are
added or removed in each iteration are chosen. It can be sequential,
In a dataset, the high dimensionality (large amount of features) can
based on performance criteria or randomized. The stop criteria can be
generate problems for data processing since a large number of irrele­
chosen according to a desired number of features, number of process
vant or misleading features do not provide significant information re­
iterations or no limit (evaluation of all possibilities). Finally, the vali­
lated with the target variable in a learning process (Khalid et al., 2014).
dation depends on the learning task that will be carried out (regression
Additionally, a large number of correlated predictors (multicollinearity)
or classification). It is commonly addressed through cross validation.
is usually associated with model overfitting (Magidson et al., 2013). To
Several elements of FS process depend on the characteristics of the
solve this, from computer and data science, the Feature Selection (FS)
dataset used and the learning task for which it will be used. Our dataset
approach was proposed. FS is based on the selection of the best features
had continuous numerical target variable and numerical features (in­
among all the features that are useful for a determined machine
cluding those encoded as dummy variables). Since the target variable
learning task (Khalid et al., 2014). The resulting reduced dataset can be
was numerical, the supervised learning task was regression.
processed more easily (because fewer features are presented and the
Each FS method generates lists of features selected for each window.
instances size is decreased), so the models obtained are more simple
New subsets were generated from these lists. Since the FS process is
and accurate (Chrysostomou, 2009). There are several FS algorithms
done in relation to the target variable (pCLRI), this one was separated
that can be classified into three categories, depending on the process
from the others for the process.
used to achieve their objective (Corrales et al., 2018): Filter, Wrapper
We describe in the following subsections each method and elements
and Embedded methods. These methods have been applied in the
of the FS process according to our needs. We applied each method in
agriculture domain in order to optimize the analysis of the most re­
order to compare the results.
levant features for a specific problem (Bocca and Rodrigues, 2016; Kale
and Sonavane, 2019; Sharif et al., 2018).
The FS methods are supported by other elements such as (Venkatesh

5
E. Lasso, et al.
2.3.1. Filter method: Pearson’s correlation coefficient
A filter method based on Pearson’s correlation coefficient was used,
given its performance in FS related to regression model building
(Rendall et al., 2019). The features were individually correlated to the
target variable by the Pearson’s correlation coefficient from the corre­
lation function available in Pandas for Python (McKinney, 2010). The
threshold to select the features was addressed by the Rule of Thumb
proposed by Krehbiel (2004), which takes into account the sample size
for statistical significance. Features with a correlation coefficient
|r| 2/ n , where n is the number of dataset features, were selected.
2.3.2. Wrapper method
The wrapper methods used were Sequential Feature Selector (SFS)
(Rückstieß et al., 2011) and Recursive Feature Elimination (RFE)
(Guyon et al., 2002). SFS sequentially implements the backward and
forward searching and RFE uses a criterion of importance assigned to
each feature in each iteration to remove the one with less value. SFS
and RFE are available for Python in mlxtend (Raschka, 2018) and Scikit-
learn (Pedregosa et al., 2011) libraries respectively. Features were
evaluated by using a learning algorithm. Cross validation was used to
estimate the accuracy of each subset of features and those that decrease
the performance in the training were removed from the dataset. We
applied SFS and RFE with two different ensemble learning algorithms:
Random Forest (RForest), based on the combination of simple decision
trees, training each tree independently, using a random sample of the
data (Liaw and Wiener, 2002), available in Scikit-learn; and Gradient
Boosting (XGBoost) library for Python, based also on a combination of
decision trees but it builds trees one at a time, where each new tree
helps to correct errors made by previously trained tree (Chen and
Guestrin, 2016).
2.3.3. Embedded method: correlation-based feature subset selection
The embedded methods were created to reduced computation time
(Guyon and Elisseeff, 2003), which is high in wrapper methods. Em­
bedded methods include the feature selection as part of the training
process (Chandrashekar and Sahin, 2014). We used the Feature Selec­
tion component from XGBoost Algorithm, and Least Absolute Shrinkage
and Selection Operator (LASSO) (Tibshirani, 1996) available in Scikit-
learn for Python.
2.4. Selection of the best data subset and learning algorithm
The process for choosing the best reduced subset of those obtained
by the FS methods is described below. We trained four supervised
learning algorithms with subset and compared their performance me­
trics to select the best one, since the chosen subset would be used to
generate a CLRI prediction model. To ensure a better result, the best
configuration for each algorithm was obtained through a pipeline with
a randomized search (available in Scikit-learn). The algorithms used
were: XGBoost, RandomForestRegressor, Suppor Vector Regression (SVR)
and DecisionTreeRegressor, also available in Scikit-learn. The randomized
search allowed to test multiple combinations of the hyper-parameters of
each algorithm. The range of hyper-parameters is described in
Appendix A. Cross-validation was used to measure the closeness of the
prediction to the eventual outcomes for each of the resulting models.
The metric used was mean absolute error (MAE). Thus, we got the best
parameters for each algorithm applied to each data subset and its MAE.
Additionally, we tested the original dataset (not reduced) with the same
algorithms and got the MAE. Finally, we selected the combination of
algorithm and data subset where the lowest MAE was obtained.
2.5. Selection of the best features from correlation analysis
We analyzed the correlations between the selected features in the
best data subset obtained. In the generation of weather windows, there
may be highly correlated variables where one explains similarly to
6
Computers and Electronics in Agriculture 176 (2020) 105640
another. For this, we took the correlations between the features selected
from the Pearson’s coefficient. In the case of finding two variables with
a moderate or high correlation (absolute value > 0.5) (Koo and Li,
2016), we removed the one that had a lower importance score, given by
the algorithm the model was trained with. This process was not done
previously since the importance and relevance of all features within
learning tasks was not known yet. In addition, we built a new dataset
with the resultant features and train a model with it, in order to com­
pare the MAE with the best one obtained in the previous section. The
testing set corresponds to data from block 2 of the CATIE experiment.
2.6. Contribution of selected features in CLRI predictions
SHAP (SHapley Additive exPlanations) was introduced by Lundberg
et al. (2017) as a tool to interpret the predictions made by a machine
learning model. Given a model, a set of test data and a set of features,
SHAP provides an interpretation of the importance of each feature for a
particular prediction. The value of importance is given according to
how each feature contributes to model outcomes. The contribution is
quantified taking as a reference the average model output over the
entire training dataset (base value). The sum of the SHAP values for all
the features is the difference of the prediction to the base value.
We used the SHAP library for Python and SHAPforxgboost library for
R (Lundberg et al., 2017) to get the SHAP values. The explainer used
was the TreeExplainer, since our model corresponded to one of parallel
tree boosting. We used the dependence and summary plots to represent
the impact of the value of each feature on the model output (Lundberg
et al., 2019). This allowed us to understand the contribution of each
feature in light of the scientific knowledge on the disease.
3. Results
3.1. Datasets preprocessing
Four datasets, 14D, 7D, 4D and 3D, were generated with 439 CLRI
measurements (instances) each. The number of features of each dataset
depends on the window: 14D had 13 features (8 related to weather), 7D
had 69 features (64 related to weather), 4D had 93 features (88 related
to weather) and 3D had 101 features (96 related to weather). The non-
weather variables are: shade, mgmt, hGrowth, cCLRI and pCLRI.
3.2. Feature selection
Table 1 shows the number of features defined as relevant and ir­
relevant by each of the feature selection methods and approaches. The
total number of features in each dataset of the windows are 12, 68, 93
and 100 for 14D, 7D, 4D and 3D respectively.
3.3. Selection of the best data subset and learning algorithm
Table 2 shows the best reduced data subset and the learning algo­
rithm that got the lowest MAE for each window. The complete results
are attached as a Supplementary File. The best results were obtained in
the dataset corresponding to window 4D and reduced by Embedded -
XGBoost method (24 features), and using XGBoost as learning algo­
rithm (MAE = 7.19). The features of this reduced dataset were: tMin14-
11, tAvg14-11, rDay14-11, tAvg13-10, tMax13-10, tAvg12-9, tMin11-8,
pre11-8, pre10-7, tMax9-6, hMin9-6, pre9-6, hMin7-4, tAvg6-3, tMax6-3,
hMin6-3, pre6-3, tMax5-2, pre5-2, tMin4-1, hGrowth, cCLRI, shade, mgmt.
3.4. Selection of the best features from correlation analysis
From features correlation and the importance values given by the
XGBoost algorithm, the final set of features considered for the CLRI
modeling were: rDay14-11, pre11-8, tMax9-6, pre6-3, tMin4-1, hGrowth,
cCLRI, shade and mgmt. The resultant MAE 6.94, was 0.25 less with the
E. Lasso, et al. Computers and Electronics in Agriculture 176 (2020) 105640

Table 1 between the features and pCLRI that was not clear in the Fig. 4 (non­
Number of features defined as relevant and irrelevant by feature selection linear relationship). The average of daily precipitation between 11 and
methods and approaches. 8 days before DP (pre11-8), until 15 mm contributes positively to in­
Dataset FS Method Approach Relevant F. Irrelevant F. cidence and above 15 mm, the reverse is seen. The average of minimum
temperatures between 4 and 1 day before DP (tMin4-1) are positively
14D Filter Pearson 2 10 related to the predicted incidence until 19° C. Above this value, the
Embedded LASSO 2 10
relationship is inverted. The number of rainy days between 14 and
XGBoost FS 12 0
Wrapper SFS Rforest 11 1 11 days before DP (rDay14-11) negatively contributes to pCLRI. No
SFS XGBoost 8 4 rainy days in this window tend to favor the incidence which decreases
RFE Rforest 7 5 after every rainy day in the window. Low maximum temperatures be­
RFE XGBoost 1 11
tween 9 and 6 days before DP (tMax9-6) increase the incidence while
7D Filter Pearson 16 52
high values have the opposite effect. The predicted incidence tends to
Embedded LASSO 2 66 increase with higher average of daily precipitation between 6 and
XGBoost FS 24 44 3 days before DP (pre6-3) until 10 mm. Above this value, pCLRI de­
Wrapper SFS Rforest 43 25 creases. For values above 19 mm, this feature has no effect in the
SFS XGBoost 11 57
predicted incidence.
RFE Rforest 49 19
RFE XGBoost 1 67

4D Filter Pearson 28 64 4. Discussion

Embedded LASSO 3 89
XGBoost FS 24 68
Wrapper SFS Rforest 12 80
We analyzed the contribution of weather and agronomic conditions
SFS XGBoost 16 76 for dispersion, germination and penetration CLR phases. The previous
RFE Rforest 61 31 incidence (cCLRI) was the feature that has the most contribution to the
RFE XGBoost 3 89 predicted incidence values, according to SHAP values. It presents a
linear behavior respect to the target variable. That was expected, since
3D Filter Pearson 30 60
Embedded LASSO 3 97 the future incidence depends largely on the current inoculum, which
XGBoost FS 22 78 under favorable conditions is maintained or increased. Conversely, if
Wrapper SFS Rforest 6 94 cCLRI is low, the incidence in DPI would not be expected to grow
SFS XGBoost 8 92 greatly. This shows that, if there is no periodic monitoring of the disease
RFE Rforest 28 72
RFE XGBoost 2 98
in the crops, it is difficult to predict a future incidence as was demon­
strated by Kushalappa et al. (1983) and Merle et al. (2019). Moreover,
the model performance and the quality of the analyzes carried out re­
reduced dataset for second time after correlation analysis than the best spond to the reliability in the field measurement process.
one found in Table 2. The features windows that we identified can be interpreted as the
sequence of weather conditions needed for disease expression in DPI
(pCLRI), from dispersal to colonization phases. Although these events
3.5. Contribution of selected features in CLRI predictions occurred before DP, they will lead to symptoms and signs visible only
after DP, due to the duration of the incubation and latent periods that
The base value of target variable pCLRI in the output of the model normally exceed 14 days (Waller, 1982). Therefore, these variables
applied to the data from block 2 was 26.79% of CLRI, which is the provide information different from that included in the CLR assessment
average of the predictions made by the model from the test set. A in DP.
summary of the SHAP values for the features, where the colors of the From our results, we deduce that, contrary to what is usually con­
points represent the value of each variable (on its own scale) is pre­ sidered, rainfall can reduces CLRI as long as its abundance is sufficient.
sented in Fig. 4. The impact on model output is related to the base value We found, for instance, that four consecutive rainy days from 14 to
and its axis shows how the prediction differs above or below its value. 11 days before DP were detrimental to CLR growth. In addition, the
From summary plot, we deduce that the relationship between pCLRI shape of the relationship between pCLRI and two of the features char­
and cCLRI is directly proportional and corresponds to contributions of acterizing precipitations (pre11-8 and pre6-3), that were retained in our
greater magnitude in the model output. This means that the amount of model, was unimodal. These results indicate that a moderate increase of
expected incidence in DPI is largely explained the previous incidence rainfall is propitious to CLR growth, possibly because free water is
present in DP. The effect of features coded as binary variables is visible necessary for germination and penetration (Nutman et al., 1963; Waller
in the summary plot: the host growth (hGrowth) and the conventional et al., 2007), but excess of rainfall is detrimental possibly due to wash-
high management (mgmt) make the predicted incidence lower. The off of uredospores by rainwater as already shown by Avelino et al.
presence of shade in crops (shade) increases the predicted incidence. (2020) and proposed Merle et al. (2020). Our results provide new
For numeric features, the contribution to model prediction (SHAP evidence of the importance of the wash-off effect to explain CLR epi­
value) according to feature value is shown in detail in Fig. 5. This demics. The peaks we found were 10 mm and 19 mm per day for pre6-3
graphical representation allows checking the type of relationship and pre11-8 respectively, which is in accordance with the 10 mm per

Table 2
Feature Selection (FS) method, learning algorithm related, minimum Mean Absolute Error (MAE) and compared MAE and number of features (No. F.) for original (O)
and reduced (R) dataset obtained from Feature Selection.
Window FS Method L. Algorithm MAE R. MAE O. No. F. R. No. F. O.

4D Embedded – XGBoost XGBoost 7.19 8.24 24 92

7D Embedded – XGBoost XGBoost 7.34 8.28 24 68
3D Wraper – RFE XGBoost 7.38 8.38 28 100
14D Wrapper – SFS XGBoost 7.43 8.5 8 12

7
E. Lasso, et al.
Fig. 4. Summary of SHAP values for the features according their values. The range of values for each feature is represented in a color gradient, where red represents
its highest value and blue the lowest. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
day reported by Merle et al. (2020) and Avelino et al. (2020).
Once the uredospores are deposited, they need water to germinate
and the temperature has a great impact. Excessively high maximum
temperatures (tMax9-6 > 29 °C) disadvantage rust development
(Ribeiro et al., 1978), while lower values around 22 °C generate op­
timal conditions for germination and penetration (Nutman et al., 1963).
The windows of the precipitation in days 6 to 3 before DP (pre6-3) and
minimum temperatures between 4 and 1 day before DP (tMin4-1) share
2 days. Rainfall around < 10 mm, that leave free water on the leaves, in
conjunction with minimum temperatures around 19 and 20 °C generate
conditions for germination and penetration phases (Nutman et al.,
1963). The minimum daily temperatures are normally reached just
before sunrise. High minimum temperatures combined with darkness
are needed for the uredospores to germinate and accomplish infection
(Leguizamon, 1985).
Even though the model was not constructed with weather data after
DP, it generates CLRI predictions with an acceptable error. However,
including weather data after DP could help improve the model. As
demonstrated by Merle et al. (2020), daily rainfall and thermal am­
plitude impact up to 11 days before the symptom appearance on the
coffee leaf.
The host growth (hGrowth equal to 1) generates a decrease in the
predicted CLRI value, where a dilution effect of the disease is verified,
as already reported in Ferrandino (2008), Merle et al. (2019), López-
Bravo et al. (2012). On the contrary, in periods of vegetative decrease
the CLRI values tend to increase due to the absence of dilution effect.
This feature appears to be essential for CLRI predictions. Any model
that would not include the effects of host growth will fail in predict
CLRI. For shaded crops (shade equal to 1), the expected incidence is
higher than for those in full sun, which is an indication of the favorable
microclimatic conditions under shade. Shade has been reported to
buffer temperatures, to increase wetness, favoring germination, infec­
tion and reducing the latent period (López-Bravo et al., 2012), to in­
tercept raindrops reducing uredospore washing (Avelino et al., 2020)
and to promote uredospore dispersal in the air due to the increased
kinetic energy of the raindrops in the understory (Avelino et al., 2020)
that heavily hit the coffee leaves (Boudrot et al., 2016). Similarly, the
8
Computers and Electronics in Agriculture 176 (2020) 105640
observed effect of management (mgmt) was expected. Proper nutrition
contributes negatively to the disease (Avelino et al., 2006) and fungi­
cides application reduces rust area and protects the plant against new
infections (Merle et al., 2019). The mgmt feature can be improved by
having the information if fungicides and fertilizers were applied in the
past month previous to DP.
The SHAP values contribute to the interpretation of the results when
applying a machine learning process. Many of these processes generate
models known as “black box” where their inputs and outputs are
known, but not the process that generates said outputs from the inputs.
SHAP values allowed us to have an idea of how the model generates a
prediction. In addition, the graphical representation facilitated the in­
terpretation of the relationships found between the features and the
target variable in light of the scientific knowledge on the disease. The
interpretation and validation was even better and easier due to the
reduction of features, according to their importance in the modeling
process and after elimination by mutual correlation analysis.
The analysis of favorable conditions for CLR was improved con­
sidering different short consecutive windows compared to a single long
duration period, where short phenomena can go unnoticed. Although,
statistically, in the shorter window the modeling task would have more
“options” for the generation of the functions in the resulting model, the
4-day window was better than the 3-day one. If the window is too short,
there is no biological response related to disease phases.
In the application of the FS methods, the Wrapper RFE method is the
one that selects the lowest proportion of relevant features. In the
Wrapper RFE method, since cCLRI has a much greater importance than
other features, the reduction of the sizes of the sets of features in each
iteration lead only to consider that feature.
5. Conclusions
We identified the favorable conditions of rain and temperature that
lead to dispersal, germination and penetration CLR phases.
Additionally, we trained a machine learning model able to estimate the
disease incidence 28 days later. The combination of the analysis of
weather variables characterized in windows of short duration, CLRI
E. Lasso, et al. Computers and Electronics in Agriculture 176 (2020) 105640

Fig. 5. Dependence plots for numeric features relating the contribution to model prediction (SHAP value) according to feature value. The red curve shows the smooth
tendency and the histograms over the axis, the values distributions. (For interpretation of the references to colour in this figure legend, the reader is referred to the
web version of this article.)

monitoring and crop properties, with feature selection methods, ma­
chine learning and a model explanation technique, allowed us to
achieve it. All the process was made with real data from a field ex­
periment. We are aware that the model performance may be over es­
timated since the used dataset corresponds only to one location. To
improve the generalization of the model, the application of the same
approach in other regions and countries is necessary. However, the
results of our study are a promising advance for CLR modeling. This
model can be used, previous validation, to identify CLR epidemics in
their early phases and facilitate institutions, organizations and in­
dividual coffee producers decision-making for timely disease control, to
avoid or reduce coffee losses.
CRediT authorship contribution statement
Emmanuel Lasso: Conceptualization, Methodology, Software,

9
E. Lasso, et al. Computers and Electronics in Agriculture 176 (2020) 105640

Investigation, Writing - original draft, Visualization. David Camilo Acknowledgements

Corrales: Conceptualization, Methodology. Jacques Avelino:
Conceptualization, Validation, Writing - review & editing. Elias de
Melo Virginio Filho: Resources. Juan Carlos Corrales: Writing - re­
view & editing, Supervision.
Declaration of Competing Interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influ­
ence the work reported in this paper.
The authors are grateful for the technical support of Telematics
Engineering Group (GIT) of the University of Cauca, the Tropical
Agricultural Research and Higher Education Center (CATIE) and the
InnovAccion Cauca project of the Colombian Science, Technology and
Innovation Fund (SGR- CTI) for PhD scholarship granted to MSc.
Emmanuel Lasso. Additionally, we would like to thank the project
“Internacionalización de Proyectos UEES del Grupo de Ingeniería
Telemática - ID 5271”, financed by “Red de formación de talento hu­
mano para la innovación social y productiva en el departamento del
Cauca InnovAcción Cauca”.

Appendix A. Tuning of learning algorithms parameters

See Table A.3

Table A.3
Tuning of learning algorithms hyper-parameters.
Algorithm Hyper-parameter Range

XGBoost Maximum depth of a tree 3–15 in steps of 1

Fraction of observations to be randomly samples for each tree 0.05–1 in steps of 0.05
Fraction of columns to be randomly samples for each tree 0.1–1 in steps of 0.05
Learning rate 0.001, 0.01, 0.1, 0.5, 0.9

RandomForestRegressor Bootstrap True, False

Maximum number of levels in each decision tree 10–100 in steps of 10
Maximum number of features Auto, Square root
Number of trees in the forest 100–1000 in steps of 100

SupportVectorRegression Kernel type to be used in the algorithm rbf, poly, sigmoid

Kernel coefficient Scale, Auto
Regularization parameter C 0.1,1, 10, 100

DecisionTreeRegressor Maximum depth of the tree 3–15 in steps of 1

Minimum number of samples required to split an internal node 2–12 in steps of 1
Function to measure the quality of a split MSE, MAE

Appendix B. Supplementary material

Supplementary data associated with this article can be found, in the online version, at https://doi.org/10.1016/j.compag.2020.105640.

References Electrical Eng. 40, 16–28. https://doi.org/10.1016/j.compeleceng.2013.11.024..

Arcila P, J., Farfán V, F., Moreno B, A., Salazar G, L., Hincapié G, E., 2009. Sistemas de
producción de café en colombia. chinchiná. Cenicafé 2007, 309.
Avelino, J., Rivas, G., 2013. La roya anaranjada del cafeto, 47. https://hal.archives-
ouvertes.fr/hal-01071036.
Avelino, J., Willocquet, L., Savary, S., 2004. Effects of crop management patterns on
coffee rust epidemics. Plant Pathol. 53, 541–547.
Avelino, J., Zelaya, H., Merlo, A., Pineda, A., Ordoñez, M., Savary, S., 2006. The intensity
of a coffee rust epidemic is dependent on production situations. Ecol. Modell. 197,
431–447.
Avelino, J., Cristancho, M., Georgiou, S., Imbach, P., Aguilar, L., Bornemann, G.,
Läderach, P., Anzueto, F., Hruska, A.J., Morales, C., 2015. The coffee rust crises in
colombia and central america (2008–2013): impacts, plausible causes and proposed
solutions. Food Security 7, 303–321. https://doi.org/10.1007/s12571-015-0446-9.
Avelino, J., Vilchez, S., Segura-Escobar, M.B., Brenes-Loaiza, M.A., Filho, E. d. M.
Virginio, Casanoves, F., 2020. Shade tree chloroleucon eurycyclum promotes coffee
leaf rust by reducing uredospore wash-off by rain. Crop Protect. 129, 105038.
Bocca, F.F., Rodrigues, L.H.A., 2016. The effect of tuning, feature engineering, and fea­
ture selection in data mining applied to rainfed sugarcane yield modelling. Comput.
Electron. Agric. 128, 67–76. https://doi.org/10.1016/j.compag.2016.08.015..
http://www.sciencedirect.com/science/article/pii/S0168169916306391.
Boudrot, A., Pico, J., Merle, I., Granados, E., Vilchez, S., Tixier, P., Filho, E. d. M.V.,
Casanoves, F., Tapia, A., Allinne, C., et al., 2016. Shade effects on the dispersal of
airborne hemileia vastatrix uredospores. Phytopathology 106, 572–580.
Bugaud, C., Joannès-Dumec, C., Louisor, J., Tixier, P., Salmon, F., 2016. Preharvest
temperature affects chilling injury in dessert bananas during storage. J. Sci. Food
Agric. 96, 2384–2390.
Chandrashekar, G., Sahin, F., 2014. A survey on feature selection methods. Comput.
40th-year commemorative issue. http://www.sciencedirect.com/science/article/pii/
S0045790613003066.
Chen, T., Guestrin, C., 2016. XGBoost: A scalable tree boosting system. In: Proceedings of
the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, KDD ’16, ACM, New York, NY, USA. pp. 785–794. doi:10.1145/2939672.
2939785.
Chrysostomou, K., 2009. Wrapper feature selection. In: Encyclopedia of Data
Warehousing and Mining, Second Edition, IGI Global, pp. 2103–2108.
Coakley, S.M., Line, R.F., McDaniel, L.R., 1988. Predicting stripe rust severity on winter
wheat using an improved method for analyzing meteorological and rust data.
Phytopathology 78, 543–550.
Corrales, D.C., Figueroa-Casas, A., Ledezma, A., Corrales, J.C., 2016. Two-level classifier
ensembles for coffee rust estimation in Colombian crops. Int. J. Agric. Environ. Inf.
Syst. (IJAEIS) 7, 41–59.
Corrales, D.C., Lasso, E., Casas, A.F., Ledezma, A., Corrales, J.C., 2018. Estimation of
coffee rust infection and growth through two-level classifier ensembles based on
expert knowledge. Int. J. Bus. Intell. Data Min. 13, 369–387.
Corrales, D.C., Lasso, E., Ledezma, A., Corrales, J.C., 2018. Feature selection for classi­
fication tasks: expert knowledge or traditional methods? J. Intell. Fuzzy Syst. 1–11.
De Jong, E., Eskes, A., Hoogstraten, J., Zadoks, J., 1987. Temperature requirements for
germination, germ tube growth and appressorium formation of urediospores of
hemileia vastatrix. Neth. J. Plant Pathol. 93, 61–71.
Ferrandino, F., 2008. Effect of crop growth and canopy filtration on the dynamics of plant
disease epidemics spread by aerially dispersed spores. Phytopathology 98, 492–503.
Guyon, I., Elisseeff, A., 2003. An introduction to variable and feature selection. Introduct.
Variable Feature Sel. 3, 1157–1182. Cited By 5.
Guyon, I., Weston, J., Barnhill, S., Vapnik, V., 2002. Gene selection for cancer classifi­
cation using support vector machines. Mach. Learn. 46, 389–422. https://doi.org/10.
1023/A:1012487302797.

10
E. Lasso, et al.
Haggar, J., Barrios, M., Bolaños, M., Merlo, M., Moraga, P., Munguia, R., Ponce, A.,
Romero, S., Soto, G., Staver, C., Virginio, E. de M.F., 2011. Coffee agroecosystem
performance under full sun, shade, conventional and organic management regimes in
central america. Agrofor. Syst. 82, 285–301. https://doi.org/10.1007/s10457-011-
9392-5.
Kale, A.P., Sonavane, S.P., 2019. Iot based smart farming: feature subset selection for
optimized high-dimensional data using improved ga based approach for elm.
Comput. Electron. Agric. 161, 225–232. https://doi.org/10.1016/j.compag.2018.04.
027.. bigData and DSS in Agriculture, http://www.sciencedirect.com/science/
article/pii/S0168169917309389.
Khalid, S., Khalil, T., Nasreen, S., 2014. A survey of feature selection and feature ex­
traction techniques in machine learning. In: 2014 Science and Information
Conference, pp. 372–378. https://doi.org/10.1109/SAI.2014.6918213.
Koo, T.K., Li, M.Y., 2016. A guideline of selecting and reporting intraclass correlation
coefficients for reliability research. J. Chiropractic Med. 15, 155–163.
Krehbiel, T.C., 2004. Correlation coefficient rule of thumb. Decision Sci. J. Innov. Educ. 2,
97–100.
Kushalappa, A.C., 1981. Linear models applied to variation in the rate of coffee rust
development. J. Phytopathol. 101, 22–30. https://doi.org/10.1111/j.1439-0434.
1981.tb03317.x.. https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1439-0434.
1981.tb03317.x.
Kushalappa, A., Chaves, G., et al., 1980. An analysis of the development of coffee rust in
the field. Fitopatologia Brasileira 5, 95–103.
Kushalappa, A.C., Eskes, A.B., 1989. Advances in coffee rust research. Annu. Rev.
Phytopathol. 27, 503–531.
Kushalappa, A.C., Ludwig, A., 1982. Calculation of apparent infection rate in plant dis­
eases: development. Phytopathology 72, 1373–1377.
Kushalappa, A., Akutsu, M., Ludwig, A., 1983. Application of survival ratio for mono­
cyclic process of hemileia vastatrix in predicting coffee rust infection rates.
Phytopathology 73, 96–103.
Kushalappa, A., Akutsu, M., Oseguera, S., Chaves, G., Melles, C., Miranda, J., Bartolo, G.,
1984. Equations for predicting the rate of coffee rust development based on net
survival ratio for monocyclic process of hemileia vastatrix [coffea arabica].
Fitopatologia Brasileira (Brazil).
Lasso, E., Thamada, T.T., Meira, C.A.A., Corrales, J.C., 2015. Graph Patterns as
Representation of Rules Extracted from Decision Trees for Coffee Rust Detection.
Springer International Publishing, Cham, pp. 405–414. doi:10.1007/978-3-319-
24129-6_35.
Lasso, E., Thamada, T.T., Meira, C.A.A., Corrales, J.C., 2017. Expert system for coffee rust
detection based on supervised learning and graph pattern matching. Int. J. Metadata
Semant. Ontol. 12, 19–27.
Lasso, E., Valencia, O., Corrales, D.C., López, I.D., Figueroa, A., Corrales, J.C., 2018. A
cloud-based platform for decision making support in colombian agriculture: A study
case in coffee rust. In: Angelov, P., Iglesias, J.A., Corrales, J.C. (Eds.), Advances in
Information and Communication Technologies for Adapting Agriculture to Climate
Change. Springer International Publishing, Cham, pp. 182–196.
Leandro-Muñoz, M.E., Tixier, P., Germon, A., Rakotobe, V., Phillips-Mora, W., Maximova,
S., Avelino, J., 2017. Effects of microclimatic variables on the symptoms and signs
onset of moniliophthora roreri, causal agent of moniliophthora pod rot in cacao. PloS
One 12, e0184638.
Leguizamon, C., et al., 1985. Contribution à la connaissance de la résistance incomplète
du caféier arabica (coffea arabica l.) à la rouille orangée (hemileia vastatrix berk. et
br.), Th se de Doctorat. ENSA, Montpellier, France.
Liaw, A., Wiener, M., 2002. Classification and regression by randomforest. R News 2,
18–22. https://CRAN.R-project.org/doc/Rnews/.
López-Bravo, D.F., Virginio-Filho, E.d.M., Avelino, J., 2012. Shade is conducive to coffee
rust as compared to full sun exposure under standardized fruit load conditions. Crop
Protect. 38, 21–29.
Lundberg, S.M., Lee, S.-I., 2017. A unified approach to interpreting model predictions. In:
Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S.,
Garnett, R. (Eds.), Advances in Neural Information Processing Systems 30. Curran
Associates Inc, pp. 4765–4774. http://papers.nips.cc/paper/7062-a-unified-
approach-to-interpreting-model-predictions.pdf.
Lundberg, S.M., Erion, G., Chen, H., DeGrave, A., Prutkin, J.M., Nair, B., Katz, R.,
Himmelfarb, J., Bansal, N., Lee, S.-I., 2019. Explainable ai for trees: From local ex­
planations to global understanding, arXiv preprint arXiv:1905.04610.
Madden, L.V., Hughes, G., van den Bosch, F., 2007. The study of plant disease epidemics.
Am. Phytopathol. Soc. 0. doi:10.1094/9780890545058. https://apsjournals.apsnet.
org/doi/abs/10.1094/9780890545058.
Magidson, J., 2013. Correlated component regression: re-thinking regression in the pre­
sence of near collinearity. In: Abdi, H., Chin, W.W., Esposito Vinzi, V., Russolillo, G.,
11
Computers and Electronics in Agriculture 176 (2020) 105640
Trinchera, L. (Eds.), New Perspectives in Partial Least Squares and Related Methods.
Springer New York, New York, NY, pp. 65–78.
McKinney, W., 2010. Data structures for statistical computing in python. In: van der Walt,
S., Millman, J. (Eds.), Proceedings of the 9th Python in Science Conference, pp.
51–56.
Meira, C.A., Rodrigues, L.H., Moraes, S.A., 2008. Análise da epidemia da ferrugem do
cafeeiro com árvore de decisão. Trop. Plant Pathol. 33, 114–124.
Merle, I., Pico, J., Granados, E., Boudrot, A., Tixier, P., Filho, E.d.M. Virginio, Cilas, C.,
Avelino, J., 2019. Unraveling the complexity of coffee leaf rust behavior and devel­
opment in different coffea arabica agro-ecosystems. Phytopathology. doi:10.1094/
PHYTO-03-19-0094-R. arXiv:https://doi.org/10.1094/PHYTO-03-19-0094-R, pMID:
31502519. doi: 10.1094/PHYTO-03-19-0094-R.
Merle, I., Tixier, P., de Melo Virginio, E., Filho, C., Cilas, J. Avelino, 2020. Forecast
models of coffee leaf rust symptoms and signs based on identified microclimatic
combinations in coffee-based agroforestry systems in costa rica. Crop Protect. 130,
105046. https://doi.org/10.1016/j.cropro.2019.105046.. http://www.sciencedirect.
com/science/article/pii/S0261219419303928.
Nutman, F., Roberts, F., Clarke, R., 1963. Studies on the biology of hemileia vastatrix
berk. & br. Trans. Br. Mycolo. Soc. 46, 27–44.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M.,
Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D.,
Brucher, M., Perrot, M., Duchesnay, E., 2011. Scikit-learn: Machine learning in
Python. J. Mach. Learn. Res. 12, 2825–2830.
Raschka, S., 2018. Mlxtend: Providing machine learning and data science utilities and
extensions to python’s scientific computing stack. J. Open Source Softw. 3https://doi.
org/10.21105/joss.00638.. http://joss.theoj.org/papers/10.21105/joss.00638.
Rayner, R., 1961. Germination and penetration studies on coffee rust (hemileia vastatrix
b. & br). Ann. Appl. Biol. 49, 497–505.
Rendall, R., Castillo, I., Schmidt, A., Chin, S.-T., Chiang, L.H., Reis, M., 2019. Wide
spectrum feature selection (wise) for regression model building. Comput. Chem. Eng.
121, 99–110.
Ribeiro, I.J.A., Monaco, L.C., Tisseli Filho, O., Sugimori, M.H., 1978. Efeito de alta
temperatura no desenvolvimento de hemileia vastatrix em cafeeiro suscetível.
Rivillas, C., Serna, C., Cristancho, M., Gaitan, A., 2011. La Roya del Cafeto en Colombia.
Impacto, manejo y costos de control, Cientific bot036, Cenicafe.
Rossi, E., Montagnini, F., de Melo Virginio Filho, E., 2011. Effects of management prac­
tices on coffee productivity and herbaceous species diversity in agroforestry systems
in costa rica. In: Agroforestry as a tool for landscape restoration. Nova Science
Publishers, New York, pp. 115–132.
Rückstieß, T., Osendorfer, C., van der Smagt, P., 2011. Sequential feature selection for
classification. In: Australasian Joint Conference on Artificial Intelligence. Springer,
pp. 132–141.
Santacreo, R., Reyes, P., Oseguera, S., et al., 1984. Estudio del desarrollo de la roya del
cafeto hemileia vastatrix berk & br. y su relación con factores biológicos y climáticos
en condiciones de campo de dos zonas cafetaleras de honduras, ca, in: 6. Simposio
Latinoamericano sobre Caficultura24-25 Nov 1983Panamá (Panamá), IICA ICCR-
340, IICA, San José (Costa Rica). PROMECAFE.
Sharif, M., Khan, M.A., Iqbal, Z., Azam, M.F., Lali, M.I.U., Javed, M.Y., 2018. Detection
and classification of citrus diseases in agriculture based on optimized weighted seg­
mentation and feature selection. Comput. Electron. Agric. 150, 220–234. https://doi.
org/10.1016/j.compag.2018.04.023.. http://www.sciencedirect.com/science/
article/pii/S0168169917306373.
Sutton, J., Gillespie, T., Hildebrand, P., 1984. Monitoring weather factors in relation to
plant disease [crop microclimate, electrical sensors, temperature and wetness gauges,
sources of error]. Plant Diseases.
Talhinhas, P., Batista, D., Diniz, I., Vieira, A., Silva, D.N., Loureiro, A., Tavares, S.,
Pereira, A.P., Azinheira, H.G., Guerra-Guimarães, L., et al., 2017. The coffee leaf rust
pathogen hemileia vastatrix: one and a half centuries around the tropics. Molecul.
Plant Pathol. 18, 1039–1051.
Tibshirani, R., 1996. Regression shrinkage and selection via the lasso. J. Roy. Stat. Soc.:
Ser. B (Methodol.) 58, 267–288.
Venkatesh, B., Anuradha, J., 2019. A review of feature selection and its methods.
Cybernet. Inf. Technol. 19, 3–26.
Villarreyna, R., Barrios, M., Vílchez, S., Cerda, R., Vignola, R., Avelino, J., 2019.
Economic constraints as drivers of coffee rust epidemics in nicaragua. Crop Protect
104980.
Waller, J., 1982. Coffee rust—epidemiology and control. Crop Protect. 1, 385–404.
Waller, J.M., Bigger, M., Hillocks, R.J., 2007. Coffee Pests, Diseases and their
Management. CABI.
Zambolim, L., 2016. Current status and management of coffee leaf rust in brazil. Trop.
Plant Pathol. 41, 1–8.