International Journal of Biometeorology

https://doi.org/10.1007/s00484-018-1587-2

ORIGINAL PAPER

Weather-based coffee leaf rust apparent infection rate modeling

Fernando Dill Hinnah 1 & Paulo Cesar Sentelhas 1 & Carlos Alberto Alves Meira 2 & Rodrigo Naves Paiva 3

Received: 3 February 2018 / Revised: 11 July 2018 / Accepted: 14 July 2018

# ISB 2018

Abstract
Brazil is the major coffee producer in the world, with 2 million hectares cropped, with 75% of this area with Coffea arabica and
25% with Coffea canephora. Coffee leaf rust (CLR) is one of the main diseases that cause yield losses by reducing healthy leaf
area. As CLR is highly influenced by weather conditions, this study aimed to determine the best linearization model to estimate
the CLR apparent infection rate, to correlate CLR infection rates with weather variables, and to develop and assess the perfor-
mance of weather-based infection rate models to be used as a disease warning system. The CLR epidemic was analyzed for 88
site-seasons, while progress curves were assessed by linear, monomolecular, logistic, Gompertz, and exponential linearization
models for apparent infection rate determination. Correlations between CLR infection rates and weather variables were conduct-
ed at different periods. From these correlations, multiple linear regressions were developed to estimate CLR infection rates, using
the most weather-correlated variables. The Gompertz growth model had the best fit with CLR progress curves. Minimum
temperature and relative humidity were the weather variables most correlated to infection rate and, therefore, chosen to compose
a CLR forecast system. Among the models developed, the one for the condition of high coffee yield at a narrow row spacing was
the best, with only 9.4% of false negative occurrences for all the months assessed.

Keywords Hemileia vastatrix . Linearization . Gompertz . Estimation models

Introduction Currently, the average coffee yield in the Brazilian fields is

Global coffee production in 2017/2018 season was about 160
million (mi) bags (60 kg) (ICO 2017). Developing countries
are the major producers, led by Brazil, with 51 mi bags in the
2016 crop season, which is followed by Vietnam, with 30 mi
bags, Colombia, with 14 mi bags, and Indonesia, with 12 mi
bags (ICO 2017). These countries have great part of their
economies associated to the coffee cultivation; once this crop
is the second most traded commodity worldwide (Jayakumar
et al. 2016). In Brazil, coffee is one of the most important
crops, being cultivated in more than 2 million hectares. It is
considered a strategic commodity for the country (Souza et al.
2011), since Brazil is the world’s main coffee producer.
* Fernando Dill Hinnah
fhinnah@usp.br
1
ESALQ, Piracicaba, São Paulo, Brazil
2
EMBRAPA Agricultural Informatics, Campinas, Brazil
3
Procafé Foundation, Varginha, Brazil
approximately 24 bags of 60 kg ha−1 (IBGE 2017).
Coffee crop plantations in Brazil differ from those cultivated
in most of other producing countries, with coffee plants being
cultivated as monoculture under full sun condition and with a
large row spacing (≈ 3.5 m), allowing mechanized operations. In
Central America and Colombia, coffee is usually cultivated un-
der shaded agroforestry systems, which makes mechanization
very difficult (Avelino et al. 2015). While in Central America
coffee fields are mostly in mountainous areas, in Brazil, coffee
crops are cultivated in large flatter areas, where mechanization
for phytosanitary treatments and harvesting is possible. Despite
the differences in the crop systems used, diseases and pests
affect coffee fields, compromising yield in all the producing
regions of Brazil and around the world (McCook 2006).
Coffee plants have as characteristic a biennial phenological
cycle divided into six phases, with bi-annual production alterna-
tion. As described by Camargo and Camargo (2001), coffee
plants in Brazil start vegetative growth and vegetative buds for-
mation in September of the first year till March of the second
year, which is the first crop phase. The second phase is also
vegetative, from April to August of year 2, when vegetative
buds of the knots are differentiated into reproductive buds.

The third phase is the flowering and grain expansion, from
September to December of year 2 (at this time, there is an over-
lap of the first and third phases). The fourth phase is the grain
formation and filling, from January to March of year 3, when,
after that, starts the grain maturation (fifth phase), and then har-
vesting, from May to July. The sixth phase is the senescence
phase, which lasts from July to August of the third year.
Among the several factors that threaten coffee plantations,
diseases are the main ones. Coffee leaf rust (CLR) caused by
Hemileia vastatrix Berk. & Br. is considered the major disease
worldwide (Avelino et al. 2015; Camargo and Pereira 1994),
since weather conditions in the production regions present
excellent conditions of air temperature and humidity for H.
vastatrix germination and invasion (Nutman et al. 1963). In
Brazil, yield losses caused by CLR can reach as high as 35%
(Zambolim et al. 2005). The most common strategy for CLR
control is the use of fungicide sprays, with interval between
them based on the residual period of each specific chemical
product and according to regional previous knowledge. This
strategy, also known as calendar-based system, is normally
less efficient than weather-based disease warning systems,
mainly because fungicide over- or underapplication is more
likely to occur. Planning sprays based on weather conditions
for different disease is a suitable strategy due to sprays only at
the right time will result in lower sprays number or better
control, increasing the efficiency. It is expected due to the
dependence of Hemileia vastatrix germination and invasion
from humidity and temperature (Nutman et al. 1963) to im-
prove control efficiency. However, this practice is not com-
monly used against CLR, and no technical criteria are taken
under consideration for fungicides sprays (Souza et al. 2011).
The overlap of two phenological cycles makes second
and fifth coffee phases happen at the same time, resulting
in a high competition for carbohydrates between vegeta-
tive and reproductive structures. This competition leads
to a variation of coffee plants susceptibility to CLR
(DaMatta et al. 2007). During the years with high yields,
most of the carbohydrates produced by photosynthesis
are used for grain filling, rather than for plant protection,
which makes CLR incidence higher, achieving up to
70%, whereas in the years of low yields, disease inci-
dence is lower, about 30%.
Considering the pressure of the coffee companies by a
more rational CLR chemical control around the world, the
adoption of weather-based warning system will increase
(Zambolim 2016), being necessary to develop studies to sup-
port that. However, there are several barriers for the adoption
of this system by growers, which are related to the difficulty to
interact with the systems, the constant change of field disease
levels, a period to adapt to the system, and the lack of confi-
dence on the models (Magarey and Isard 2017). Even consid-
ering these barriers, there is a consensus that weather-based
warning systems, associated with internet of things (IoT), will
Int J Biometeorol
be expanding in the next years in order to make disease control
a next step of precision agriculture.
Based on the abovementioned, the hypothesis of this study
is that a well-adapted CLR infection rate model, simple to
understand, easy to use, accurate, and precise, can be a useful
tool to rationalize CLR chemical control in Brazil. For testing
it, this study aimed to assess and adjust CLR progress curves
to determine the best linearization model to estimate disease
apparent infection rate for different plant populations and fruit
loads (yields) in the main coffee-producing regions of Brazil;
to correlate CLR infection rates with weather variables, con-
sidering different latent periods for each condition; to develop
weather-based CLR infection rate models, considering the
most correlated weather variables for each condition; and to
test these models with independent data to evaluate their suit-
ability for being used as a disease warning system.
The remainder of this paper is organized as follows:
BMaterial and methods^ section defines the study areas, dis-
ease and weather data collection, linearization equations ap-
plied to calculate disease infection rates removing already
infected tissues effect, and how and which weather variables
were used to CLR progress rates correlation. BResults^ section
presents CLR intensity comparison due to high and low fruit
loads, analysis of linearization models applied to CLR
pathosystem, developed infection rate estimation models,
and their performance with statistical assessment and indepen-
dent data. BDiscussion^ section comprises considerations
about Brazilian coffee fields management, the linearization
models, how the weather variables are related among them,
the period of antecedence of the infection rate estimation
models, and how it can allow decisions based on a forecast
system. Finally, BConclusions^ section merges the main con-
siderations of the study.
Material and methods
Disease assessments
The CLR incidence data used in this study were obtained from
experiments conducted from 1998 to 2017 in Varginha (21°
34′ 00″ S, 45° 24′ 22″ W, 940 m a.s.l.), from 2006 to 2017 in
Carmo de Minas (22° 10′ 31″ S, 45° 09′ 03″ W, 1080 m a.s.l.),
and from 2010 to 2017 in Boa Esperança (21° 03′ 59″ S, 45°
34′ 37″ W, 830 m a.s.l.), all in southern Minas Gerais, a tra-
ditional coffee growing region, with more than 600,000 ha
cultivated with Arabica coffee.
CLR incidence on leaves was monthly assessed on a non-
fungicide-treated plot. CLR incidence scores were determined
on leaves from the third or fourth knot of branches at the
middle third of the plants, in a non-destructive procedure.
The plants were randomly chosen into a zigzag walking pat-
tern in the area, as recommended by (Chalfoun 1997).
Int J Biometeorol
CLR assessments were divided in different conditions, tak-
ing into account plant population (represented by row spac-
ing) and yield (hereafter named as fruit load). The plant pop-
ulation was 4000 or 8000 plants ha−1, which represents plants
spacing of 3.5 m × 0.7 m and 2.5 m × 0.5 m, respectively. The
higher distance between the rows (3.5 m) was named Bwide,^
whereas the lower distance (2.5 m) was named Bnarrow.^ In
Carmo de Minas, only narrow row spacing was available,
while in Boa Esperança, experiments were conducted only
under wide spacing. In Varginha, the experiments presented
all conditions. In each field, due to coffee biennial phenology,
two fruit loads were possible: high and low. BHigh^ represents
more than 30 bags of 60 kg ha−1, while Blow^ is lower than 10
bags of 60 kg ha−1. The disconnected range between this
classification is due to the difference observed in the field on
subsequently high and low seasons. Two cultivars, Mundo
Novo and Catuaí, both susceptible to CLR and the most com-
mon genotypes used in Brazil (Conceição et al. 2005), were
assessed in the present study.
Considering the combination of two plant populations and
two fruit loads, four conditions were evaluated: high fruit load,
wide space (HW); high fruit load, narrow space (HN); low
fruit load, wide space (LW); and low fruit load, narrow space
(LN), which are normally observed in the Brazilian coffee
plantations. The fruit load was assessed since it is an important
factor for CLR incidence—the higher the fruit load, the great-
er is the disease incidence, since under these conditions the
plant has its defense system weakened (Kushalappa et al.
1986; Meira et al. 2009).
Disease progress curves
CLR incidence and the time of each assessment were used to
build the disease progress curves (CLRPC), which allow to
describe the disease evolution over time (Nutter 1997 and
Bergamin Filho 2011). Eighty-eight site-seasons CLRPC
were built, comprising 16 seasons in Varginha (times four
field conditions), eight in Carmo de Minas (times two field
conditions), and four in Boa Esperança (times two field con-
ditions). First, CLR incidence was linearized by different
mathematical models, in order to correct infection rates due
to the remaining healthy leaf area available, and then plotted
against time for later test and selection of the most suitable
model to determine apparent infection rate (r). CLR incidence
was linearized by the following models:
Linear ¼ y  
1 The correlations were made considering three different
Monomolecular ¼ ln
  1−y
y
Logistic ¼ ln
1−y
Gompertz ¼ −lnð−lnðyÞÞ
Exponential ¼ ln y
where y is the proportion of leaves sampled with rust
symptoms.
Once linearized, CLR incidence was correlated by linear
regression with time. Through this regression model, initial
inoculum (linearized) and monthly apparent infection rate
were determined, being, respectively, the linear (a) and the
angular (b) coefficients of the regression. With these linear
regression models, linearized CLR incidence was estimat-
ed and retransformed in CLR incidence (proportion). After
that, the estimated CLR incidence was compared with ob-
served values in order to define the best model to represent
the epidemic. For choosing the best fit model, the follow-
ing statistical indices and errors were determined, as
recommended by Nutter (1997) and Bergamin Filho
(2011): mean bias error (MBE), mean absolute error
(MAE), root mean square error (RMSE), coefficient of de-
termination (R 2 ), Willmott’s Agreement Index Bd^
(Willmott 1981), and Confidence Index Bc^ (Camargo
and Sentelhas 1997).
Correlation between coffee leaf rust apparent
infection rate and weather variables
The weather variables recorded during the experimental
periods in each location were obtained from automatic
weather stations installed close to the fields where CLR
was monitored. The following weather variables were ob-
tained from these stations: maximum and minimum air
temperatures, relative humidity, and rainfall. As leaf wet-
ness duration sensor was not available, such variable was
estimated based on the number of hours with relative hu-
midity above 80 and 90% (Rowlandson et al. 2015;
Sentelhas et al. 2008).
From the abovementioned weather data, specific disease
driving variables were derived (Table 1). The weather data
were organized on a monthly basis to be compatible with
the CLR infection rate. Correlations between CLR infec-
tion rates, determined by the model with the best fit
(Gompertz) and weather variables were assessed through
the Pearson’s coefficient of correlation (P < 0.01), consid-
ering CLR apparent infection rate as dependent variable
and weather variables from Table 1, as independent ones.
The correlation significance was performed using the func-
tion Bcor.test^ in R. The same procedure was used to cor-
relation between weather variables.
time periods: 0–30d (from 0 to 30 days before incidence as-
sessment), 30–60d (from 30 to 60 days before incidence as-
sessment), and 60–90d (from 60 to 90 days before incidence
assessment). These intervals were chosen in accordance with
the latent period of Hemileia vastatrix, which lasts from 25 to
40 days (Kushalappa et al. 1983).

Table 1 Weather variables used
in the correlations with coffee leaf Variable
rust apparent infection rate, in a
monthly basis Tmed
Tmin
Tmax
Rainfall
NDR≥1mm
NDR≥20mm
RH (%)
NhRH≥90%
NhRH≥80%
T_RH≥90%
T_RH≥80%
ND_NhRH≥90%≥6h
ND_NhRH≥80%≥6h
Coffee leaf rust infection rate weather-based models
After obtaining the Pearson’s correlation coefficient for the
relationship between CLR infection rate and weather variables
for the three intervals of time, the weather-based models were
developed for each of the four conditions of fruit load and
rows space (HW, HN, LW, and LN). The stepwise procedure
was applied to select the best variables to compose the simple
or multiple regression models for estimating CLR apparent
infection rate.
The data of each condition was divided in two sets, being
one for models development and another for models valida-
tion. The number of months for models development and
validation varied according to the data availability for each
condition, respectively, as follows:
HN ¼ 259; 44; LN ¼ 259; 44; HW ¼ 215; 63; LW ¼ 215; 63:
Statistical analysis
The performance of the CLR infection rate models was
evaluated by means of contingency tables, similar to what
was done by Duttweiler et al. (2008) and Meira et al.
(2009). In addition, the observed and estimated values
were statistically compared and RMSE, MAE, MBE and
the coefficient of determination (R2) were calculated. In the
contingency tables, four different classes were considered:
true positive (TP)—the model predicts infection rate above
certain threshold and it occurs, true negative (TN)—the
model predicts a infection rate below a certain threshold
and it occurs, false positive (FP)—the model predicts in-
fection rate above certain threshold but it does not occur,
false negative (FN)—the model predicts infection rate be-
low a certain limit but it does not occur. Precision, recall,
Int J Biometeorol
Definition
Average temperature
Minimum temperature average
Maximum temperature average
Total rainfall
Number of days with rainfall ≥ 1 mm
Number of days with rainfall ≥ 20 mm
Average relative humidity
Number of hours with relative humidity ≥ 90%
Number of hours with relative humidity ≥ 80%
Average temperature on periods with relative humidity ≥ 90%
Average temperature on periods with relative humidity ≥ 80%
Number of days with NhRH≥90% equal or above 6 h
Number of days with NhRH≥80% equal of above 6 h
and f-measures were also calculated, according to (Huang
et al. 2005), being:
 
nTP
Precision ¼
ðnTP þ nFPÞ
 
nTP
Recall ¼
ðnTP þ nFNÞ
 
2* nTP
F‐measure ¼
2* nTP þ nFP þ nFN
where n is the total of events under such condition.
Precision express the proportion of predicted positive cases
that are correctly real positives, Recall is the proportion of
real positive cases that are correctly predicted positive, and
F-measure is a measure to quantify the tradeoff between
precision and recall.
In the contingency table, the following infection rate
thresholds were considered: 5 and 10 percental points (pp),
which were also used by Meira et al. (2009) for testing empir-
ical CLR models performance in Brazil. Also, 5 and 10 pp are
stated by Zambolim et al. (1997) and Kushalappa et al. (1984),
respectively, to use as risk limit on CLR control. The statistical
indices and contingency table were also used to determine the
model with the best performance to estimate CLR infection
rate for all conditions together, considering high and low fruit
loads, with a total of 107 months of data for each.
Results
Coffee leaf rust progress curves
The CLR presented different intensities according to fruit
loads (Fig. 1), beginning in October/November, increasing
Int J Biometeorol
during the end of spring, summer, and early autumn, and fi-
nally decreasing during the late autumn and winter, which
corresponds to the dry season when harvest is done.
In Varginha, besides fruit loads, Fig. 1a, b present the dif-
ferences of CLR incidence between wide and narrow rows. A
very small difference on the CLR during the first months was
detected when comparing these two conditions. The median
was higher on January, February, and March on narrow space;
however, after this period, the wide space had the highest
medians, with the final incidence very similar. The larger
Fig. 1 Coffee leaf rust progress in four different field conditions in
Varginha: a high fruit load, wide space; b high fruit load, narrow space;
c low fruit load, wide space; d low fruit load, narrow space; in Carmo de
healthy tissues available in wide spaces are the possible cause
of these variations. Moreover, inoculum source was abundant.
In these two conditions, median final incidence was above
80% (Fig. 1a, b).
For low fruit load conditions (Fig. 1c, d), both wide and
narrow spaces presented low CLR incidence in comparison to
high fruit loads. The median was lower than 35% and the third
quartile reached 40% only in June (Fig. 1c). The epidemic
onset was in December for both wide and narrow spaces.
Similarly to high fruit load conditions (Fig. 1a, b), the
Minas: e high fruit load, narrow space; f low fruit load, narrow space; in
Boa Esperança: g high fruit load, wide space; h low fruit load, wide space,
in the state of Minas Gerais, Brazil

maximum CLR incidence was reached firstly under narrow
space, but with similar intensities till the end of the season.
In Carmo de Minas and Boa Esperança, the comparison of
row spaces (wide x narrow) was not possible, since in these
locations the experiments only had one of these conditions.
Both sites present high incidence values for high fruit loads
(Fig. 1e, g), similar to Varginha, however with median CLR
incidence lower than 80%, which shows a less favorable en-
vironment condition to the disease than in Varginha. In all
these locations, for high fruit load, the maximum incidence
was in June, whereas for low fruit load there was a variation of
the maximum values between May and July (Fig. 1c, d, f, h).
Despite the considerations of Bock (1962) and Kushalappa
(1989) of higher CLR intensity on lower altitudes, the lowest
incidence values for high and low fruit loads are found in Boa
Esperança, despite the lower altitude, which leads to higher
temperatures.
Analysis of coffee leaf rust growth models
Statistical indexes measuring accuracy (RMSE, MBE, MAE,
and d index), precision (R2 and R*2), and the general perfor-
mance (c index) of the different models used to linearize CLR
progress curve were assessed (Table 2). The highest R2 was
obtained by the linear model; however, when comparing ob-
served and estimated CLR incidence proportion estimated by
the linear correlation between linearized CLR and time, the
highest fit (R*2) was obtained with the Gompertz model,
followed by logistic. Consequently, RMSE and MAE were
lower for the estimates with the Gompertz, as well as MBE.
Also, d index was close-to-perfection accuracy between ob-
served and estimated data, while the c index showed a BVery
Good^ confidence of the estimates, according to Camargo and
Sentelhas (1997). On the other hand, the estimates obtained by
the exponential model presented the lowest accuracy and pre-
cision values, and, therefore, the highest errors.
Figure 2 presents the regression analysis between CLR
incidence estimated by means of each model of linearization.
Table 2 Statistical indexes and errors for performance of five different linearization models for estimating coffee leaf rust infection rate, considering the
overall dataset from Varginha, Carmo de Minas, and Boa Esperança, state of Minas Gerais, Brazil
Model R2 R*2
Linear 0.900 0.856
Monomolecular 0.835 0.765
Logistic 0.834 0.937
Gompertz 0.868 0.958
Exponential 0.780 0.706
R2 = coefficient of determination, between time and disease proportion transformed values; R*2 = coefficient of determination between observed and
estimated disease proportion with the different linearization models; RMSE = root mean square error (percental points, pp); MBE = mean bias error (pp);
MAE = mean absolute error (pp); d index = Willmott’s Agreement Index; and c index: Confidence Index
Int J Biometeorol
As described before, the best precision and accuracy were
obtained by Gompertz model, which showed low dispersion
(R*2 = 0.958) and high accuracy, with regression linear coef-
ficient of 0.007 and the angular coefficient of 0.973.
Weather variables vs CLR apparent infection rate
The Pearson’s correlation coefficients obtained for the rela-
tionship between CLR infection rate and weather variables
are presented in Table 3. For HN conditions, RH, NhRH90%,
NhRH80%, NDRH90%>6h, T_RH90%, T_RH80%, and Tmin
had the overall highest correlation values for all periods; how-
ever, the highest ones were mainly for −30–60d period. For
this period of antecedence, the highest correlation was with
RH for HN (r = 0.465) and HW (r = 0.492), and with number
of days with rainfall above or equal to 1 mm (NDR≥1mm) for
LN (r = 0.383) and for LW (r = 0.346).
Despite the variables most correlated to CLR infection rate,
other ones also presented high coefficient of correlations. For
LN condition, Rainfall, T_RH80%, RH, T_RH90%, and Tmin
had high correlations, mainly for −30–60d period. For LW
condition, RH, Rainfall, NhRH80%, and NhRH90% also pre-
sented high correlations.
In general, CLR infection rate under low fruit load condi-
tions (LN and LW) presented lower correlations with weather
variables than under high fruit load, which is probably asso-
ciated to the highest plants’ tolerance to CLR when the com-
petition for carbohydrates is lower.
Coffee leaf rust infection rate estimation models
Models for estimating CLR infection rate were developed for
each coffee plantation condition, using the overall highest
correlation weather variables (Tmed, Tmin, RH, Rainfall,
and NDR≥1mm). The stepwise procedure was applied for
selecting the best variables, among the five listed above, to
compose the models. After the stepwise, multiple linear re-
gressions were performed with the remained variables for
RMSE MBE MAE d index c index
0.091 0.022 0.072 0.929 0.689
0.096 0.040 0.066 0.936 0.776
0.079 −0.010 0.045 0.928 0.795
0.050 0.000 0.037 0.957 0.848
0.200 0.025 0.111 0.842 0.619
Int J Biometeorol

Fig. 2 Relationship between 100 Logistic 100

a b Monomolecular

Estimated Incidence (%)

observed and estimated incidence
of coffee leaf rust by logistic (a), 80
monomolecular (b), exponential 80
(c), Gompertz (d), and linear (e)
models. Dashed line represents 60 60
the trend (linear regression) and
straight line the 1:1 line 40 40

20 20
y = 0.987x - 0.009 y = 0.9926x + 1.939
R*² = 0.937 R² = 0.765
0 0
0 20 40 60 80 100 0 20 40 60 80 100

100 100 Gompertz

c Exponential d
Estimated Incidence (%)

80 80

60 60

40 40

20 20
y = 1.374x - 7.862 y = 0.973x + 0.007
R*² = 0.706 R*² = 0.958
0 0
0 20 40 60 80 100 0 20 40 60 80 100

100 Linear
e
Estimated Incidence (%)

80

60

40

20 y = 0.929x + 4.124
R*² = 0.856
0
0 20 40 60 80 100
Observed Incidence (%)

each condition, and their significance was tested with a p
value below of 0.05. For HN and LN conditions, CLR infec-
tion rate models considered Tmin and RH as significant inde-
pendent variables. For HW condition, Tmed and RH were the
significant variables selected, whereas for LW condition,
NDR≥1mm was the only variable with significance to de-
scribe CLR infection rate. The CLR infection rate models
generated were the following:
rHN ¼ −1:293 þ 0:019 Tminð30−60dÞ þ 0:017 RHð30−60dÞ
conditions; Tmin is the average minimum temperature for
the period considered; RH the average relative humidity
for the period considered; Tmed the average mean temper-
ature for the period considered; NDR≥1mm the number of
days with rainfall ≥1 mm for the period considered; 30–
60d is the period between 30 and 60 days before CLR
incidence assessment; and the number of observations in
the dataset. All models were significant (*) at the P ≤ 0.01
level.
r ¼ 0:516* n ¼ 259

rLN ¼ −0:635 þ 0:018 Tminð30−60dÞ þ 0:006 RHð30−60dÞ r ¼ 0:382* n ¼ 259 Performance of CLR infection rate models
*
rHW ¼ −1:842 þ 0:027 Tmedð30−60dÞ þ 0:020 RHð30−60dÞ r ¼ 0:532 n ¼ 215

rLW ¼ 0:013 þ 0:0155 NDR≥ 1mmð30−60dÞ r ¼ 0:363* n ¼ 215 Both CLR infection rate models for HN and HW condi-
tions had similar performance (Table 4). The model for
where rHN, rLN, rHW, and rLW are the CLR infection rate, HW condition presented a small tendency of underestima-
respectively, for HN, LN, HW, and LW coffee plantation tion (MBE = −1.4 pp), whereas the model for HN showed
 Int J Biometeorol

Table 3 Pearson’s correlation coefficient between monthly coffee leaf rust (CLR) apparent infection rate and weather variables

Field conditions

Weather variables Period assessed before rust symptoms

0–30d −30–60d −60–90d 0–30d −30–60d −60–90d

HN LN
Tmed 0.064 0.239* 0.389* 0.209* 0.244* 0.269*
Tmin 0.154 0.344* 0.440* 0.260* 0.322* 0.302*
Tmax − 0.019 0.119 0.294* 0.138 0.144 0.205*
Rainfall 0.221* 0.321* 0.360* 0.260* 0.365* 0.199*
NDR≥1mm 0.202* 0.354* 0.394* 0.257* 0.383* 0.241*
NDR≥20mm 0.161* 0.264* 0.341* 0.194* 0.304* 0.186*
RH (%) 0.396* 0.465* 0.334* 0.285* 0.337* 0.169*
NhRH90% 0.362* 0.432* 0.355* 0.268* 0.309* 0.154
NhRH80% 0.388* 0.455* 0.333* 0.272* 0.322* 0.162*
T_RH90% 0.194* 0.381* 0.453* 0.272* 0.334* 0.294*
T_RH80% 0.213* 0.394* 0.452* 0.284* 0.345* 0.295*
NDRH90%>6 h 0.384* 0.431* 0.307* 0.253* 0.302* 0.131
NDRH80%>6 h 0.351* 0.390* 0.246* 0.264* 0.280* 0.094
HW LW
Tmed − 0.054 0.200* 0.411* 0.138 0.189* 0.249*
Tmin 0.051 0.318* 0.481* 0.223* 0.271* 0.263*
Tmax − 0.139 0.075 0.296* 0.050 0.089 0.197*
Rainfall 0.143 0.316* 0.368* 0.268* 0.302* 0.124
NDR≥1mm 0.085 0.293* 0.393* 0.277* 0.346* 0.156
NDR≥20mm 0.064 0.194* 0.270* 0.134 0.222* 0.066
RH (%) 0.399* 0.492* 0.392* 0.303* 0.271* 0.054
NhRH90% 0.386* 0.452* 0.368* 0.289* 0.233* 0.010
NhRH80% 0.388* 0.480* 0.371* 0.299* 0.266* 0.035
T_RH90% 0.094 0.358* 0.502* 0.241* 0.270* 0.261*
T_RH80% 0.119 0.384* 0.515* 0.255* 0.290* 0.247*
NDRH90%>6h 0.391* 0.432* 0.295* 0.246* 0.192* − 0.029
NDRH80%>6h 0.381* 0.400* 0.194* 0.217* 0.189* − 0.014

Period assessed before CLR assessments: 0–30d is from 0 to 30 days before incidence assessment; −30–60d is from 30 to 60 days before incidence
assessment; −60–90d is from 60 to 90 days before incidence assessment. Dataset from Varginha, Carmo de Minas, and Boa Esperança, state of Minas
Gerais, Brazil. Asterisk indicates that the correlations were (*) significant at P ≤ 0.01
HN high fruit load, narrow space; LN low fruit load, narrow space; HW high fruit load, wide space; LW low fruit load, wide space

no tendency, with MBE very close to zero. However, the
overall performance of these two models were very sim-
ilar, with MAE between 8.4 and 8.9 pp and R2 of 0.81
and 0.82, respectively, for HN and HW conditions. Due
to the highest bias, the model for HW resulted in more
false negative conditions (Table 5), underestimating the
CLR incidence. The CLR infection rate models for low
fruit load conditions (LN and LW) presented higher
RMSE and MBE than those for high fruit load (HN and
HW); however, MAE and R2 varied between the models
for LN and LW, with a poorer performance (lower
precision and accuracy) for the model of LN condition
(Table 4).
The lowest false estimates (false positive plus false
negative) for 5 pp threshold were observed for the HN
model, followed by HW (Table 5). On the other hand,
models for low fruit load condition presented the poorest
performances. Whereas the models for HN and HW pre-
sented true estimates in more than 70% of the situations,
the models for LN and LW only had true estimates below
65% (Table 5). Such performance resulted in higher
Precision, Recall, and F-measure for the model that
Int J Biometeorol

describes CLR infection rate under high fruit load (H) Table 5 Contingency table of the coffee rust infection rate estimation
models’ performance during the validation phase, for four different field
condition (Table 5), for both thresholds assessed.
conditions
Based on the results from the contingency table
(Table 5), the model for HN condition was the one chosen Models Threshold on 5 pp (%) 10 pp (%)
for estimating for CLR infection rate for both high and low
HN (n = 44 months)
fruit load field conditions, disregarding row spacing. The
True negative 29.55 36.36
results of the performance of the model to that are present-
False positive 20.45 22.73
ed in Tables 6 and 7. The MAE of 8.60 pp for high load
True positive 45.45 36.36
and 9.50 for low load (Table 6) and small percentages of
False negative 4.55 4.55
false negatives (Table 7) demonstrates the good perfor-
mance of the chosen model under different field condi- Precision 68.97 61.54
tions. The R2 is also high, 0.839 and 0.785 for high and Recall 90.91 88.89
low fruit loads, respectively. F-measure 78.43 72.73
However, for low fruit load conditions, there is a high LN (n = 44)
MBE, overestimating CLR infection rate, similar to what True negative 20.45 38.64
was presented in Tables 4 and 5. This poor performance is a False positive 47.73 45.45
consequence of the high false positives observed for this field True positive 25.00 15.91
condition. False negative 6.82 0.00
For both high and low fruit loads, the HN model presented Precision 34.38 25.93
the same trend (Fig. 3) when used for estimating disease in- Recall 78.57 100.00
tensity, overestimating low values and underestimating the F-measure 47.83 41.18
high one. For high fruit load, the trend line crosses 1:1 line HW (n = 63)
around 40% of disease intensity, whereas for low fruit load, it True negative 26.98 44.44
happens around 70% of incidence. False positive 19.05 17.46
The CLR infection rate model for HN condition resulted in True positive 44.44 23.81
high estimates precision for both fruit load condition False negative 9.52 14.29
(Table 6). However, for low fruit load condition, the model Precision 70.00 57.69
for HN condition was not as accurate, with lower Precision, Recall 82.35 62.50
Recall, and F-measure (Table 7) than for high fruit load con- F-measure 75.68 60.00
dition, for both thresholds assessed. As the main problems LW (n = 63)
with CLR is during the growing seasons with plants under True negative 49.21 68.25
high fruit load, the poor performance of the model for low False positive 12.70 3.17
fruit load conditions is not a major problem when thinking True positive 14.29 1.59
about to use this model for operational purposes, in a disease False negative 23.81 25.40
warning system. Precision 52.94 33.33
Recall 37.50 5.88
F-measure 43.90 10.00
Table 4 Errors and coefficient of determination for the validation of the Thresholds of 5 and 10 percental points (pp) were chosen based on
models for coffee leaf rust infection rate estimation, for different field Zambolim et al. (1997) and Kushalappa et al. (1984), respectively, to
conditions use as risk limit on CLR control, also used on Meira et al. (2009) to model
performance assessment. Dataset from Varginha, Carmo de Minas, and
Field RMSE MBE MAE R2
Boa Esperança, state of Minas Gerais, Brazil
condition (pp) (pp) (pp)
HN high fruit load, narrow space; LN low fruit load, narrow space; HW
HN 13.40 0.08 8.40 0.810 high fruit load, wide space; LW low fruit load, wide space
LN 14.50 9.50 10.90 0.729
HW 12.40 − 1.40 8.90 0.820 Discussion
LW 14.45 2.07 8.32 0.828
Other studies showed the high correlation between CLR inci-
Dataset from Varginha, Carmo de Minas, and Boa Esperança, state of dence and fruit load conditions (Japiassú et al. 2007; López-
Minas Gerais, Brazil
Bravo et al. 2012; Neto et al. 2013). According to the CLR
HN high fruit load, narrow space; LN low fruit load, narrow space; HW
progress curves in Fig. 1, high fruit load condition is consid-
high fruit load, wide space; LW low fruit load, wide space; pp percental
points; RMSE root mean square error; MBE mean bias error; MAE mean ered the most important, due to several aspects, such as (1) the
absolute error; R2 coefficient of determination management of CLR is pivotal in this condition to keep
 Int J Biometeorol

Table 6 Errors and coefficient of determination for validation of the making a specific pruning, known as Bframework skeleton^
best performance model (HN) for coffee leaf rust infection rate
that results in no harvest and production in the next season
estimation, under the two possible field conditions: High = high fruit
load and Low = low fruit load (Reis et al. 2018). The outcome is more profitable fields, as a
consequence of the low investments done in the seasons with
Field condition RMSE (pp) MBE (pp) MAE (pp) R2 low fruit load; (3) even in the cases when growers keep the
High 12.70 0.30 8.60 0.839
coffee plants producing during low fruit load seasons, the
Low 14.20 5.30 9.50 0.785
CLR management is easier, since its incidence is lower
(Fig. 1), mainly when it is considered that the fungicides have
Dataset from Varginha, Carmo de Minas, and Boa Esperança, state of a residual period between 60 and 90 days, which is enough to
Minas Gerais, Brazil keep CLR under control in low fruit load seasons.
HN high fruit load, narrow space; RMSE root mean square error The growth models expected to fit better CLR epidemic
(percental points, pp); MBE mean bias error (pp); MAE mean absolute
are logistic and Gompertz due to the polycyclic character-
error (pp); R2 coefficient of determination
istics of H. vastatrix pathogen (Berger 1981). Gompertz
sigmoidal shape with inflection point around 37% (1/e)
healthy leaves in the present season and to the next one. In the of disease intensity differs from the perfect sigmoidal
cases of high defoliation (high fruit load without CLR con- shape of logistic (inflection point on 50%). The higher
trol), blooming efficiency is negatively affected (Godoy et al. CLR intensity in narrow row spacing is probably related
1997). In this case, photosynthesis decreases as a consequence to the larger land surface covered by leaves, increasing the
of reduced leaf area, decreasing the number of active flowers, spores deposition area and making the disease to progress
which reduces potential yield of the next season; (2) currently, faster, which explains the best fit of the CLR progress
there is a practice being used in Brazil known as Bzero-crop curves by Gompertz model.
season.^ This method is applied when growers focus on The Gompertz model was selected to describe CLR prog-
obtaining maximum yields during the high fruit load seasons, ress due to the highest coefficient of determination (R*2) and
the smallest errors presented (Table 2, Fig. 2), which are the
main parameters for defining the linearization models to de-
scribe diseases (Campbell and Madden 1990, Nutter 1997,
Table 7 Contingency table of the HN model for coffee rust infection
rate estimation, on the two fruit loads conditions: High = high fruit load Bergamin Filho 2011). As reported by Kushalappa and
and Low = low fruit load Eskes (1989), the best performance of Gompertz model to
describe CLR progress curve can also be related to the higher
Fruit load conditions Threshold
susceptibility of young leaves on the season onset to this
5 pp (%) 10 pp (%) disease.
The most important variables that affect the CLR infection
High (n = 107 months) rate are the ones with interference on pathogen germination,
True negative 27.10 39.25 infection, and colonization process. Similar to Wallin (1962)
False positive 19.63 21.50 and Madden et al. (1978), leaf wetness duration (LWD) (esti-
True positive 46.73 29.91 mated based on RH) and the temperature during LWD period
False negative 6.54 9.35 are the most important variables to this disease. Different au-
Precision 70.42 58.18 thors found similar weather influence on CLR epidemic pat-
Recall 87.72 76.19 terns. According to Kushalappa and Chaves (1980), rainfall
F-measure 78.13 65.98 and minimum temperature were the meteorological variables
Low (n = 107) that most explained the CLR variation. Rainfall had high cor-
True negative 26.17 42.06 relation with long time before the symptoms, at 56 and
False positive 38.32 35.51 42 days, and minimum temperature showed high correlation
True positive 29.91 17.76 in the same period, however, not as high as rainfall. In the
False negative 5.61 4.67 present study, rainfall showed overall lower importance than
Precision 43.84 33.33 Tmin, RH, T_RH90%, and T_RH80%, mainly for high loads
Recall 84.21 79.17 (Table 3).
F-measure 57.66 46.91 Kushalappa et al. (1983) used periods of free water on
leaves and temperature to predict CLR infection rates. Meira
Dataset from Varginha, Carmo de Minas, and Boa Esperança, state of et al. (2009) found that temperature during periods of high RH
Minas Gerais, Brazil. Thresholds of 5 and 10 percental points (pp) were
chosen based on Zambolim et al. (1997) and Kushalappa et al. (1984),
(≥ 95%) and RH as the main weather variables for predicting
respectively, to use as risk limit on CLR control, also used on Meira et al. CLR infection rates. These variables are related to the temper-
(2009) to model performance assessment ature during the leaf wetness period and the length of this
Int J Biometeorol
Fig. 3 Performance of the best model (HN) when applied for the two possible fruit load conditions: High = high fruit load; Low = low fruit load field
conditions; HN = high fruit load, narrow space. Dashed line represents the trend line and straight line the 1:1 line
period are usually correlated to disease progress, due to their
importance on the infection process (germination and penetra-
tion). These variables are also recognized as important for
other diseases, such as Alternaria solani in tomato (Madden
et al. 1978), Pyricularia oryzae in rice (Kim et al. 1988), and
Phytophtora infestans in potato (Wallin 1962).
The period of antecedence in which the weather variables
were considered for modeling CLR infection rate was from 30
to 60 days before disease symptoms. It is related to the CLR
latency period, which varies from 25 to more than 30 days
(Moraes et al. 1976; Zambolim et al. 2005). Such long period
of antecedence for estimating the CLR allows to make deci-
sions before the disease symptoms appear. This period for
management comprehends acquiring weather data, organizing
them, running the model, and sending a warning to growers to
plan sprays under suitable weather conditions for that (without
rain and with low wind speed).
Tmin, RH, and Tmed are present in three of the CLR
infection rate models (HN, HW, and LN), whereas
NDR≥1mm is the variable considered for one of them
(LW). The models based on temperature and relative hu-
midity can be better extrapolated to other locations, since
they have good spatial correlation and are possible to be
estimated for other sites with data from surrounding weath-
er stations (Xavier et al. 2015). On the other hand, the
model based on rainfall is limited to the location where
this variable is measured, since rainfall has an expressively
spatial variability (Xavier et al. 2015).
As LWD and rainfall have high spatial variability (Thomas
et al. 2002), their use in disease warning systems can be an
issue. Moreover, LWD-related variables are not directly mea-
sured at standard weather stations, and when they are, there is
a lack of standard for measurements (Rowlandson et al. 2015).
Therefore, when the estimation capability is not diminished,
these variabilities can be neglected, resulting in simpler infec-
tion rate estimation models.
The weather variables presented strong correlation with
CLR infection rates (Table 8), grouped in temperature related
(1), rainfall related (2), and relative humidity related (3).
Following the same order, the correlations showed (1) temper-
atures are related with temperature during the periods of high
relative humidity (T_RH90% and T_RH80%; (2) NDR≥1mm is
also related with temperature during high relative humidity
periods (T_RH90% and T_RH80%); (3) relative humidity is
highly related to the number of hours with high relative hu-
midity (NhRH80% and NhRH90%), a variable used to estimate
leaf wetness duration (Rowlandson et al. 2015; Sentelhas et al.
2008), as well as the number of days with NDRH80%>6h and
NDRH90%>6h.
Among the infection rate models developed, that one for
HN condition was the best, considering its small errors and
high precision and accuracy. The use of only this model can
simplify CLR infection rate estimation, avoiding running dif-
ferent models, with different inputs, and not requiring infor-
mation from the growers related to their fields. Also, this
model is easy to understand and interpret, which are important
aspects for its adoption by the growers (Gillespie et al. 1993;
Gleason et al. 1995; Magarey and Isard 2017).
A model with reduced false negative frequency is more
conservative (Duttweiler et al. 2008), since false negative
means non-spray when it is necessary. In the field, false neg-
ative results in disease progress, while the objective of using a
disease forecast system is to keep low disease intensity as well
as to avoid over spraying. CLR incidence under lower values
has a key hole to keep the epidemic under control during the
season, as shown by Kushalappa and Chaves (1980),
 Int J Biometeorol

Table 8 Pearson correlation coefficients among weather variables used for coffee leaf rust infection rate estimation

Tmed = average temperature; Tmin = minimum temperature average; Tmax = maximum temperature average; RH (%) = average relative humidity;
Rainfall = total rainfall; NDR1 = number of days with rainfall ≥ 1 mm; NDR20 = number of days with rainfall ≥ 20 mm; NhRH≥90 = number of hours
with relative humidity ≥ 90%; NhRH≥80 = number of hours with relative humidity ≥ 80%; T_RH≥90 = average temperature on periods with relative
humidity ≥ 90%; T_RH≥80 = average temperature on periods with relative humidity ≥ 80%; NDRH90.6h = number of days with NhRH≥90 equal of
above 6 h; NDRH80.6h = number of days with NhRH≥80 equal of above 6 h. Values not shown were not significant (P ≤ 0.01)

considering that low inoculum source tends to result in low
disease progress.
The CLR infection rate model presented small percentage
of errors, caused when false negative occurred in conditions
(low and high fruit loads). It can be compensated by the
threshold used. The lower the threshold used (5%) more con-
servative becomes the system (Duttweiler et al. 2008). As for
CLR the residual period of the fungicides is from 60 to
90 days, the possibility of lack of control is reduced. It occurs
because if a spray is missed by a false negative estimation, the
plants will be still protected once the fungicide will be active
in the residual period. The precision and F-measure of the
CLR model for HN for high fruit load conditions were higher
than 70%, for 5% threshold, and from 58 to 65% for 10%. The
Recall was 87.72 and 76.19% for the same thresholds. The
Recall was sharply superior than that obtained by Meira et al.
(2009), presenting high capacity of true positives and true
negatives (Table 7).
Meira et al. (2008) reached an accuracy of 73% forecasting
CLR infection rates, with 79% of the events of more than 5 pp
increase correctly estimated. These authors found temperature
during leaf wetness periods, expected yield, mean of maxi-
mum temperatures during the incubation period and relative
humidity as the most important weather variables for CLR
progress. Using the HN model from the present model, we
obtained an accuracy of 73.83% of accuracy for 5 pp, similar
to what was found by Meira et al. (2008).
While Meira et al. (2008) used simply disease intensity
values, in the present study, it was considered the healthy
tissue available for infection, which is different according to
the incidence level. For example, an evolution from 2 to 40%
is different than an evolution from 20 to 58%, although both
presented an absolute difference of 38%. In the second case
(20 to 58%), there is less healthy leaf area available for infec-
tion, which means that for this second case, the weather con-
ditions were more favorable to disease progress than in first
Int J Biometeorol
case. Although the use of these epidemiological techniques,
the main advance in this study is the use of simple weather
data to predict the disease, allowing CLR management with-
out values of disease intensity from field measurements.
The results obtained in the present study showed that
weather conditions drive CLR epidemics more intensively
under high fruit loads, as can be observed in Tables 3 to 7.
In addition to the environment, another important factor that
was not considered in our model was the availability of inoc-
ulum. It is an important vertex of the disease triangle
(Bergamin Filho and Amorim 1996); however, it is not com-
monly measured in the coffee fields in Brazil. The necessity of
inoculum information could make forecast system inapplica-
ble. However, once available, inoculum data could help to
develop disease-weather-integrated models, what would im-
prove the precision and accuracy of the estimates, as also
mentioned by Kushalappa et al. (1984).
Conclusions
CLR has higher intensity values during seasons with high fruit
load. The Gompertz growth model was the best to describe
CLR epidemics accurately. Monthly minimum air temperature
and relative humidity were the main weather variables to es-
timate CLR apparent infection rate. These variables are highly
correlated to temperature during high RH period and to the
leaf wetness period, respectively. The model developed to
estimate CLR apparent infection rate for HN conditions can
also be used for other fields conditions (fruit load and rows
space), with false negative occurrences below 10%, and accu-
racy above 73% for a 5% threshold. For high fruit load, which
represent the season when growers intend to have higher in-
comes, the CLR control is pivotal. The forecast system pre-
sented here showed to be promising for recommending sprays
for coffee plantations in Brazil; however, more field trials are
essential to evaluate its performance in other coffee regions
and under different fungicide management situations.
Acknowledgements This research was supported by funds from
FAPESP (São Paulo Research Foundation), Brazil, under the process
number 2014/17781-8. The disease incidence and weather data were
gently provided by Fundação Procafé.
References
Avelino J, Cristancho M, Georgiou S et al (2015) The coffee rust crises in
Colombia and Central America (2008–2013): impacts, plausible
causes and proposed solutions. Food Secur 7:303–321. https://doi.
org/10.1007/s12571-015-0446-9
Bergamin Filho A (2011) Curvas de progresso das doenças. In: Amorim
L, Rezende JAM, Bergamin Filho A (eds) Manual de Fitopatologia
– Princípios e Conceitos, 4th edn. São Paulo, pp 647–666
Bergamin Filho A, Amorim L (1996) Doenças de plantas tropicais:
epidemiologia e controle econômico, 1st edn. Editora Agronômica
Ceres, São Paulo
Berger RD (1981) Comparison of the Gompertz and logistic equations to
describe plant disease progress. Phytopathology 71:716–719
Bock KR (1962) Seasonal periodicity of coffee leaf rust and factors
affecting the severity of outbreaks in Kenya Colony. Trans Br
Mycol Soc 45:289–300. https://doi.org/10.1016/S0007-
1536(62)80068-0
de Camargo AP, de Camargo MBP (2001) Definição e esquematização
das fases fenológicas do cafeeiro arábica nas condições tropicais do
Brasil. Bragantia 60:65–68. https://doi.org/10.1590/S0006-
87052001000100008
de Camargo AP, Pereira AR (1994) Agrometeorology of the coffee crop.
Switzerland, Geneva
Camargo AP de, Sentelhas PC (1997) Performance evaluation of different
potential evapotranspiration estimating methods in the state of São
Paulo, Brazil. Rev Bras Agrometeorol 5:89–97
Campbell CL, Madden LV (1990) Introduction to plant disease epidemi-
ology. John Wiley & Sons, New York
Chalfoun SM (1997) Doenças do cafeeiro: importância, identificação e
métodos de controle, First. FAEPE, Lavras
Conceição CHC, Guerreiro-Filho O, Gonçalves W (2005) Flutuação
populacional do bicho-mineiro em cultivares de café arábica
resistentes á ferrugem. Bragantia 64:625–631. https://doi.org/10.
1590/S0006-87052005000400012
DaMatta FM, Ronchi CP, Maestri M, Barros RS (2007) Ecophysiology of
coffee growth and production. Brazilian J Plant Physiol 19:485–
510. https://doi.org/10.1590/S1677-04202007000400014
Duttweiler KB, Gleason ML, Dixon PM et al (2008) Adaptation of an
apple sooty blotch and flyspeck warning system for the Upper
Midwest United States. Plant Dis 92:1215–1222. https://doi.org/
10.1094/PDIS-92-8-1215
Gillespie TJ, Srivastava B, Pitblado RE (1993) Using operational weather
data to schedule fungicide sprays on tomatoes in southern Ontario.
Canada Am Meteorol Soc 22:567–573
Gleason ML, MacNab AA, Pitblado RE et al (1995) Disease-warning
systems for processing tomatoes in eastern North America: are we
there yet? Plant Dis 79:113. https://doi.org/10.1094/PD-79-0113
Godoy CV, Bergamin Filho A, Salgado CL (1997) Doenças do cafeeiro.
In: Kimati H, Amorim L, Bergamin Filho A et al (eds) Manual de
Fitopatologia – Doenças das Plantas Cultivadas, 3rd edn. Editora
Agronômica Ceres Ltda., São Paulo, pp 178–192
Huang YJ, Powers R, Montelione GT (2005) Protein NMR recall, preci-
sion, and F-measure scores (RPF scores): structure quality assess-
ment measures based on information retrieval statistics. J Am Chem
Soc 127:1665–1674. https://doi.org/10.1021/ja047109h
ICO (2017) International coffee organization. In: Total Prod. by all
Export. Ctries. http://www.ico.org/prices/po-production.pdf.
Accessed 15 Dec 2017
Instituto Brasileiro de Geografia e Estatística (IBGE) (2017) Levantamento
Sistemático da Produção Agrícola - Pesquisa Mensal de Previsão e
Acompanhamento das Safras Agrícolas no Ano Civil
Japiassú LB, Garcia AWR, Miguel AE et al (2007) Influência da carga
pendente, do espaçamento e de fatores climáticos no
desenvolvimento da ferrugem do cafeeiro. Simpósio Pesqui dos
Cafés do Bras:171–175
Jayakumar M, Rajavel M, Surendran U (2016) Climate-based statistical
regression models for crop yield forecasting of coffee in humid
tropical Kerala, India. Int J Biometeorol 60:1943–1952. https://doi.
org/10.1007/s00484-016-1181-4
Kim CH, MacKenzie DR, Rush MC (1988) Field testing a computerized
forecasting system for rice blast disease. Phytopathology 78:931–
934
Kushalappa AC (1989) Rust management: an epidemiological approach
and chemical control. In: Kushalappa AC, Eskes AB (eds) Coffee

rust: epidemiology, resistance, and management, first. CRC Press,
Boca Raton, Florida, pp 81–140
Kushalappa AC, Akutsu M, Ludwig A (1983) Application of survival
ratio for monocyclic process of Hemileia vastatrix in predicting
coffee rust infection rates. Phytopathology 73:96. https://doi.org/
10.1094/Phyto-73-96
Kushalappa AC, Akutsu M, Oseguera SH et al (1984) Equations for
predicting the rate of coffee rust development based on net survival
ratio for monocyclic process of Hemileia vastatrix. Fitopatol Bras 9:
255–271
Kushalappa AC, Chaves GM (1980) An analysis of the development of
coffee rust in the field. Fitopatol Bras 5:95–103
Kushalappa AC, Eskes AB (1989) Coffee rust: epidemiology, resistance
and management, 1st edn. CRC Press, Boca Raton, Florida
Kushalappa AC, Hernandez TA, Lemos HG (1986) Evaluation of simple
and complex coffee rust forecasts to time fungicide application.
Fitopatol Bras 11:515–526
López-Bravo DF, de Virginio-Filho E, M, Avelino J (2012) Shade is
conducive to coffee rust as compared to full sun exposure under
standardized fruit load conditions. Crop Prot 38:21–29. https://doi.
org/10.1016/j.cropro.2012.03.011
Madden L, Pennypacker SP, MacNab AA (1978) FAST, a forecast system
for Alternaria solani on tomato. Phytopathology 68:1354–1358.
https://doi.org/10.1094/Phyto-68-1354
Magarey RD, Isard SA (2017) A troubleshooting guide for mechanistic
plant pest forecast models. J Integr Pest Manag 8:1–7. https://doi.
org/10.1093/jipm/pmw015
McCook S (2006) Global rust belt: Hemileia vastatrix and the ecological
integration of world coffee production since 1850. J Glob Hist 1:
177–195. https://doi.org/10.1017/S174002280600012X
Meira CAA, Rodrigues LHA, de MSA (2009) Modelos de alerta para o
controle da ferrugem-do-cafeeiro em lavouras com alta carga
pendente. Pesqui Agropecu Bras 44:233–242. https://doi.org/10.
1590/S0100-204X2009000300003
Meira CAA, Rodrigues LHA, Moraes SA (2008) Análise da epidemia da
ferrugem do cafeeiro com árvore de decisão. Trop Plant Pathol 33:
114–124
Moraes SA de, Sugimori MH, Ribeiro IJA, et al. (1976) Incubation period
of Hemileia vastatrix B. et Br. in three regions of São Paulo State.
Summa Phytopathol 2:32–38
Neto CDG, Rodrigues LHA, Thamada T, Meira CAA (2013) Warning
models for coffee rust (Hemileia vastatrix Berkeley & Broome) by
data mining techniques. 6
Int J Biometeorol
Nutman FJ, Roberts FM, Clarke RT (1963) Studies on the biology of
Hemileia vastatrix Berk. & Br. Trans Br Mycol Soc 46:27–44.
https://doi.org/10.1016/S0007-1536(63)80005-4
Nutter FWJ (1997) Quantifying the temporal dynamics of plant virus
epidemics: a review. Crop Prot 16:603–618. https://doi.org/10.
1016/S0261-2194(97)00055-0
Reis EAC, Freitas T, de Carvalho ML, F et al (2018) Characterization of
coffee cultivares leaf-resistant subjected to framework pruning.
Coffee S 13:63–70
Rowlandson T, Gleason M, Sentelhas PC et al (2015) Reconsidering leaf
wetness duration determination for plant disease management. Plant
Dis 99:310–319. https://doi.org/10.1094/PDIS-05-14-0529-FE
Sentelhas PC, Dalla Marta A, Orlandini S et al (2008) Suitability of
relative humidity as an estimator of leaf wetness duration. Agric
For Meteorol 148:392–400. https://doi.org/10.1016/j.agrformet.
2007.09.011
Souza AGC, Rodrigues FÁ, Maffia LA, Mizubuti ESG (2011) Infection
process of cercospora coffeicola on coffee leaf. J Phytopathol 159:
6–11. https://doi.org/10.1111/j.1439-0434.2010.01710.x
Thomas CS, Skinner PW, Fox AD et al (2002) Utilization of GIS/GPS-
based information technology in commercial crop decision making
in California. J Nematol 34:200–206
Wallin JR (1962) Summary of recent progress in predicting late blight
epidemics in United States and Canada. Am Potato J 39:306–312
Willmott CJ (1981) On the validation of models. Phys Geogr 2:184–194
Xavier AC, King CW, Scanlon BR (2015) Daily gridded meteorological
variables in Brazil (1980–2013). Int J Climatol 36:2644–2659.
https://doi.org/10.1002/joc.4518
Zambolim L (2016) Current status and management of coffee leaf rust in
Brazil. Trop Plant Pathol 41:1–8. https://doi.org/10.1007/s40858-
016-0065-9
Zambolim L, Vale FXR, Pereira AA, Chaves GM (1997) Café (Coffea
Arabica L.): controle de doenças – doenças causadas por fungos,
bactérias e virus. In: Vale FXR, Zambolim L (eds) Controle de
doenças de plantas: grandes culturas. pp 83–140
Zambolim L, Vale FXR, Zambolim EM (2005) Doenças do Cafeeiro. In:
Kimati H, Amorim L, Rezende JAM et al (eds) Manual de
Fitopatologia – Doenças das Plantas Cultivadas, 2nd edn. Editora
Agronômica Ceres Ltda, São Paulo, pp 165–180