Earth Science Informatics

https://doi.org/10.1007/s12145-023-01188-0

RESEARCH

Appropriate statistical rainfall distribution models for the computation

of standardized precipitation index (SPI) in Cameroon
A.R. Gamgo Fotse1 · G.M. Guenang1,2 · A.J. Komkoua Mbienda1,2,3 · Derbetini A. Vondou1

Received: 4 September 2023 / Accepted: 2 December 2023

© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2023

Abstract
The performance of the Standardized Precipitation Index (SPI) is affected by the choice of an incorrect probability distribution
function, which can skew the values of the index, exaggerating or minimizing the severity of drought. This study aims to
test data fitability of ten statistical distribution functions (gamma, weibull, exponential, lognormal, gumbel, cauchy, logistic,
chi-square, burr, pareto) for SPI computation at time scales (TSs) of 3, 6, 9, 12, 15, 18, 21 and 24 months, and to quantify
the errors made if the gamma function were used by default as is the case in general. Monthly precipitation data collected
at 24 meteorological stations distributed in the five Agro-Ecological Zones (AEZs) of Cameroon were used for the period
1951-2005. The parameters of the distribution functions were estimated with the Maximum Likelihood (ML) method, and
the Kolmogorov-Smirnov (K-S) test was applied to assess how well these functions fit the data. The results show that the
logistic and burr distributions provide for several stations of the five AEZs the best data fits. A comparative study between
the SPIs from the appropriate distribution and the gamma functions shows a significant qualitative and quantitative difference
in several stations and the root mean square error (RMSE) increases with the TS and with the severity of drought.

Keywords Standardized precipitation index · Cumulative distribution functions · Kolmogorov-smirnov test ·

Maximum likelihood method · Time scale

Introduction (Maia et al. 2015; Gebremeskel Haile et al. 2019). Accord-

ing to climatologists and meteorologists, it is the state of
Drought, like other natural phenomena closely linked to cli- an environment facing a significantly long and severe lack
mate change, is increasingly affecting the four corners of the of water, less than normal with negative impacts on flora,
globe (Bhaga et al. 2020). It is one of the costliest natural dis- fauna and societies (Quenum et al. 2019). Although occa-
asters in the world, affecting more people than other forms of sional droughts have always been part of the earth’s natural
disasters (Zarei et al. 2021). Indeed it is a natural hazard that phenomena, higher temperatures, greater water evaporation
begins slowly so that we often speak of a slowly evolving phe- and less vegetation cover all contribute to exacerbating the
nomenon (Sylla et al. 2016). As a result, its manifestations phenomenon (Ojha et al. 2021). Since drought can be ana-
may take longer to make themselves felt (Han et al. 2019). lyzed and interpreted from different angles and different
The impacts of drought vary according to regions, needs, and perceptions (Liu et al. 2018), there is no single definition of
disciplinary perspectives (Liu et al. 2012; Dai et al. 2020). drought accepted worldwide (Wilhite and Glantz 1985). In
They depend on the socio-economic environment in which it general, it is defined according to the situation experienced
occurs, since each region has its own climatic characteristics from one area to another (Qin et al. 2015). Depending on
the manifestations observed, droughts are classified into four
Communicated by: H. Babaie types: meteorological, hydrological, agricultural and socio-
A.R. Gamgo Fotse, G.M. Guenang, A.J. Komkoua Mbienda and Der-
economic (Wang et al. 2016). The three first types refer to
betini A. Vondou contributed equally to this work the deficits in precipitation, in soil moisture and in stream-
flow, respectively (Dai et al. 2020) while socio-economic
B G.M. Guenang drought refers to the insufficiency of water resources systems
merlin.guenang@yahoo.fr
to meet the water demand (Zhao et al. 2019). The current
Extended author information available on the last page of the article study focuses on meteorological drought which, according

123
 Earth Science Informatics

to Wang et al. (2016) is the starting phase for other types of respectively (Stagge et al. 2015). Other studies around the
droughts. world and in particular in Africa, have shown some distri-
Since each region has its own climatic characteristics, bution functions more appropriate than the gamma function
drought monitoring involves many different methods because which is in most cases used by default (Angelidis et al. 2012;
the amount of precipitation, the seasonal cycle and the nature Guenang and Mkankam Kamga 2014; Okpara and Tarhule
of the precipitation vary from region to region (Wilhite et al. 2015; Pieper et al. 2020; Zhang and Li 2020). However, all
2007; Park et al. 2017; Bhaga et al. 2020). This complex- these studies are limited to a reduced number of distribu-
ity in accurately describing the phenomenon has prompted tion functions and the quantification of errors made by using
researchers to define drought indices ranging from the sim- inappropriate distributions remains a real challenge.
plest to the most complex. These indices make it possible The objective of this study is to find appropriate statisti-
to characterize the drought by intensity, duration, spatial cal rainfall distribution models for the computation of SPI
extent, probability of recurrence (Spinoni et al. 2014; Huang and to quantify the errors made on the SPI values if inap-
et al. 2017), and its detection at different stages of its evo- propriate statistical models of precipitation distribution were
lution (location, time of appearance and end) (He et al. used beforehand for its computation. Given that the study
2018; Santé et al. 2019; Zhang and Li 2020). There are area has a high rainfall variability, it was necessary to go fur-
several drought indicators, the choice of which depends on ther by increasing the number of distribution functions to be
the type of impact to be taken into account as part of the tested in order to increase the probability of finding the most
mechanism for monitoring and understanding changes in appropriate leading to a more accurate SPI. Subsequently,
the vulnerability of the phenomenon (Huang et al. 2019; the error made by using the default gamma function as is
Bae et al. 2019). Of these different indices, we can cite most often the case was quantified. The next section (“Mate-
the Palmer Drought Index (PDSI: Palmer (1965)), the Stan- rials and methods”) shows materials and methods. “Results”
dardized Precipitation and Evapotranspiration Index (SPEI: shows the results and “Discussion and conclusion” provides
Vicente-Serrano et al. (2010)), the Standardized precipitation the discussion and conclusion.
index (SPI: McKee et al. (1993)), the rainfall anomaly index
(RAI: Van Rooy (1965)) and the Reconnaissance Drought
Index (RDI: Tsakiris and Vangelis (2005)). The SPI is rec- Materials and methods
ommended by the World Meteorological Organization as a
standard for characterizing meteorological droughts (Hayes Study area
et al. 2011) because of the particular advantages: it offers
good flexibility of use for multiple TSs (Hayes et al. 1999; Located in the heart of Central Africa, between 1.40◦ N and
Gidey et al. 2018; Guenang et al. 2019), it applies to all cli- 13◦ N of latitude and 8.30◦ E–16.10◦ E of longitude, at the
matic regimes and has good spatial consistency which allows bottom of the Gulf of Guinea, Cameroon covers an area of
comparisons between different areas subject to different cli- 475, 650km 2 , including 466, 050km 2 of continental area and
mates (Hayes et al. 1999; Pieper et al. 2020), its probabilistic 9, 600km 2 of maritime area with a 402 km long maritime
nature places it in a historical context which is well suited facade. The country belongs to the junction area between
to decision-making (Gebremichael et al. 2022). Due to these equatorial Africa in the south and tropical Africa to the north
exceptional advantages, the index have shown effectiveness (Net 2019). Cameroon is characterized by an extraordinarily
in detecting various historical drought events in many regions contrasting relief where high and low lands alternate. It is
of the world (Dogan et al. 2012; Ndayiragije et al. 2022). traditionally divided into five AEZs which are defined on the
Motivated by all these strengths of the SPI and the fact that it basis of their ecological, edaphic and climatic characteristics
only depends on the precipitation for which data were avail- (Nfornkah et al. 2021). Details on geographical and climatic
able, it was used as a drought indicator in this study. characteristics of these AEZs are presented in Table 1.
SPI proponents have suggested using the gamma distri-
bution to fit cumulative precipitation in the calculation of this Data used
index, but many studies show the limitations of this distribu-
tion (Stagge et al. 2015; Touma et al. 2015; Blain et al. 2018), Monthly precipitation data ranging from 1951 to 2005 were
and researchers have indicated that the applicability of theo- obtained from the database of the National Meteorological
retical distributions to describe cumulative precipitation was Service of Cameroon. They are from 24 measuring meteoro-
inconsistent between different regions and climates (Raziei logical stations and were successfully used in many studies
2021). So, some findings point at the gamma and weibull Penlap et al. (2004); Guenang and Mkankam Kamga (2014).
distributions to be the best suited for long periods (larger The geographical positions of these stations and the topogra-
than 3 months) and for short periods (smaller than 3 months) phy of the domain are shown in Fig. 1. The stations on which

123
Earth Science Informatics

Table 1 Geographical and climatic characteristics of the AEZs of and 24 months) and the adjusted distribution is transformed
Cameroon (Vondou et al. 2021) into a normal standardized distribution, so that the average
AEZ Location Mean Annual SPI is equal to zero (Raziei 2021). The ML estimation method
rainfall (mm) was used to find the optimal parameters of the distribution
AEZ1 Sudano-Sahelian Zone 885.8
functions to be tested and the K-S test was then performed to
choose the best fit distribution from the following ten func-
AEZ2 High Savannah Zone 1604.0
tions: gamma, weibull, ex ponential, lognor mal, gumbel,
AEZ3 Western Highlands Zone 2239.1
cauchy, logistic, chi −squar e, burr and par eto. The low-
AEZ4 Dense humid forest zone 2458.6
est K-S statistics determines the best fit distribution. This was
with monomodal rainfall
afterwards used in the data generation to calculate the cumu-
AEZ5 Dense humid Forest Zone 1640.1
lative distribution function (CDF) which were transformed
with bimodal rainfall
into normalized random variables, and then into SPI. The
same procedure is applied for each station and all TSs.
The period covered by the SPI varies according to the type
the study were focused are distinguished by a different color- of drought subject to the analysis and applications envisaged
ing from the others. The selection of representative stations (Gebremichael et al. 2022). Thus, the interpretation of the
by zone was made on the basis of the minimum missing values. SPI indicates the anomalies, which are deviations from the
average of the total precipitation observed for any period.
Computation of the SPI However, precipitation with high positive values corresponds
to very wet periods (positive SPI) while high negative val-
The SPI is computed by fitting an appropriate probability ues correspond to periods of extreme drought (negative SPI).
density function to the frequency distribution of precipita- McKee et al. (1993), uses the SPI classification indicated in
tion, summed over a considered TS (1, 3, 6, 9, 12, 15, 18, 21 Table 2 to define different drought categories.

Fig. 1 Study area with the

geographical location of the 24
precipitation stations (indicated
by numbers) in Cameroon

123
 Earth Science Informatics

Table 2 Classification of drought according to SPI values (Awchi and satisfactory. It has the desirable properties of a good esti-
Kalyana 2017) mator; in fact it is correct (it tends in probability towards
SPI value Classification the true value θ ), asymptotically unbiased (the mathematical
expectation of the estimator θ̂ is equal to the true value of the
2.0 and above Extreme wet
parameter θ ) and asymptotically efficient (Horváth 1993).
1,5 to 1,99 Severe wet
1,0 to 1,49 Moderate wet Law of statistical distributions used to fit data
0 to 0,99 Mild wet
-0,99 to 0 Mild drought The gamma law
-1,0 to -1,49 Moderate drought
-1,5 to -1,99 Severe drought Several researches have been made on the gamma law, in par-
-2.0 and below Extreme drought ticular (Choi and Wette 1969) deal in detail with the gamma
law. The random variable X follows a gamma distribution if
its probability density function (PDF) is:
The ML method
1 α
f (x) = x (α−1) exp (− ) (4)
The ML method makes it possible to estimate the parameters β (α) (α) β
of a regression model, under the assumption that the true law
To obtain the gamma cumulative function, we proceed as
of distribution of said parameters is known (Streit and Lug-
follows:
inbuhl 1994). It consists, for a given sample, in maximizing
 x  x
the likelihood function (joint density function) with respect 1 α
to the parameters. It seeks to find the parameter capable (with F(x) = f (x) = (α) x (α−1) exp (− )d x (5)
0 β (α) 0 β
a high probability) of reproducing the true values of the sam-
ple (those actually observed), i.e. to find the most likely value with
of the parameter of a population starting from a given sample ⎧
(Horváth 1993). Applied to a set of data, it provides values ⎨ α > 0 is the shape parameter
of the distribution parameter which maximize the likelihood β > 0 is the scale parameter
⎩
function (Meng et al. 2014).  is the mathematical gamma function
Either a random sample X 1 , X 2 , X 3 , .., X n from a distri-
α and β are obtained by the ML method as follows:
bution F(x; θ1 , θ2 , ....θ p ). When they exist, the estimators
obtained by the ML method are the solutions θˆ1 , θˆ2 , ....θˆp of ⎧ 
⎪
⎨ α̂ = 1
(1 + 1 + 43A )
the system of p equations: 4A

⎪ β̂ = −x
α̂
(6)
∂ L(θ1 , θ2 , ....θ p ) ⎩ ln (x)
=0 (1) A = ln (−x) − n
∂θr
With n the number of observation years. We also note
with r=1,2,..,p; where the likelihood function is defined by: that for x = 0, this function is not defined, and its modified
cumulative function is in the form:

n
L(θ1 , θ2 , ....θ p ) = f (X i , θ1 , θ2 , ....θ p ) (2) H (x) = q + (1 − q)F(x) (7)
i

With q the probability at each station of having zero pre-

It is often easier to maximize the logarithm of the likeli-
cipitation over the entire considered period.
hood function than the likelihood itself. Either method leads
to the same maximum because the logarithmic function is a
The weibull law
monotonically increasing function.


n The PDF of a random positive variable X distributed accord-
ln L(θ1 , θ2 , ....θ p ) = ln f (X i , θ1 , θ2 , ....θ p ) (3) ing to the Weibull law (Panahi and Asadi 2011) is:
i
f (x, α, β) = αβx α−1 exp (−βx α ) (8)
The ML method is considered to be a very efficient estima-
tor because it generates results with a lower variance value. Where α and β are respectively the shape and scale param-
Moreover, in long series (n > 100) the results are even more eters which are obtained by the ML method seen above and

123
Earth Science Informatics

which are presented in detail by Wu (2002). There is no the shape of the distribution. They can be determined by the
closed-form expressions of the parameters α and β, and there- ML estimator method as follows:
fore they are estimated by maximizing the log-likelihood
expression of the equation (Panahi and Asadi 2011). Its com- μ̂ = n1 n
i=1 ln x i
plementary cumulative distribution function is a stretched (15)
σˆ2 = n1 n
i=1 (ln x i − μ̂)2
ex ponential function so its explicit form is given by:

x
F(x) = 1 − exp (−( )β ) (9) The gumbel law
α
Also called doubly exponential law or law of extreme values,
The exponential law
a random variable X is distributed according to a Gumbel law
(Cooray 2010) if its PDF is given by :
A random variable X is distributed according to an exponen-
tial law if its PDF is given by:
1 x −μ x −μ
f (x) = exp [− exp (− )] exp (− ) (16)
1 exp [−(x − μ)] β β β
f (x) = (10)
β β
with
with x ≥ μ and β > 0, where μ is the location parameter
and β the scale parameter (Rahman and Pearson 2007). The ⎧
⎨μ > 0 is the position or mode parameter
scale parameter is often denoted λ = β1 and is called constant β>0 is the non-zero scale parameter, positive or negative
⎩
failure rate. The PDF of the ex ponential law can therefore −∞ < x < +∞
be written:

The terms μ and β are estimated using the ML method.

f (x) = λ exp [−(x − μ)]λ (11)
Its cumulative distribution function is of the form:
Its distribution function is of the form:
x −u
F(x) = exp [− exp (− )] (17)
β
F(x) = 1 − exp(−(x − μ))λ (12)

The parameters μ and λ are estimated from a random Gumbel’s law represents the maximum and minimum of a
and independent sample. The ML estimator is determined number of samples of normally distributed data.
by canceling the derivative of the logarithm of the likelihood
function of the ex ponential law, which leads to: The cauchy law

1 A random variable X follows a Cauchy law or even a Lorentz

λ̂ = (13)
x̄ law if its PDF depending on the two parameters μ > 0 and
n
β > 0 (Schuster 2012) is defined by:
with x̄ = 1
n i=1 x i

1 β
The lognormal law f (x) = [ ] (18)
π (x − μ)2 + β 2
A positive random variable x follows a lognormal distri-
bution if the logarithm of the random variable is normally with −∞ < x < +∞. The particularity of this law is that
distributed. The PDF of a lognor mal distribution is defined it has neither expectation nor variance. The term μ is the
as (Mage and Ott 1984): position parameter and the term β the scale parameter, that is
the spread parameter. Likewise the term μ represents both the
1 −(ln x − μ)2 mode and the median. These two parameters are estimated
f (x) = √ exp [ ] (14) by the ML method. Its cumulative distribution function is of
xσ 2π 2σ 2
the form:
with x > 0, σ > 0 and −∞ < μ < +∞
The term μ is the scale parameter that stretches or shrinks 1 (x − μ) 1
F(x) = tan−1 ( )+ (19)
the distribution, and σ 2 is the shape parameter that affects π β 2

123
 Earth Science Informatics

The logistic law with

⎧
A random variable X follows a logistic law if its PDF is ⎪ α > 0 is the shape parameter
⎪
⎨
given by Pérez-Sánchez and Senent-Aparicio (2018): γ > 0 is the shape parameter
⎪
⎪ λ > 0 is the scale parameter
⎩
x >0
exp −(x−α)
β
f (x) = (20)
(α)(1 + exp −(x−α)
β )2 The estimation of these parameters with the ML method is
the most common (Ghitany and Al-Awadhi 2002). Its cumu-
−∞ < x < +∞, with α the shape parameter and β the scale lative distribution function is:
parameter non-zero and positive. Its cumulative distribution x
function is given by: F(x) = 1 − (1 + ( )α )−γ (25)
λ
1 The pareto law
F(x) = (21)
1 + exp ( −(x−α)
β )
A random variable X follows a Pareto law if its PDF is defined
The parameters α and β are estimated by the ML method by the relation (Pérez-Sánchez and Senent-Aparicio 2018):
and were considered as starting value for the program α = 0
αβ α
and β = 1. f (x) = (26)
x α+1

The chi-square law with β ≤ x ≤ ∞, α > 0 and β > 0. The terms α and β
are respectively the shape and scale parameters, which are
It is a continuous distribution with k degrees of freedom, estimated by the ML method as follows:
used to describe the distribution of a sum of squared random
variables (Robertson 1969). Similarly, its importance also β̂ = min1≤i≤n xi
(27)
comes from its usefulness for independent data sets to test α̂ = ln x n−n ln β
i
the goodness of fit of a data distribution (Canal 2005). A
random variable X follows a chi − squar e distribution if its Its cumulative distribution function is :
PDF is given by:
β
F(x) = 1 − ( )α (28)
1 ( k2 −1) 1 x
f (x) = x exp (− x) (22)
2k
2 ( 2 )
k 2
The K-S fit test
with x ≥ 0. Its cumulative distribution function is : As mentioned by Stephens (1970), this test is inspired by
the statistics proposed by Kolmogorov (1933) for fitting to
γ ( k2 , x2 ) a distribution. It determines to what extent the data Xi (i=1,
F(x) = (23)
( k2 ) ...n) follow a specific distribution law F(X). The K-S test
is a nonparametric test that can be used to compare a sam-
with γ ( k2 , x2 ) the lower incomplete gamma function ple with a reference probability distribution or to compare
two samples (Mitchell 1971). The idea is to calculate the
maximum difference, in absolute value, between the empir-
The burr law
ical cumulative distribution and the theoretical cumulative
distribution under the null hypothesis for the running sum
The burr type XII distribution is a continuous and widely
of the chosen TS. Under the H0 hypothesis, this difference
known distribution because it includes the characteristics
is small and the distribution of observations fits well into
of various well-known distributions such as for example
a given distribution (Berger and Zhou 2014). For a specific
the weibull and gamma distributions (Pérez-Sánchez and
data set and distribution, the better the law fits the data, the
Senent-Aparicio 2018). A random variable X follows a burr
weaker the K-S test will be. So, for a law to be the best,
or burr type XII distribution if its PDF is:
its K-S test must be considerably weaker than the others. It
αγ x α−1 x quantifies a distance between the empirical distribution func-
f (x) = ( ) (1 + ( )α )−γ −1 (24) tion of the sample and the cumulative distribution function
λ λ λ

123
Earth Science Informatics

of the reference distribution or between the empirical dis- The results for all 24 stations and for eight TSs (3, 6, 9, 12,
tribution functions of two samples. Therefore, the smaller 15, 18, 21 and 24 months) are presented in Table 6. In general,
the D statistic, the closer the theoretical distribution is to the the logistic and burr distributions are the most suitable for
empirical distribution (Massey Jr 1951; Ramachandran and most stations except for 9-month TS where weibull followed
Tsokos 2015). For a distribution function cumulative F(x) by burr outperform the others. In Table 6, it is observed that
given, Stephens (1970) defined the statistic (K-S) by: for short (3-month) and long (> 6-month) TSs , the logistic
and burr distributions are the most appropriates respectively.
Dn = max x Fn (x) − F(x) (29) From the statistical point of view and at all TSs, the function
burr is the most representative followed by logistic and then
with −∞ < x < +∞, and by Glivenko-Cantelli theorem gamma. Table 7 summarizes the best fits for all AEZs. It is
(DeHardt 1971), observed a few cases where functions that better fit the data
are in equal numbers.
n
1
Fn (x) = I(−∞,x) (X i ) (30) Analysis of computed SPIs with adequate and
n
i=1 default gamma distributions
where Time series of SPIs computed with adequate distributions
⎧
⎪
⎪ n is the observation parameter in population x
⎨ SPI time series were calculated using the best fit distribu-
Fn (x) is the empirical cumulative distribution function
tion at each station and results are shown in Figs. 5, 6 and 7.
⎪
⎪ F(x) is the theoretical cumulative distribution function
⎩ The SPIs on a 3-month TS (Fig. 5) show for each station,
I(−∞,x) is the indicator function of the event x
a high frequency of drought events ranging from mild to
extreme categories. For 12-month SPI (Fig. 6), each sta-
tion shows at least one extremely dry episode. Throughout
Results the study period, the six stations differ markedly in the fre-
quency of extreme drought periods (4 in Bafia and Poli, 3
Determination of the adequate distribution in Nkongsamba, 2 in Ngaoundéré and Koundja, and none in
functions Douala). For 24-month SPI (Fig. 7) and during the first 30
years, Ngaounderé, Douala and Nkongsamba stations only
Figures 2, 3 and 4 show comparative results of the CDF for recorded very few drought but the following years recorded
historical precipitation data and for each of the ten trial distri- more frequent drought events. The dramatic drought events
bution functions. The results are shown for six target stations of the 1970s and 1980s are highlighted in each station and the
of the AEZs (Poli, Ngaounderé, Koundja, Bafia, Douala and magnitude and duration of the drought increased, especially
Nkongsamba) and for 3, 12 and 24 months TS. The follow- from the mid-1970s.
ing abbreviations were adopted for the functions: gamma
(g), weibull (w), ex ponential (e), lognor mal (ln), gumbel Evaluating the shift in SPI values due to the use of gamma
(gu), cauchy (c), logistic (lo), chi −squar e (ch), burr (bu) function instead of the appropriate functions
and par eto ( p). The K-S test was applied and the results are
presented in Tables 3, 4 and 5 for 3, 12 and 24 months TSs Figures 5, 6 and 7 show SPIs computed using both gamma
respectively. and the best-fit distributions. At 3-month TS (Fig. 5), the val-
The results show that at 3-month TS (Table 3), the logistic ues of SPIs obtained with the gamma distribution (S P Ig )
distribution is the best fit in the four stations namely Poli, are in general smaller than those obtained with the best dis-
Ngaoundere, Bafia and Koundja. For the two other stations tribution (S P Ibd ), which means that gamma leads to an
Douala and Nkongsamba, burr and weibull are the best underestimation of extreme humidity and an overestimation
fit respectively. At 12-month TS (Table 4), data from the of severe and extreme drought events. For the stations of
stations of Poli, Nkongsamba and Douala fit better with the AEZ4 (Douala and Nkongsamba) in the Littoral area,
the burr distribution; Koundja shows the gamma as the the SPIs are less sensitive to the choice of the distribu-
best fit, while data from Ngaoundere and Bafia are better tion function. Similar patterns are observed at 12-month TS
suited to logistic distribution. At 24-month TS (Table 5), (Fig. 6), but the differences between both SPIs (S P Ig and
the burr distribution is the best choice at Poli, Nkongsamba S P Ibd ) are higher and more depicted. As for 24-month TS
and Bafia while Ngaoundere, Koundja and Douala show a (Fig. 7), results are similar to those obtained at 3 and 12
preference to logistic, gamma and gumbel as the best fits months in the AEZs 1 and 2 (Poli and Ngaoundéré respec-
respectively. tively) while the reverse situation is observed in the AEZ4

123
 Earth Science Informatics

Fig. 2 Cumulative distribution a− AEZ1:Poli b− AEZ2:Ngaoundere

functions for 3-month aggregate

1.0

1.0
precipitation showing empirical
cumulative distribution function

0.8

0.8
and gamma, weibull,

3 month CDF(x)

3 month CDF(x)
ex ponential, lognor mal,

0.6

0.6
Empirical CDF Empirical CDF
gumbel, cauchy, logistic, Gamma CDF Gamma CDF
chi − squar e, burr , and Weibull CDF
Exponential CDF
Weibull CDF
Exponential CDF

0.4

0.4
par eto distributions fitted to Lognormal CDF Lognormal CDF
Gumbel CDF Gumbel CDF
station data from a) Poli Cauchy CDF Cauchy CDF

0.2

0.2
(AEZ1), b) Ngaoundere Logistic CDF Logistic CDF
Chi−square CDF Chi−square CDF
(AEZ2), c) Koundja (AEZ3), d) Burr CDF Burr CDF

0.0

0.0
Pareto CDF Pareto CDF
Bafia (AEZ5), e) Douala
(AEZ4), f) Nkongsamba (AEZ4) 0 2 4 6 8 10 12 0 2 4 6 8 10 12
x (dm) x (dm)

c− AEZ3:Koundja d− AEZ5:Bafia
1.0

1.0
0.8

0.8
3 month CDF(x)

3 month CDF(x)
0.6

0.6
Empirical CDF Empirical CDF
Gamma CDF Gamma CDF
Weibull CDF Weibull CDF
Exponential CDF Exponential CDF
0.4

0.4
Lognormal CDF Lognormal CDF
Gumbel CDF Gumbel CDF
Cauchy CDF Cauchy CDF
0.2

0.2
Logistic CDF Logistic CDF
Chi−square CDF Chi−square CDF
Burr CDF Burr CDF
0.0

0.0
Pareto CDF Pareto CDF

0 2 4 6 8 10 12 0 2 4 6 8
x (dm) x (dm)

e−AEZ4:Douala f− AEZ4:Nkongsamba
1.0

1.0
0.8

0.8
3 month CDF(x)

3 month CDF(x)
0.6

0.6

Empirical CDF Empirical CDF

Gamma CDF Gamma CDF
Weibull CDF Weibull CDF
Exponential CDF Exponential CDF
0.4

0.4

Lognormal CDF Lognormal CDF

Gumbel CDF Gumbel CDF
Cauchy CDF Cauchy CDF
0.2

0.2

Logistic CDF Logistic CDF

Chi−square CDF Chi−square CDF
Burr CDF Burr CDF
0.0

0.0

Pareto CDF Pareto CDF

0 5 10 15 20 25 30 0 5 10 15
x (dm) x (dm)

(Douala). However, in the AEZs 4 and 5 (Nkongsamba and lowest, sometimes equal to zero because gamma is the most
Bafia respectively), the gamma distribution leads to an over- suitable and match the default function (gamma) or equal to
estimation of extremely humid and drought intensity. low values due to the fact that the appropriate distribution
Figure 8 shows the root mean square error (RMSE) function found has flexibility similar to the gamma function.
between the SPI values computed from the appropriate dis-
tribution function and the gamma function for 3, 6, 9, 12,
15, 18, 21 and 24 months TSs at each station. Results are Discussion and conclusion
shown for the four drought categories. In general, the RMSE
increases with the severity of drought (from mild to extreme Discussion
drought) for each TS and in any AEZ; likewise, it increases
with TS for the AEZs 1, 2 and 5 (Poli, Ngaoundere and Bafia In most studies on SPI, gamma is chosen by default as the
respectively), but for other areas (AEZs 3 and 4) no consis- best fit without any comparison with other distributions. It
tent increase with TS is observed and the RMSEs are the was shown that the SPI with the commonly used gamma

123
Earth Science Informatics

Fig. 3 Same as in Fig. 2, but for a− AEZ1:Poli b− AEZ2:Ngaoundere

12-month aggregated

1.0

1.0
precipitation

0.8

0.8
12 month CDF(x)

12 month CDF(x)
0.6

0.6
Empirical CDF Empirical CDF
Gamma CDF Gamma CDF
Weibull CDF Weibull CDF
Exponential CDF Exponential CDF

0.4

0.4
Lognormal CDF Lognormal CDF
Gumbel CDF Gumbel CDF
Cauchy CDF Cauchy CDF

0.2

0.2
Logistic CDF Logistic CDF
Chi−square CDF Chi−square CDF
Burr CDF Burr CDF

0.0

0.0
Pareto CDF Pareto CDF

10 12 14 16 18 20 8 10 12 14 16 18 20
x (dm) x (dm)

c− AEZ3:Koundja d− AEZ5:Bafia
1.0

1.0
0.8

0.8
12 month CDF(x)

12 month CDF(x)
0.6

0.6
Empirical CDF Empirical CDF
Gamma CDF Gamma CDF
Weibull CDF Weibull CDF
Exponential CDF Exponential CDF
0.4

0.4
Lognormal CDF Lognormal CDF
Gumbel CDF Gumbel CDF
Cauchy CDF Cauchy CDF
0.2

0.2
Logistic CDF Logistic CDF
Chi−square CDF Chi−square CDF
Burr CDF Burr CDF
0.0

0.0
Pareto CDF Pareto CDF

15 20 25 10 12 14 16 18 20
x (dm) x (dm)

e−AEZ4:Douala f− AEZ4:Nkongsamba
1.0

1.0
0.8

0.8
12 month CDF(x)

12 month CDF(x)
0.6

0.6

Empirical CDF Empirical CDF

Gamma CDF Gamma CDF
Weibull CDF Weibull CDF
Exponential CDF Exponential CDF
0.4

0.4

Lognormal CDF Lognormal CDF

Gumbel CDF Gumbel CDF
Cauchy CDF Cauchy CDF
0.2

0.2

Logistic CDF Logistic CDF

Chi−square CDF Chi−square CDF
Burr CDF Burr CDF
0.0

0.0

Pareto CDF Pareto CDF

25 30 35 40 45 50 55 15 20 25 30 35
x (dm) x (dm)

distribution leads to shortcomings in evaluating ensemble function depends on the geographical location of the station
simulations (Pieper et al. 2020). For West Africa, Okpara and the TS considered. This is confirmed by the results of the
and Tarhule (2015) verified that the type two gamma distri- current study that show different distribution functions for
bution was a better model for adjusting precipitation over different areas. The results also corroborate those of Cindric
the Niger basin. So, it is clear that several functions can and Pasaric (2012), who suggested that it is not possible to
override the gamma distribution as the best fit in many recommend a single, optimal distribution because the ratio of
stations and provide better SPI values. From this current skewness and the coefficient of variation of data precipitation
study on a larger number of distribution functions, we found could be the indicator for the choice of the most appropriate
that new functions (burr and logistic) are able to better distribution for a particular region. Moreover, Angelidis et al.
fit the data in some stations as compared to the findings of (2012) and Stagge et al. (2015) suggested that the suitable
Guenang and Mkankam Kamga (2014) where only four func- probability distribution is related with the TS of precipitation
tions (gamma, weibull, lognor mal and ex ponential) were data to be fitted.
tested. Therefore, the choice of the appropriate distribution

123
 Earth Science Informatics

Fig. 4 Same as in Fig. 2, but for a− AEZ1:Poli b− AEZ2:Ngaoundere

24-month aggregated

1.0

1.0
precipitation

0.8

0.8
24 month CDF(x)

24 month CDF(x)
0.6

0.6
Empirical CDF Empirical CDF
Gamma CDF Gamma CDF
Weibull CDF Weibull CDF
Exponential CDF Exponential CDF

0.4

0.4
Lognormal CDF Lognormal CDF
Gumbel CDF Gumbel CDF
Cauchy CDF Cauchy CDF

0.2

0.2
Logistic CDF Logistic CDF
Chi−square CDF Chi−square CDF
Burr CDF Burr CDF

0.0

0.0
Pareto CDF Pareto CDF

25 30 35 20 25 30 35
x (dm) x (dm)

c− AEZ3:Koundja d− AEZ5:Bafia
1.0

1.0
0.8

0.8
24 month CDF(x)

24 month CDF(x)
0.6

0.6
Empirical CDF Empirical CDF
Gamma CDF Gamma CDF
Weibull CDF Weibull CDF
Exponential CDF Exponential CDF
0.4

0.4
Lognormal CDF Lognormal CDF
Gumbel CDF Gumbel CDF
Cauchy CDF Cauchy CDF
0.2

0.2
Logistic CDF Logistic CDF
Chi−square CDF Chi−square CDF
Burr CDF Burr CDF
0.0

0.0
Pareto CDF Pareto CDF

30 35 40 45 50 24 26 28 30 32 34 36
x (dm) x (dm)

e−AEZ4:Douala f− AEZ4:Nkongsamba
1.0

1.0
0.8

0.8
24 month CDF(x)

24 month CDF(x)
0.6

0.6

Empirical CDF Empirical CDF

Gamma CDF Gamma CDF
Weibull CDF Weibull CDF
Exponential CDF Exponential CDF
0.4

0.4

Lognormal CDF Lognormal CDF

Gumbel CDF Gumbel CDF
Cauchy CDF Cauchy CDF
0.2

0.2

Logistic CDF Logistic CDF

Chi−square CDF Chi−square CDF
Burr CDF Burr CDF
0.0

0.0

Pareto CDF Pareto CDF

50 60 70 80 90 100 40 45 50 55 60 65
x (dm) x (dm)

Stagge et al. (2015) compared the seven probability distri- respectively, perform better than the correspondents two-
butions and concluded that the gamma distribution produces parameter distributions, which is in agreement with the
the best fit for precipitation with long period (>6 months), results obtained in this paper where the three-parameter burr
while weibull is consistently the best for precipitation with distribution gave the best results in most cases. Some distri-
short period (1 to 3 month). In the present study, the logistic bution functions such as chi −squar e, cauchy, ex ponential
distribution produces the best fit for precipitation with short and par eto generally show poorer fit in the study area.
accumulation (3 months TS), while for long periods (> 6 Guenang and Mkankam Kamga (2014) also found that the
months TS) the burr distribution performs the best. Pieper ex ponential distribution is the least suitable for the domain.
et al. (2020) and Zhang and Li (2020) estimated that the Considering the effects of different probability distribu-
appropriate probability distribution is related to the number tions on SPI characteristics in comparison with the default
of parameters of the PDF of the distribution to be fitted to gamma distribution, it is observed that SPIs for mild and
rainfall data. They demonstrated that distributions with three moderate droughts are less sensitive to the distribution func-
parameters such as ex ponential weibull and log −logistic tions used, than those corresponding to severe and extreme

123
Earth Science Informatics

Table 3 Values of the K-S fit

No Stations 3 month TS
test for the ten distribution
g w e ln gu c lo ch bu p
functions and for 3-month TS
1 Maroua 0.109 0.110 0.108 0.146 0.099 0.216 0.135 0.179 0.109 0.107
2 Kaelé 0.102 0.099 0.093 0.157 0.109 0.213 0.136 0.176 0.099 0.093
3 Garoua 0.129 0.110 0.140 0.169 0.093 0.157 0.090 0.160 0.111 0.139
4 Poli 0.147 0.136 0.141 0.186 0.103 0.159 0.099 0.257 0.136 0.141
5 Ngaoundéré 0.156 0.146 0.157 0.185 0.109 0.146 0.089 0.247 0.145 0.158
6 Meiganga 0.127 0.119 0.132 0.159 0.098 0.159 0.089 0.210 0.118 0.131
7 Tibati 0.139 0.125 0.146 0.171 0.093 0.148 0.079 0.228 0.125 0.146
8 Koundja 0.110 0.097 0.140 0.163 0.082 0.148 0.075 0.197 0.097 0.141
9 Yoko 0.154 0.127 0.184 0.191 0.096 0.123 0.072 0.187 0.127 0.184
10 Nkongsamba 0.075 0.065 0.106 0.122 0.071 0.168 0.096 0.191 0.066 0.106
11 Bafia 0.151 0.114 0.228 0.199 0.100 0.099 0.079 0.150 0.114 0.227
12 Nang-Eboko 0.114 0.085 0.196 0.156 0.082 0.109 0.056 0.111 0.085 0.196
13 Bertoua 0.127 0.086 0.218 0.167 0.089 0.098 0.054 0.119 0.086 0.218
14 Batouri 0.118 0.085 0.219 0.155 0.095 0.098 0.056 0.109 0.085 0.218
15 Ngambé 0.078 0.053 0.125 0.131 0.069 0.155 0.088 0.171 0.053 0.125
16 Douala 0.070 0.064 0.083 0.086 0.074 0.177 0.099 0.227 0.064 0.83
17 Abong-mbang 0.111 0.076 0.226 0.146 0.085 0.099 0.047 0.100 0.076 0.227
18 Yaoundé 0.121 0.081 0.207 0.156 0.095 0.104 0.052 0.107 0.082 0.206
19 Akonolinga 0.118 0.086 0.213 0.141 0.098 0.099 0.060 0.099 0.086 0.213
20 Eséka 0.114 0.072 0.228 0.155 0.088 0.098 0.052 0.113 0.072 0.229
21 Yokadouma 0.124 0.089 0.222 0.152 0.105 0.092 0.059 0.101 0.089 0.223
22 Lomié 0.114 0.076 0.225 0.139 0.098 0.089 0.053 0.118 0.076 0.225
23 Kribi 0.039 0.036 0.223 0.073 0.033 0.121 0.061 0.047 0.026 0.222
24 Sangmélima 0.102 0.059 0.244 0.132 0.088 0.092 0.039 0.135 0.060 0.245
Bold values indicate the lowest K-S statistics determining the best fit distribution

Table 4 Values of the K-S fit

No Stations 12 month TS
test for the ten distribution
g w e ln gu c lo ch bu p
functions and for 12-month TS
1 Maroua 0.050 0.086 0.494 0.041 0.060 0.091 0.059 0.287 0.047 0.495
2 Kaelé 0.063 0.084 0.473 0.069 0.071 0.111 0.067 0.254 0.061 0.473
3 Garoua 0.062 0.096 0.499 0.052 0.045 0.095 0.063 0.265 0.043 0.500
4 Poli 0.044 0.064 0.502 0.053 0.083 0.082 0.048 0.263 0.043 0.502
5 Ngaoundéré 0.062 0.040 0.507 0.072 0.117 0.084 0.029 0.279 0.030 0.506
6 Meiganga 0.037 0.077 0.531 0.035 0.062 0.102 0.047 0.243 0.040 0.530
7 Tibati 0.035 0.059 0.507 0.042 0.065 0.077 0.030 0.249 0.025 0.507
8 Koundja 0.042 0.059 0.534 0.048 0.066 0.091 0.045 0.259 0.044 0.534
9 Yoko 0.045 0.077 0.521 0.039 0.042 0.099 0.054 0.249 0.052 0.522
10 Nkongsamba 0.044 0.049 0.506 0.052 0.075 0.077 0.031 0.217 0.028 0.505
11 Bafia 0.049 0.088 0.503 0.056 0.097 0.072 0.032 0.264 0.037 0.503
12 Nang-Eboko 0.079 0.038 0.472 0.089 0.101 0.099 0.056 0.218 0.038 0.472
13 Bertoua 0.027 0.079 0.495 0.024 0.057 0.081 0.035 0.227 0.030 0.494

123
 Earth Science Informatics

Table 4 continued
No Stations 12 month TS
g w e ln gu c lo ch bu p

14 Batouri 0.048 0.047 0.510 0.056 0.079 0.091 0.052 0.241 0.048 0.510
15 Ngambé 0.071 0.077 0.522 0.074 0.075 0.101 0.068 0.234 0.076 0.521
16 Douala 0.077 0.075 0.495 0.081 0.080 0.102 0.077 0.133 0.074 0.495
17 Abong-mbang 0.041 0.084 0.497 0.044 0.081 0.073 0.030 0.214 0.032 0.497
18 Yaoundé 0.029 0.051 0.506 0.037 0.068 0.081 0.026 0.241 0.022 0.505
19 Akonolinga 0.051 0.045 0.504 0.060 0.091 0.086 0.041 0.256 0.039 0.504
20 Eséka 0.038 0.085 0.535 0.030 0.045 0.103 0.037 0.226 0.033 0.535
21 Yokadouma 0.056 0.810 0.512 0.061 0.106 0.086 0.041 0.233 0.045 0.512
22 Lomié 0.039 0.093 0.516 0.035 0.072 0.075 0.025 0.269 0.030 0.517
23 Kribi 0.031 0.059 0.490 0.031 0.057 0.096 0.047 0.114 0.045 0.491
24 Sangmélima 0.035 0.098 0.524 0.027 0.044 0.096 0.038 0.257 0.032 0.524
Bold values indicate the lowest K-S statistics determining the best fit distribution

Table 5 Values of the K-S fit

No Stations 24 month TS
test for the ten distribution
g w e ln gu c lo ch bu p
functions and for 24-month TS
1 Maroua 0.033 0.080 0.529 0.031 0.067 0.087 0.035 0.281 0.034 0.529
2 Kaelé 0.094 0.118 0.504 0.092 0.073 0.105 0.102 0.232 0.107 0.505
3 Garoua 0.040 0.061 0.541 0.043 0.075 0.094 0.047 0.269 0.047 0.540
4 Poli 0.061 0.057 0.535 0.066 0.114 0.086 0.045 0.258 0.039 0.535
5 Ngaoundéré 0.058 0.054 0.527 0.066 0.132 0.104 0.046 0.267 0.054 0.526
6 Meiganga 0.068 0.106 0.543 0.071 0.069 0.106 0.078 0.222 0.082 0.544
7 Tibati 0.079 0.053 0.542 0.085 0.109 0.098 0.060 0.238 0.058 0.543
8 Koundja 0.054 0.065 0.556 0.058 0.076 0.100 0.065 0.269 0.070 0.556
9 Yoko 0.039 0.088 0.565 0.038 0.055 0.102 0.049 0.239 0.039 0.566
10 Nkongsamba 0.048 0.058 0.538 0.055 0.095 0.080 0.049 0.177 0.045 0.539
11 Bafia 0.060 0.083 0.556 0.065 0.094 0.081 0.060 0.261 0.057 0.555
12 Nang-Eboko 0.074 0.032 0.491 0.081 0.097 0.102 0.050 0.167 0.032 0.491
13 Bertoua 0.090 0.038 0.534 0.095 0.107 0.114 0.070 0.253 0.039 0.535
14 Batouri 0.053 0.091 0.541 0.054 0.065 0.107 0.063 0.234 0.077 0.542
15 Ngambé 0.054 0.049 0.551 0.056 0.097 0.102 0.063 0.200 0.054 0.552
16 Douala 0.092 0.106 0.521 0.093 0.075 0.107 0.100 0.084 0.098 0.522
17 Abong-mbang 0.051 0.034 0.536 0.057 0.094 0.089 0.045 0.248 0.036 0.537
18 Yaoundé 0.068 0.060 0.536 0.074 0.094 0.088 0.061 0.233 0.056 0.537
19 Akonolinga 0.090 0.092 0.553 0.093 0.093 0.129 0.084 0.289 0.092 0.552
20 Eséka 0.062 0.101 0.562 0.057 0.053 0.100 0.065 0.242 0.062 0.563
21 Yokadouma 0.028 0.060 0.542 0.035 0.094 0.086 0.025 0.223 0.029 0.543
22 Lomié 0.045 0.080 0.549 0.050 0.097 0.073 0.034 0.278 0.039 0.549
23 Kribi 0.063 0.092 0.522 0.057 0.055 0.116 0.067 0.101 0.070 0.523
24 Sangmélima 00.056 0.090 0.566 0.052 0.075 0.107 0.060 0.259 0.055 0.567
Bold values indicate the lowest K-S statistics determining the best fit distribution

123
Earth Science Informatics

Table 6 Best distribution functions that better fit station precipitation data at different TSs
area No Stations TS
3 month 6 month 9 month 12 month 15 month 18 month 21 month 24 month

AEZ1 1 Maroua gu lo lo ln gu g w ln
2 Kaele p lo bu bu gu bu lo gu
3 Garoua lo lo w bu ln g bu g
4 Poli lo lo w bu g g bu bu

AEZ2 5 Ngaoundere lo lo w lo g g lo lo
6 Meiganga lo w bu ln g g g g
7 Tibati lo lo g bu g g g w

AEZ3 8 Koundja lo bu w g g lo g g

AEZ4 9 Douala bu bu w w ch g ch gu
10 Nkongsamba w w w bu ln g bu bu
11 Ngambe bu w w lo ln g g w

AEZ5 12 Yoko lo w g ln bu w g ln
13 Bafia lo bu g lo bu bu g bu
14 Nang-Eboko lo bu bu w bu bu bu w
15 Yaounde lo bu g bu bu g bu bu
16 Eseka lo g g ln g ln ln gu
17 Akonolinga lo g g bu g bu g lo
18 Bertoua lo bu ln ln bu bu lo w
19 Batouri lo lo w ln w g g ln
20 Abong-mbang lo bu bu lo lo bu g w
21 Yokadouma lo w bu lo lo bu g lo
22 Lomie lo lo lo g lo g g g
23 Kribi bu bu ln g ln ln ln gu
24 Sangmelima lo g bu ln g ln ln ln

droughts. This statement agrees with that of Angelidis et al. change, particularly due to global warming, strong fluctu-
(2012) who concludes that the consistency of the calculated ations in average precipitation have a severe effect on the
SPI with different distributions is good for normal periods, occurrence of drought. The current study also found that the
while becoming poor for very dry or very wet periods. We magnitude and duration of drought increased with time for
also found that different probability distributions lead to a both short and long TSs. This may be the consequence of
large difference in severe and extreme droughts. In fact, the reduced precipitation resulting from climate change as sug-
SPI time series patterns obtained with gamma lead to an gested by Vicente-Serrano et al. (2010) and Tirivarombo et al.
underestimation of extreme humidity and an overestimation (2018) because temperature is also an important factor that
of severe and extreme drought events as compare to that can influence the availability of water as it controls the rates
obtained with the best distribution. In a context of climate of evapotranspiration.

Table 7 Best distribution functions by AEZs at different TSs

AEZs T́S
3 month 6 month 9 month 12 month 15 month 18 month 21 month 24 month

AEZ1 lo lo w bu gu g bu bu, gu, g, ln

AEZ2 lo lo bu, w, g bu, lo, ln g g g lo, g, w
AEZ3 lo bu w g g lo g g
AEZ4 bu w w bu, lo, w ln g bu, g, ch bu, gu, w
AEZ5 lo bu g ln bu bu g bu, lo, w, g, gu, ln

123
 Earth Science Informatics

AEZ1: AEZ2:
Poli Ngaoundere

3
logistic Fit gamma Fit logistic Fit gamma Fit

2
1

1
3−months SPI

3−months SPI
0

0
−1

−1
−2

−2
−3

−3
1950 1960 1970 1980 1990 1950 1960 1970 1980 1990 2000

Years Years

AEZ3: AEZ5:
Koundja Bafia
3

3
logistic Fit gamma Fit logistic Fit gamma Fit
2

2
1

1
3−months SPI

3−months SPI
0

0
−1

−1
−2

−2
−3

−3

1950 1960 1970 1980 1990 2000 1950 1960 1970 1980 1990 2000

Years Years

AEZ4: AEZ4:
Douala Nkongsamba
3

burr Fit gamma Fit weibull Fit gamma Fit

2

2
1

1
3−months SPI

3−months SPI
0

0
−1

−1
−2

−2
−3

−3

1950 1960 1970 1980 1990 2000 1950 1960 1970 1980 1990 2000

Years Years

Fig. 5 SPIs at 3-month TS. It is overlaid SPI computed using appropriate distribution function to that from default gamma distribution function
at the stations of a) Poli (AEZ1), b) Ngaoundere (AEZ2), c) Koundja (AE3), d) Bafia (AEZ5), e) Douala (AEZ4), f) Nkongsamba (AEZ4)

123
Earth Science Informatics

AEZ1: AEZ2:
Poli Ngaoundere

3
burr Fit gamma Fit logistic Fit gamma Fit

2
1

1
12−months SPI

12−months SPI
0

0
−1

−1
−2

−2
−3

−3
1950 1960 1970 1980 1990 2000 1950 1960 1970 1980 1990 2000

Years Years

AEZ3: AEZ5:
Koundja Bafia
3

3
gamma Fit gamma Fit logistic Fit gamma Fit
2

2
1

1
12−months SPI

12−months SPI
0

0
−1

−1
−2

−2
−3

−3

1950 1960 1970 1980 1990 2000 1950 1960 1970 1980 1990

Years Years

AEZ4: AEZ4:
Douala Nkongsamba
3

burr Fit gamma Fit burr Fit gamma Fit

2

2
1

1
12−months SPI

12−months SPI
0

0
−1

−1
−2

−2
−3

−3

1950 1960 1970 1980 1990 2000 1950 1960 1970 1980 1990 2000

Years Years

Fig. 6 Same as in Fig. 5, but for 12-month TS

123
 Earth Science Informatics

AEZ1: AEZ2:
Poli Ngaoundere

3
burr Fit gamma Fit logistic Fit gamma Fit

2
1

1
24−months SPI

24−months SPI
0

0
−1

−1
−2

−2
−3

−3
1950 1960 1970 1980 1990 1950 1960 1970 1980 1990 2000

Years Years

AEZ3: AEZ5:
Koundja Bafia
3

3
gamma Fit gamma Fit burr Fit gamma Fit
2

2
1

1
24−months SPI

24−months SPI
0

0
−1

−1
−2

−2
−3

−3

1950 1960 1970 1980 1990 2000 1950 1960 1970 1980 1990

Years Years

AEZ4: AEZ4:
Douala Nkongsamba
3

gumbel Fit gamma Fit burr Fit gamma Fit

2

2
1

1
24−months SPI

24−months SPI
0

0
−1

−1
−2

−2
−3

−3

1950 1960 1970 1980 1990 2000 1950 1960 1970 1980 1990 2000

Years Years

Fig. 7 Same as in Fig. 5, but for 24-month TS

123
Earth Science Informatics

Fig. 8 RMSE between the SPI values of the appropriate distribution function and the gamma distribution function for each station and for 3, 6, 9,
12, 15, 18, 21 and 24 months TSs

Conclusion for different TSs (3, 6, 12, 15, 18, 21 and 24 months). The
ML method was used to estimate the parameters of the dis-
This study was undertaken to contribute to the improve- tribution functions. The K-S statistic was used to select the
ment of mathematical tools for modeling drought which is distribution functions that better fit station data, which were
a dangerous phenomenon and whose adaptation is difficult then used to calculate the SPI. The results were used to
in developing countries such as Cameroon. The SPI used as study drought occurrence and quantify the errors made if
drought indicator was studied in this paper by examining the non appropriate distribution functions were used.
relevance of using probability distribution functions differ- The appropriate distribution function for precipitation data
ent from those commonly used to fit and describe observed was found to depend on the location of the station and the
precipitation data, as preliminary step for SPI computation. number of months in TS. The gamma distribution usually
Ten statistical distribution functions were tested to find the used as default is not always the best fit for SPI compu-
best fit in each of the 24 observation stations belonging to tation if many functions were tested. It was found that the
the five AEZs of Cameroon over the period 1951–2005 and logistic probability distribution remains the best choices in

123
 Earth Science Informatics

most cases for 3-month TS while above, burr shows the best Blain GC, de Avila AMH, Pereira VR (2018) Using the normality
fit. For 12 months and more, gamma, burr and logistic are assumption to calculate probability-based standardized drought
indices: selection criteria with emphases on typical events. Int J
found to be the best fits for many different stations. In all Climatol 38:e418–e436. https://doi.org/10.1002/joc.5381
cases, a significant difference was found between the SPIs Canal L (2005) A normal approximation for the chi-square distribution.
calculated with the best fit function and those calculated with Comput Stat Data Anal 48(4):803–808. https://doi.org/10.1016/j.
the default gamma distribution; the differences between both csda.2004.04.001
Choi SC, Wette R (1969) Maximum likelihood estimation of the param-
SPIs are more significant from 12-month TS with higher val- eters of the gamma distribution and their bias. Technometrics
ues of the RMSEs. 11(4):683–690. https://doi.org/10.2307/1266892
This study raises the importance and the necessity of a Cindric JJK, Pasaric Z (2012) Statistical distributions for the spi com-
preliminary study consisting in finding the best distribution putation. EMS Annual Meeting AbstractsBerlin: EMS (Berlin) pp
GF48: EMS2012–316
functions fitting the data and using them for the calculation
Cooray K (2010) Generalized gumbel distribution. J. Appl Stat
of SPI in order to reduce errors and increase the accuracy of 37(1):171–179. https://doi.org/10.1080/02664760802698995
the results. Dai M, Huang S, Huang Q et al (2020) Assessing agricultural drought
risk and its dynamic evolution characteristics. Agric Water Manag
Acknowledgements The authors are very grateful to the National 231(106):003. https://doi.org/10.1016/j.agwat.2020.106003
Meteorological Service of Cameroon for providing observed precip- DeHardt J (1971) Generalizations of the glivenko-cantelli theorem.
itation data. Annals Math Stat 42(6):2050–2055. https://www.jstor.org/stable/
2240133
Author Contributions “Material preparation, data collection and analy- Dogan S, Berktay A, Singh VP (2012) Comparison of multi-monthly
sis were performed by A.R. Gamgo Fotse and G.M. Guenang. The first rainfall-based drought severity indices, with application to semi-
draft of the manuscript was written by A.R. Gamgo Fotse. The authors arid konya closed basin, turkey. J Hydrol 470:255–268. https://doi.
A.R. Gamgo, G. M. Guenang, A.J. Komkoua Mbienda and D.A. Vondou org/10.1016/j.jhydrol.2012.09.003
commented and approved the final manuscript.” Gebremeskel Haile G, Tang Q, Sun S et al. (2019) Droughts in east
africa: Causes, impacts and resilience. Earth-Sci Rev 193:146–
Funding Not applicable. 161. https://doi.org/10.1016/j.earscirev.2019.04.015
Gebremichael HB, Raba GA, Beketie KT et al. (2022) Temporal and
Availability of data and materials The datasets generated and/or ana- spatial characteristics of drought, future changes and possible
lyzed during the current study are available on request by contacting drivers over upper awash basin, ethiopia, using spi and spei. Env-
the corresponding author. iron, Develop Sustain 1–39. https://doi.org/10.1007/s10668-022-
02743-3
Ghitany M, Al-Awadhi S (2002) Maximum likelihood estimation of burr
Declarations
xii distribution parameters under random censoring. J Appl Stat
29(7):955–965. https://doi.org/10.1080/0266476022000006667
Gidey E, Dikinya O, Sebego R et al. (2018) Modeling the spatio-
Ethical Approval Not applicable. temporal meteorological drought characteristics using the stan-
dardized precipitation index (spi) in raya and its environs, northern
Competing interests The authors declare no competing interests. ethiopia. Earth Syst Environ 2:281–292. https://doi.org/10.1007/
s41748-018-0057-7
Guenang GM, Mkankam Kamga F (2014) Computation of the stan-
References dardized precipitation index (spi) and its use to assess drought
occurrences in cameroon over recent decades. J Appl Meteorol Cli-
matol 53:2310–2324. https://doi.org/10.1175/JAMC-D-14-0032.
Angelidis P, Maris F, Kotsovinos N et al. (2012) Computation of
1
drought index spi with alternative distribution functions. Water
Guenang GM, Komkoua Mbienda A, Pokam Mba W et al. (2019) Sen-
Resour Manag 26:2453–2473. https://doi.org/10.1007/s11269-
sitivity of spi to distribution functions and correlation between its
012-0026-0
values at different time scales in central africa. Earth Syst Environ
Awchi TA, Kalyana MM (2017) Meteorological drought analysis in
3:203–214. https://doi.org/10.1007/s41748-019-00102-3
northern iraq using spi and gis. Sustainable Water Resour Manag
Han Z, Huang S, Huang Q et al (2019) Propagation dynamics from mete-
3:451–463. https://doi.org/10.1007/s40899-017-0111-x
orological to groundwater drought and their possible influence
Bae H, Ji H, Lim YJ et al (2019) Characteristics of drought propagation
factors. J Hydrol 578(124):102. https://doi.org/10.1016/j.jhydrol.
in south korea: Relationship between meteorological, agricultural,
2019.124102
and hydrological droughts. Natural Hazards 99(1):1–16. https://
Hayes M, Svoboda M, Wall N et al. (2011) The lincoln declaration on
doi.org/10.1007/s11069-019-03676-3
drought indices: universal meteorological drought index recom-
Berger VW, Zhou Y (2014) Kolmogorov-smirnov test: Overview.
mended. Bullet Am Meteorolog Soc 92(4):485–488. http://www.
Statistics reference online, Wiley statsref. https://doi.org/10.1002/
jstor.org/stable/26226865
9781118445112.stat06558
Hayes MJ, Svoboda MD, Wiihite DA et al (1999) Monitoring the
Bhaga TD, Dube T, Shekede MD et al (2020) Impacts of climate vari-
1996 drought using the standardized precipitation index. Bul-
ability and drought on surface water resources in sub-saharan africa
let Am Meteorolog Soc 80(3):429–438. https://doi.org/10.1175/
using remote sensing: A review. Remote Sens 12(24):4184. https://
1520-0477(1999)080<0429:MTDUTS>2.0.CO;2
doi.org/10.3390/rs12244184

123
Earth Science Informatics

He Z, Liang H, Yang C et al. (2018) Temporal-spatial evolution of Okpara J, Tarhule A (2015) Evaluation of drought indices in the niger
the hydrologic drought characteristics of the karst drainage basins basin, west africa. J Geogr Earth Sci 3:1–32. https://doi.org/10.
in south china. Int J Appl Earth Observation Geoinfor 64:22–30. 15640/jges.v3n2a1
https://doi.org/10.1016/j.jag.2017.08.010 Palmer W (1965) Meteorological drought. research paper no.
Horváth L (1993) The maximum likelihood method for testing changes 45. US Weather Bureau: Washington, DC. https://www.
in the parameters of normal observations. Annals Stat 671–680. droughtmanagement.info/literature/USWB_Meteorological_
https://doi.org/10.1214/aos/1176349143 Drought_1965.pdf
Huang S, Li P, Huang Q et al. (2017) The propagation from meteoro- Panahi H, Asadi S (2011) Estimation of the weibull dis-
logical to hydrological drought and its potential influence factors. tribution based on type-ii censored samples. Appl Math
J Hydrol 547:184–195. https://doi.org/10.1016/j.jhydrol.2017.01. Sci 5(52):2549–2558. http://www.m-hikari.com/ams/ams-2011/
041 ams-49-52-2011/panahiAMS49-52-2011.pdf
Huang S, Wang L, Wang H et al. (2019) Spatio-temporal characteris- Park S, Im J, Park S et al. (2017) Drought monitoring using high resolu-
tics of drought structure across china using an integrated drought tion soil moisture through multi- sensor satellite data fusion over
index. Agric Water Manag 218:182–192. https://doi.org/10.1016/ the korean peninsula. Agric Forest Meteorol 237:257–269. https://
j.agwat.2019.03.053 doi.org/10.1016/j.agrformet.2017.02.022
Kolmogorov AN (1933) Sulla determinazione empirica di una legge Penlap K, Matulla C, Storch H et al. (2004) Downscaling of gcm sce-
didistribuzione. Giorn Dell’inst Ital Degli Att 4:89–91 narios to assess precipitation changes in the little rainy season
Liu D, You J, Xie Q et al (2018) Spatial and temporal characteristics of (march-june) in cameroon. Climate Res 26:85–96. https://doi.org/
drought and flood in quanzhou based on standardized precipitation 10.3354/cr026085
index (spi) in recent 55 years. J Geosci Environ Protection 6(8):25– Pérez-Sánchez J, Senent-Aparicio J (2018) Analysis of meteorolog-
37. https://doi.org/10.4236/gep.2018.68003 ical droughts and dry spells in semiarid regions: a comparative
Liu L, Hong Y, Bednarczyk CN et al. (2012) Hydro-climatological analysis of probability distribution functions in the segura basin
drought analyses and projections using meteorological and hydro- (se spain). Theoretical Appl Climatol 133:1061–1074. https://doi.
logical drought indices: a case study in blue river basin, okla- org/10.1007/s00704-017-2239-x
homa. Water Resour Manag 26:2761–2779. https://doi.org/10. Pieper P, Düsterhus A, Baehr J (2020) A universal standardized precip-
1007/s11269-012-0044-y itation index candidate distribution function for observations and
Mage DT, Ott WR (1984) An evaluation of the methods of frac- simulations. Hydrol Earth Syst Sci 24(9):4541–4565. https://doi.
tiles, moments and maximum likelihood for estimating parameters org/10.5194/hess-24-4541-2020
when sampling air quality data from a stationary lognormal dis- Qin Y, Yang D, Lei H et al. (2015) Comparative analysis of drought
tribution. Atmospheric Environ 18(1):163–171. https://doi.org/10. based on precipitation and soil moisture indices in haihe basin of
1016/0004-6981(84)90239-7 north china during the period of 1960-2010. J Hydrol 526:55–67.
Maia R, Vivas E, Serralheiro R et al. (2015) Socioeconomic evaluation https://doi.org/10.1016/j.jhydrol.2014.09.068, drought processes,
of drought effects. main principles and application to guadiana and modeling, and mitigation
algarve case studies. Water Resour Manag 29:575–588. https://doi. Quenum GML, Klutse NA, Dieng D et al (2019) Identification of
org/10.1007/s11269-014-0883-9 potential drought areas in west africa under climate change
Massey FJ Jr (1951) The kolmogorov-smirnov test for goodness of fit. J and variability. Earth Syst Environ 3:429–444. https://doi.org/10.
Am Stat Assoc 46(253):68–78. https://doi.org/10.1080/01621459. 1007/s41748-019-00133-w
1951.10500769 Rahman M, Pearson LM (2007) Estimation in two-parameter exponen-
McKee T, Doesken N, Kliest J (1993) The relationship of drought tial distributions. J Stat Comput Simulat 70(4):371–386. https://
frequency and duration to time scales. in proceedings of the doi.org/10.1080/00949650108812128
8th conference of applied climatology, 17-22 january, ana- Ramachandran KM, Tsokos CP (2015) Mathematical statistics with
heim, ca. Am Meterolog Soc, Boston, Massachusetts pp 179– applications in R, Second Edition, Chapter 7. Academic Press.
184 https://doi.org/10.1016/C2012-0-07341-3
Meng XP, Zhao CQ, Huo L (2014) Maximum likelihood method for Raziei T (2021) Performance evaluation of different probability distri-
parameter estimation of weibull distribution model based on fruit bution functions for computing standardized precipitation index
fly optimization algorithm. In: Applied Mechanics and Materi- over diverse climates of iran. Int J Climatol 41(5):3352–3373.
als, Trans Tech Publ 3508–3511. https://doi.org/10.4028/www. https://doi.org/10.1002/joc.7023
scientific.net/AMM.602-605.3508 Robertson G (1969) Computation of the noncentral chi-square distri-
Mitchell B (1971) A comparison of chi-square and kolmogorov- bution. Bell Syst Technical J 48(1):201–207. https://doi.org/10.
smirnov tests. Area 237–241. https://www.jstor.org/stable/ 1002/j.1538-7305.1969.tb01111.x
20000590 Santé N, N’Go YA, Soro GE et al (2019) Characterization of meteoro-
Ndayiragije JM, Li F, Nkunzimana A (2022) Assessment of two drought logical droughts occurrences in côte d’ivoire: case of the sassandra
indices to quantify and characterize drought incidents: A case study watershed. Climate 7(4):60. https://doi.org/10.3390/cli7040060
of the northern part of burundi. Atmosphere 13(11):1882. https:// Schuster S (2012) Parameter estimation for the cauchy distribution. In:
doi.org/10.3390/atmos13111882 2012 19th International conference on systems, signals and image
Net F (2019) Cameroon livelihood zone map and descriptions. FEWS processing (IWSSIP), pp 350–353
NET, Washington, DC Spinoni J, Naumann G, Carrao H et al (2014) World drought frequency,
Nfornkah BN, Enongene K, Kaam R, et al. (2021) Growth potential and duration, and severity for 1951–2010. Int J Climatol 34(8):2792–
sustainability of economically important rattan species in agro- 2804. https://doi.org/10.1002/joc.3875
ecological zones of cameroon. INBAR Working Paper. https:// Stagge JH, Tallaksen LM, Gudmundsson L et al (2015) Candidate dis-
www.inbar.int/wp-content/uploads/2021/06/June-2021_Growth- tributions for climatological drought indices (spi and spei). Int J
Potential-and-Sustainability-of-Economically-Important- Climatol 35(13):4027–4040. https://doi.org/10.1002/joc.4267
Rattan-Species-in-Agro-Ecological-Zones-of-Cameroon.pdf Stephens MA (1970) Use of the kolmogorov-smirnov, cramer-von
Ojha SS, Singh V, Roshni T (2021) Comparison of meteorological mises and related statistics without extensive tables. J Royal Stat
drought using spi and spei. Civ Eng J 7:2130–2149. https://doi. Soc: Series B (Methodological) 32(1):115–122. https://doi.org/10.
org/10.28991/cej-2021-03091783 1111/j.2517-6161.1970.tb00821.x

123
 Earth Science Informatics

Streit RL, Luginbuhl TE (1994) Maximum likelihood method for prob- Wilhite DA, Glantz MH (1985) Understanding: the drought phe-
abilistic multihypothesis tracking. In: Signal and data processing nomenon: the role of definitions. Water Int 10(3):111–120. https://
of small targets 1994, SPIE, pp 394–405, https://doi.org/10.1117/ doi.org/10.1080/02508068508686328
12.179066 Wilhite DA, Svoboda MD, Hayes MJ (2007) Understanding the com-
Sylla MB, Nikiema PM, Gibba P et al. (2016) Climate change over west plex impacts of drought: A key to enhancing drought mitigation
africa: Recent trends and future projections. Adaptation to climate and preparedness. Water Resour Manag 21:763–774. https://doi.
change and variability in rural West Africa pp 25–40. https://doi. org/10.1007/s11269-006-9076-5
org/10.1007/978-3-319-31499-0_3 Wu SJ (2002) Estimations of the parameters of the weibull distribution
Tirivarombo S, Osupile D, Eliasson P (2018) Drought monitoring with progressively censored data. J Japan Stat Soc 32(2):155–163.
and analysis: standardised precipitation evapotranspiration index https://doi.org/10.14490/jjss.32.155
(spei) and standardised precipitation index (spi). Physics and Zarei AR, Shabani A, Moghimi MM (2021) Accuracy assessment of the
Chemistry of the Earth, Parts A/B/C 106:1–10. https://doi.org/10. spei, rdi and spi drought indices in regions of iran with different
1016/j.pce.2018.07.001 climate conditions. Pure Appl Geophys 178:1387–1403. https://
Touma D, Ashfaq M, Nayak MA et al (2015) A multi-model and multi- doi.org/10.1007/s00024-021-02704-3
index evaluation of drought characteristics in the 21st century. Zhang Y, Li Z (2020) Uncertainty analysis of standardized precipitation
J Hydrol 526:196–207. https://doi.org/10.1016/j.jhydrol.2014.12. index due to the effects of probability distributions and parameter
011 errors. Front Earth Sci 8:76. https://doi.org/10.3389/feart.2020.
Tsakiris G, Vangelis H (2005) Establishing a drought index incorporat- 00076
ing evapotranspiration. European Water 9(10):3–11 Zhao M, Huang S, Huang Q et al (2019) Assessing socio-economic
Van Rooy M (1965) A rainfall anomally index independent of time and drought evolution characteristics and their possible meteorological
space, notos. Notos 14:43–48 driving force. Geomatics, Natural Hazards and Risk 10(1):1084–
Vicente-Serrano SM, Beguería S, López-Moreno JI (2010) A multi- 1101. https://doi.org/10.1080/19475705.2018.1564706
scalar drought index sensitive to global warming: the standardized
precipitation evapotranspiration index. J Climate 23(7):1696–
1718. https://doi.org/10.1175/2009JCLI2909.1 Publisher’s Note Springer Nature remains neutral with regard to juris-
Vondou DA, Guenang GM, Djiotang TLA et al (2021) Trends and dictional claims in published maps and institutional affiliations.
interannual variability of extreme rainfall indices over cameroon.
Sustainability 13(12):6803. https://doi.org/10.3390/su13126803 Springer Nature or its licensor (e.g. a society or other partner) holds
Wang W, Ertsen MW, Svoboda MD et al (2016). Propagation of exclusive rights to this article under a publishing agreement with the
drought: from meteorological drought to agricultural and hydro- author(s) or other rightsholder(s); author self-archiving of the accepted
logical drought. https://doi.org/10.1155/2016/6547209 manuscript version of this article is solely governed by the terms of such
publishing agreement and applicable law.

Authors and Affiliations

A.R. Gamgo Fotse1 · G.M. Guenang1,2 · A.J. Komkoua Mbienda1,2,3 · Derbetini A. Vondou1

A.R. Gamgo Fotse 1 Department of Physics, University of Yaounde 1, Yaounde,

gamgoraissa94@gmail.com Cameroon
A.J. Komkoua Mbienda 2 Department of Physics, University of Dschang, Dschang,
kombiend@yahoo.fr Cameroon
Derbetini A. Vondou 3 Earth System Physics Section, The Abdus Salam ICTP,
derbetini@yahoo.fr Trieste 34151, Italy

123