Download as pdf or txt
Download as pdf or txt
You are on page 1of 23

GeoJournal (2023) 88:S85–S107

https://doi.org/10.1007/s10708-021-10543-x (0123456789().,-volV)
( 01234567
89().,-volV)

Long-term forecasting of tropical cyclones over Bay


of Bengal using linear and non-linear statistical models
Sweta Sen . Narayan Chandra Nayak . William Kumar Mohanty

Accepted: 2 November 2021 / Published online: 18 November 2021


 The Author(s), under exclusive licence to Springer Nature B.V. 2021

Abstract Forecasting tropical cyclones with climate models is to improve the forecasting accuracy of
and physical variability and observed cyclonic distur- tropical cyclones. We have shown that the intensifi-
bances has been developed over the years for all the cation rate of the severe cyclonic storms over the Bay
ocean basins successfully and is still one of the of Bengal has been significant and increasing over the
priorities for disaster risk reduction policymaking. years. Results show that the ANN-NARX model with
This study attempts to forecast seasonal cyclonic sea surface temperature and near-surface wind speed
disturbances and severe cyclonic storms over the Bay as predictors is the best performance model for long-
of Bengal, where about 80% of the tropical cyclones of term forecasting of cyclonic disturbances. Hence, the
the North Indian Ocean are formed. We have used distribution of cyclonic disturbances is non-linear. The
three time-series models, namely, the seasonal autore- correlations between observed and predicted occur-
gressive integrated moving average with exogenous rences are 0.80 and 0.85 for cyclonic disturbances and
variables (SARIMAX) model, artificial neural net- severe cyclonic storms, respectively, corroborating,
work-nonlinear autoregressive with exogenous vari- by and large, the forecasting accuracies of some
ables (ANN-NARX) model, and the hybrid model. previous studies. The forecasting of cyclonic distur-
The basic purpose of considering three different bances indicates that they will vary from 5 to 13
annually and there will be, on average, one severe
cyclonic storm per year. The likelihood of occurrence
S. Sen (&) of severe cyclonic storms is most significant in the
Advanced Technology Development Centre, Indian post-monsoon season. This forecast till 2050 would
Institute of Technology Kharagpur, Kharagpur, help the scientific community and policymakers
West Bengal, India
e-mail: sensweta@iitkgp.ac.in
significantly for applications and good disaster risk
governance.
N. C. Nayak
Department of Humanities and Social Sciences, Indian Keywords Tropical cyclones  Seasonal
Institute of Technology Kharagpur, Kharagpur,
West Bengal, India
forecasting  SARIMAX  Artificial Neural Network 
e-mail: ncnayak@hss.iitkgp.ac.in Hybrid model  Predictors  Bay of Bengal

W. K. Mohanty
Department of Geology and Geophysics, Indian Institute
of Technology Kharagpur, Kharagpur,
West Bengal, India
e-mail: wkmohanty@gg.iitkgp.ac.in

123
S86 GeoJournal (2023) 88:S85–S107

Introduction disaster risk reduction, focusing on multi-sectoral


policies to make a resilient planet for future genera-
The forecasting of cyclonic disturbances has improved tions. Needless to say, to understand the risk, it is
significantly over the years due to the application of essential to forecast the disaster occurrences. The
modern modeling techniques. Several studies (Bengts- present study, thus, attempts to predict the tropical
son et al., 1995; Haarsma et al., 1993) have shown that cyclones in the Bay of Bengal for the long run. While
simulating tropical storms from the observed tropical doing so, we examine which statistical forecasting
storms using different climatological and physical method is appropriate to employ for long-term fore-
characteristics gives close resemblance. Statistical casting of tropical cyclone counts. In addition, we also
forecasting of tropical cyclones has been performed analyze the long-term trends and intensification rates
globally for frequency forecasting in different ocean of cyclonic disturbances and severe cyclonic storms
basins such as the Australian basin (Nicholls, 1992), over the past 128 years (1891–2018). The findings of
the western North Pacific (Chan et al., 1998), the this study would help the planners and policymakers
Atlantic basin (Gray et al., 1993, 1994) and so on. for adaptation and preparedness against the devastat-
About 80% of the tropical cyclones in the North ing effects of tropical cyclones.
Indian Ocean are formed in the Bay of Bengal with an This study is distinct in the following ways. First,
approximate ratio of 4:1 in its occurrences compared based on comparison across linear, non-linear and
to the Arabian Sea (Neumann, 1993; IMD, 2011). hybrid models, we eventually perform statistical
Only 10% of cyclones formed over the Bay of Bengal forecasting using non-linear models whose application
dissipate in the basin (Bhardwaj & Singh, 2020). The in forecasting cyclonic activities is limited. Second,
Bay of Bengal has witnessed some extremely devas- we consider three critical exogenous variables for
tating cyclones over the years, causing a global forecasting cyclonic disturbances. This makes it easier
casualty rate of more than 75% in the adjoining to compare the contribution of each exogenous
countries (Chowdhury, 2002). The two most recent variable to the cyclonic disturbances over the Bay of
cyclones over the Bay of Bengal were FANI and Bengal. However, the various predictors used for
AMPHAN, which resulted in an economic loss of forecasting cyclonic disturbances have satellite record
about $1.81 billion and $15.38 billion and over 90 and data from 1951, so our forecasting is carried out using
84 casualties, respectively. The coasts of West Bengal, the 1951–2018 dataset.
Odisha, Andhra Pradesh, and Tamil Nadu in India, Accordingly, the remainder of the paper is orga-
entire Bangladesh and the adjacent Myanmar coast nized as follows. Section 2 presents the literature
remain at high risk due to tropical cyclones in the Bay review with particular emphasis on forecasting tech-
Bengal (Bhardwaj et al., 2020). niques used over the Bay of Bengal. Section 3
The advantage of long-run forecasting is that with describes the data and methodology thoroughly. In
its help, we can determine the cyclonic disturbances Section 4, we present the results of tropical cyclones
beforehand for better preparedness. Long-term fore- over the Bay of Bengal with discussion. Section 5
casting is gaining importance due to its societal and provides the summary and concludes the study.
scientific significance for achieving disaster risk
reduction goals. Several world organizations are
known for their actions towards disaster risk reduction Overview of literature
and climate change effects by forecasting for the long-
term using simulations and assuming scenarios. The There have been studies in the past (Villarini et al.,
prominent among them are the Sendai framework of 2019; Vitart et al., 1997, 1999; Zhang & Villarini,
disaster risk reduction under the United Nations Office 2019) on tropical cyclone forecasting with good fit
of Disaster Risk Reduction (UNDRR), the Sustainable models and high correlations with the observed
Development Goals (SDGs), and the Intergovernmen- occurrences. Statistical and dynamic forecasting mod-
tal Panel for Climate Change (IPCC). els such as European Centre for Medium-Range
The UNDRR raises the importance of disaster risk Weather Forecast’s (ECMWF) dynamical system
governance for all countries. Its primary thrust is to (Vitart & Stockdale, 2001), global atmospheric gen-
develop appropriate local and national strategies for eral circulation models (AGCMs) (LaRow et al.,

123
GeoJournal (2023) 88:S85–S107 S87

2010), and general circulation model (GCM) coupled forecast model for storm surges generated by a tropical
with the ocean–atmosphere-land–sea ice model (Sun cyclone using a dynamic non-linear simulation model
& Ahn, 2011) are quite common. However, some with oceanographic and meteorological parameters as
recent studies have observed that GCM has a relatively predictors, which can predict the landfall. Similarly,
lesser ability to predict the inter-annual variability due Kutty et al. (2020) have provided a statistical evalu-
to its incorrect response of the observed remote ation of tropical cyclone forecasts over the Bay of
response (Nair et al., 2013; Rajeevan et al., 2012). Bengal using the limited area model. Comparing
Consequently, the forecasters tend to explore the different models, they found that in performing
multi-model ensemble approach (Acharya et al., forecasts of tropical cyclone tracks, Kalman filter
2014). three-dimensional variational data assimilation sys-
Some other models, which are found to have strong tem is superior to the three-dimensional variational
abilities to forecast tropical cyclonic occurrences and (3DVAR) data assimilation system in WRF. Srinivas
landfalls, are the threshold autoregressive model–a et al. (2013) used the ARW model to predict tropical
hybrid statistical model combining the regression cyclones in the Bay of Bengal, considering 21
model and the time-series autoregressive model– cyclones with a depth of 9 km resolution. Bhardwaj
(Feng & Luo, 2014), the auto-regressive integrated et al. (2020) applied the extreme value theory to
moving average (ARIMA) model (Geetha & Narisa, forecast the probabilities of wind speed of tropical
2016), projection pursuit regression with smooth cyclones over the Bay of Bengal. They predicted that
multiple additive regression technique (SMART) at least one tropical cyclone of wind speed C 64 kts
generalization (Chan et al., 1998), and the least sum can occur every year and that the return period of a
of absolute deviations regression (Gray et al., super cyclone can be over an interval of 6 years.
1993, 1994). Studies have also been carried out on Artificial neural network (ANN) models have been
long-term forecasting of tropical cyclones using applied extensively for non-linear meteorological
support vector regression model (Richman & Leslie, events (Chattopadhyay, 2007; Chattopadhyay & Chat-
2012; Wijnands et al., 2014), multilayer perceptron topadhyay, 2008; Gardner & Dorling, 1998). A non-
(Nath et al., 2016) and feed-forward neural networks linear ANN with a single-hidden-layer feed-forward
(FFNN) (Yip & Yau, 2012). Recently, Mitchell and neural network was used to predict post-monsoon
Camp (2021) have applied the Conway–Maxwell– rainfall in the southern peninsular region of India by
Poisson distribution to model seasonal tropical Acharya et al. (2014), which resulted in significantly
cyclone counts for East China and forecasted its higher skill than individual GCM. Exponential regres-
tropical cyclones over a few months ahead. The sion and ANN to forecast post-monsoon rainfall were
Conway–Maxwell–Poisson denotes the under-disper- applied by Chattopadhyay et al. (2010) while using
sion and improves the predictive skill. However, it rainfall amounts and SST anomalies as predictors. Ali
seems to have the disadvantages of forecasting only et al. (2007) predicted the positions of cyclones for
for the short term. 24 h in advance using 12-h location at the 6-hourly
Over the years, the Indian Meteorological Depart- intervals by employing the single-hidden layer ANN
ment (IMD) has developed models like regional non- model over the North Indian Ocean. Chaudhuri et al.
hydrostatic mesoscale modeling system–weather (2017) used swarm intelligence followed by neural
research forecasting data assimilation-weather networks to forecast the maximum sustained wind
research and forecasting-advance weather research speed of the cyclones over the Bay of Bengal and
and forecasting system (WRFDA-WRF-ARW), hur- found their methods giving lesser error percentage
ricane WRF and the numerical weather prediction than other methods like operational tropical cyclone
method–to forecast different components over the Bay intensity forecasts and high-resolution ARW model.
of Bengal namely, cyclone genesis potential, cyclone Among all the models, evidently, ANN seems to
track, cyclone intensity, rapid intensification index, have provided better prediction results covering both
and decaying intensity after the landfall. However, monsoon prediction (Acharya et al., 2014) and
they are also purely short-term in nature (IMD, 2019). cyclone intensity prediction (Baik & Paek, 2000).
Among other forecasting studies over the Bay of Recently, the ANN models like the Multilayer
Bengal, Dube et al. (1997) have developed a 48-h perceptron model, Radian basis function network

123
S88 GeoJournal (2023) 88:S85–S107

model and Generalized regression neural network provides details about frequencies, probabilities, gen-
model were employed for forecasting pressure drop esis and impacts of cyclonic disturbances over the Bay
and maximum sustained wind speed of tropical of Bengal. According to the current classification,
cyclones over the Arabian Sea (Sarkar et al., 2021). tropical cyclones can be divided into eight categories
The Cooperation search algorithm approach of the starting with the low-pressure area (Bhardwaj &
extreme learning method model (an AI model) is yet Singh, 2020). However, till 2015, IMD’s classifica-
another technique that seems to have a novel multi- tions were limited to three categories only, namely,
scale nonstationary hydrological prediction tool (Feng depression, cyclonic storm, and severe cyclonic storm,
et al., 2021). as shown in Table 1. Accordingly, we analyze the data
It is clear from the extant literature that though in two different ways. First, we take all the seven
several studies have been carried out over the Bay of categories (except low-pressure area) from depression
Bengal and different techniques have been applied to (greater than 31 kmph) to super cyclonic storms
forecast tropical cyclones, the application of non- (greater than 222 kmph) together and denote them as
linear statistical models with ANN seems limited. cyclonic disturbances. Then, we consider cyclones
Moreover, a comparison of the linear model with the having maximum sustained wind speed greater than
non-linear one and applying a hybrid model combin- 89 kmph (i.e., severe cyclonic storms, very severe
ing both are hardly found in the forecasting literature, cyclonic storms, extremely severe cyclonic storms and
especially on tropical cyclones. The present study, super cyclonic storms) and denote them as severe
thus, attempts to employ both linear, non-linear and cyclonic storms. Accordingly, we run models for
hybrid statistical models and finds the superiority of cyclonic disturbances for all seven categories fol-
one over the other while forecasting the occurrences of lowed by severe cyclonic storms (Table 1).
tropical cyclones over the Bay of Bengal for a long- To analyze trends, we consider the cyclonic distur-
term period. bances and severe cyclonic storms formed during
1891–2018 over the Bay of Bengal lying between 77
and 97E and 5 and 25N. We collected data from the
Data and methodology Tracks of Storms and Depressions over the Bay of
Bengal published by the IMD. Data on the three
Our study aims to find a statistical model for exogenous variables used for forecasting were
forecasting tropical cyclones using predictor(s), which extracted from the CORDEX dataset available with
have a strong relationship with a tropical cyclone’s the Indian Institute of Tropical Meteorology for
formation to dissipation. We introduce three statistical 1951–2018. According to cyclonic seasons, we
models, namely, (1) the seasonal autoregressive divided the yearly cyclones into four seasons: Jan-
integrated moving average with exogenous variable uary–February (JF)—winter season, March–April-
(SARIMAX) model, (2) the ANN non-linear autore- May (MAM)—pre-monsoon season, June-July–Au-
gressive with exogenous variable (ANN-NARX) gust-September (JJAS)—monsoon season, and Octo-
model, and (3) the hybrid model comprising both the ber–November-December (OND)—post-monsoon
above linear and non-linear components, and compare season. However, we have eliminated the JF season
them to find the best model that can forecast long-term from our analysis to avoid noise in the forecasting
tropical cyclones. The details about the data are process as the cyclonic disturbances were very few in
described below. the winter seasons.

Sources of data Method of analysis

Two leading institutions provide information on Selection of predictors


tropical cyclones for the North Indian Ocean, namely,
the IMD and the Joint Typhoon Warning Centre Tropical cyclonic activities are mainly dependent on
(JTWC). While JTWC has archived data from 1972, the characteristics of large-scale circulation, which are
IMD has it from 1861. We use IMD dataset as we need the fundamental reasons behind the seasonal forecast-
a time-series dataset for a longer period. The IMD ing of these disturbances (Gray, 1979). The latter seem

123
GeoJournal (2023) 88:S85–S107 S89

Table 1 Classification of cyclonic disturbances in the North Indian Ocean (Arabian Sea and Bay of Bengal)
Category of cyclonic Max sustained surface Max sustained surface
disturbances winds (kmph) winds (knots)

Low pressure area \ 31 \ 17


Tropical depression 31–49 17–27 Cyclonic
Deep depression 50–61 28–33 Disturbances
(CDs)
Cyclonic storm 62–88 34–47
Severe cyclonic storm 89–118 48–63 Severe Cyclonic
Very severe cyclonic 119–221 64–119 Storms (SCSs)
storm
Extremely severe 167–221 90–119
cyclonic storm
Super cyclonic storm [ 222 [ 119
The eight categories of cyclonic disturbances have been referred from Bhardwaj & Singh (2020)

to have strong relationships with sea surface temper- together as predictors, which have the lowest corre-
ature, surface wind speed, and sea surface pressure. lation. To make the time-series stationery, we have
The positive relationship between sea surface temper- differentiated each exogenous variable by order 3. A
ature and tropical cyclone intensity has been verified conceptual framework of the forecasting system of
in several studies (Emanuel, 1988; Merrill, 1987; cyclonic disturbances and severe cyclonic distur-
Miller, 1958) and the former is proven as a predom- bances is shown in Fig. 1. The analysis of SARIMAX,
inant predictor for the increased intensity of tropical ANN-NARX, and the hybrid model is programmed in
cyclonic disturbances (Evans, 1993). Tropical ocean MATLAB. The detailed formulations of these three
basins witness peak occurrences of cyclonic distur- models are given as follows.
bances when sea surface temperature is high (Gutzler
et al., 2013; Zhao & Raga, 2015). It is further observed Model Formulation
that the lower the central pressure of the storms, the
more will be the intensity of the storms. The minimum The SARIMAX model The SARIMAX is a
sustained pressure is strongly correlated with the conditional mean time-series model. In this model, a
intensities of the strongest tropical cyclones, which variable’s future value is assumed to be a linear
can be explained by thermodynamics and the effects of function of its past observations, exogenous variables,
water substance on density (Emanuel, 1988, 1995). and random errors. The performance of the
Surface wind speed and sea-surface pressure affect air- SARIMAX model is considered better than that of
sea interactions in the Bay of Bengal (Emanuel, 1995; the classical ARIMA model (Vagropoulos et al.,
Reddy et al., 2018). This leads to a very active bed for 2016), though, in the past, natural disasters like
cyclonic disturbances and consequently, maximum earthquakes and hydroclimatic events have also been
tropical cyclones are formed in the Bay of Bengal forecasted by the latter (Amei et al., 2012; Islam et al.,
region (Bhardwaj & Singh, 2020). 2021). The existing studies have highlighted the
The pairwise correlation coefficients across the effectiveness of SARIMAX in fields like electricity,
above-mentioned three exogenous variables (see rainfall, traffic, and sales, to name a few (Elamin &
Table 2) indicate strong correlations between sea- Fukushige, 2018; Farajzadeh & Alizadeh, 2018).
surface pressure and sea-surface temperature, and sea- However, for forecasting cyclones, so far, its
surface pressure and near-surface wind speed. So, to application seems to have been limited.
avoid spurious results, apart from estimating the A SARIMAX (p,d,q) model of time-series cyclonic
models with individual predictors, we consider sea- disturbances using two exogenous variables (X1 and
surface temperature and near-surface wind speed X2) applies the following equation:

123
S90 GeoJournal (2023) 88:S85–S107

Table 2 Correlation between exogenous variables (predictors)


Exogenous variables (predictors) Sea-surface temperature (TS) Near-surface wind speed (WIND) Sea-surface pressure (PSL)

Sea-surface temperature (TS) 1.0000


Near-surface wind speed (WIND) 0.0780 1.0000
Sea-surface pressure (PSL) - 0.3268* - 0.5999* 1.0000
*Means significant at 0.05 level

Fig. 1 The framework of the model formulations and forecast- Note M1, M2, …, M12 are model numbers which are run with
ing system used in this study for forecasting tropical cyclonic combinations of predictors
disturbances and severe cyclonic storms over the Bay of Bengal.

 
1  /1 L  . . .  /p Lp ð1  U1 L  . . .  Ua La Þ While employing SARIMAX, the cyclonic distur-
   bance counts of each season are taken as response
1  Ld yt ¼ c þ X1 b1 þ X2 b2 þ 1 þ h1 L þ h2 L2
  variables and the exogenous variables become the
þ . . . þ hq Lq Þ 1 þ H1 L þ . . . þ Hb Lb et
predictors. As the data are found to be non-stationary,
ð1Þ the differentiation (i.e., d-degree of integration) of the
The underlying process has p autoregressive, d response variable is considered to make the time-
degrees of integration and q moving average terms, series stationery. To decide on the number of times the
and seasonal AR(a) and MA(b) that generate the time- series has to be differentiated to achieve stationarity,
series of the form as in Eq. 1. yt and et are the actual the augmented dickey fuller (ADF) test is employed.
values and random errors at time t, respectively. b1, b2, The model selection is based on the autocorrelation
/i (i = 1, 2, …., p) and hj (j = 0, 1, 2, …., q) are model function (ACF) and partial autocorrelation function
(PACF) plots to determine the p-autoregressive and q-
parameters. Random errors,et , are assumed to be
moving average. Further, the chosen model satisfies
independently and identically distributed with mean
the lowest Akaike information criterion (AIC) and
zero and constant variance r2.

123
GeoJournal (2023) 88:S85–S107 S91

Bayesian information criterion (BIC) (Bhaumik, the FFNN. The weights are linked together into the
2015). parameter vector h.
The FFNN has two advantages. First, the input is
The ANN-NARX model In recent years, the ANN more accurate as a series–parallel architecture is
model as a forecasting tool seems to have gained created during the training of the network in which the
momentum, thanks to its pattern recognition and true output is used instead of feeding back the
classification capabilities. The ANN is a non-linear estimated output. Second, the resulting network has
model, considered an alternative technique in a pure feed-forward architecture, and static back-
forecasting natural hazards (Sahoo et al., 2009). The propagation can be used for training. Our study applies
NARX is a recurrent dynamic network model based on the close loop function for converting NARX net-
the linear autoregressive with exogenous variable(s). works from the series–parallel architecture, which is
It has been applied for various time-series modeling to then implemented for multi-step predictions for
forecast wind speed, rainfall, droughts, electricity, 32 years (2019–2050).
cash demand, exchange rate, etc. (Cadenas et al., We employ the FFNN NARX model to estimate
2016; Morid et al., 2007; Vaz et al., 2016). This and forecast tropical cyclonic disturbances and severe
method carries a more efficient learning process than cyclonic storms. Along with the past cyclonic distur-
other neural networks as it converges and generalizes bances yt, we consider other predictors xt. et is the error
better. The training of the NARX model follows Feed- term:
Forward Neural Networks (FFNNs), which are
yt ¼ f ðxt1 . . .; xtd ; yt1 ; . . .; ytd Þ þ et ð4Þ
universal function approximators of a class of linear
regression models (Georga et al., 2017), as shown in We choose the Levenberg–Marquardt optimiza-
Eq. (2): tion, a back-propagation algorithm, as the training
X
m function updates the weights and biases. Figure 2
f ðx; hÞ ¼ hj / j ð x Þ ð2Þ shows the neural network architecture of our study.
j¼0 The model is trained and the errors are visualized in
graphs to see the fitting. The model having the least
where h ¼ ðh0 ; h1 ; . . .; hm ÞT represents model param- error and the best fitting is chosen for forecasting.
eters and / ¼ ð/0 ; /1 ; . . .; /m ÞT represents a vector of
non-linear parametric basis function, with h0 as the The hybrid model A hybrid model combines both
bias parameter and /0 (x) is equal to unity. Each basis linear and non-linear models (SARIMAX and ANN-
function /j (x) constitutes itself a non-linear function NARX models) and consequently, it may carry the
of a linear combination of the input x 2 Rd , which is advantages of both. It has been employed earlier for
Pd
ð1Þ predicting various hydrological and wind speed
represented as /j ðxÞ ¼ hð wji xi Þ. Here, the coeffi- parameters (Liu et al., 2012; Pektaş & Cigizoglu,
i¼0
ð1Þ 2013), including drought (Mishra et al., 2007) and
cients wji with i = 1, …, d represent the weights,
ð1Þ
economic and financial variables (Yang et al., 2017;
wj0 (x0 = 1) represents the biases corresponding to the Zhang, 2003). The past empirical studies have proven
links between the input and the jth node of the hidden that the forecasting accuracy can often be improved
layer of the FFNN, and h is a non-linear differentiable under a combined model (Clemen, 1989; Newbold &
activation function, which typically follows the Granger, 1974).
sigmoidal function. Accordingly, the output of an For applying the hybrid model, we assume that the
FFNN can be mathematically represented as: distribution of cyclonic disturbances has both linear
and non-linear components (Fig. 3). The cyclonic
X
m X
d
yl ðx; wÞ ¼ w2lj h w1ji xi ð3Þ disturbances are first regressed with the SARIMAX
j¼0 i¼0 model. The residuals are then modeled with the ANN-
NARX model. As it is necessary to check whether the
where the coefficients w2lj with l = 1, …, L represent residuals of the SARIMAX model have non-linearity,
the weights and w2l0 are the biases of the links between we perform the Brock, Dechert, and Scheinkman
the hidden layer and the lth node of the output layer of (BDS) test. It helps one detect the general stochastic

123
S92 GeoJournal (2023) 88:S85–S107

Fig. 2 An artificial neural network architecture showing a non- framework has 9 hidden layers for each variable and 9 lags. It
linear autoregressive model (ANN-NARX) with exogenous passes through the non-linear function and generates output
variable–x(t) and input variable–y(t). Note The above y(t) and forecasts the series

Fig. 3 The hybrid model estimation framework

non-linearity in a time-series (Mann, 1945). The null et ¼ yt  L^t ð6Þ


hypothesis of the BDS test is that the series is
independent and identically distributed with an where L^t is the forecast value for time t. Hence, the
asymptotically normal distribution. Once the null hybrid model forecast will be:
hypothesis is rejected i.e., the presence of non-
y^t ¼ L^t þ N^t þ 2 ð7Þ
linearity in the series is approved, we apply ANN-
NARX to the time-series. Accordingly, the hybrid where 2 is the error of the hybrid model.
model can be represented by the following equations
(Eq. 5–7). Measuring forecasting performances
The time-series
yt ¼ Lt þ Nt ð5Þ We employ four different forecasting consistency
measures to compare the performances of SARIMAX,
where Lt represents the linear component and Nt ANN-NARX, and the hybrid models. They are the root
represents the non-linear component. The residuals mean square error (RMSE), the mean absolute error
contain the Nt, and are modeled as: (MAE), the Nash–Sutcliffe coefficient of efficiency
(NSC), and mean absolute scaled error (MASE). The

123
GeoJournal (2023) 88:S85–S107 S93

assessment of the model performance necessitates a cyclonic storms. Further, we also estimate the Pear-
goodness of fit measure like NSC and an absolute error son’s correlation coefficient of the best fit models
measure like RSME (Legates & McCabe, 1999). between the predicted ðPi Þ and the observed ðOi Þ for
RMSE and MAE are the most commonly used scale- testing the accuracy (LaRow et al., 2010; Vitart et al.,
dependent measures, dependent on absolute errors or 2007).
squared errors. Since the time-series are of the same
scale, RMSE and MAE are considered good indica-
tors. They estimate the error between the actual and Results and discussion
the estimated. The lesser the value, the more efficient
is the estimated model. Contrarily, the NSC (Nash & Temporal trends and intensification rates
Sutcliffe, 1970) is a relative measure of testing
forecasting accuracy, ranging between - ? and 1. To examine the trends of tropical cyclones over the
More the value closer to 1, it would be nearer to the Bay of Bengal, we take the yearly counts of cyclonic
perfect fit. If the NSC has negative values, the storms and severe cyclonic storms. The seasonal
forecasting is less accurate. Many earlier studies have severe cyclonic storms and cyclonic disturbance
employed NSC to test the accuracy and efficiency of counts are shown in Fig. 4. It is observed that the
the forecasting models (Barzegar et al., 2016; Hwang cyclonic disturbances tend to be maximum in mon-
et al., 2012). Another measure is MASE, which puts a soon seasons followed by post-monsoon seasons. Still,
thrust on scaled errors for forecasting. It is useful when severe cyclonic storms are the highest in the post-
the series has values close to zero or negative and has a monsoon seasons, followed by the pre-monsoon
lesser influence on the outliers (Hyndman & Koehler, seasons. Most of the cyclonic disturbances formed in
2006). All the above measures are expressed as the Bay of Bengal during the pre-monsoon seasons
follows (Eq. 8–11): result in severe cyclonic storms. Post-monsoon sea-
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sons record the highest number of occurrences of
1X n
severe cyclonic storms. As each season has different
RMSE ¼ ð Pi  O i Þ 2 ð8Þ
n i¼1 characteristics, we undertake a seasonal analysis.
Over time, there has been an increase in the
where Pi and Oi are the predicted and observed frequency of tropical cyclone counts in the Bay of
cyclonic disturbances, respectively and n represents Bengal. We analyze changes in the trends using the
the number of observations. intensification rate. The intensification rate is defined
1X n as the ratio between severe cyclonic storm occurrences
MAE ¼ jPi  Oi j ð9Þ and the total occurrences of cyclonic disturbances. The
n i¼1
results suggest that only 5% of the cyclonic distur-
Pn bances have intensified to severe cyclonic storms in
ð Pi  O i Þ 2
NSC ¼ 1  Pi¼1
n  2 ð10Þ monsoon seasons. However, the frequency of severe
i¼1 O  Oi cyclonic storms has increased through the past
decades as there has been at least one severe cyclonic
where O represents the average of the observed
storm in the Bay of Bengal every year. The intensi-
occurrences.
fication rate has increased from 1961 till 2000, with
MASE ¼ meanðjqi jÞ ð11Þ the highest intensification rate being the 22nd pentad.
Having noticed a decreasing trend since 2000, a rising
where qi is the scaled error, independent of the scale of trend has been recorded since 2011 (Fig. 5).
the data, and To test the significance and the movement of the
Pi  Oi trend, the non-parametric Mann–Kendall trend test
qi ¼ 1
Pn (Mann, 1945; Yue et al., 2002) is conducted. We reject
n1 i¼2 jOi  Oi1 j
the null hypothesis, which assumes that the intensifi-
The accuracy test is conducted to select the best fit cation rates and years are independent, implying no
model. Consequently, the chosen model is employed trend. There is an increasing trend in the intensifica-
for forecasting cyclonic disturbances and severe tion rates (Table 3). Further, there is a rising

123
S94 GeoJournal (2023) 88:S85–S107

Cyclonic Disturbances Severe Cyclonic Storms

1217
610

447

235
142
136

59

31

PRE-MONSOON MONSOON POST-MONSOON ANNUAL


OCCURANCES

Fig. 4 Seasonal distribution of tropical cyclonic disturbances and severe cyclonic storms over the Bay of Bengal from 1891 to 2018

0.35
0.33
0.3 0.30 0.31
INTENSIFICATION RATE SCS

0.28 0.28 0.29


0.25 0.26
0.25 0.24 0.25 0.24
0.22
0.2 0.2 0.19 0.19
0.18 0.19

0.15 0.16 0.15 0.15


0.14

0.1 0.10 0.1 0.11

0.05 0.05 0.04

Fig. 5 Intensification rate of severe cyclonic storms during 1891–2018. Note The dotted line represents the trend which clearly shows
an increasing trend due to climate change

123
GeoJournal (2023) 88:S85–S107 S95

Table 3 Mann–Kendall test results to show the existence of trend of severe cyclonic storms intensifiction rate during 1891–2018 for
pre-monsoon, monsoon and post-monsoon seasons and 5-year pentads
Cyclonic seasons No. of obs Kendall’s tau-a Kendall’s tau-b Kendall’s score SE of score Prob [|z|

Pre-monsoon 96 0.0496 0.0617 226 291.752 0.441


Monsoon 124 - 0.0964 - 0.1580 - 735 328.892 0.026**
Post-monsoon 124 0.3184 0.3289 2428 460.654 0.000***
Annual 26 0.3354 0.3354 109 45.369 0.017***
1. ** significant at 5%; *** significant at 1%; 2. The null hypothesis is H0: 5 years and intensification rates of severe cyclonic storms
over the Bay of Bengal are independent

probability of cyclonic disturbances turning to severe increase in sea-surface temperature, there is an


cyclonic storms. There also exists an increasing trend increase in the cyclonic disturbance counts by 0.01
of intensification in the post-monsoon but a declining times and 0.14 times, respectively. The M10 shows
trend in the monsoon seasons. However, the intensi- that with a one-unit increase in sea-surface tempera-
fication rates in the pre-monsoon seasons do not turn ture, there is an increase in the cyclonic disturbances
out to be significant. This may be attributed to the by 0.73 times. With a unit increase in near-surface
increase in temperature and warming of the North wind speed, there is a 0.03 times increase in the
Indian Ocean Tropical Easterly Jet, more on the occurrences of cyclonic disturbances. The residual
equatorial side than on the northern side during the autocorrelation functions are well settled within
summer monsoon seasons (Krishna, 2009). confidence limits with none or very few individual
correlations just crossing the confidence limits, which
The estimated forecasting models is acceptable among 20 lags. For residual checks, the
histograms and quantile–quantile plots of residuals
Before we compare the performance of the three (Fig. 6) are drawn and they satisfy the goodness of fit
different forecasting models, we present the estimates criterion of the models.
of all three models. For the SARIMAX model, the For the ANN-NARX model, a three-layer feed-
Box-Jenkins methodology is employed on time-series forward neural network model is used for estimating
data comprising the following four different stages: cyclonic disturbances and severe cyclonic storms with
model identification, estimation of the chosen model, an optimized Levenberg–Marquardt algorithm. Each
diagnostic checking, and forecasting (Bhaumik, dataset is divided into a 14:3:3 ratio with 70% for
2015). The parameters (p,d,q) of the best fit SAR- training, 15% for validation, and 15% for testing. The
IMAX model is ARIMAX (6,1,3) with Seasonal training of the time-series was continued until the
AR(6) and MA(9) for sea-surface pressure; ARIMAX validation error failed to decrease. The error his-
(6,1,3) with Seasonal AR(9) and MA(9) for near- togram, error autocorrelation, time-series plot, and
surface wind speed; and ARIMAX (6,1,3) with regression are considered to validate the optimum
Seasonal AR(9) and MA(9) for surface temperature. network.
Lastly, the model with two exogenous variables – To determine the optimum neurons and lags, the
near-surface wind speed and sea-surface temperature – trial and testing method is adopted with a minimum of
is the ARIMAX (6,1,3) model with Seasonal AR(9) 1 and a maximum of 10. To select the optimum model,
and MA(9) parameters. training performance, training state, error autocorre-
All the constant and exogenous terms are found to lation, error histogram, the correlation between input
be significant. The results indicate that with one unit and error, and the response of output time-series are
(pascal) decrease in sea-surface pressure, there is an compared between models. The network chosen for
increase in cyclonic disturbance counts by (-)0.07 cyclonic disturbances and sea-surface pressure as
times. With one unit (meters per second) increase in predictors is (9,9) with nine hidden neurons and nine
near-surface wind speed and one unit (Kelvin) lags. For cyclonic disturbances and sea-surface

123
S96 GeoJournal (2023) 88:S85–S107

Fig. 6 Residual quantile–quantile plot of SARIMAX models with different predictor(s) of cyclonic disturbances

temperature as predictors, ANN-NARX (9,9) is cho- rejected, which proves the existence of a non-linear
sen. The cyclonic disturbances with near-surface wind pattern in the residual data of SARIMAX (Table 4).
speed have the optimum model of ANN-NARX The application of the hybrid model comprises two
(12,12). For M11 with two exogenous variables – steps. First, to estimate the linear part where the
near-surface wind speed and sea-surface temperature– SARIMAX model is employed. Second, to calculate
the optimum model is ANN-NARX (12, 12). Figure 7 the residuals from the SARIMAX model (the non-
shows the ANN-NARX model estimation with the linear part) by using the ANN-NARX model. By
errors of all four models. summing up the SARIMAX estimates and the ANN-
As mentioned earlier, before estimating the hybrid NARX residual estimates, we get the estimates of the
model, we employ the Brock, Dechert, and Scheink- hybrid model.
man (BDS) test on the residuals of the SARIMAX
estimates (Lee et al., 1993). The null hypothesis is

123
GeoJournal (2023) 88:S85–S107 S97

Fig. 7 Estimated cyclonic disturbances (CD) response of output and error time-series plot of ANN-NARX models with predictor(s)

Performance comparison across forecasting differences of 0.02 between M2 and M3; 0.12 between
models M5 and M6; 0.1 between M8 and M9; and 0.14
between M11 and M12.
Tables 5 and 6 show the details of the accuracy tests of Among all, evidently, the best performing model is
different models for cyclonic disturbances and severe the ANN-NARX with sea-surface temperature and
cyclonic storms. We observe that there exist signifi- near-surface wind speed together as predictors for
cant error differences between all the models. First, cyclonic disturbances and also for severe cyclonic
SARIMAX is a relatively least performing model for storms, which is M11. The performance results
forecasting tropical cyclones. Further, the MASE and suggest that the time-series distribution of cyclonic
NSC estimate also find that it is not appropriate for disturbances and severe cyclonic storms are non-linear
forecasting cyclonic disturbances. For predicting sev- in nature.
ere cyclonic storms, the hybrid model also doesn’t The previous studies of seasonal tropical cyclone
seem to show better performance. However, we forecasting, mostly modeled over the Atlantic and the
observe a close resemblance in the performances Pacific Ocean, had reported correlations of 0.81
between the ANN-NARX and the hybrid model with (Vitart et al., 2007) and 0.74 (LaRow et al., 2010)

123
S98 GeoJournal (2023) 88:S85–S107

Table 4 Results of Brock, Dechert and Scheinkman (BDS) test to examine the existence of non-linearity
Model combinations Dimension (m) BDS Statistic z-Statistic Normal Prob

Cyclonic disturbances with PSL 3 0.012 2.263 0.024


4 0.069 10.91 0.000
5 0.124 18.739 0.000
6 0.167 26.228 0.000
Cyclonic disturbances with WIND 3 0.04 7.256 0.000
4 0.1 15.165 0.000
5 0.152 22.377 0.000
6 0.194 29.682 0.000
Cyclonic disturbances with TS 3 0.044 8.009 0.000
4 0.118 18.136 0.000
5 0.171 25.302 0.000
6 0.212 32.668 0.000
Cyclonic disturbances with TS and WIND 3 0.049 7.917 0.000
4 0.124 16.982 0.000
5 0.176 23.154 0.000
6 0.215 29.344 0.000
1. There is presence of non-linearity in all the distributions as Prob. is less than 0.05 for all; 2. TS represents sea-surface temperature,
WIND represents near surface wind speed and PSL represents sea-surface pressure

Table 5 Forecasting model Models Estimation process RMSE MAE NSC MASE
performances of cyclonic
disturbances Sea Surface Pressure (PSL)
M1 SARIMAX 3.5964 3.0887 - 2.1525 1.253
M2 ANN-NARX 1.4384 1.1109 0.4903 0.4534
M3 HYBRID 1.4041 0.8291 0.5122 0.3381
Near Surface Wind Speed (WIND)
M4 SARIMAX 3.6954 3.2550 - 2.3285 1.321
M5 ANN-NARX 1.409 0.977 0.5088 0.3984
M6 HYBRID 1.5706 1.1297 0.3896 0.4607
Surface Temperature (TS)
M7 SARIMAX 3.4399 3.1488 - 1.8841 1.277
M8 ANN-NARX 1.3457 1.0555 0.5539 0.4308
M9 HYBRID 1.4914 1.1416 0.4521 0.4659
The best performance
Near Surface Wind Speed (WIND) and Surface Temperature (TS)
model under each
exogenous variable is given M10 SARIMAX 4.1948 3.8532 - 3.2887 1.563
in bold and the best model M11 ANN-NARX 1.2330 0.9446 0.6255 0.3855
among the entire set of M12 HYBRID 1.446 0.9946 0.4849 0.4059
models is in italics

between observed and predicted occurrences. Inter- forecasting results for both cyclonic disturbances and
estingly, our estimates indicate correlations of 0.80 severe cyclonic storms are presented in the following
and 0.85 with the observed annual cyclonic distur- two subsections (Table 7).
bances and severe cyclonic storms, respectively. The

123
GeoJournal (2023) 88:S85–S107 S99

Table 6 Forecasting model performances of severe cyclonic storms with Near-Surface Wind Speed (WIND) and Surface tem-
perature (TS) as predictors
Models Estimation process RSME MAE NSC MASE

M10 SARIMAX 1.4944 1.0127 - 1.5963 1.083


M11 ANN-NARX 0.8272 0.7008 0.2261 0.731
M12 HYBRID 1.9167 1.7005 - 1.1678 1.765
The best performance model values are given in bold

Table 7 Forecasts of cyclonic disturbances and severe cyclonic disturbances over the Bay of Bengal till 2050
Year Cyclonic Severe cyclonic Year Cyclonic Severe cyclonic
disturbances-annual storms- annual disturbances-annual storms- annual

2019 4.96 1.39 2035 12.60 2.10


2020 8.08 1.51 2036 6.76 1.48
2021 9.05 1.44 2037 4.73 2.10
2022 6.31 1.69 2038 10.42 0.55
2023 6.56 0.06 2039 10.84 1.73
2024 7.39 1.93 2040 7.48 2.50
2025 5.27 1.64 2041 7.71 0.94
2026 4.32 0.34 2042 8.65 1.54
2027 4.66 0.92 2043 7.07 2.69
2028 7.45 0.23 2044 11.74 1.00
2029 9.73 3.36 2045 11.98 1.41
2030 8.98 1.16 2046 5.28 2.03
2031 6.31 -0.11 2047 7.71 1.22
2032 10.19 1.61 2048 11.92 1.76
2033 6.89 2.38 2049 10.10 0.78
2034 6.80 1.79 2050 7.05 1.07

Forecasting cyclonic disturbances U-shaped curve. Overall, there has been an increasing
trend of cyclonic disturbances over the years.
Comparing across models, we observe that the model The forecasts suggest that during 2019–2050, on
with exogenous variables as sea-surface temperature average, there would be eight cyclonic disturbances
and near-surface wind speed is most accurate for per year, with three cyclonic disturbances in pre-
forecasting cyclonic disturbances over the Bay of monsoon and monsoon season each and 2 in the post-
Bengal. The observed and predicted cyclonic distur- monsoon season. Cyclonic disturbances are expected
bances are shown in Fig. 8, where panel-WIND and to vary from 5 to 13. Over a period of 12 different
TS have the closest resemblance. years, the cyclonic occurrences are expected to be
Further, the forecasted occurrences of cyclonic greater than the average (e.g., 2021, 2029, 2030, 2032,
disturbances are fitted in a polynomial curve with 95% 2035, 2038, 2039, 2042, 2044, 2045, 2048, and 2049),
prediction bounds (Fig. 9). We observe a steep upward registering increasing trend in 35% cases. The season-
trend of cyclonic disturbances during pre-monsoons wise forecast indicates that in pre-monsoon seasons,
and a wave-like but upward trend during monsoons, the cyclonic occurrences are expected to rise post-
whereas the post-monsoon season forms a downward 2029. In monsoon seasons, there would be an increase
in the occurrences from 2035, with the highest

123
S100 GeoJournal (2023) 88:S85–S107

Fig. 8 Observed and predicted cyclonic disturbance plots with TS as exogenous variables. M11 has strong correlation of 0.8
all the predictors of cyclonic disturbances (CD). Note 1. The best with observed cyclonic disturbances; 2. TS represents sea-
performance model is M11 (bottom left corner) with WIND and surface temperature; WIND represents near surface wind speed

occurrences in 2039. However, in post-monsoon We fit a polynomial curve for the forecasted severe
seasons, the cyclonic occurrences would remain the cyclonic storms with 95% prediction bounds (Fig. 12).
same as observed in the last two decades. It is observed that the pre-monsoon seasons have an
inverted-U-shaped curve, while the monsoon seasons
Forecasting severe cyclonic storms have a slightly upward trend. However, the post-
monsoons have a wave-like structure–first downward
Instead of carving out all the 12 models to find the best sloping and then increasing over the years. The overall
fit and best performances, we have taken a set of picture indicates that the severe cyclonic storms
models (namely, M10, M11, M12) exhibiting the best decrease initially, then increase and again start
combination of predictors. Similar to the case for decreasing. The forecasts from 2019 to 2050 suggest
cyclonic disturbances, the ANN-NARX model with that, on average, there would be about 1.44 severe
sea-surface temperature and near-surface wind speed cyclonic storms per year, and the likelihood of severe
turns out to be the best performance model for cyclonic storms in the post-monsoon seasons is
forecasting severe cyclonic storms. Figure 10 shows relatively high. We further observe that while in the
the diagnostics of the best performance model for pre-monsoon and monsoon seasons, severe cyclonic
forecasting severe cyclonic storms. Figure 11 presents activities are expected to increase, the occurrences in
the observed and the predicted annual severe cyclonic the post-monsoon seasons will decrease or remain
storm counts. constant.

123
GeoJournal (2023) 88:S85–S107 S101

Fig. 9 Seasonal forecasts of cyclonic disturbances with 95% prediction bounds of the period from 2019 to 2050

Fig. 10 Quantile–quantile and response of output element plots predictors is the best performing model; 2. TS represents sea-
for the best performance model of severe cyclonic storms (SCS). surface temperature and WIND represents near surface wind
Note 1. The ANN-NARX model with TS and WIND as speed

Summary and conclusion close resemblance with observed cyclonic distur-


bances and severe cyclonic storm counts over the
The main aim of this study was to find out a statistical Bay of Bengal. This study is an advancement over the
model for long-term forecasting that could establish a existing ones as it performed and compared—linear,

123
S102 GeoJournal (2023) 88:S85–S107

Fig. 11 Observed and predicted severe cyclonic storm (SCS) correlation of 0.85 with observed severe cyclonic storms; 2. TS
plot of the best performance model—M11 with WIND and TS represents sea-surface temperature and WIND represents near
predictors. Note 1. M11 has the most precise estimates, with a surface wind speed

Fig. 12 Seasonal forecasts of severe cyclonic storm occurrences with 95% prediction bounds of the period from 2019 to 2050

123
GeoJournal (2023) 88:S85–S107 S103

non-linear, and hybrid models with predictors to find It is a long-term forecasting study where we have
out the best fit of the distribution of cyclonic considered only the tropical cyclone counts over the
disturbances over the Bay of Bengal. It contributes years. However, other intrinsic details of tropical
significantly to the current literature on tropical cyclones like genesis or six hourly interval data have
cyclone forecasting methods, which will assist the not been considered, which remains a limitation of this
scientific community and academicians alike. study. Future studies may focus on these aspects,
The trend analysis indicated an increase in cyclonic including intensity, landfall, etc. Further, it may be
disturbances, which may be attributed to climate significant to note that the ANN-NARX model is not
change (Knutson et al., 2010; Nordhaus, 2013). We free from its limitations. One of the shortcomings is
need to control or cut our emissions to subjugate the that the connecting non-linear functions through
intensification rate of cyclonic disturbances in the Bay which the weights and biases are generated are not
of Bengal. The intensification of severe cyclonic known. This problem can be addressed by other
storms had an upward trend from 1961 to 2000 and methods of artificial intelligence or by integration with
again from 2011 to 2018. However, when we tested for numerical models. Nonetheless, the ANN-NARX
detecting trends seasonally, post-monsoon seasons model generally applies the sigmoid function (Can-
showed increasing probabilities of severe cyclonic denas et al., 2016).
storms, while monsoon seasons showed a negative Suffice to state that the devastating consequences of
intensification trend. This conforms to the previous tropical cyclones are not unknown. Extreme sea-level
findings (Balaguru et al., 2014; Krishna, 2009). rise and coastal flooding due to tropical cyclones are
Overall, the modeling approach with different the leading causes of damages to lives, properties, and
exogenous variables appear promising. With the help loss of lives (Dube et al., 1997). Cyclonic disturbances
of the observed cyclonic disturbances and severe are on the rise and will grow faster in the future if
cyclonic storms, we introduced twelve models for different countries continue to have inefficient adap-
forecasting with different exogenous variables and tive capacity coupled with low preparedness and
their combinations (as predictors). It was observed that ‘fading memory syndrome’ (Raghavan &
for forecasting cyclonic disturbances, the ANN- Rajesh, 2003). Though there has been a significant
NARX model performed, by and large, equally as fall in the casualties associated with the cyclones in
the hybrid model. However, for severe cyclonic recent years due to the improved warning system, the
storms, the ANN-NARX performed the best. The damages to infrastructure, trees, and loss of livestock
performance of the SARIMAX model was relatively continue to be a severe problem.
lower compared to the hybrid and ANN-NARX The primary cause for the increase in tropical
models. Near-surface wind speed and sea surface cyclones is the rising environmental pollution. Hence,
temperature together as predictors provided the best fit strict emission laws need to be put in place in the form
with minimum errors compared to the cases when each of carbon taxing or carbon compensation by govern-
predictor was considered individually. Consequently, ment intervention. This would tamper with the
it can be inferred that the cyclonic disturbances and increase in sea surface temperature and other influ-
severe cyclonic storms are very close to non-linear encing factors in the long term. Meanwhile, the
distributions. policymakers and planners should give special atten-
The ANN-NARX models used for forecasting tion to the vulnerable coastal areas by building
cyclonic disturbances and severe cyclonic storms over mangrove forests as a natural buffer and adequate
the Bay of Bengal seemed to have much better embankments to protect them from the fury of tropical
performances of 0.8 and 0.85, respectively. Our results cyclones. Green infrastructure (mangroves and wet-
were persistent with the assumption that there would lands) along with gray infrastructure will deliver a
be a rise in cyclonic disturbances and severe cyclonic ‘‘triple win’’ (Browder et al., 2019), benefitting the
storms over the years till 2050. The occurrences of economy, communities, and the environment.
cyclonic disturbances would vary from 5 to 13, with an
average of 8 annually. Severe cyclonic storms will Acknowledgements The authors are thankful to the Indian
Meteorological Department and Regional Specialized
have an average occurrence of 1.44 per year, corrob-
Meteorological Centre (http://www.rsmcnewdelhi.imd.gov.in/
orating the study by Bhratwaj and Singh (2020).

123
S104 GeoJournal (2023) 88:S85–S107

index.php?lang=en) for providing reports and records of tropical Bengtsson, L., Botzet, M., & Esch, M. (1995). Hurricane-type
cyclones. We also acknowledge the Centre for Climate Change vortices in a general circulation model. Tellus A, 47(2),
Research, Indian Institute of Tropical Meteorology, Pune, India, 175–196. https://doi.org/10.1034/j.1600-0870.1995.t01-1-
for supplying the CORDEX dataset (http://cccr.tropmet.res.in/ 00003.x
home/cordexsa_datasets.jsp). We also thank the anonymous Bhardwaj, P., & Singh, O. (2020). Climatological characteris-
reviewers for giving their valuable suggestions. tics of Bay of Bengal tropical cyclones: 1972–2017. The-
oretical and Applied Climatology, 139(1), 615–629.
Funding No funding has been received for this study. https://doi.org/10.1007/s00704-019-02989-4
Bhardwaj, P., Singh, O., & Yadav, R. B. S. (2020). Probabilistic
Data availability All data is available online in respective assessment of tropical cyclones’ extreme wind speed in the
sources as provided in the data and methodology section. Bay of Bengal: Implications for future cyclonic hazard.
Natural Hazards, 101(1), 275–295. https://doi.org/10.
1007/s11069-020-03873-5
Code availability This study has been carried out in
Bhaumik, S. K. (2015). Principles of econometrics: A modern
MATLAB Version R2019b for data analysis. Codes will be
approach using eviews. Oxford University Press.
provided upon contacting the authors.
Browder, G., Ozment, S., Rehberger Bescos, I., Gartner, T., &
Lange, G. (2019). Integrating Green and Gray Creating
Declarations Next Generation Infrastructure. Washington, DC: World
Bank and World Resources Institute. World Bank and
Conflict of interest There are no known financial or non- World Resources Institute. https://openknowledge.
financial potential conflicts of interest among the authors worldbank.org/handle/10986/31430
regarding this paper. We followed all the research ethics in the Cadenas, E., Rivera, W., Campos-Amezcua, R., & Cadenas, R.
submitted work. (2016). Wind speed forecasting using the NARX model,
case: La Mata, Oaxaca México. Neural Computing and
Human and animal rights This research is not an experiment Applications, 27(8), 2417–2428. https://doi.org/10.1007/
or survey-based study, thus there is no involvement of Humans s00521-015-2012-y
or Animals in the study. Chan, J. C., Shi, J. E., & Lam, C. M. (1998). Seasonal fore-
casting of tropical cyclone activity over the western North
Pacific and the South China Sea. Weather and Forecasting,
References 13(4), 997–1004. https://doi.org/10.1175/1520-
0434(1998)013%3c0997:SFOTCA%3e2.0.CO;2
Chattopadhyay, G., Chattopadhyay, S., & Jain, R. (2010).
Acharya, N., Shrivastava, N. A., Panigrahi, B. K., & Mohanty, Multivariate forecast of winter monsoon rainfall in India
U. C. (2014). Development of an artificial neural network using SST anomaly as a predictor: Neurocomputing and
based multi-model ensemble to estimate the northeast statistical approaches. Comptes Rendus Geoscience,
monsoon rainfall over south peninsular India: An appli- 342(10), 755–765. https://doi.org/10.1016/j.crte.2010.06.
cation of extreme learning machine. Climate Dynamics, 004
43(5–6), 1303–1310. https://doi.org/10.1007/s00382-013- Chattopadhyay, S. (2007). Feed forward artificial neural net-
1942-2 work model to predict the average summer-monsoon
Ali, M. M., Kishtawal, C. M., & Jain, S. (2007). Predicting rainfall in India. Acta Geophysica, 55(3), 369–382. https://
cyclone tracks in the north Indian Ocean: An artificial doi.org/10.2478/s11600-007-0020-8
neural network approach. Geophysical Research Letters, Chattopadhyay, S., & Chattopadhyay, G. (2008). Comparative
34(4), L04603. https://doi.org/10.1029/2006GL028353 study among different neural net learning algorithms
Amei, A., Fu, W., & Ho, C. H. (2012). Time-series analysis for applied to rainfall time-series. Meteorological Applica-
predicting the occurrences of large scale earthquakes. In- tions: A Journal of Forecasting, Practical Applications,
ternational Journal of Applied Science and Technology, Training Techniques and Modelling, 15(2), 273–280.
2(7), 64–75. https://doi.org/10.1002/met.71
Baik, J. J., & Paek, J. S. (2000). A neural network model for Chaudhuri, S., Basu, D., Das, D., Goswami, S., & Varshney, S.
predicting typhoon intensity. Journal of the Meteorologi- (2017). Swarm intelligence and neural nets in forecasting
cal Society of Japan. Ser. II, 78(6), 857–869. https://doi. the maximum sustained wind speed along the track of
org/10.2151/jmsj1965.78.6_857 tropical cyclones over Bay of Bengal. Natural Hazards,
Balaguru, K., Taraphdar, S., Leung, L. R., & Foltz, G. R. (2014). 87(3), 1413–1433. https://doi.org/10.1007/s11069-017-
Increase in the intensity of post-monsoon Bay of Bengal 2824-4
tropical cyclones. Geophysical Research Letters, 41(10), Chowdhury, K.M.M.H. (2002). Cyclone preparedness and
3594–3601. https://doi.org/10.1002/2014GL060197 management in Bangladesh. In: BPATC (ed.) Improvement
Barzegar, R., Adamowski, J., & Moghaddam, A. A. (2016). of early warning system and responses in Bangladesh
Application of wavelet-artificial intelligence hybrid mod- towards total disaster risk management approach.
els for water quality prediction: A case study in Aji-Chay BPATC, Dhaka.
River Iran. Stochastic Environmental Research and Risk Clemen, R. T. (1989). Combining forecasts: A review and
Assessment, 30(7), 1797–1819. https://doi.org/10.1007/ annotated bibliography. International Journal of
s00477-016-1213-y

123
GeoJournal (2023) 88:S85–S107 S105

Forecasting, 5(4), 559–583. https://doi.org/10.1016/0169- Gutzler, D. S., Wood, K. M., Ritchie, E. A., Douglas, A. V., &
2070(89)90012-5 Lewis, M. D. (2013). Interannual variability of tropical
Dube, S. K., Rao, A. D., Sinha, P. C., Murty, T. S., & Bahulayan, cyclone activity along the Pacific coast of North America.
N. (1997). Storm surge in the Bay of Bengal and Arabian Atmósfera, 26(2), 149–162. https://doi.org/10.1016/S0187-
Sea the problem and its prediction. Mausam, 48(2), 6236(13)71069-5
283–304. Haarsma, R. J., Mitchell, J. F., & Senior, C. A. (1993). Tropical
Elamin, N., & Fukushige, M. (2018). Modeling and forecasting disturbances in a GCM. Climate Dynamics, 8(5), 247–257.
hourly electricity demand by SARIMAX with interactions. https://doi.org/10.1007/BF00198619
Energy, 165(2), 257–268. https://doi.org/10.1016/j.energy. Hwang, S. H., Ham, D. H., & Kim, J. H. (2012). A new measure
2018.09.157 for assessing the efficiency of hydrological data-driven
Emanuel, K. A. (1988). The maximum intensity of hurricanes. forecasting models. Hydrological Sciences Journal, 57(7),
Journal of the Atmospheric Sciences, 45(7), 1143–1155. 1257–1274. https://doi.org/10.1080/02626667.2012.
https://doi.org/10.1175/1520-0469(1988)045%3c1143: 710335
TMIOH%3e2.0.CO;2 Hyndman, R. J., & Koehler, A. B. (2006). Another look at
Emanuel, K. A. (1995). Sensitivity of tropical cyclones to sur- measures of forecast accuracy. International journal of
face exchange coefficients and a revised steady-state model forecasting, 22(4), 679–688. https://doi.org/10.1016/j.
incorporating eye dynamics. Journal of Atmospheric Sci- ijforecast.2006.03.001
ences, 52(22), 3969–3976. https://doi.org/10.1175/1520- Indian Meteorological Department (IMD). (2011). Tracks of
0469(1995)052%3c3969:SOTCTS%3e2.0.CO;2 cyclones and depressions over North Indian Ocean (from
Evans, J. L. (1993). Sensitivity of tropical cyclone intensity to 1891 onwards). Cyclone Warning and Research Centre,
sea surface temperature. Journal of Climate, 6(6), India Meteorological Department, Regional Meteorologi-
1133–1140. https://doi.org/10.1175/1520- cal Centre, Chennai. http://www.rsmcnewdelhi.imd.gov.in
0442(1993)006%3c1133:SOTCIT%3e2.0.CO;2 Indian Meteorological Department (IMD) report. (2019). Report
Farajzadeh, J., & Alizadeh, F. (2018). A hybrid linear–non- on cyclonic disturbances over the north Indian Ocean
linear approach to predict the monthly rainfall over the during 2018, No. ESSO/IMD/CWD Report-01(2019)/09.
Urmia Lake watershed using wavelet-SARIMAX-LSSVM Retrieved 21 April, 2020, from http://www.rsmcnewdelhi.
conjugated model. Journal of Hydroinformatics, 20(1), imd.gov.in
246–262. https://doi.org/10.2166/hydro.2017.013 Islam, A. R. M. T., Karim, M. R., & Mondol, M. A. H. (2021).
Feng, L., & Luo, G. (2014). Application of a non-linear model in Appraising trends and forecasting of hydroclimatic vari-
landfall number forecasting for tropical cyclones in China. ables in the north and northeast regions of Bangladesh.
Natural Hazards, 73(3), 1475–1482. https://doi.org/10. Theoretical and Applied Climatology, 143(1), 33–50.
1007/s11069-014-1146-z https://doi.org/10.1007/s00704-020-03411-0
Feng, Z. K., Niu, W. J., Tang, Z. Y., Xu, Y., & Zhang, H. R. Knutson, T. R., McBride, J. L., Chan, J., Emanuel, K., Holland,
(2021). Evolutionary artificial intelligence model via G., Landsea, C., & Sugi, M. (2010). Tropical cyclones and
cooperation search algorithm and extreme learning climate change. Nature Geoscience, 3(3), 157–163. https://
machine for multiple scales nonstationary hydrological doi.org/10.1038/ngeo779
time series prediction. Journal of Hydrology, 595(1), Krishna, K. M. (2009). Intensifying tropical cyclones over the
126062. https://doi.org/10.1016/j.jhydrol.2021.126062 North Indian Ocean during summer monsoon—Global
Gardner, M. W., & Dorling, S. R. (1998). Artificial neural net- warming. Global and Planetary Change, 65(1–2), 12–16.
works (the multilayer perceptron)—A review of applica- https://doi.org/10.1016/j.gloplacha.2008.10.007
tions in the atmospheric sciences. Atmospheric Kutty, G., Gogoi, R., Rakesh, V., & Pateria, M. (2020). Com-
Environment, 32(14–15), 2627–2636. https://doi.org/10. parison of the performance of HYBRID ETKF-3DVAR
1016/S1352-2310(97)00447-0 and 3DVAR data assimilation scheme on the forecast of
Geetha, A., & Nasira, G. M. (2016). Time-series modeling and tropical cyclones formed over the Bay of Bengal. Journal
forecasting: Tropical cyclone prediction using ARIMA of Earth System Science, 129(1), 1–14. https://doi.org/10.
model. In 2016 3rd International Conference on Comput- 1007/s12040-020-01497-8
ing for Sustainable Global Development (INDIACom), LaRow, T. E., Stefanova, L., Shin, D. W., & Cocke, S. (2010).
3080–3086. IEEE. Seasonal Atlantic tropical cyclone hindcasting/forecasting
Gray, W. M. (1979). Hurricanes: Their formation, structure and using two sea surface temperature datasets. Geophysical
likely role in the tropical circulation. Meteorology over the Research Letters, 37(2), L02804. https://doi.org/10.1029/
tropical oceans, 155(1), 218. 2009GL041459
Gray, W. M., Landsea, C. W., Mielke, P. W., Jr., & Berry, K. J. Lee, T. H., White, H., & Granger, C. W. (1993). Testing for
(1993). Predicting Atlantic basin seasonal tropical cyclone neglected non-linearity in time-series models: A compar-
activity by 1 August. Weather and Forecasting, 8(1), ison of neural network methods and alternative tests.
73–86. https://doi.org/10.1175/1520-0434(1993)008% Journal of Econometrics, 56(3), 269–290. https://doi.org/
3c0073:PABSTC%3e2.0.CO;2 10.1016/0304-4076(93)90122-L
Gray, W. M., Landsea, C. W., Mielke, P. W., Jr., & Berry, K. J. Legates, D. R., & McCabe, G. J., Jr. (1999). Evaluating the use
(1994). Predicting Atlantic basin seasonal tropical cyclone of ‘‘goodness-of-fit’’ measures in hydrologic and hydro-
activity by 1 June. Weather and Forecasting, 9(1), climatic model validation. Water Resources Research,
103–115. https://doi.org/10.1175/1520-0434(1994)009% 35(1), 233–241. https://doi.org/10.1029/1998WR900018
3c0103:PABSTC%3e2.0.CO;2

123
S106 GeoJournal (2023) 88:S85–S107

Liu, H., Chen, C., Tian, H. Q., & Li, Y. F. (2012). A hybrid Journal of Hydrology, 500(1), 21–36. https://doi.org/10.
model for wind speed prediction using empirical mode 1016/j.jhydrol.2013.07.020
decomposition and artificial neural networks. Renewable Raghavan, S., & Rajesh, S. (2003). Trends in tropical cyclone
Energy, 48(1), 545–556. https://doi.org/10.1016/j.renene. impact: A study in Andhra Pradesh, India: A study in
2012.06.012 Andhra Pradesh, India. Bulletin of the American Meteo-
Mann, H. B. (1945). Non-parametric tests against rological Society, 84(5), 635–644. https://doi.org/10.1175/
trend. Econometrica: Journal of the econometric society, BAMS-84-5-635
https://doi.org/10.2307/1907187. Rajeevan, M., Unnikrishnan, C. K., Bhate, J., Niranjan Kumar,
Merrill, R. T. (1987). An experiment in statistical prediction of K., & Sreekala, P. P. (2012). Northeast monsoon over
tropical cyclone intensity change (Vol. 34). US Depart- India: Variability and prediction. Meteorological Appli-
ment of Commerce, National Oceanic and Atmospheric cations, 19(2), 226–236. https://doi.org/10.1002/met.1322
Administration, National Weather Service, National Hur- Reddy, B. N. K., Venkatesan, R., Osuri, K. K., Mathew, S.,
ricane Center. Kadiyam, J., & Joseph, K. J. (2018). Comparison of
Miller, B. I. (1958). On the maximum intensity of hurricanes. AMSR-2 wind speed and sea surface temperature with
Journal of Atmospheric Sciences, 15(2), 184–195. https:// moored buoy observations over the Northern Indian Ocean.
doi.org/10.1175/1520-0469(1958)015%3c0184: Journal of Earth System Science, 127(1), 1–11. https://doi.
OTMIOH%3e2.0.CO;2 org/10.1007/s12040-017-0902-3
Mishra, A. K., Desai, V. R., & Singh, V. P. (2007). Drought Richman, M. B., & Leslie, L. M. (2012). Adaptive machine
forecasting using a hybrid stochastic and neural network learning approaches to seasonal prediction of tropical
model. Journal of Hydrologic Engineering, 12(6), cyclones. Procedia Computer Science, 12(1), 276–281.
626–638. https://doi.org/10.1061/(ASCE)1084- https://doi.org/10.1016/j.procs.2012.09.069
0699(2007)12:6(626) Sahoo, G. B., Schladow, S. G., & Reuter, J. E. (2009). Fore-
Mitchell, T. D., & Camp, J. (2021). The use of the conway– casting stream water temperature using regression analysis,
maxwell–poisson in the seasonal forecasting of tropical artificial neural network, and chaotic non-linear dynamic
cyclones. Weather and Forecasting, 36(3), 929–939. models. Journal of Hydrology, 378(3–4), 325–342. https://
https://doi.org/10.1175/WAF-D-20-0160.1 doi.org/10.1016/j.jhydrol.2009.09.037
Morid, S., Smakhtin, V., & Bagherzadeh, K. (2007). Drought Sarkar, I., Chaudhuri, S., & Pal, J. (2021). Artificial intelligence
forecasting using artificial neural networks and time-series in forecasting central pressure drop and maximum sus-
of drought indices. International Journal of Climatology: A tained wind speed of cyclonic systems over Arabian Sea:
Journal of the Royal Meteorological Society, 27(15), skill comparison with conventional models. Meteorology
2103–2111. https://doi.org/10.1002/joc.1498 and Atmospheric Physics, 1–20.
Nair, A., Acharya, N., Singh, A., Mohanty, U. C., & Panda, T. C. Srinivas, C. V., Bhaskar Rao, D., Yesubabu, V., Baskaran, R., &
(2013). On the predictability of northeast monsoon rainfall Venkatraman, B. (2013). Tropical cyclone predictions over
over south peninsular India in general circulation models. the Bay of Bengal using the high-resolution advanced
Pure and Applied Geophysics, 170(11), 1945–1967. research weather research and forecasting (ARW) model.
https://doi.org/10.1007/s00024-012-0633-y Quarterly Journal of the Royal Meteorological Society,
Nash, J. E., & Sutcliffe, J. V. (1970). River flow forecasting 139(676), 1810–1825. https://doi.org/10.1002/qj.2064
through conceptual models part I—A discussion of prin- Sun, J., & Ahn, J. B. (2011). A GCM-based forecasting model
ciples. Journal of Hydrology, 10(3), 282–290. https://doi. for the landfall of tropical cyclones in China. Advances in
org/10.1016/0022-1694(70)90255-6 Atmospheric Sciences, 28(5), 1049–1055. https://doi.org/
Nath, S., Kotal, S. D., & Kundu, P. K. (2016). Seasonal pre- 10.1007/s00376-011-0122-8
diction of tropical cyclone activity over the north Indian Vagropoulos, S. I., Chouliaras, G. I., Kardakos, E. G., Simoglou,
Ocean using three artificial neural networks. Meteorology C. K., & Bakirtzis, A. G. (2016). Comparison of SAR-
and Atmospheric Physics, 128(6), 751–762. https://doi.org/ IMAX, SARIMA, modified SARIMA and ANN-based
10.1007/s00703-016-0446-0 models for short-term PV generation forecasting. IEEE
Neumann, C. J. (1993). Global Overview. In: Global guide to International Energy Conference (ENERGYCON), 1–6.
tropical cyclone forecasting, WMO/TC-No. 560. Report IEEE. https://doi.org/10.1109/ENERGYCON.2016.
No. World Meteorological Organization, Geneva. 7514029
Newbold, P., & Granger, C. W. (1974). Experience with fore- Vaz, A. G. R., Elsinga, B., Van Sark, W. G. J. H. M., & Brito, M.
casting univariate time-series and the combination of C. (2016). An artificial neural network to assess the impact
forecasts. Journal of the Royal Statistical Society: Series A of neighbouring photovoltaic systems in power forecasting
(general), 137(2), 131–146. https://doi.org/10.2307/ in Utrecht, the Netherlands. Renewable Energy, 85(1),
2344546 631–641. https://doi.org/10.1016/j.renene.2015.06.061
Nicholls, N. (1992). Recent performance of a method for fore- Villarini, G., Luitel, B., Vecchi, G. A., & Ghosh, J. (2019).
casting Australian seasonal tropical cyclone activity. Aus- Multi-model ensemble forecasting of North Atlantic trop-
tralian Meteorological Magazine, 40(2), 105–110. ical cyclone activity. Climate Dynamics, 53(12),
Nordhaus, W. (2013). The climate casino: Risk, uncertainty, and 7461–7477. https://doi.org/10.1007/s00382-016-3369-z
economics for a warming world. Yale University Press. Vitart, F., Anderson, J. L., & Stern, W. F. (1997). Simulation of
Pektaş, A. O., & Cigizoglu, H. K. (2013). ANN hybrid model interannual variability of tropical storm frequency in an
versus ARIMA and ARIMAX models of runoff coefficient. ensemble of GCM integrations. Journal of Climate, 10(4),

123
GeoJournal (2023) 88:S85–S107 S107

745–760. https://doi.org/10.1175/1520-0442(1997)010% Yip, Z. K., & Yau, M. K. (2012). Application of artificial neural
3c0745:SOIVOT%3e2.0.CO;2 networks on North Atlantic tropical cyclogenesis potential
Vitart, F., Anderson, J. L., & Stern, W. F. (1999). Impact of index in climate change. Journal of Atmospheric and
large-scale circulation on tropical storm frequency, inten- Oceanic Technology, 29(9), 1202–1220. https://doi.org/10.
sity, and location, simulated by an ensemble of GCM 1175/JTECH-D-11-00178.1
integrations. Journal of Climate, 12(11), 3237–3254. Yue, S., Pilon, P., & Cavadias, G. (2002). Power of the Mann-
https://doi.org/10.1175/1520-0442(1999)012%3c3237: Kendall and Spearman’s rho tests for detecting monotonic
IOLSCO%3e2.0.CO;2 trends in hydrological series. Journal of Hydrology,
Vitart, F., Huddleston, M. R., Déqué, M., Peake, D., Palmer, T. 259(1–4), 254–271. https://doi.org/10.1016/S0022-
N., Stockdale, T. N., & Weisheimer, A. (2007). Dynami- 1694(01)00594-7
cally-based seasonal forecasts of Atlantic tropical storm Zhang, G. P. (2003). Time-series forecasting using a hybrid
activity issued in June by EUROSIP. Geophysical ARIMA and neural network model. Neurocomputing,
Research Letters, 34(16), L16815. https://doi.org/10.1029/ 50(1), 159–175. https://doi.org/10.1016/S0925-
2007GL030740 2312(01)00702-0
Vitart, F., & Stockdale, T. N. (2001). Seasonal forecasting of Zhang, W., & Villarini, G. (2019). Seasonal forecasting of
tropical storms using coupled GCM integrations. Monthly western North Pacific tropical cyclone frequency using the
Weather Review, 129(10), 2521–2537. https://doi.org/10. North American multi-model ensemble. Climate Dynam-
1175/1520-0493(2001)129%3c2521:SFOTSU%3e2.0. ics, 52(9), 5985–5997. https://doi.org/10.1007/s00382-
CO;2 018-4490-y
Wijnands, J. S., Shelton, K., & Kuleshov, Y. (2014). Improving Zhao, H., & Raga, G. B. (2015). On the distinct interannual
the operational methodology of tropical cyclone seasonal variability of tropical cyclone activity over the eastern
prediction in the Australian and the South Pacific Ocean North Pacific. Atmósfera, 28(3), 161–178. https://doi.org/
regions. Advances in Meteorology, 2014(1), 838746. 10.20937/ATM.2015.28.03.02
https://doi.org/10.1155/2014/838746
Yang, Z., Ce, L., & Lian, L. (2017). Electricity price forecasting
Publisher’s Note Springer Nature remains neutral with
by a hybrid model, combining wavelet transform, ARMA
regard to jurisdictional claims in published maps and
and kernel-based extreme learning machine methods. Ap-
institutional affiliations.
plied Energy, 190(1), 291–305. https://doi.org/10.1016/j.
apenergy.2016.12.130

123

You might also like