Ocean Modelling 136 (2019) 1–12

Contents lists available at ScienceDirect

Ocean Modelling
journal homepage: www.elsevier.com/locate/ocemod

Statistical simulation of ocean current patterns using autoregressive logistic T

regression models: A case study in the Gulf of Mexico
Helios Chiria, , Ana Julia Abascala, Sonia Castanedob, José Antonio A. Antolínezb, Yonggang Liuc,
⁎

Robert H. Weisbergc, Raul Medinaa

a
Environmental Hydraulics Institute, Universidad de Cantabria, Avda. Isabel Torres, 15, Parque Científico y Tecnológico de Cantabria, 39011 Santander, Spain
b
Departamento de Ciencias y Técnicas del Agua y del Medio Ambiente, Universidad de Cantabria, 39005 Santander, Spain
c
College of Marine Science, University of South Florida, St. Petersburg, FL, USA

ARTICLE INFO ABSTRACT

Keywords: Autoregressive logistic regression models have been demonstrated to be a powerful tool for statistical simulation
Statistical modelling of spatial patterns in climate and meteorology fields. In this paper we introduce a statistical framework for the
Autoregressive logistic regression simulation of ocean current patterns based on the autoregressive logistic regression models, and apply it to the
Ocean current patterns Gulf of Mexico Loop Current. The statistical model is forced by three autoregressive terms, the wind stress curl in
Gulf of Mexico
the Gulf of Mexico and in the Caribbean Sea, and the sea level pressure anomalies over the North Atlantic. It is
Loop Current
used to replicate the bi-weekly historical sequence of 8 Loop Current patterns, obtained from a 24-year altimetry
derived dataset. The model reproduces the inter-annual and intra-annual variability of the original time series,
showing notable fitting capacity. A point-by-point comparison between the actual and simulated pattern series
confirms the capability of the model in analysing the evolution of ocean current patterns. The predictive skill of
the model is also explored, and the preliminary forecast (up to 3 months) results are encouraging. The presented
statistical framework may find more practical applications in the future, such as the generation of statistically
sound climate-based oceanographic scenarios for risk analyses, and the mid-term probabilistic prediction of
ocean current patterns.

1. Introduction databases can result from numerical oceanographic reanalysis (e.g.,

Ocean currents have a key influence in many different fields like
navigation (e.g., Lo and McCord, 1998), fishery management (e.g.,
Roberts and Van den Berg, 2002), oil and gas industry (e.g., Abascal
et al., 2017), transport of floating objects in the ocean (e.g., Mínguez
et al., 2012), climate studies (e.g., Gent, 2017) and ecological analyses
(e.g., Rogers et al., 2017), among others. Nevertheless, analysing and
understanding the mechanism and behaviour of the currents in a given
area of interest is generally not an easy task (e.g., Alves et al., 2018).
Indeed ocean circulation is one of the most complex features of the
Earth science, being the response, either globally or locally, to the in-
teraction of several phenomena like tides, winds, water density gra-
dients, Earth rotation, and being influenced also by bathymetric fea-
tures and coast orientation (e.g., Kamenkovich, 1977).
Available from around the early 2000's, several ocean current da-
tabases with broad spatial coverage and long (progressively increasing)
temporal extension established the basis for an in-depth and compre-
hensive analysis of ocean currents in different areas. These current
Saha et al., 2010; Chassignet et al., 2009) or from satellite altimetry
campaigns (e.g., Sudre et al., 2013). In situ deployed instruments like
ADCP (e.g., Liu and Weisberg, 2005) or HF radar (e.g., Solabarrieta
et al., 2015; Liu et al., 2007b) can also be sources of long time series of
ocean current data. However, it is challenging to analyse the data and
extract valuable information from a large dataset. In order to address
this issue, a common practice is to summarize the variability of the
current in a given study area using a limited number of characteristic
spatial patterns or groups. This can be achieved through different
techniques, e.g., the unsupervised learning algorithms of Self-Orga-
nizing Map (SOM) (Kohonen, 1982; Kohonen et al., 2001) (e.g. Liu and
Weisberg, 2005; Iskandar et al., 2008) and K-means (Lloyd, 1957;
McQueen, 1967) (e.g., Hsieh et al., 2004; Solabarrieta et al., 2014).
Once representative spatial patterns are identified, the original time
series can be expressed in terms of these patterns, associating each time
stamp to the most similar pattern (the Best Matching Units, BMUs). This
approximation provides a simplified and useful way to analyse the
currents in a specific region. It allows, for example, estimating the

⁎
Corresponding author.
E-mail address: chirih@unican.es (H. Chiri).

https://doi.org/10.1016/j.ocemod.2019.02.010
Received 26 June 2018; Received in revised form 15 January 2019; Accepted 16 February 2019
Available online 22 February 2019
1463-5003/ © 2019 Elsevier Ltd. All rights reserved.
H. Chiri, et al.
probability of occurrence of the different patterns, investigating the
presence of any recurrent temporal sequence of current states or
checking if some specific situation tends to persist more than others
(e.g., Jin et al., 2010; Liu et al., 2016b; Weisberg and Liu, 2017). In
some studies, the analysis of pattern evolution is used to detect seasonal
cycles through the data series (e.g., Liu et al., 2008; Solabarrieta et al.,
2015). Recently, some authors have tried to link the inter-annual
variability of some currents to well-known climate indexes and atmo-
spheric conditions (e.g., Tsui and Wu, 2012; Zeng et al., 2015b).
However, there is still a lack of methodologies to deal simultaneously
with the memory of a system and its connection with external para-
meters or forcings. In order to address this issue, we propose to use
autoregressive logistic regression (ALR) models as a new statistical
framework for investigating ocean currents by taking into account the
memory of the system, and the effect of local or external factors at the
same time.
The approach of studying environmental conditions through the
analysis of representative pattern series using statistical methods de-
rives from climatology and meteorology. In these fields, logistic re-
gression models are proving to be a particularly powerful tool capable
of modelling the dynamics by considering simultaneously the influence
of factors of different spatial and temporal scales. These factors af-
fecting the investigated dynamic represent the predictors or covariates
of the model. Prasad et al. (2010) apply logistic regression to pre-
cipitation anomalies for all India and for two homogeneous meteor-
ological subdivision on both sides of the country, considering 36 cov-
ariates. Guanche et al. (2014) analyse daily atmospheric circulation
patterns over the Northeastern Atlantic by fitting an ALR model con-
sidering as predictors monthly sea level pressure anomalies over the
study domain, long-term trend effects, and including a Markov Chain in
the model to consider the autoregressive component of the analysed
system. An ALR model with a similar structure is applied by Antolínez
et al. (2016) to European Atlantic wave climate, with a complementary
model to better estimate the persistence probability of the patterns.
Cárdenas et al. (2017) are able to simulate daily wind states for
Northeastern Atlantic by applying an ALR model that considers sea-
sonality, pressure anomalies over the North Atlantic and a second-order
Markov Chain. These studies demonstrated ALR as a powerful tool
capable of uncovering the most important features of the analysed
variable, such as its dependence on different scale covariates or the
limits of its memory.
Despite those proved capabilities, logistic regression has never been
explored in the field of ocean currents. In this paper, we introduce the
use of ALR models as a tool for statistical simulation of ocean current
patterns considering the influence of different temporal and spatial
scale predictors. To show the capability of ALR application in ocean
currents, we set-up and fit a model to emulate the dynamics of the Gulf
of Mexico (GoM) Loop Current (LC). This current presents a very par-
ticular spatial variability and multiple oceanographic studies exist in
the literature analysing how its dynamics are affected by several factors
of different spatial and temporal scales. These two aspects make of the
Loop Current a perfect example for a logistic model application.
The rest of this paper is organized as follows. In Section 2 we pro-
vide a description of the study area and a brief summary of the results of
previous studies analysing the effect of different scale factors over the
Loop Current variability. In Section 3 we present the data used in the
study. The ALR methodology and its application to the Loop Current are
described in Section 4. Section 5 shows the analysis of the results and
Section 6 summarizes the main conclusions of this paper.
2. Study site
Ocean circulation in the GoM is dominated by the LC system. The LC
is a part of the North Atlantic Ocean's western boundary current, be-
tween the Yucatan Current and the Florida Current, leading to the Gulf
Stream. After entering in the GoM through the Yucatan Strait, the LC
2
Ocean Modelling 136 (2019) 1–12
flows following an anticyclonic looping path before exiting the GoM
through the Strait of Florida. The main feature of the LC is the position
of this looping path, which can alternatively penetrate northwestward
toward the Mississippi River Delta or retract toward Cuba (e.g. Sturges
and Lugo-Fernández, 2005). Additionally, it has been observed that
large intrusions generally produce the shedding of anticyclonic eddies
that start travelling westward in the GoM, at the same time that the LC
began to retract back to the south toward a more direct entrance-exit
path (e.g. Lugo-Fernández, 2016).
Multiple studies based on different kinds of analysis share a similar
conclusion that determined LC events tend to be affected by analogous
previous LC states. Alvera-Azcárate et al. (2009) analyse altimetry-de-
rived LC geostrophic component over a period of approximately
13 years. They observe that large eddy shedding tend to produce large
LC retraction, followed by a long period before the LC penetrates en-
ough for a new shedding and vice versa, the shedding of a small eddy
results in a limited LC retreat and in a shorter time until the next in-
trusion and eddy shedding. Lugo-Fernández (2007) applies different
statistical analyses on observational databases from published sources
in his aim to understand if the LC dynamics represents a chaotic system
or not. He obtains that the system have a memory of about
10–18 months, the time it takes for the LC to complete an entire cycle
consisting of inrushing, shedding, and returning to its initial position.
Liu et al. (2016b), analysing the temporal evolution of 16 LC patterns
extracted from a 23 years altimetry-derived database, identify some
characteristic pattern sequences recurrent over the analysed historical
period.
In order to analyse inter-annual variability of the LC, some recent
studies tried to link the LC variability with large-scale climate indices.
Lugo-Fernández (2007) argues that LC presents oscillations in time
scale of 3–5 years which agrees with the time scale of North Atlantic
Oscillation (NAO) and of El Niño-Southern Oscillation (ENSO). Muller-
Karger et al. (2015) analyse monthly Atlantic Multidecadal Oscillation
(AMO) index (Enfield et al., 2001) and the monthly Multivariate ENSO
Index (MEI) (Wolter and Timlin, 2011) with several oceanographic
variables in the GoM but do not find significant correlation between the
indices and any of the considered variables. Zeng et al. (2015b) after
extracting three LC pattern through a SOM application, compare the
frequency of occurrence of a specific pattern considered as an eddy
shedding situation with various climate indices, including Oceanic Niño
Index (ONI), NAO, Southern Oscillation Index (SOI) and Pacific Decadal
Oscillation (PDO). These authors also analyse the influence on the
frequency of occurrence of the selected pattern of the wind stress curl in
the GoM, Caribbean Sea and Bahamas area. They find that the corre-
lation of the frequency of occurrence of the considered LC pattern is 0.6
with the six-month moving averaged ONI considering a 90-day lag and
0.83 with the wind stress curl in the Caribbean Sea. Oey et al. (2003)
also investigate the affection on the LC of the upstream conditions in
the Caribbean Sea and even further in the North Atlantic. Through
numerical experiments, they find that the wind stress curl in the Car-
ibbean Sea spin up eddies that, drifting later toward the GoM, highly
affect the LC eddy shedding capacity. Their results also reveal a sig-
nificant influence on the LC eddy shedding frequency of the interactions
between the North Atlantic and the Caribbean Sea through the Greater
and Lesser Antilles Passages.
In the present study, taking advantage of this previous knowledge of
the LC and its variability, we will consider as possible covariates of the
ALR model for the LC: the memory of the system, the impact of the
mentioned climate indices (NAO, AMO, ENSO, MEI, ONI, SOI and PDO)
considering a temporal lag up to 1 year, the local effect of the wind
stress curl in the GoM and the possible upstream affection of the wind
stress curl in the Caribbean Sea. In order to consider a possible con-
nection to the North Atlantic variability, we will also include in the
analysis the sea level pressure anomalies over the North Atlantic. This
variable is often used when dealing with the variability of the North
Atlantic (e.g. Guanche et al., 2014; Klotzbach and Gray, 2008).
H. Chiri, et al. Ocean Modelling 136 (2019) 1–12

3. Data description 4. Methods

Current data in the GoM are a derived product of altimetry data 4.1. Logistic regression general framework
from the Archiving, Validation and Interpretation of Satellite
Oceanographic Data (AVISO+). Several recent works have shown that Logistic regression is a special case of generalized linear model
this represents a good product to study the geostrophic component of (GLM). GLMs are statistical models widely used to investigate effects of
the LC (Liu et al., 2011, 2016a, 2016b; Weisberg and Liu, 2017). AVISO explanatory variables on a response or dependent variable. In practice,
+ is a multimission gridded sea level anomaly data set produced by the these models link a linear combination of a group of predictor variables
Ssalto/Duacs with support from the Cnes (http://www.aviso.altimetry. to a function of the response variable. Logistic regression models are
fr/duacs/) and distributed by the Copernicus Marine and Environment widely used when the response variable to be modelled is categorical,
Monitoring Service (CMEMS, http://www.marine.copernicus.eu). It is a i.e., it can take exclusively one from a set of possible values. In the case
global product with horizontal resolution of 1/4° and daily temporal of logistic regression, the function of the response variable linked to the
resolution. The delayed-time data are used for the period between 1 covariates is the logit function. This function represents the natural
January 1993 through 6 May 2016, and the near real-time version are logarithm of the odds of each possible outcome with respect to an ar-
used for the remaining time (7 May 2016 to 31 December 2016). The bitrary baseline category. Based on this definition, if Y ϵ {1, … , n} is the
anomaly fields are combined with the mean dynamic topography set of n ocean current patterns, and the last category J is the baseline, a
(MDT_CNES-CLS13, Rio et al., 2014) produced by CLS Space Oceano- generic multinomial logistic regression model for the j-th pattern can be
graphy Division and distributed by Aviso, with support from Cnes. expressed as follows:
Surface geostrophic currents were then calculated as the gradient of the
absolute sea level, similar to that in the work of Liu et al. (2008). The P (Yt = j | Xt )
log = j + j Xt , j = 1, …, n 1
geostrophic current data used in this work are limited to the GoM re- P (Yt = J | Xt ) (1)
gion, within latitude and longitude ranges of [22.8, 31.8] North and
[98.8, 81.8] West, respectively, and with the omission of coastal ocean where Xt is the vector of covariates at time t, βj refers to the vector of
regions with water depth < 100 m where conventional altimetry data coefficients for each of the values of the covariates at time t, and αj
are not as reliable as in the open oceans due to a number of factors represents the intercept of the model. The left-hand-side of Eq. (1) is the
explained by Vignudelli et al. (2011) (see Fig. 2). logit of a generic category of the response variable, the right-hand-side
The climate indices data were obtained from Climate Explorer is the linear combination of the predictors.
website of the Royal Netherlands Meteorological Institute (http:// As mentioned in Section 2, the influences of the LC variability have
climexp.knmi.nl/). As mentioned in Section 2, based on previous stu- been studied in previous studies, including the memory of the LC
dies of the LC variability, in this paper we consider the following cli- system, possible relationship with some climate index phases and the
mate indices: NAO, AMO, ENSO, MEI, ONI, SOI and PDO. response to atmospheric variations. Here we set up an ALR model
Wind stress curl was calculated data from the National Center for taking all these features into account (see Fig. 1).
Environmental Prediction (NCEP) North American Regional Reanalysis The influence of the memory of the system is explored by adding to
(NARR, 2005; Mesinger et al., 2006) wind data. NARR database in- the model autoregressive terms as a Markov Chain of an order equal to
cludes several meteorological variables from 1979 to present with a the number of previous states to be considered. The implementation in
horizontal resolution of approximatively 32 km, 45 vertical layers, and an ALR model is achieved by using a contrast matrix, such as the
3-hourly frequency, covering the whole North American Region. NARR
is a result of the NCEP Eta model with the Regional Data Assimilation
System (RDAS) and significantly improves the accuracy of temperature,
winds and precipitation compared to the NCEP-DOE Global Reanalysis
2. In this work we analyse the local effect of the wind in the GoM and
the upstream effect in the Caribbean Sea. As shown by other authors,
from the point of view of the hydrodynamic flow, two different sub-
areas of the Caribbean Sea can be identified: the Venezuela-Colombia
Basin and the Cayman Basin, separated by a shallow ridge extending
from Honduras and Nicaragua to Hispaniola, called the Nicaraguan Rise
(e.g. Alvera-Azcárate et al., 2009). Following Zeng et al. (2015b), with
the aim of considering only the most direct effect of the upstream at-
mospheric situation, the wind over the Cayman Basin is considered in
this study. Hence, for the GoM region wind stress curl was calculated
within latitude and longitude ranges of [22.8, 31.8] North and [98.8,
81.8] West, respectively, and for the Caribbean Sea within latitude and
longitude ranges of [17.0, 21.4] North and [89.0, 78.3] West, respec-
tively. In order to reduce the dimensionality of the data, we apply
Principal Component Analysis (PCA) to the data and keep the principal
components (PCs) retaining the 95% of the total variance, resulting in
68 and 41 PCs for the GoM and the Caribbean Sea domains, respec-
tively.
Sea level pressure data are provided by the NCEP Climate Forecast
System Reanalysis (CFSR, Saha et al., 2010). In this study we use CFSR
monthly mean data with 1° spatial resolution, within latitude and
longitude ranges of [0, 65] North and [−65, 15] East, respectively. In
this domain we calculated the monthly anomalies with respect to the
average months. Similarly to that for the wind stress curl data, we re-
duce the dimensionality of the data by keeping the first 10 PCs of the Fig. 1. Diagram of the proposed methodology to obtain an autoregressive lo-
obtained anomaly series, which account for 95% of the total variance. gistic regression (ALR) model for the loop current patterns (LCPs) simulation.

3
H. Chiri, et al.
Helmert matrix (de Vries et al., 1998). A Markov Chain of order d can
be implemented as follow:
d n 1
XtARd = Zkt i tween 14 days (red horizontal line in the wavelet spectral analysis pa-
ik
i=1 k=1 (2)
where Zkt−i
is the dummy variable obtained through the Helmert ma-
trix transformation, and represents the previous Loop Current pattern
(LCP) at time t-i. γik is the coefficient associated with LCP k at time t-i.
Note that the autoregressive component adds a set of (n − 1) ∗ d new
parameters for each category of the model.
The influence of climate indices or atmospheric conditions on the
logit of the LCPs can be included in the model as follows:
nc
XtC = Ci C
i=1
i
(3)
where Ci is the vector of the values of the nC climate indices considered
or of the nC principal components of the atmospheric field considered,
at time t, and βiC is the vector of associated coefficients.
Eqs. (2) and (3) show how to introduce the effect of the considered
variables in a logistic regression model. However, the actual sig-
nificance of these variables must be carefully assessed. After defining
the response variable of our model, in Section 4.3 a statistical based
method is described to assess the significance of each of these compo-
nents. An outline of the methodology followed to obtain the LCPs and to
define the components of the ALR model is presented in Fig. 1.
4.2. Response variable characterization
The response variable in an ALR model is categorical, i.e., it is made
up of different categories that the dependent variable can exclusively
assume. In our case, the LC is the response variable and the categories
are represented by a set of its typical spatial patterns (LCPs).
In order to select these representing LCPs we follow a three-step
clustering methodology which aims to extract patterns that homo-
genously span the space of the original data considering the most sig-
nificant temporal scale for the pattern selection. The three phases of this
clustering methodology are presented next (see Fig. 1).
4.2.1. Principal component analysis
In order to avoid spatially correlated variables that may disturb the
LC data clustering, the PCA is applied to the altimetry-derived daily
data of geostrophic current components. Through this analysis, original
data are redefined with respect to new orthogonal dimensions, or em-
pirical orthogonal functions (EOF; the eigenvectors of the covariance
matrix of that dataset). PCs are the time evolution of the weights of the
EOFs. The obtained EOFs and the associated PCs are ordered based on
the variance of the data explained by each component. The first 84 PCs,
accounting for 95% of the total variance are then considered. As ex-
amples, the first two and last two EOFs and the associated PCs obtained
are shown in Fig. 2. The percentage of the total variance accounted for
by each component is shown in the individual EOF plot. As expected,
the first two EOFs show patterns coherent with the LC system, while the
last modes represent smaller scale spatial variations and higher fre-
quencies affecting quite homogeneously the whole basin.
4.2.2. Wavelet spectrum analysis
As a second step in the clustering process, a joint PC-wavelet ana-
lysis similar to that in Liu et al. (2016b) is performed over the selected
leading PCs. The aim of this step of the clustering methodology, con-
sisting in a rectified wavelet spectral analysis (Liu et al., 2007a) on the
obtained PCs, is to identify the temporal scale under which no major
variability of the GoM geostrophic currents occurs. In Fig. 2 we show, as
examples, the results of this analysis for the first two and last two
considered PCs. The rectified wavelet spectra for the first two PCs
(accounting for 15.9% and 12.3% of the total variance, respectively)
4
Ocean Modelling 136 (2019) 1–12
show that variations of these GoM current components are mainly in
the seasonal to annual time scales (2 months to 1 year). The variations
of the last two PCs are generally faster, being the most energetic be-
nels of Fig. 2) and 1 year. We have performed this analysis for the total
of the 84 considered PCs and found that, in all cases, the major varia-
bility for each PC occurs at a time scale greater than or equal to 14 days.
Therefore, in order to reduce the source of unexpected noise that
could negatively affect the statistical model, the 84 daily PCs of the LC
are rescaled to bi-weekly frequency by calculating the average values
from the daily series.
Please note that in this work the application of PCA has a dual
purpose: first, to allow the joint PC-wavelet analysis with the aim of
exploring the temporal scales of the GoM geostrophic current varia-
bility and eventually choosing a consistent time scale for the pattern
extraction and, second, to reduce the current data dimensionality and
simplify the clustering process.
4.2.3. K-means
As a last step of the pattern extraction, the K-means algorithm
technique is applied to the bi-weekly LC PCs. The Maximum
Dissimilarity algorithm was used to pre-select initial centroids, ensuring
K-means technique to correctly describe the diversity of the data
(Camus et al., 2011; Antolínez et al., 2018). Several tests were per-
formed in order to find an equilibrium between the spatial variability of
the current and the number of groups to be considered. As a result of
the clustering, 8 centroids are selected representing 8 LCPs (Fig. 3).
These typical LC conditions extracted from the 24-year geostrophic
current dataset, will represent the 8 categories in which the ALR model
will classify the LC. The order of the current patterns have been ar-
ranged following a similarity criteria, but only for an easier inter-
pretation of the analysis of the results. From the perspective of the ALR
model, the order of the categories of the response variable does not
affect the results. The patterns show a broad variability of the LC and
cover well different situations between the maximum retraction (LCP 1)
and the deepest intrusion (LCP 10). It is interesting to see how for LCP
1, LCP 2 and LCP 3, representing retraction or mid intrusion stages, the
LC path stays close to the southwestern part of the west Florida shelf,
while the loop tends to curve westward as its path advances deeper in
the GoM (LCP 4, LCP 5 and LCP 6), in accordance with the findings of
Weisberg and Liu (2017). The obtained spatial LCPs also agree well
with the results of Liu et al. (2016b) that extract 16 spatial pattern with
a 4 × 4 SOM application for sea level and current data over the eastern
part of the GoM. Alvera-Azcárate et al. (2009) describe extreme LC
intrusion as an intrusion that extends up to 92°W or 27°N. LCP 7 and
LCP 8 show these situations, highlighting that the used clustering al-
gorithm spans the space of the original data enough to represent these
extreme conditions. Once the characteristic patterns are obtained, the
original 24-year series of GoM geostrophic current is expressed in a
sequence of BMUs, in which each time step of the GoM geostrophic
current is expressed by the most similar LCP extracted. The frequencies
of occurrence of the LCPs in the historical series is, in order from LCP 1
to LCP 8, the following: 6.5%, 12.6%, 13.6%, 14.1%, 15.7%, 14.5%,
11.8% and 11.2%.
4.3. Model set-up
Building an ALR model is the hardest part of applying logistic re-
gression. Although in some applications the inclusion of certain pre-
dictors can be easily guessed, i.e., the wind effect in shallow water areas
or tidal harmonic components in narrow channel flows, in general it
may be challenging to select the set of explanatory variables that ac-
tually influence the current of interest. Moreover, it must be considered
that even if a model with several well chosen features is likely to fit the
data better, as the number of feature increases, possible effects and
interactions among some covariates increase as well (Agresti, 2007).
H. Chiri, et al. Ocean Modelling 136 (2019) 1–12

Fig. 2. Top: first 2 and last 2 EOFs of the geostrophic current in the Gulf of Mexico. The variance accounted for by each mode is shown in each graph. Current units in
m/s. Bottom: first 2 and last 2 PCs plots and the associated wavelet power spectra. The PC variance is normalized to be 0.5. Rectified wavelet power spectra are
shown as filled contours in base 2 logarithm. The regions of > 90% confidence are shown with black contours. Cross-hatched regions on either end indicate the “cone
of influence”, where edge effects become important.
5
H. Chiri, et al. Ocean Modelling 136 (2019) 1–12

Fig. 3. Selected loop current patterns representing the categories of the response variable in the ALR model. Current speed units in m/s.

Table 1
Statistical inference for different nested model (C: constant value, AR#: autoregressive term of order #, WSC GoM: wind stress curl in the GoM, WSC CAR: wind stress
curl in the Caribbean Sea; #wl: temporal lag of # weeks). For each examined model, covariates, number of parameters, degrees of freedom (df), deviance increment
(ΔDev) chi-square distribution value associated and significance are shown. The final selected model highlighted in bold.
Model Covariates N. param df ΔDev X95%2 Significant

0 C 1 5000
I C + AR1 8 4944 181.7 74.5 Yes
II C + AR1 + AR2 15 4888 99.7 74.5 Yes
III C + AR1 + AR2 + AR3 22 4832 88.6 74.5 Yes
IV C + AR1 + AR2 + AR3 + AR4 29 4776 25.7 74.5 No
III C + AR1 + AR2 + AR3 22 4832
V C + AR1 + AR2 + AR3 + WSC GoM 90 4288 652.6 599.4 Yes
VI C + AR1 + AR2 + AR3 + WSC GoM + WSC CAR 131 3960 333.0 371.2 No
V C + AR1 + AR2 + AR3 + WSC GoM 90 4288
VII C + AR1 + AR2 + AR3 + WSC GoM + WSC CAR 6wl 131 3960 375.1 371.2 Yes
VIII C + AR1 + AR2 + AR3 + WSC GoM + WSC CAR 6wl + MSLPAs 4wl 141 3880 128.3 101.9 Yes
IX C + AR1 + AR2 + AR3 + WSC GoM + WSC CAR 6wl + MSLPAs 4wl + ONI 6wl 142 3872 9.8 15.5 No

6
H. Chiri, et al. Ocean Modelling 136 (2019) 1–12

Generally, it is preferable to keep the model as simple as possible for a (ΔDevIX = 9.8 < X95%, 1422). Based on these results, we choose Model
better interpretation of the results. VIII as the optimal ALR model for the statistical simulation of the LCP
Statistical inference-based algorithms can be used to build the series.
model. One particularly useful when studying ocean currents is the Hence, considering as covariates three autoregressive terms, the PCs
forward stepwise method. It consists of starting from the simplest model of the wind stress curl in the GoM, the PCs of the wind stress curl in the
possible, i.e., considering just the constant value, and then adding se- Caribbean Sea with a 6-week lag and the PCs of the North Atlantic
quentially new explanatory variables by testing if they produce a sig- MSLPAs, based on Eqs. (1), (2) and (3) the final equation for our ALR
nificant improvement in the model fitting quality or not. Several sig- model can be written as follows:
nificance tests can be used like Wald's test or Aikake's information
criteria. Here, following Guanche et al. (2014), we propose the use of P (Yt = j | Yt 1 , Yt 2 , Yt 3 , Xt
WG
, XtWC , XtSlpaNA )
the likelihood ratio statistic. This method compares the deviance ratio exp ( j +
3 7
z t i ik +
68
C WG iC
WG
+
41
C WC iC
WC
i=1 k=1 k i=1 i i=1 i
ΔDev (the change of fitting quality for two nested models) and the chi- 10 SlpaNA
square distribution with Δdf = Δnp ∗ (n-1) degree of freedom. Δnp is the + C SlpaNA iC
i=1 i )
=
number of parameters added in the more complex model. Assuming a 7
exp +
3 7 t i 68 WG CWG
k=1 ( j i=1 k=1 zk ik + i = 1 Ci i
confidence level α = 95%, when ΔDev > X0.95, Δdf2 the increasing of 41 WC 10 SlpaNA
fitting quality of the model including the new covariates is statistically + C WC iC
i=1 i
+ C SlpaNA iC
i=1 i )
significant. (5)
As mentioned earlier in Section 4.1, the aim of this analysis is to
where Yt represent the LCP representing the LC state at the time t, Xt
assess the influence of autoregressive terms and inter-annual variability
represents the conditions of a covariate at the time t, Ci specifically
on the LCP evolution. The results of the forward stepwise method based
refers to the i-th PC of the considered covariate, superscript acronyms
on likelihood ratio statistic are shown in Table 1. Model 0, the Null
specify the covariate: wind stress curl in the GoM (WG), wind stress curl
model, only considers the intercept value. Nesting to Model 0, Model I
in the Caribbean Sea (WC) and MSLPAs over the North Atlantic
includes the effect of an autoregressive term, i.e., it checks if the
(SlpaNA).
manifestation of a current pattern is influenced by the previous current
The proposed model has a total of 141 coefficients: one for the in-
spatial situation or not. The increment in deviance obtained by adding
tercept, 7 for each autoregressive component, 68 corresponding to the
this effect ΔDevI = 1781.7 > X95%, 82 allows to accept the hypothesis
number of PCs of the wind stress curl in the GoM, 41 for the PCs of the
that the first autoregressive term is significant in the 24-year series of
wind stress curl in the Caribbean Sea and 10 corresponding to the PCs
LCPs. The autocorrelation of the series is further explored by progres-
of the North Atlantic MSLPAs. These coefficients are adjusted to max-
sively adding a second, third and fourth autoregressive term in Model
imize the fit of the model to the historical data. Several available sta-
II, Model III and Model IV, respectively. As shown in Table 1, the second
tistical software allow for an easy implementation and fit of a logistic
and third term are significant as well, even if the improvement achieved
regression model. In this work scikit-learn library (Pedregosa et al.,
in the model fit decreases as the order of the autoregressive terms in-
2011) of Python is used. We apply the one-vs-rest scheme, l2 penalty
creases. Based on the statistical test, the effect of including a fourth
option, a tolerance default value of 1e-4 and the liblinear solver algo-
autoregressive term is not significant, being the improvement in the
rithm.
deviance between Model IV and Model III ΔDevIV = 25.7 < X95%, 292.
Once the predictors have been selected and the corresponding
This finding, that three previous steps at a bi-weekly interval are sig-
coefficients have been adjusted, we use the fitted model to simulate the
nificant to determine the next LC status, is in concordance with the
sequence of LCPs for the whole 24-year period. Considering the prob-
results of the study by Zeng et al. (2015a) in which the authors develop
abilistic nature of the statistical model, we run the model 100 times and
a model based on an autoregressive artificial neural network to predict
obtain the average values of the probabilities.
the GoM sea surface height and find that six steps, at weekly scale, is the
optimal delay for their model.
Once the autoregressive component of the model has been assessed, 5. Results
we test if any of the atmospheric variables and climate indices in-
vestigated in the previous studies on the LC variability can improve the 5.1. Model fit diagnosis
goodness-of-fit of the ALR model. Since the LC patterns are calculated
based on bi-weekly averaged current data, all the predictors are first In order to check the capabilities of the selected model and also to
resampled at the same time scale for a concordance among LCP series show the improvement brought by the different predictors, we present
and predictor time series. In Model V, we add to Model III the wind in Fig. 4 the comparison among the historical series of LCP annual
stress curl in the GoM, and obtain ΔDevV = 652.6 > X95%, 902; hence probability of occurrence (panel a) and the ones obtained by the si-
the contribution of the new predictor is significant. In the next model mulations provided by Model V (in panel b), Model VII (in panel c) and
we include the effect of the wind stress curl in the Caribbean Sea, which by the finally selected Model VIII (panel d). In each bar of the figure, the
results not significant (ΔDevVI = 330.0 < X95%, 1312). However, con- probabilities of occurrence of the 8 LCPs are stacked for every year so
sidering a delay between the LCP series and the series of wind stress the yearly sum of probabilities is one. As can be seen, the results of the
curl in the Caribbean, we find that the effect of the wind in the selected model (panel d) compare favorably with the historical series of
Caribbean Sea with a 6-week delay significantly improves the goodness- LCPs (panel a). To check how the model capabilities improve by pro-
of-fit of the ALR model, as shown by Model VII in Table 1. In the same gressively including these covariates, in panel b we show the results of
way, spanning different time lag of the North Atlantic MSLPAs with Model V, considering the three autoregressive terms and only the wind
respect to the LCP series, we find that Model VIII, including North stress curl in the GoM. The model is capable of reproducing some of the
Atlantic MSLPAs with a 4-week time lag significantly improves the variability found in the historical series but, in general, a lot of incon-
model (ΔDevVIII = 128.3 > X95%, 1412). To further improve the cap- sistencies exist with respect to the historical series. Model VII, con-
ability of the model, we consider the previously introduced climate sidering also the effect of the wind stress curl in the Caribbean Sea with
indices, each one with a possible time lag up to 1 year. However, none a 6-week lag, has greatly improved skills to reproduce the LCP series
of these indices with none of the considered lags is found to be statis- variability. This is clear, for example, during the period 1998–2001 that
tically significant at all (not shown here). As an example, in Table 1 we the Model V is unable to replicate and that is instead simulated quite
show Model IX which includes ONI with 6-week lag as a covariate. favorably by the Model VII. Clear improvements can be seen also for the
Following the test, the increment in the deviance is not significant periods 2004–2005, 2009–2010 and for the year 2014. By also

7
H. Chiri, et al.
Fig. 4. Panel a: evolution of historical annual frequency of occurrence of the selected LCPs. Panel b: evolution of annual probability of occurrence of the selected LCPs
based on Model V simulation. Panel c: evolution of annual probability of occurrence of the selected LCPs based on Model VII simulation. Panel d: evolution of annual
probability of occurrence of the selected LCPs based on Model VIII simulation.
including the variability provided by the MSLPAs over the North
Atlantic with a 4-week lag, Model VIII is able to further expand the
capability of the statistical model to reconstruct the LCPs original series,
as shown in panel d.
To explore the intra-annual variability of the LCP evolution and to
check for the skills of the selected ALR model, in Fig. 5 we show the
comparison between historical (left panel) and modelled (right panel)
annual average probability of occurrence of LCPs at monthly scale. The
main features in the intra-annual scale are replicated quite well by the
model. Historical frequency of retraction stage patterns, like LCP 1 and
LCP 2, are higher in summer and autumn, with the maximum retraction
(LCP 1) more likely between June and November. This feature is well
reflected by the model results. Both historical and modelled LCP
probabilities show that weak intrusion stages (LCP 3 and LCP 4) are
common situations from autumn to early winter and in early spring,
with LCP 3 more likely in autumn and LCP 4 in early spring. The his-
torical frequency of occurrence of the intermediate stages LCP 5 and
LCP 6 is higher in autumn and winter, with a minimum in late summer.
Again, the model is shown to be capable of reproducing such variations.
As represented by both panels, the deeper intrusions (LCP 7 and LCP 8)
have a bimodal behaviour. They are more likely to occur in winter and
in summer and less so in spring and autumn. The periodicity of these
8
Ocean Modelling 136 (2019) 1–12
LCP agrees with the frequencies found by other authors for eddy
shedding events, often associated with deep intrusion states (Chang and
Oey, 2012; Hall and Leben, 2016). The comparison with the seasonality
of the LCPs extracted by the historical data highlights the capability of
the model in properly simulating the intra-annual variability as well.
The previous analysis of the model performance suggests that the
proposed ARL model is capable of reproducing the probability of oc-
currence of the extracted LCPs. Next, in order to quantify this cap-
ability, a series of indices generally used when assessing a statistical
classifier are considered: Sensitivity, Precision, F-score and Accuracy.
Sensitivity, or True Positive Rate (TPR), is the proportion of real posi-
tive cases that are correctly predicted as positive (true positives).
Precision, or Positive Predictive Value (PPV) is defined as the propor-
tion of predicted positive cases, correctly (true positives) or incorrectly
(false positives), that are actually true positives. F-score can be inter-
preted as a weighted average of the Precision and Sensitivity. Accuracy
(ACC) is the percentage of real positive and real negative correctly
predicted. The values obtained for these parameters are shown in
Table 2 and confirm the general good capabilities of the model (average
values > 0.75). At the same time these parameters allow for a detailed
assessment of the skills of the model in reproducing each singular LCP.
The model performs better in simulating LCP 1, LCP 2 and LCP 3,
H. Chiri, et al. Ocean Modelling 136 (2019) 1–12

Fig. 5. Left: annual average probability of occurrence of the 8 extracted loop current patterns, based on historical data (left) and on the model simulation (right).

Table 2 correctly predicts the pattern (true positives) 68 times, but it misses the
Sensitivity (TPR), precision (PPV) and accuracy (ACC) for each category and correct prediction 17 times (false negatives), assigning 7 times to LCP 1,
average values. once to LCP 2 and so on. An interesting feature extractable from Fig. 6 is
Index LCP1 LCP2 LCP3 LCP4 LCP5 LCP6 LCP7 LCP8 Avg that when the model get “confused”, i.e., it does not predict the true
pattern, it tends to pick up some other spatial pattern not very different
TPR 1.00 1.0 0.80 0.70 0.73 0.79 0.58 0.73 0.79 to the actual one. An exception to this is given by LCP 3 that, when
PPV 0.82 0.90 0.81 0.74 0.76 0.69 0.80 0.76 0.78
missed by the model, is in most cases replaced by LCP 1 or LCP 8,
F-score 0.90 0.95 0.80 0.72 0.75 0.74 0.67 0.74 0.78
ACC 0.99 0.99 0.95 0.92 0.92 0.92 0.93 0.94 0.94 actually quite different to LCP 3 (see Fig. 2).

5.2. Exploratory prediction exercises

corresponding with retraction and weak intrusion LC stages (see F-
score, for example). In particular, TPR values equal to 1 for LCP 1 and
Although it is not the main goal of this paper, we perform here a set
LCP 2 remark that the model is able to find all the samples from the
of prediction exercises for a first exploration of the model skills in
historical series corresponding to these two patterns. On the other side,
predicting future stages of the LC. Efforts have been made in the pre-
the model weakest performance occurs with LCP 7, however with the
diction of the LC system evolution using statistical data analyses. Lugo-
values of the parameters > 0.5 in any case. ACC values are higher than
Fernández and Leben (2010) presented a linear model for the prediction
0.90 for all categories. These results confirm that the proposed ALR
of the time between two consecutive LC eddy shedding events based on
model is capable of simulating the LCP sequence, particularly for the
the retreat latitude of the LC at the end of the first event. Zeng et al.
patterns of LC retraction phase.
(2015a) developed an autoregressive artificial neural network for LC
These four examined parameters are obtained from the confusion
evolution forecast that is demonstrated to be reliable up to 4 to 6 weeks
matrix. This matrix breaks down the overall model results compared to
ahead.
actual samples for each category. So it is useful to analyse it in detail.
In order to test the forecasting skills of the proposed ALR metho-
The confusion matrix for our case is shown in Fig. 6, left panel as ab-
dology, we recalibrate the model using the historical LCP series from
solute counts and right panel as percentage. The confusion matrix re-
1993 to 2015, and we simulate the year 2016. We execute two simu-
lates the categories predicted by the models to the actual ones. Take
lations of the ALR model to predict 6 month ahead for each run. In the
LP3 (third row) for an example, the matrix indicates that the model
first simulation, we try to predict from January 2016 to June 2016, and

Fig. 6. Original (left) and normalized (right) confusion matrix from the ALR model simulation.

9
H. Chiri, et al. Ocean Modelling 136 (2019) 1–12

Table 3 during the first 5 steps of simulation (10 weeks), with LCP 5 the most
Evolution of actual and simulated LCP for the first 6 months of 2016. Dates refer likely predicted pattern during this period. During the next step the
to the first day of each 14-day time step. The predicted LCPs correspond to the transition from LCP 5 to LCP 6 occurs. The model is capable to predict
most likely LCP simulated by the model and the associated simulated prob- both this change and the persistence of the new pattern. As shown in
ability is shown in parenthesis. Simulated LCP matching the actual ones are
Table 3, starting from April the forecast of the model becomes less
highlighted in bold.
accurate. Instead of LCP 1, the model wrongly predicts two weeks of
Date Actual LCP Predicted LCP further intrusion as expressed by LCP 8, followed by 6 weeks of LCP 4
eventually switching to LCP 5. Although missing the retraction from
2016 1 4 5 5 (87%)
2016 1 18 5 5 (93%)
LCP 6 to LCP 1 between March and April, it is interesting to see how the
2016 2 1 5 5 (95%) model still forecasts a retraction of the LC (but from LCP 8 to LCP 4).
2016 2 15 5 5 (68%) However, neither the intensity nor the timing of the retraction is cor-
2016 2 29 5 5 (78%) rectly predicted.
2016 3 14 6 6 (54%)
For the second prediction exercise, the three steps prior to the first
2016 3 28 6 6 (68%)
2016 4 11 1 8 (63%) simulated time are a sequence of three LCP 1. As shown in Table 4, the
2016 4 25 1 8 (40%) model correctly predicts LCP 1 as the most likely pattern to occur. The
2016 5 9 1 4 (61%) model predicts that this pattern would persist for the next 6 months.
2016 5 23 1 4 (89%) Actually, the pattern only persisted from July to September, and
2016 6 6 1 4 (56%)
2016 6 20 1 5 (85%)
changed to LCP 2 starting from October. Even if the model misses to
predict LCP 2, it must be observed that both LCP 1 and LCP 2 refers to
retraction LC stages (see Fig. 2), so the predicted evolution of the LC for
in the other, from July 2016 to December 2016. We then use the his- the last 3 months of 2016 does not diverge excessively from the actual
torical LCP series of 2016 to evaluate the prediction results. Similar to one.
what described in Section 5.1, we run the model 100 times for each test Both forecast exercises show that the model has been capable to
and the results refer to the average of the probabilities obtained with predict the true pattern up to 3 months ahead. However, these are ex-
the 100 simulations. ploratory experiments. A deeper investigation should be carried on to
First of all we describe the actual historical evolution of LCPs for the better assess its predictive skills, based on a longer validation period
year 2016, as obtained by the historical sequence of BMUs (see Tables 3 and a more detailed analysis. Moreover, due to the limited length of the
and 4). The year begins with the LC in a mid-intrusion phase (LCP 5). available historical database of the currents in the GoM, especially in
This configuration persists until approximately mid-March, i.e. during 5 light of the complicated variability of the studied phenomenon, we
steps in the series. Then the situation starts to evolve, represented by suggest that longer historical databases should be used for a proper
LCP 6 for the next two steps (4 weeks), which still shows a mid-intru- prediction assessment study.
sion phase but is characterized as a tightened bottom of the “loop”.
Next, starting from around mid-April, the LC experiences a strong re- 6. Conclusions
traction, represented by LCP 1 for about 6 months until mid-October.
During the rest of the year, the LC is represented by LCP 2, corre- In this paper we introduce the Autoregressive Logistic Regression
sponding to a first step toward a new retraction-intrusion-retraction (ALR) models as the core of a statistical framework for ocean current
cycle. pattern simulation. An ALR model allows statistically simulating cur-
Since the ALR model considers the effect of three autoregressive rent patterns considering simultaneously the influence of different
terms, the prediction model is fed, besides the values of the predictors spatial and temporal scale factors (predictors/covariates).
for the whole simulation period, by the three LCPs prior to the first time In order to check the capability of these models in ocean current
stamp of simulation. For the first exercise, corresponding to the pre- simulation, we set-up, run and validate an ALR model for the Gulf of
diction of the first 6 months of 2016, the three LCPs prior to the first Mexico Loop Current as a case study. Firstly, we extract 8 re-
simulation step are LCP 6, LCP 4 and LCP 5, in this temporal sequence. presentative patterns at bi-weekly interval from a 24-year (1993 to
The results of the first prediction exercise are shown in Table 3. It is 2016) altimetry-derived database of the geostrophic current in the Gulf
encouraging that the model predicts well the persistence of LCP 5 of Mexico, using a combination of PCA, wavelet spectral analysis and K-
means techniques. In order to define a model capable of reproducing
the historical evolution of these patterns we investigate the capacity of
Table 4 different factors as predictors of the LC variability. Based on previous
Evolution of actual and simulated LCP for the last 6 months of 2016. Dates refer studies, we consider the following parameters: autoregressive terms of
to the first day of each 14-day time step. The predicted LCPs correspond to the different order, wind stress curl over the GoM and the Caribben Sea, the
most likely LCP simulated by the model and the associated simulated prob-
North Atlantic MSLPAs and various climate indices with different time
ability is shown in parenthesis. Simulated LCP matching the actual ones are
lag. We find that autoregressive terms of order 1, 2 and 3 (i.e. the
highlighted in bold.
patterns occurring on the six previous weeks), the wind stress curl over
Date Actual LCP Predicted LCP the GoM with no lag and over the Caribbean Sea with 6-week lag and
the North Atlantic MSLPAs with 4-week lag are statistically significant
2016 7 4 1 1 (85%)
2016 7 18 1 1 (97%) for our ALR model. None of the climate indices is significant as pre-
2016 8 1 1 1 (78%) dictors for the LCP sequence. The final model combines 141 coeffi-
2016 8 15 1 1 (69%) cients: one for the intercept, 7 for each autoregressive component, 68
2016 8 29 1 1 (96%)
for the PCs of the wind stress curl in the GoM, 41 for the PCs of the wind
2016 9 12 1 1 (61%)
2016 9 26 1 1 (94%) stress curl in the Caribbean Sea and 10 corresponding to the PCs of the
2016 10 10 2 1 (94%) North Atlantic MSLPAs. These coefficients are optimized to achieve the
2016 10 24 2 1 (79%) best fit between the model and the historical pattern series.
2016 11 7 2 1 (94%) In order to check for the goodness-of-fit of the model, we compare
2016 11 21 2 1 (54%)
the simulated and historical Loop Current pattern series at both intra-
2016 12 5 2 1 (48%)
2016 12 19 2 1 (91%) annual and inter-annual scales. Despites the strong variability of the
selected current, we find that the model is capable of reproducing its

10
H. Chiri, et al.
behaviour at both scales. A point-by-point comparison between actual
and simulated pattern series is analysed through the confusion matrix
and four derived indexes: Sensitivity, Precision, F-score and Accuracy.
We find that the overall performance of the model is good (average
values of the three coefficients > 0.75) and that it is particularly effi-
cient in simulating the patterns representative of retraction stages
(LCP1 and LCP2). We also find that in the fitting process, when the
model does not select the true pattern, it tends to pick up some other
spatial pattern close to the actual one.
We also perform two forecast exercises for a preliminary assessment
of the predictive skills of the model and found that, in both cases, the
model has been capable of correctly predicting the actual LCP during
the first 3 months of forecast. For the rest of the simulated period, al-
though the predicted patterns do not match the historical LCPs, the
model looks, to some extent, capable of predicting the general future
behaviour of the LC. Even if the model forecast capabilities look pro-
mising, at least a few months ahead, a more complete study is required
to better analyse this interesting application of the ALR model.
The results of this study show the effectiveness of the proposed
statistical framework in analysing the evolution of ocean current pat-
terns. This methodology could be implemented in practical applications
like, for example, to emulate statistically sound oceanographic sce-
narios, through climate-based Monte Carlo simulations, and even for
probabilistic mid-term forecasts of ocean current patterns.
Acknowledgments
The authors acknowledge the support of the Spanish Ministry of
Economy, Industry and Competitiveness (MINECO) under PLVMA3D
(TRA2014-59570-R) and OILHAZARD3D (TRA2017-89164-R) National
research projects. J.A.A. Antolínez was funded by the Spanish Ministry
of Education, Culture and Sport (MECD) FPU (Formación del
Profesorado Universitario) studentship BOE-A-2013-12235. Y. Liu and
R.H. Weisberg are partially supported by the Gulf of Mexico Research
Initiative (GOMRI) grant G-231804. The authors would like to thank
Prof. Fernando J. Mendez from the University of Cantabria for fruitful
discussions about logistic regression and clustering techniques, and
three anonymous reviewers for key suggestions.
References
Abascal, A.J., Sanchez, J., Chiri, H., Ferrer, M.I., Cárdenas, M., Gallego, A., Castanedo, S.,
Medina, R., Alonso-Martirena, A., Berx, B., Turrell, W.R., Hughes, S.L., 2017.
Operational oil spill trajectory modelling using HF radar currents: a northwest
European continental shelf case study. Mar. Pollut. Bull. 119 (1), 336–350.
Agresti, A., 2007. An Introduction to Categorical Data Analysis. John Willey and Sons.
Inc., Hoboken.
Alvera-Azcárate, A., Barth, A., Weisberg, R.H., 2009. The surface circulation of the
Caribbean Sea and the Gulf of Mexico as inferred from satellite altimetry. J. Phys.
Oceanogr. 39 (3), 640–657.
Alves, J.M.R., Peliz, A., Caldeira, R.M.A., Miranda, P.M.A., 2018. Atmosphere-ocean
feedbacks in a coastal upwelling system. Ocean Model. 123, 55–65.
Antolínez, J.A.A., Méndez, F.J., Camus, P., Vitousek, S., González, E.M., Ruggiero, P.,
Barnard, P., 2016. A multiscale climate emulator for long-term morphodynamics
(MUSCLE-morpho). J. Geophys. Res. Oceans 121, 775–791.
Antolínez, J.A.A., Murray, A.B., Méndez, F.J., Moore, L.J., Farley, G., Wood, J., 2018.
Downscaling changing coastlines in a changing climate: the Hybrid approach. J.
Geophys. Res. Earth 123 (2), 229–251.
Camus, P., Mendez, F.J., Medina, R., Cofiño, A.S., 2011. Analysis of clustering and se-
lection algorithms for the study of multivariate wave climate. Coast. Eng. 58 (6),
453–462.
Cárdenas, M., Abascal, A.J., Castanedo, S., Chiri, H., Núñez, P., Antolínez, J.A.A., Mellor,
A., Clements, A., Perez-Diaz, B., Medina, R., 2017. Short-term and medium-term
forecast of oil spill trajectories: application to local and regional scales. In: 2017
International Oil Spill Conference.
Chang, Y.L., Oey, L.Y., 2012. Why does the Loop Current tend to shed more eddies in
summer and winter? Geophys. Res. Lett. 39 (5).
Chassignet, E.P., Hurlburt, H.E., Metzger, E.J., Smedstad, O.M., Cummings, J., Halliwell,
G.R., Bleck, R., Baraille, R., Wallcraft, A.J., Lozano, C., Tolman, H., Srinivasan, A.,
Hankin, S., Cornillon, P., Weisberg, R.H., Barth, A., He, R., Werner, C., Wilkin, J.,
2009. U.S. GODAE: global ocean prediction with the HYbrid Coordinate Ocean Model
(HYCOM). Oceanography 22, 48–59.
de Vries, S.O., Fidler, V., Kuipers, W.D., Hunink, M.G., 1998. Fitting multistate transition
11
Ocean Modelling 136 (2019) 1–12
models with autoregressive logistic regression: supervised exercise in intermittent
claudication. Med. Decis. Mak. 18 (1), 52–60.
Enfield, D.B., Mestas-Nuñez, A.M., Trimble, P.J., 2001. The Atlantic multidecadal oscil-
lation and its relation to rainfall and river flows in the continental US. Geophys. Res.
Lett. 28 (10), 2077–2080.
Gent, P.R., 2017. A commentary on the Atlantic meridional overturning circulation sta-
bility in climate models. Ocean Model. 122, 57–66.
Guanche, Y., Mínguez, R., Méndez, F.J., 2014. Autoregressive logistic regression applied
to atmospheric circulation patterns. Clim. Dyn. 42 (1–2), 537–552.
Hall, C.A., Leben, R.R., 2016. Observational evidence of seasonality in the timing of loop
current eddy separation. Dyn. Atmos. Oceans 76, 240–267.
Hsieh, C.H.H., Chiu, T.S., Shih, C.T., 2004. Copepod diversity and composition as in-
dicators of intrusion of the Kuroshio Branch Current into the Northern Taiwan Strait
in Spring 2000. Zool. Stud. 43 (2), 393–403.
Iskandar, I., Tozuka, T., Masumoto, Y., Yamagata, T., 2008. Impact of Indian Ocean di-
pole on intraseasonal zonal currents at 90°E on the equator as revealed by self-or-
ganizing map. Geophys. Res. Lett. 35 (14).
Jin, B., Wang, G., Liu, Y., Zhang, R., 2010. Interaction between the East China Sea
Kuroshio and the Ryukyu current as revealed by the self-organizing map. J. Geophys.
Res. Oceans 115 (C12).
Kamenkovich, V.M., 1977. Fundamentals of Ocean Dynamics. Elsevier Sci. Publ. Co,
Amsterdam.
Klotzbach, P.J., Gray, W.M., 2008. Multidecadal variability in North Atlantic tropical
cyclone activity. J. Clim. 21 (15), 3929–3935.
Kohonen, T., 1982. Self-organized formation of topologically correct features maps. Biol.
Cybern. 43, 59–69.
Kohonen, T., Schroeder, M.R., Huang, T.S., 2001. Self-organizing Maps, third ed.
Springer-Verlag, New York.
Liu, Y., Weisberg, R.H., 2005. Patterns of ocean current variability on the West Florida
Shelf using the self-organizing map. J. Geophys. Res. Oceans 110 (C6).
Liu, Y., Liang, X.S., Weisberg, R.H., 2007a. Rectification of the bias in the wavelet power
spectrum. J. Atmos. Ocean. Technol. 24 (12), 2093–2102.
Liu, Y., Weisberg, R.H., Shay, L.K., 2007b. Current patterns on the West Florida Shelf from
joint self-organizing map analyses of HF radar and ADCP data. J. Atmos. Ocean.
Technol. 24 (4), 702–712.
Liu, Y., Weisberg, R.H., Yuan, Y., 2008. Patterns of upper layer circulation variability in
the South China Sea from satellite altimetry using the Self-organizing Map. Acta
Oceanol. Sin. 27 (Suppl), 129–144.
Liu, Y., Weisberg, R.H., Hu, C., Kovach, C., Riethmuller, R., 2011. Evolution of the Loop
Current system during the Deepwater Horizon oil spill event as observed with drifters
and satellites. In: Liu, Y. (Ed.), Monitoring and Modelling the Deepwater Horizon Oil
Spill: A Record-breaking Enterprise. Geophysical Monograph Series 195. American
Geophysical Union, Washington, DC, pp. 91–101.
Liu, Y., Weisberg, R.H., Lenes, J.M., Zheng, L., Hubbard, K., Walsh, J.J., 2016a. Offshore
forcing on the “pressure point” of the West Florida shelf: anomalous upwelling and its
influence on harmful algal blooms. J. Geophys. Res. Oceans 121, 5501–5515.
Liu, Y., Weisberg, R.H., Vignudelli, S., Mitchum, G.T., 2016b. Patterns of the Loop Current
system and regions of sea surface height variability in the eastern Gulf of Mexico
revealed by the Self-organizing Maps. J. Geophys. Res. Oceans 121, 2347–2366.
Lloyd, S.P., 1957. Least squares quantization in PCM. In: Technical Report RR-5497, Bell
Lab, September 1957.
Lo, H.K., McCord, M.R., 1998. Adaptive ship routing through stochastic ocean currents:
general formulations and empirical results. Transp. Res. A Policy Pract. 32 (7),
547–561.
Lugo-Fernández, A., 2007. Is the Loop Current a chaotic oscillator? J. Phys. Oceanogr. 37
(6), 1455–1469.
Lugo-Fernández, A., 2016. The Loop Current dynamics experiment (2009–2011) in the
Gulf of Mexico. Dyn. Atmos. Oceans 76, 155.
Lugo-Fernández, A., Leben, R.R., 2010. On the linear relationship between Loop Current
retreat latitude and eddy separation period. J. Phys. Oceanogr. 40 (12), 2778–2784.
McQueen, J.B., 1967. Some methods for classification and analysis of multivariate ob-
servations. In: Le Cam, L.M., Neyman, J. (Eds.), Proceedings of the Fifth Berkeley
Symposium on Mathematical Statistics and Probability. 1. University of California
Press, California, pp. 281–297.
Mesinger, F., DiMego, G., Kalnay, E., Mitchell, K., Shafran, P.C., Ebisuzaki, W., Jović, D.,
Woollen, J., Rogers, E., Berbery, E.H., Ek, M.B., Fan, Y., Grumbine, R., Higgins, W.,
Li, H., Lin, Y., Manikin, G., Parrish, D., Shi, W., 2006. North American regional re-
analysis. Bull. Am. Meteorol. Soc. 87, 343–360.
Mínguez, R., Abascal, A.J., Castanedo, S., Medina, R., 2012. Stochastic Lagrangian tra-
jectory model for drifting objects in the ocean. Stoch. Env. Res. Risk A. 26 (8),
1081–1093.
Muller-Karger, F.E., Smith, J.P., Werner, S., Chen, R., Roffer, M., Liu, Y., Muhling, B.,
Lindo-Atichati, D., Lamkin, J., Cerdeira-Estrada, S., Enfield, D.B., 2015. Natural
variability of surface oceanographic conditions in the offshore Gulf of Mexico. Prog.
Oceanogr. 134, 54–76.
NARR. National Centers for Environmental Prediction/National Weather Service/NOAA/
U.S. Department of Commerce, 2005. NCEP North American Regional Reanalysis
(NARR). Research Data Archive at the National Center for Atmospheric Research,
Computational and Information Systems Laboratory Updated monthly. http://rda.
ucar.edu/datasets/ds608.0/, Accessed date: 8 November 2017.
Oey, L.Y., Lee, H.C., Schmitz Jr., W.J., 2003. Effects of winds and Caribbean eddies on the
frequency of Loop Current eddy shedding: a numerical model study. J. Geophys. Res.
108, 3324.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M.,
Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D.,
Brucher, M., Perrot, M., Duchesnay, É., 2011. Scikit-learn: machine learning in
H. Chiri, et al.
Python. J. Mach. Learn. Res. 2825–2830.
Prasad, K., Dash, S.K., Mohanty, U.C., 2010. A logistic regression approach for monthly
rainfall forecasts in meteorological subdivisions of India based on DEMETER retro-
spective forecasts. Int. J. Climatol. 30, 1577–1588.
Rio, M.H., Mulet, S., Picot, N., 2014. Beyond GOCE for the ocean circulation estimate:
synergetic use of altimetry, gravimetry, and in situ data provides new insight into
geostrophic and Ekman currents. Geophys. Res. Lett. 41, 8918–8925.
Roberts, M.J., Van den Berg, M., 2002. Recruitment variability of chokka squid (Loligo
vulgaris reynaudii)-role of currents on the Agulhas Bank (South Africa) in paralarvae
distribution and food abundance. B. Mar. Sci. 71 (2), 691–710.
Rogers, J.S., Monismith, S.G., Fringer, O.B., Koweek, D.A., Dunbar, R.B., 2017. A coupled
wave-hydrodynamic model of an atoll with high friction: mechanisms for flow,
connectivity, and ecological implications. Ocean Model. 110, 66–82.
Saha, S., Moorthi, S., Pan, H., Wu, X., Wang, J., Nadiga, S., Tripp, P., Kistler, R., Woollen,
J., Behringer, D., Liu, H., Stokes, D., Grumbine, R., Gayno, G., Wang, J., Hou, Y.,
Chuang, H., Juang, H., Sela, J., Iredell, M., Treadon, R., Kleist, D., Van Delst, P.,
Keyser, D., Derber, J., Ek, M., Meng, J., Wei, H., Yang, R., Lord, S., van den Dool, H.,
Kumar, A., Wang, W., Long, C., Chelliah, M., Xue, Y., Huang, B., Schemm, J.,
Ebisuzaki, W., Lin, R., Xie, P., Chen, M., Zhou, S., Higgins, W., Zou, C., Liu, Q., Chen,
Y., Han, Y., Cucurull, L., Reynolds, R.W., Rutledge, G., Goldberg, M., 2010. The NCEP
climate forecast system reanalysis. Bull. Am. Meteorol. Soc. 91, 1015–1057.
Solabarrieta, L., Rubio, A., Castanedo, S., Medina, R., Charria, G., Hernández, C., 2014.
Surface water circulation patterns in the southeastern Bay of Biscay: new evidences
from HF radar data. Cont. Shelf Res. 74, 60–76.
12
Ocean Modelling 136 (2019) 1–12
Solabarrieta, L., Rubio, A., Cárdenas, M., Castanedo, S., Esnaola, G., Méndez, F.J.,
Medina, R., Ferrer, L., 2015. Probabilistic relationships between wind and surface
water circulation patterns in the SE Bay of Biscay. Ocean Dyn. 65, 1289–1303.
Sturges, W., Lugo-Fernández, A. (Eds.), 2005. Circulation in the Gulf of Mexico:
Observations and Models. Washington DC American Geophysical Union Geophysical
Monograph Series 161.
Sudre, J., Maes, C., Garçon, V., 2013. On the global estimates of geostrophic and Ekman
surface currents. In: Limnology and Oceanography, Fluid and Environment. 3.
Tsui, I.F., Wu, C.R., 2012. Variability analysis of Kuroshio intrusion through Luzon Strait
using growing hierarchical self-organizing map. Ocean Dyn. 62 (8), 1187–1194.
Vignudelli, S., Kostianoy, A.G., Cipollini, P., Benveniste, J. (Eds.), 2011. Coastal
Altimetry, first ed. Springer, New York 566 pp.
Weisberg, R.H., Liu, Y., 2017. On the Loop Current penetration into the Gulf of Mexico. J.
Geophys. Res. Oceans 122, 9679–9694.
Wolter, K., Timlin, M.S., 2011. El Niño/Southern Oscillation behaviour since 1871 as
diagnosed in an extended multivariate ENSO index (MEI. ext). Int. J. Climatol. 31 (7),
1074–1087.
Zeng, X., Li, Y., He, R., 2015a. Predictability of the loop current variation and eddy
shedding process in the Gulf of Mexico using an artificial neural network approach. J.
Atmos. Ocean. Technol. 32 (5), 1098–1111.
Zeng, X., Li, Y., He, R., Yin, Y., 2015b. Clustering of Loop Current patterns based on the
satellite-observed sea surface height and self-organizing map. Remote Sens. Lett. 6
(1), 11–19.