129136

Transportation Research Part E 161 (2022) 102722
Contents lists available at ScienceDirect
Transportation Research Part E

journal homepage: www.elsevier.com/locate/tre
Applications of machine learning methods in port operations – A

systematic literature review
Siyavash Filom a, *, Amir M. Amiri b, Saiedeh Razavi a
a
Civil Engineering Department & McMaster Institute for Transportation and Logistics, McMaster University, Hamilton, ON, Canada
b
McMaster Institute for Transportation and Logistics, McMaster University, Hamilton, ON, Canada
A R T I C L E I N F O A B S T R A C T
Keywords: Ports are pivotal nodes in supply chain and transportation networks, in which most of the existing
Seaport data remain underutilized. Machine learning methods are versatile tools to utilize and harness the
Port hidden power of the data. Considering ever-growing adoption of machine learning as a data-
Machine learning
driven decision-making tool, the port industry is far behind other modes of transportation in
Data analytics
Systematic literature review
this transition. To fill the gap, we aimed to provide a comprehensive systematic literature review
Container terminals on this topic to analyze the previous research from different perspectives such as area of the
application, type of application, machine learning method, data, and location of the study. Results
showed that the number of articles in the field has been increasing annually, and the most
prevalent use case of machine learning methods is to predict different port characteristics.
However, there are emerging prescriptive and autonomous use cases of machine learning
methods in the literature. Furthermore, research gaps and challenges are identified, and future
research directions have been discussed from method-centric and application-centric points of
view.
1. Introduction
Maritime transportation, as the backbone of global trade, is responsible for approximately 90% of global transportation (UNCTAD,
2018). Seaports (hereinafter ports) are the main nodes in the maritime transportation network which are connected through shipping
routes. Likewise, in the modern global supply chain concept, ports’ function has shifted from ordinary operation centers (loading,
unloading, and storage) to pivotal nodes of the global supply chain which orchestrate the whole supply chain (Han, 2018). This
evolution has considerably augmented port demand in recent decades which has been also resonated through the emergence of the
container idea that integrated global trade and facilitates the connection between different transport modes (Rashed et al., 2018).
Furthermore, most of the world’s major ports are geographically surrounded by cities, which constrains their physical expansion.
Therefore, ports are enforced to increase their efficiency by external and internal measures with considerations given to decrease the
overall cost of logistics (Wu and Goh, 2010).
Considering port operations scale, even a slight enhancement at strategical, tactical, and operational decision levels will lead to a
considerable cost decrease collectively (Shankar et al., 2019). In order to make more reliable and appropriate decisions, at all decision
levels, to enhance productivity and operational efficiency, investments in automation and digitalization of ports have been increased
noticeably (de la Peña Zarzuelo et al., 2020). The developments resulting from such investments have led to the recent generation of
* Corresponding author.
E-mail addresses: Filoms@mcmaster.ca (S. Filom), Amiria7@mcmaster.ca (A.M. Amiri), Razavi@mcmaster.ca (S. Razavi).
https://doi.org/10.1016/j.tre.2022.102722
Received 8 September 2021; Received in revised form 11 April 2022; Accepted 21 April 2022
Available online 28 April 2022
1366-5545/© 2022 Elsevier Ltd. All rights reserved.
S. Filom et al. Transportation Research Part E 161 (2022) 102722
digital transformation in the context of the port – the “Smart Port” (Heilig et al., 2019). On the other hand, Industry 4.0 is the latest
transformation wave, based on cyber-physical systems (Xu et al., 2018), which covers an eclectic range of new technologies namely the
Internet of Things (IoT), Blockchain technology, Big Data, robotics, augmented reality, and Artificial Intelligence (AI). Although the
implementation of Industry 4.0 gained momentum in recent years amongst port stakeholders, there are still considerable efforts to
make to reach the ultimate potential of these so-called “blue oceans” (de la Peña Zarzuelo et al., 2020). By executing such technologies
in ports, an immense amount of wide-ranging data streams are becoming available in ports. This empowers both industry and
academia to harness the power of the data to make port operations more efficient.
Similar to many other industries, increasing efficiency in ports is highly dependent on the quality of the decisions made in port
operations and investment. Different decision-making paradigms have emerged over time, which aim to improve the quality of de
cisions. According to Mortenson et al. (2015), the evolution of decision-making paradigms can be categorized under six distinct time
intervals from 1910 to the present. On the other hand, information has been the cornerstone of enhancing the quality of the decisions.
Information system in seaports, which is also known as Port Community System (PCS), has evolved through four generations (Heilig
et al., 2019). An integrated overview of the evolution of decision making and PCS is presented in Fig. 1. It can be deduced from Fig. 1
that we are experiencing the era of “Business Analytics”, which can be supported by data from the fourth-generation PCS, a cloud-
based digitalized platform that enables port stakeholders, actors, and customers to share relevant information and collaborate elec
tronically (Moros-Daza et al., 2020). To this end, and to have a more structured theoretically founded approach, the term “Business
Analytics” is coined to capture a more general and insightful overview of the concept (Fig. 1). There are several definitions for the term,
but one of the most widely accepted definitions is provided by Davenport and Harris (2007) as “The extensive use of data, statistical
and quantitative analysis, explanatory and predictive models, and fact-based management to drive decisions and actions.”.
Accordingly, business analytics could be defined by its three main definite components (Evans, 2012):
• Descriptive Analytics: “What happened?” and/or “Why it has happened?”

• Predictive Analytics: “What will happen next?” and/or “Why will it happen next?”
• Prescriptive Analytics: “What should be done next?”
Predictive and prescriptive analytics provide more insight and foresight than descriptive analytics, which is more commonly used
in the port industry to inform port stakeholders. According to Lepenioti et al. (2020), predictive analytics covers methods such as
Machine learning (ML), statistics, and probabilistic models. Although predictive analytics has received more attention in recent years,
making appropriate decisions based on reliable predictions is still a challenge (Bertsimas and Kallus, 2014). This is where prescriptive
analytics come to the scene to bridge the gap from a prediction to a decision. Prescriptive analytics entails ML methods, Operations
Research (OR) methods, simulation, and logic-based models. Besides, there is another avenue for ML methods applications in the port
Fig. 1. Evolution of decision-making paradigms and port community systems (Source: Authors’ own elaboration based on Decision Making Par
adigms: (Mortenson et al., 2015) & Port Community Systems: (Heilig et al., 2019) and (Moros-Daza et al., 2020)).
2
operations identified in the literature which could be called “Autonomous” applications. In this type of application, ML methods are
deployed to streamline and automate daily operations in ports, with different degrees of autonomy. Throughout this research, the
articles are categorized based on the predictive, prescriptive, and autonomous as types of application.
According to the Association for the Advancement of Artificial Intelligence (AAAI), AI is defined as “advancing the scientific un
derstanding of the mechanisms underlying thought and intelligent behavior and their embodiment in machines” (“Association for the
Advancement of Artificial Intelligence,” 2018). ML is the most important subset of AI through which a system is able to automatically
learn and improve from data without being explicitly programmed (Bhavsar et al., 2017). Despite AI’s range of impact, its applications
are still very limited due to its limitations (i.e., sensitivity to training data volume and quality and lengthy offline training). The
majority of AI’s recent progress relates to generating some simple response using some input data, and there is still no clear path to
make higher levels of intelligence possible. However, in data-driven decision-making, ML methods depicted promising results in many
industries (Liang and Liu, 2018). Categorizing the applications of ML and AI in port operations could shed light on previous studies and
harmonize future research directions in the field.
This study is designed to answer the following research questions by conducting a systematic literature review: 1) How have ML
methods been applied to port operations? 2) Which ML methods have been used for different port operations? 3) What are the main
data types used for ML applications in port operations? 4) What is the overall trend of analytics used in the literature? 5) What are the
underlying implications in the state-of-the-art literature? Furthermore, the paper aims to discuss the limitations and propose future
research directions.
To provide a more structured overview, the authors divided the reviewed articles into five application areas (1) demand prediction;
(2) landside operations; (3) seaside operations; (4) safety; and (5) other applications. ML methods are gradually substituting previous
approaches used to study port- relevant problems in all five groups and are creating new avenues for further research. Prior approaches
in port operations were mainly based on statistical and OR methods. In demand prediction, statistical methods have historically been
used with reasonably acceptable results. However, ML methods have recently gained more momentum to forecast the demand since
they are able to capture non-linear behavior of the historical demand time-series and address non-stationarity, seasonality, mutability,
cyclicity, randomicity, and complexity in the data (Geng et al., 2015; Niu et al., 2018). ML is a powerful tool for both predictive and
prescriptive analytics while OR methods (mathematical programming, evolutionary computation, and heuristics) are more suitable for
prescriptive analytics (Crainic et al., 2009). By comparing the ML with OR and statistical approaches, the following four advantages for
ML-based solutions could be deduced; however, it is best to harness the potential strengths of all approaches (Cheimanoff et al., 2021;
Lepenioti et al., 2020; Olafsson et al., 2008).
i. OR and statistical approaches mainly deploy abstraction assumptions to streamline the causality between inputs and outputs of
the system and therefore, may not be capable to address the unforeseen aspects of real-world problems. This might lead to
deficient decision-making procedures based on fault model outputs (Alpaydin, 2020; Ozkarahan et al., 2005). The ML-based
approach is not rooted in prior assumptions or knowledge.
ii. A port is the main hub in the supply chain, in which a voluminous number of interactions occur on a daily basis (i.e., ships, cargo
handling facilities, workforce, trucks, and rail). Inevitably, there are numerous unrevealed relations between actors in the real-
world or variables in the extracted data. A major advantage of ML algorithms is to reveal previously undistinguished relations
(Wuest et al., 2016). This ability of the ML-based approach is noticeably important for business managers and decision-makers,
through which the competitiveness of the overall system could be increased (Buczak and Guven, 2016; Jordan and Mitchell,
2015; Kraus et al., 2020).
iii. Ports are inherently large and complex environments including a broad range of activities and actors. Therefore, this will result
in high-dimensional data which might contain multi-collinearity between variables. This feature will cause significant sensi
tivity in the model’s coefficients in statistical or approaches, resulting in less reliable outcomes. On the flip side, the ML
approach is more capable to handle multi-collinearity among input variables and increase the model’s applicability and prevent
the solution from unstable coefficients (Alpaydin, 2020; Kang et al., 2015; Mo et al., 2018).
iv. Considering the scope, size, and complexity of port-related problems, exploiting OR lens demands large-scale and complex
models. For example, many of the port-related decision-making problems are combinatorial optimization problems (i.e., berth
allocation problem and container stowage plan), which are also NP-Hard problems (Cheimanoff et al., 2021). In such problems,
solution time exponentially increases with the number of inputs (Karimi-Mamaghan et al., 2021). In the same vein, such models
need to be dynamic in order to be efficient while they are going to be deployed in the port’s dynamic business environment
(Barua et al., 2020). Computational complexity will significantly downgrade the dynamic feature of OR models and usability for
decision-makers in dynamic environments. Consequently, solving such dramatically dynamic large-scale models are expensive
in term of computational complexity (Chen et al., 2022; Hottung et al., 2020). ML methods, which are based on the model-free
philosophy, depict promising performance for combinatorial optimizations (Bengio et al., 2021).
While ML has gained considerable momentum and depicts favorable results in many fields of port operations, it is important to note
that ML is not an all-purpose tool and cannot solve all port-related problems. Although implementing ML methods in the maritime
transportation and port industry have been gradually attracting scholars, only a few studies offered to review, categorize, and find
research directions in the relevant literature (Barua et al., 2020; Heilig et al., 2019; Munim et al., 2020). Barua et al. (2020) reviewed
ML-based applications in international freight transport, in which the main focus was on the shipping industry. Munim et al. (2020)
investigated big data and AI in the maritime industry. The study deployed a bibliometric approach to review the literature and offered
a less in-depth technical review. The study by Heilig et al. (2019) focused on data-driven decision-making only in container terminals
3
and did not follow a systematic literature selection approach to gather relevant studies. Moreover, a slight improvement in the port
operations could initiate several direct and indirect economic, environmental, and social improvements. Nevertheless, there remains a
lack of review and synthesis of the existing literature to provide a comprehensive overview of the state-of-the-art addressed topics,
applied ML methods, research orientations, and delineate future research roadmap. This study strives to fill this gap by establishing a
systematic literature review. Therefore, the present study is unique due to its scope, structure, and comprehensiveness. This study
provides a systematic literature review on the applications of ML for business analytics in port operations. Fig. 2 provides a schematic
overview of this study, which is the nexus between Data (PCS), Business (Port Operations), and Methods.
The purpose of this paper is to provide a bird eye compendium of the growing applications of ML in port operations. Therefore, it is
essential to provide a clear overview of research motivations and highlight its differences from the shipping industry. The port industry
is more innovative-taker rather than innovative-maker (Carlan and Sys, 2016) and most of the port-related gathered data is under-
utilized and under-analyzed (Heilig et al., 2019). In the port industry, as one of the most long-established industries, decision-
making procedures are more based on human perception rather than on a data-driven approach (Brouer et al., 2016). In addition,
maritime transportation is beyond other transportation modes in deploying big data analytics and AI solutions (Yau et al., 2020). On
the other hand, the context of AI and big data in the maritime industry is a standalone research discipline according to Lotka’s law with
a high degree of authorship concentration. This fact indicates the point that the number of researchers in the field is limited (Munim
et al., 2020). This could be caused by rich Automatic Identification System (AIS) data for vessels which paves the way for more so
phisticated and in-depth studies mainly in the shipping industry (Yang et al., 2019a). Many of the current AI applications in the field
are developed for “vessel operation” and are based on AIS data; however, there are many other research opportunities to utilize the
available “port operations” data. Furthermore, in the transition path towards ports’ digitalization, opportunities exist for scholars to
advance the state of applied research in the digitized ports (Moros-Daza et al., 2020). The major contributions of this study to the
existing body of knowledge in this domain include the following:
• 70 studies (out of more than 2,000) are identified that applied a ML method to port operations
• A comprehensive synthesis of the state-of-the-art studies has been presented which includes three main categorization themes: Area
of the port operation (demand prediction, landside operation, seaside operation, safety), type of the application (predictive,
prescriptive, autonomous), and type of the ML method.
• Discussing the emerging role of ML applications to port operations apart predictive use cases which are prescriptive (decision-
making) and autonomous
• Underlying trends and implications are further provided
Fig. 2. Overview of the business analytics in port operations.
4
• Challenges, limitations, and future research direction are identified in applying ML methods to port operations
The rest of this paper is organized as follows: Section 2 presents the research questions and implemented methodology to address
them. Section 3 provides a brief overview of ML methods. In Section 4, ML applications in port operations are discussed. In Section 5
concluding remarks and research implications are presented and discussed. Finally, in Section 6, future research directions and
challenges are identified.
2. Review methodology
In this study, a systematic literature review has been performed to study published research articles and conference papers
comprehensively and accurately while maintaining an unbiased opinion to provide a summary of the state-of-the-art research and
propose future research directions of ML application in port operations. The methodology is influenced by (Colicchia and Strozzi,
2012; Kitchenham et al., 2010) to explain, extract, combine, and integrate all the present academic research related to the domain of
inquiry. An overview of the literature retrieval procedure is presented in Fig. 3.
Fig. 3. Literature retrieval procedure.
5
To answer the research questions, and to build a robust, systematic, and reproducible literature selection framework, this study
combines four groups of keywords to collect relevant materials from Scopus and Web of Science (WoS) databases. The first group
contains “Port”, “Seaport”, “Smart Port”, “Container Port”, “Shipping”, “Port Automation”, or “Port Operations” aiming to cover port
industry materials. The second group contains “Maritime Transportation”, “Intermodal Transportation”, “Freight Transportation”, or
“Multimodal Transportation” to capture any transportation which might cross the ports. The third group includes “Machine Learning”,
“Deep Learning”, or “Artificial Intelligence” to capture the ML-related methods. Finally, the fourth group entails “Data Analytics”,
“Business Analytics”, “Big Data”, “Data Science”, or “Digitalization”.
The first and second groups are combined using OR operator to capture the overall port-related side of the search query. The third
and fourth groups are also combined by OR operator which is responsible to capture the ML-related aspect of the search query. Then,
the two new groups are combined using AND operator as a final search query to ensure about reflection of the articles which applied
ML method to port operations. and the words in each group are mixed by ‘OR’ operator. Then, two groups of keywords are combined
by the ‘AND’ operator to shape the final search query.
In the next step, Scopus and Web of Science databases are selected to collect the relevant papers using the search query from the
previous step. First, Scopus and WoS yielded 2347 and 1196 results for the query, respectively. Then, without any filtration based on
the publication year, by removing duplicates, 2586 results remained and after trimming them by title, to realize the main topic, 976
articles were filtered. In the next step, to further reduce the size of the database, retained papers’ abstracts were analyzed to build the
final database. The main target in this step was to remove the shipping-related studies and keep the port-related research articles. After
completing these steps, 243 studies were collected for a detailed review. By examining the full text of the 243 papers from various
standpoints (method, scope, and quality), 70 articles were selected for the final database of this research. A significant drop in the
number of articles from 2586 to 70 results has been occurred in the literature retrieval procedure. Since a few cross-disciplinary
keywords were included in the search query (i.e., “port”, “Machine Learning”, Business Analytics”, and “Digitalization”), the very
initial database entailed numerous articles from those cross-disciplinary fields. Therefore, after a careful review of the abstracts, the
number of articles significantly dropped from 2586 to 70 results, in the literature retrieval procedure.
The selected references are validated in terms of quality before making the final review. All authors re-examined all the 70 selected
Table 1
List of selected journals and conferences.
Journal Number Journal Number
Transportation Research Part E: Logistics and 4 Journal of Transportation Engineering 1

Transportation Review
Ocean Engineering 3 Journal of Urban Planning and Development 1
Research in Transportation Economics 3 Knowledge-based Systems 1
Applied Sciences 2 Marine Policy 1
International Journal of Logistics Research and 2 International Journal of Production Research 1
Applications
Journal of Advanced Transportation 2 Maritime Transport Research 1
Journal of Cleaner Production 2 Mathematical and Computer Modelling 1
Journal of the Transportation Research Board 2 Mathematics 1
Maritime Economics & Logistics 2 Neurocomputing 1
Research in Transportation Business & Management 2 Operational Research 1
The Asian Journal of Shipping and Logistics 2 Polish Maritime Research 1
European Journal of Operational Research 2 Sensors (Switzerland) 1
Maritime Policy & Management 2 Transportation Research Procedia 1
Applied Mathematical Modelling 1 Conference
Computers & Industrial Engineering 1
Discrete Dynamics in Nature and Society 1 The 4th International Conference on Big Data Applications and Services 1
Environmental Modelling & Software 1 2018 International Joint Conference on Neural Networks (IJCNN) 1
Applied Soft Computing 1 31st Canadian Conference on Artificial Intelligence 1
European Journal of Transport and Infrastructure 1 51st Annual Transportation Research Forum 1
Research
Expert Systems with Applications 1 AAMAS ’19: Proceedings of the 18th International Conference on Autonomous 1
Agents and Multiagent Systems
Industrial Management & Data Systems 1 Advances in Computational Intelligence 1
International Journal of Distributed Sensor 1 IEEE International Conference on Systems, Man and Cybernetics 1
Networks
International Journal of Wireless and Mobile 1 IEEE, International Conference on Mobile Data Management 1
Computing
Journal of Computational Science 1 International Conference on Artificial Intelligence and Electromechanical 1
Automation (AIEA)
Computers and Industrial Engineering 1 International Conference on Intelligent Human-Machine Systems and Cybernetics, 1
IHMSC
Journal of Marine Science and Engineering 1 International Conference on Intelligent Transportation, Big Data & Smart City 1
(ICITBS)
Journal of Marine Science and Technology 1 Total 70
Journal of Physics: Conference Series (JPCS) 1
Journal of the Transportation Research Board 1
6
articles to ensure the unbiasedness and reliability of the framework. An overview of the sources of the selected publications is
demonstrated in Table 1. It is crystal clear that based on the journals, the application of ML methods in port operations is well-
recognized amongst high-quality peer-reviewed journals. Moreover, the list of journals indicates that the topic is published in
various topics which highlights the fact that the topic has cross-disciplinary nature.
3. Overview of machine learning methods
Machine Learning (ML) is the study of methods that can learn from data by experience. ML extracts knowledge from data and makes
informed predictions and decisions based on what has been learned, without the need for prior knowledge of the data and context. In
contrast to the earlier definition of intelligent systems, such as data mining or expert systems, which were based on pre-determined
rules to analyze the data, ML does not rely on present rules or equations as a model.
3.1. Machine learning categories
Machine Learning methods can be classified into three categories of Supervised (SL), Unsupervised (UL), and RL (Géron, 2019;
Goodfellow et al., 2016; Murphy, 2012; Russell and Norvig, 2002; Shalev-Shwartz and Ben-David, 2014; Sutton and Barto, 2018).
What distinguishes these algorithms from each other is the existence of labeled data. The following will discuss these algorithms in
more detail.
3.1.1. Supervised learning

In supervised learning, the aim is to learn from a labeled dataset (training set), to make accurate predictions for the response to new,
unseen data (testing set). Supervised learning algorithms can be classified into classification and regression, depending on whether the
goal is to predict a class label or a numeric value. There are various types of supervised learning methods most of which can be used for
both regression and classification problems. K-Nearest Neighbors (KNN), linear models, Naive Bayes (NB), ANN, Support Vector
Machines (SVM), and tree-based algorithms are the most common examples of SL methods.
3.1.2. Unsupervised learning

In contrast to supervised ML, in unsupervised learning, the algorithm seeks to make sense of data by extracting features, co-
occurrence, and underlying patterns without any labeled data. Unsupervised learning is used for various purposes, including clus
tering, anomaly detection, association, autoencoders, and data transformations. A common application of unsupervised data trans
formations is dimensionality reduction (e.g., Principal Component Analysis (PCA)), to generalize data and distill the relevant
information from a high-dimensional representation of data. Clustering algorithms however try to partition the data in such a way that
data points in a single cluster are very similar, but different from points in other clusters. k-Means, Agglomerative, and Density-based
spatial clustering (DBSCAN) are three examples of clustering algorithms that take different approaches for grouping similar data.
3.1.3. Reinforcement learning

RL is used when there is no labeled or predefined data. RL relies on the learning agent to achieve the best possible solution out of a
set of possible solutions, using a reward mechanism. In RL, the agent tries to solve a task in an environment. The agent is not told which
action to take but instead must determine which action can result in the maximum reward. Evolutionary RL (ERL) is a hybrid algorithm
that combines evolutionary algorithms with RL. Policy-based and Value-based are the main two types of RL algorithms. The former
explicitly generates a representation of a policy and keeps it in memory during learning, while the latter only stores a value function,
not any explicit policy (Nachum et al., 2017).
3.2. Conventional machine learning methods vs. deep learning
Conventional ML methods suffer from their limitations confronting high dimentional natural data in their raw format (Lecun et al.,
2015). Deep Learning Methods (DL) transcended conventional approaches so that DL models are capable to distinguish intricate
hidden structures in high-dimensional which is common in real-world applications. DL is developed based on ANNs and its goal is to
imitate the human way of decision making by using data. Although classic ML algorithms are able to solve many real-world problems,
they may fail to provide desirable results when working with high-dimensional data that has a large number of inputs and outputs. DL
is capable of dealing with high-dimensional data, unlike many other ML algorithms. Moreover, deep learning could be considered as
Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) which are two main DL algorithms that are frequently
used for such applications. CNN is best suited for analyzing spatial data such as images, while RNN is mostly applied to sequential data
such as time series or text data. RNN can retain information about the input previously received, due to their internal hidden state
vector that acts as a memory. DL is a part of a broader family of ML methods, and can be regarded as SL, UL, or RL, depending on the
nature of the problem.
4. Machine learning applications in port operations
In this study, ML applications in port operations have been thoroughly reviewed and are presented in this section. To provide a
more structured review, the applications are divided into five areas (1) demand prediction; (2) landside operations; (3) seaside
7
Table 2
Summary of the ML applications in port operations.
Article Input* Method Output Location** Type of
Application***
Demand
Prediction
(C.H. Wei, 1999) Container Throughput ANN Container Kaohsiung, Taiwan P
Throughput
(Lam et al., 2004) Container Throughput, regional GDP, ANN Container Port of Hong Kong P
import, export, retained import, re- Throughput
exports, domestic export, electricity
consumption, population,
expenditure on building and
construction (1983–2001)
(Peng and Chu, Container Throughput (2003–2006) CD, Trigonometric model, Container Taiwan three main P
2009) Seasonal dummy Throughput ports
regression, Grey Forecast,
Hybrid Grey, SARIMA
(Gosasang et al., GDP, World GDP, Exchange Rate (US MLPNN, LR Number of Bangkok port P
2011) Dollar), Population, Inflation Rate, Containers, Import, (Thailand)
Interest Rate, Fuel Price (1999–2010) and Export
(Ping and Fei, Guangdong GDP, Total import and BPNN-GA Container Guangdong (China) P
2013) export, Industry output, Hong Kong Throughput
port Throughput (2000–2011)
(Geng et al., Container Throughput, Total MRSVR-CASPSO Container Shanghai port (China) P
2015) investment in fixed assets, total Throughput
imports and exports, industrial
output, first industrial value, second
industry value, tertiary industry
value, population, total retail sales of
consumer goods, freight volume,
highway freight volume, and railway
volume (1978–2013)
(Milenković et al., Container Throughput (2010–2016) GAFANN, SA-FANN, Total, Loaded, Port of Barcelona P
2019) ARIMA Unloaded, Empty, (Spain)
and Transit
Container Flow
(Gökkuş et al., GDP, Population, Export, Container ANN-ABC, ANN-LM, Container Istanbul, Izmir, Mersin P
2017) throughput (1989–2015) Multiple Regression with throughput (Turkey)
GA, LSSVM
(Xie et al., 2017) Container Throughput (1995–2017) SARMA, LSSVR, EMD- Container Port of Singapore, Port P
LSSVR, CD-LSSVR, X12-A- of Los Angeles (USA)
LSSVR,X12-M− LSSVR,
X12-A-SAL,X12-M− SAL
(Niu et al., 2018) Container Throughput (1995–2016 VMD-ARIMA-GA-SVR, Container Singapore, Shanghai P
Singapore, 2001–2016 Shanghai) VMD-ARIMA-PSO-SVR, Throughput (China)
VMD-ARIMA-GWO-SVR
(Mo et al., 2018) Container Throughput (2001–2015) GMDHNN-SARIMA-SVR, Shanghai, Xiamen P
GMDHNN-SARIMA-GP (China)
(Chan et al., Container Throughput (2004–2015) MA, MARS, ARIMA, GM, Container Port of Ningbo- P
2019) ANN, SVR Throughput Zhoushan (China)
(Du et al., 2019) Container Throughput (2001–2017) ECS, SARMA, ARIMA, Container Shanghai port, China’s P
LSSVM,ELM Throughput above-scale ports
(China)
(Shankar et al., Container Throughput (1995–2018) LSTM, DL Container Port of Singapore P
2019) Throughput
(J. J. Ruiz-Aguilar Container Throughput (2010–2014) SOM, SARIMA,SVR volume of Port of Algeciras P
et al., 2020) containers passing (Spain)
through the port
(Yang and Chang, Container Throughput (2001–2019) CNN-LSTM Container 5 Taiwanese ports P
2020) Throughput
(Cuong et al., Container Throughput (10 years) ANN Container Vietnam Ports P
2022) Throughput
Landside
Operations
(Al-Deek, 2001) Truck Traffic BPNN Truck Traffic Port of Miami, Port of P
Jacksonville (USA)
(Sarvareddy et al., truck counts and vessel freight data, BPNN Truck Traffic Port of Canaveral P
2005) 422 individual vessel records, 1 year, (USA)
2001–2002,
SVM Truck Traffic Port of Houston (USA) P
(continued on next page)
8
Table 2 (continued )
Application***
(Xie and Huynh, Daily total discharged containers,

2010) number of loaded containers, number
of truck drop-offs, number of truck
pick-ups, seven months 2008–2009
(Fotuhi et al., Truck arrival times and locations Multi-agent Simulation + Truck Turnaround Port of Houston - Port S
2013) RL (Q-Learning) Time of Charleston (USA)
(Kourounioti Terminal data- Containers pick up ANN Container Dwell A port in middle east P
et al., 2016) data, one year Time
(Mi et al., 2016) 1650 Container Number Images HOG + SVM Container Keyhole Taicang Port (China) A
Position
(Shen et al., 2017) 19 slots, 19 containers from 4-yard DQN Stowage Plan Ningbo Port (China) S
bays in 2 blocks
(Bu et al., 2018) 3000 Container Number Images Deep CNN Container Number – A
(Saikia et al., 23ship slots, 15 need to be filled, 49 Evolutionary Deep RL Stowage Plan – S
2018) containers in 7 stacks
(Verma et al., 1391 slots, number of containers RL Stowage Plan – S
2019) 25,000 to 7,000,000
(Gao et al., 2019) 5 years daily volumes of containers LSTM Truck Traffic – P
entering the storage area
(Li et al., 2020) 1000 Container Images YOLO Container Keyhole Zhenjiang Port A
Position (China)
(Lee, 2019b) Container Video CNN Container Keyhole Busan New Port A
Position (South Korea), Port of
Singapore
(Adi et al., 2020) artificially generated data based on Deep RL Truck Traffic Busan New Port S
Busan New Port properties with five (South Korea)
container terminals
(Hottung et al., 900,000 instances from (Tierney DL Pre-marshaling plan S
2020) et al., 2017)
(X. Q. Feng et al., 26,000 Container Number Images EAST + CNN Container Number – A
2020)
(X. Feng et al., 26,000 Container Number Images YOLO + CRNN Container Number – A
2020)
(Luo and Huang, vessel freight data, ship unloading Wavelet NN Truck Traffic Guangzhou Bulk P
2020) operation records, and truck data for Terminal (China)
20 days
(Zhang et al., 3,000 Container Image Localization: ResNet + U- Container Keyhole Port of Miami, Port of A
2021) net, Recognition: CRNN Position Jacksonville (USA)
(Fahdi et al., 11 RTG’s fuel consumption details MLR RTG’s CO2 Port of Casablanca P
2021) emissions (Morocco)
(Hu et al., 2021) Port Layout MADDPG RL AGV’s anti-conflict Yangshai port (China) S
path planning
(Zhang et al., Port Layout DLR Truck Routing – S
2022)
Seaside
Operations
(Lokuge and crane productivity, available trucks at ANN Berth assignment Port of Colombo (Sri S
Alahakoon, berth, number of cranes, human Lanka)
2007) resource quality, and operational
delay
(Fancello et al., Name of ship, previous port, number FFNN ETA – P
2011) of loading and unloading crews, and
ETA month, day, and hour
(Pani et al., 2015) vessel properties, vessel owner and LOR, RF, CT ETA Port of Caligari (Italy), P
service, vessel position, and weather Port of Antwerp
conditions (Belgium)
(de León et al., Previous Experiments KNN Best OR solution for Japan S
2017) Bulk-BPA
(Abualhaol et al., AIS (2015) K-Means ++ Port Congestion Port of Halifax P
2018) Indicator (Canada), Port of
Singapore, Port of
Hong Kong
(Kim and Lee, AIS, Port MIS DNN Ship destination Port of Yeosu (South P
2019) within the harbor Korea)
area
(Li and He, 2020) Historical Liner Berth Time (Four DNN Liner Berth Time Eastern Coastal area of
Years) China
Historical Port call data (11 Years) Turnaround time P
9
Table 2 (continued )
Application***
(Stepec et al., Gradient Boosting - Port of Bordeaux

2020b) CatBoost (France)
(Liu et al., 2020) AIS K-means clustering Approach Channel Dagusha channel (Port P
Capacity of Tianjin) (China)
(Kolley et al., AIS (2016–2017) KNN, LR, DT ETA Port of Miami (USA) P
2021)
(Park et al., 2021) AIS RL + Metropolis Hastings ETA Port of Busan (South P
algorithm Korea)
(Fuentes, 2021) AIS (2013–2019) DBSCAN Ranking ports in Mediterranean Port P
bunkering
operations
Safety
(Ozturk et al., Navigation simulator + experts’ RF Collision risk in port – P
2019) interview basins
(Lee et al., 2020) Berthing velocities, berth’s number, DCT, RFC, ETC, GBC, SVM, Range of safe A Tanker Terminal P
state of the ship (Ballast, Half, Laden), and MLP berthing velocities (South Korea)
pilot class, Deadweight Tonnages
(DWT), berthing angle, tugboat
horsepower, wind speed, and wave
height
(Xiao et al., 2020) 125,259 inspection records from more Binary Logistic Regression, Ship Detention Risk, 20 countries P
than 20 countries DT Effectiveness of the
inspection regime
(Yan et al., 2021) 1600 inspection records BRF Ship Detention Risk Port of Hong Kong P
(Kim et al., 2021) operations, accident, and weather DNN, RF, GBDT Accident Port of Busan (South P
Korea)
(Alvarellos et al., weather, recorded movements of 46 ANN, GBDT Moored vessel Port of Punta P
2021) ships movements Langosteria
(Atak and vessel capacity, operation time, XGBoost, Light GBM, SVM, container terminal Two Ports in Turkey P
Arslanoğlu, temperature, wind speed, humidity, Naïve Bayes, and KNN accidents
2021) and frequency of accidents
Other
(Panchapakesan 1.6 million container voyages RF, DT, NB, SVM Damage or not – P
et al., 2018b)
(Urda Muñoz border inspection time-series, DNN amount of Port Algeciras (Spain) P
et al., 2019) 2010–2018 inspection
(Ruiz-Aguilar border inspection time-series, BRNN, RF, SVM, and KNN amount of Port Algeciras (Spain) P
et al., 2020) 2010–2012 inspection
(Peng et al., 2020) ship characteristics, arrival time, GBR, RF, BPNN, LR, and Ship Energy Jingtang Port (China) P
efficiency of facilities, handling KNN consumption at port
volume, type of trade, and weekday,
2015–2016
(Fabregat et al., weather, road traffic, air traffic, cruise GBR, SVM, RF, MARS, and Air Pollutants Port of Barcelona P
2021) ships traffic information, and wind FFNN emanated from port (Spain)
(Peng et al., 2022) AIS (March 2017) DBSCAN + LSTM Port Congestion Ningbo, Shanghai, and P
rates Singapore
*
Numbers in parentheses show the data length.
**
Unreported cases: The exact location was not mentioned in the article due to the confidentiality of the research.
***
Predictive(P), Prescriptive(S), and Autonomous (A).
operations; (4) safety; and (5) other applications. A summary overview of ML applications in these five categories is presented in Table.
2.
4.1. Demand forecast
Application of ML in port’s overall demand forecast is somehow different from other ML applications for port operations reviewed
in this study. Demand forecast applications focus on the prediction of the port throughput, which is not considered “operation” in
nature but significantly impacts port operations. Reliable demand forecasting can be a vital and essential asset for port operators and
stakeholders, which brings both short- and long-term benefits. In the short term, port decision-makers, terminal operators, hinterland
service providers, and virtually all port actors will be able to more precisely and efficiently plan their tasks. Better planning leads to
smoother port operations and a higher level of efficiency. On the other hand, reliable forecasts of demand pave the way for port
stakeholders and decision-makers to develop more effective long-term strategic plans, such as port expansions. However, physical
expansions of ports are constrained by budgets and availability of lands and can be considered as irreversible financial decisions;
therefore, require to be based on accurate and concrete port demand forecast. Appropriate decision-timing will hinder the potential
under or over-utilization of the port facilities which eventually prevents financial loss and improves efficiency. An overview of the
10
literature shows that most of the studies conducted in this area aim to forecast the demand for containerized cargoes instead of other
indicators such as cargo type. This might be rooted in (1) the comparability and uniqueness of containerized cargo units compared to
those of bulk and general cargo, or (2) shifting toward containerization as a global trend in maritime transportation (Steenken et al.,
2005).
Container throughput forecasting is an inherently complex and dynamic process due to numerous socio-economic factors affecting
the outcomes, namely Gross Domestic Product (GDP), location, seasonality patterns, fuel price, population, policies, and political
tensions. Most of the demand forecast research studies are built upon historical port’s throughput time series data (Peng and Chu,
2009; Shankar et al., 2019; Xie et al., 2017; Yang and Chang, 2020). Traditionally, time series-related problems have been resolved by
deploying statistical models which are mainly based on linear assumptions (Chen and Leung, 2004; Niu et al., 2018). Recently, in lieu
of the traditional linear statistical models, ML algorithms have been successfully established (Shankar et al., 2019; Wu and Pan, 2010;
Yang and Chang, 2020). Although ML and statistical models are strongly overlapping sub-disciplines of the field of statistics and
applied mathematics, they are two different concepts that have their own foci and preferences. For example, Regression is a standard
tool in statistics while it is also categorized as a supervised ML method since it allows to learn relationships between variables and to
extrapolate these relationships into the future (Barua et al., 2020). A detailed overview of this contentious debate about different
definitions is conducted by (Boelaert and Ollion, 2018). In using regression, ML can yield more accurate predictions in numerous fields
in comparison to statistical models, especially in dealing with big and unstructured data (Gately, 1995; Ghoddusi et al., 2019; Moscoso-
López et al., 2020; Varian, 2014). Briefly to conclude, in the port demand forecast context, regression is mostly exploited as a statistical
model (Gosasang et al., 2011; Peng and Chu, 2009). Accordingly, in the following, regression is categorized as a statistical method.
ML-based models are able to conduct free-model philosophy, which paves the way to capture the inherently dynamic nature of the
data, and also the model will be more noise-tolerant (Kaushik and Giri, 2020; Tay and Cao, 2001). ML models are frequently deployed
due to their ability to learn from the data without any prior knowledge about the problem (Hwarng and Ang, 2001). Also, one of the
most important advantages of ML-based models is that in many cases, the forecasting process could be automated; on the other hand,
statistical models, such as Linear Regression (LR) and ARIMA, must be re-conducted periodically since the input data change
constantly (Önsel Ekici et al., 2016).
There are three main approaches to forecast port demand in the literature. First, statistical models are mainly built up on linear
assumptions to streamline the underlying process (Xie et al., 2017). Since the container demand forecast process is complex and
nonlinear, statistical models could not perform well to reveal complex patterns (Xie et al., 2017). This approach includes various
methods amongst them ARIMA, Seasonal ARIMA (SARIMA), and LR are the most widely used. Second, ML-based models are frequently
deployed due to their ability to learn from the data without any prior knowledge about the problem. The third approach in the
literature is to combine statistical and ML-based algorithms to build a hybrid model, through which the advantages of both models are
aggregated. The two latter approaches are explored in the literature in the rest of this section.
4.1.1. ML methods for demand forecast

Utilizing ML in container demand forecasting can be traced back to a study conducted by C.H. Wei (1999) in which the authors
employed an ANN model to forecast transshipment container throughput at Port of Kaohsiung in Taiwan. Afterward, Lam et al. (2004)
modeled the throughput of Port of Hong Kong using ANN, wherein the Input variables were national GDP, import and export volume,
retained import, re-exports, domestic export, population, expenditure on building and construction, and electricity consumption from
1983 to 2001. According to the results, all the developed ANN models outperformed regression analysis, comparing the obtained Mean
Absolute Error (MAE) and R2 measures. Gosasang et al. (2011) conducted a study to forecast container throughput at Port of Bangkok,
Thailand. To do so, the researchers compared the performance of Multi-Layer Perceptron Neural Network (MLPNN) and LR methods
based on seven inputs, including GDP, world GDP, US dollar exchange rate, population, inflation rate, interest rate, and fuel price. The
utilized data covered 12 years of information from 1999 to 2010, wherein the outputs were container number, import amount, and
export amount. Based on the results, the MLPNN method outperformed LR, comparing their Root Mean Square Error (RMSE) and MAE.
In another study, Chan et al. (2019) compared the performance of six forecasting methods using a 12-year container throughput
dataset (from 2004 to 2015) for Port of Ningbo, China. The comparison was done among Moving Average (MA), Multivariate Adaptive
Regression Splines (MARS), ARIMA, Grey Model (GM), ANN, and Support Vector Regression (SVR), using Mean Error (ME), RMSE,
MAE, Mean Percentage Error (MPE), and Mean Absolute Percentage Error (MAPE). SVR showed the best performance followed by
ARIMA, MARS, ANN, GM, and MA. In alignment with Gosasang et al. (2011), ML-based models depict a better forecasting performance
in comparison to traditional statistical methods.
Gökkuş, Yildirim, & Aydin (2017) conducted thorough research to forecast container traffic at three major Turkish seaports,
including Istanbul, Izmir, and Mersin. This study employed four different models, including Artificial Neural Network with Artificial
Bee Colony (ANN-ABC), Artificial Neural Network with Levenberg-Marquardt (ANN-LM), Multiple Nonlinear Regression with Genetic
Algorithm (MNR-GA), and Least Square Vector Machine (LSSVM). The utilized dataset includes GDP, population, exports, and
container throughput from 1989 to 2015. Comparing the amount of obtained RMSE, MAPE, and R2 , for Port of Izmir and Port of
Istanbul, LSSVM showed the best performance, while in the case of Port of Mersin, ANN-ABC and ANN-LM outperformed the other
techniques.
In a very novel attempt, Cuong et al. (2022) deployed ANN to forecast throughput of Vietnam ports in which the model is used to
forecast container throughput which is then used for effective decision making strategy considering cooperation and competition
dynamics amongst the ports. 10 years of container volume data is used to train the model. The novelty is to use demand prediction as
the input of decision-making model.
11
4.1.2. Hybrid models for demand forecast

What can be clearly seen in the container demand forecast literature is the continual growth of using hybrid models. By combining
ML and statistical models, the hybrid model is able to harness the advantages of methods that leads to better prediction performance.
(Ping and Fei, 2013) predicted Port of Guangdong throughput in China through developing a Back Propagation Neural Network
(BPNN) model in combination with Genetic Algorithm (GA). Port throughput, regional GDP, total imports and exports, and regional
industry output were the inputs of the GA-BPNN model for the years 2000 to 2011. Geng et al. (2015) utilized a specific type of Robust
SVR (RSVR) method for container demand forecast, in which the input features are selected using MARS method (MRSVR), while
Chaotic Simulated Annealing Particle Swarm Optimization (CSAPSO) is responsible for determining the optimized values of param
eters of MRSVR model. The suggested approach was implemented using Port of Shanghai throughput data (for the years 1978 to 2012),
including investment in fixed assets, imports and exports, industrial output, GDP, industrial value, population, total retail sales of
consumer goods, freight volume, highway freight volume, and railway freight volume as inputs. The proposed MRSVR-CSAPSO
method showed the best performance, compared to MRSVR-PSO, RSVR-CSAPSO, MBPNN, and ARIMA models.
Xie et al. (2017) proposed a novel decomposition-ensemble method to decompose container throughput time-series into different
components. Afterward, different Data Characteristic Analysis (DCA) methods (X12-ARIMA, Empirical Model Decomposition (EMD),
and Classical Decomposition (CD)), were deployed to analyze each component to capture seasonality, mutability, stationarity, and
complexity of the data. Eventually, based on the DCA results, several prediction methods were utilized for each component, including
LSSVR for non-stationary components, and SARIMA and ARIMA for stationary components. The hybrid approach was implemented on
two datasets from Port of Los Angeles and Port of Singapore from the years 1995 to 2017. Results showed the superiority of X12-
ARIMA-SAL method which, compared to CD-LSSVR and X12-ARIMA-LSSVR methods.
Niu et al. (2018) built a hybrid decomposition-ensemble model using Variational Mode Decomposition (VMD) to decompose the
data into different components. This study applied the ARIMA model to predict low-frequency components, and SVR coupled with
Hybrid Grey Wolf Optimization (HGWO) method to forecast high-frequency components. The proposed hybrid VMD-ARIMA-HGWO-
SVR model was experimented by two datasets of container throughput from the Port of Singapore (1995–2016) and the Port of
Shanghai (2001–2016). The model showed the best performance in comparison to VMD-ARIMA-GA-SVR, VMD-ARIMA-PSO-SVR, and
VMD-ARIMA-GWO-SVR.
Mo et al. (2018) utilized the Group Method of Data Handling Neural Network (GMDHNN) to decompose the data. Then, the authors
used SARIMA for linear component and a selective combination of SVR, BPNN, and Genetic Programming (GP) for non-linear com
ponents. The model is implemented on datasets of the Port of Xiamen and the Port of Shanghai container throughput for the years 2001
to 2015. For the case of the Port of Xiamen, GMDHNN-SARIMA-SVR was the most accurate model, while for the Port of Shanghai,
GMDHNN-SARIMA-GP showed the best performance. Du et al. (2019) proposed a hybrid model to predict short-term container de
mand using VMD to split the Port of Shanghai dataset (2001 to 2017) into different components. Then, Butterfly optimization-based
Extreme Learning Machine (BELM) algorithm was deployed to predict the outputs. Finally, Training Error Series (TES) was used as
Error Correction Strategy (ECS) to eliminate the long-run effect of the data and to provide better short-term prediction performance.
The authors compared the VMD-BELM-ECS model to VMD-BELM, SARIMA, ARIMA, and LSSVM models, and the former showed the
best performance.
Milenković et al. (2019) combined two heuristic methods, Simulated Annealing (SA) and GA, to optimize the structure of Fuzzy
ANN (FANN). Moreover, these two non-parametric approaches were compared with the traditional parametric ARIMA approach, using
the container throughput dataset of the Port of Barcelona for the years 2010 to 2016. The utilized dataset contains five container flows,
including loaded, unloaded, empty, transit, and total flow. In all cases, non-parametric methods outperformed the parametric method.
More precisely, for total, transit, and unloaded flows, GA-FANN performed slightly better. For loaded and empty container flows, SA-
FANN resulted in lower error values.
In another study, Ruiz-Aguilar et al. (2020) developed a model to forecast container volume in Port of Algeciras, Spain, for the years
2010 to 2014. The method contains a Self-Organizing Map (SOM) to decompose time-series into clusters as input for the SARIMA
model. The output of SARIMA is fed to a SVR model in combination with the historical data of each cluster. Eventually, the hybrid
SOM-SARIMA-SVR model predicts the result (container volume) for each cluster. By deploying the hybrid model, the advantages of
both linear models (SARIMA) and nonlinear models (SVR) are captured, which is evident by comparing its prediction performance to
the performance of SVR, SOM-SVR, and SARIMA-SVR models.
Shankar et al. (2019) implemented deep-learning-based LSTM networks to predict container throughput using the Port of
Singapore monthly container dataset from 1995 to 2018. LSTM is a type of RNN that is able to capture temporal relationships between
time lags. LSTM results were compared to ARIMA, Simple Exponential Smoothing (SES), Error Trend Seasonality (ETS), Holt-Winter’s
(HW), and ARIMA-ANN models. LSTM outperformed all the other methods based on four error measures of RME, RAE, RMSE, and
MAPE. Yang and Chang (2020) combined LSTM model with CNN to forecast the demand of five ports in Taiwan for the years 2001 to
2019. The model was able to capture sequential information by its inherent special storage system through LSTM architecture in
combination with CNN’s ability to extract high-level features. The model was then compared to results of Random Forest Regression
and SVR and showed superiority, comparing their MAPE, SMAPE, RMSE, and MASE.
4.1.3. Summary of demand forecast

Overall, the overview of the relevant literature showed that ML-based methods outperformed statistical models. Accordingly, ANN
and SVM seem to be the most popular methods when it comes to forecasting port demands. What stands out in the recent studies is the
striking shift toward exploiting hybrid models since they are capable to simulate both linear and non-linear characteristics of the input
data. There are two points to indicate here. First, although, more than 50 percent of the cargo volume in maritime transportation is
12
related to bulk cargo (UNCTAD, 2020), the majority of the previous studies have focused on containerized cargo, and to the best of our
knowledge, no study has been done on bulk cargo demand forecast using ML methods so far. Second, transport is intrinsically demand-
driven, and this is more important for maritime transportation as several socio-economic factors are affecting the outcome. Therefore,
prediction of ports’ throughput solely based on container throughput information could lead to biased and misleading results.
Although there are some research studies that take GDP, industrial output, import and export, and population into account (Geng et al.,
2015; Gökkuş et al., 2017; Lam et al., 2004; Ping and Fei, 2013), there are still several exogenous factors to consider in the field.
4.2. Landside operations
Port’s landside operations start when the cargo is unloaded or loaded for a ship at berth and end when the ship is released from the
gate. Operations such as loading/unloading, storage, and transportation are core activities at any port. Due to uncertainties in demand
forecast, and short time intervals between arriving ships, are inevitable, in which demand is more than the terminal capacity. The
peaks are exacerbated by emerging gigantic ships (Caballini et al., 2020). The mentioned flaw in the terminal operations creates a
bottleneck for overall port operations which might impact the port competitiveness and induce financial loss.
Terminal operations include various types of activities in a port ranging from the interface of sea and land, which is known as a
quay, to the final release of cargo. Seaside operations are more significantly aided by information technologies in comparison with
terminal (landside) operations (Heilig and Voß, 2017). In the context of landside operations, most of the studies have been focused on
container ports. The terminal part in a container port includes three main divisions, quayside, yard, and landside area. Quayside is an
interface between sea and land in which containers are loaded and unloaded by ship-to-shore cranes to. In the yard area, which is a
buffer between seaside and landside, main operations involve loading, unloading, storage, and stacking activities while in the landside
area, operations entail internal transportation (by truck or rail) and container storage (Steenken et al., 2004). Several applications of
ML methods in landside operations were found in the literature which are described below.
4.2.1. Stowage planning

Loading containers from a port comprises an important sequence of decisions to be taken by port operators and that affects ship
turnaround time. If the port operators could not meet the pre-agreed time window for the operations, the ship charterer must pay
demurrage to the ship owner which will diminish the port credit and increase the cost of logistics. Therefore, an optimized “stowage
plan” is salient for port operators. Stowage planning is an instance of combinatorial optimization which could be solved by heuristics in
practice (Ding and Chou, 2015). Moreover, appropriate stowage planning leads to smoother yard and seaside operations considering
the fact that redundant container relocation movements (i.e., reshuffling, rehandling) cause major inefficiencies in container terminals
(Ku and Arthanari, 2016). The overall aim is to minimize container reshuffles and yard crane shifts while adhering to safety constraints
such as staircase shape sequencing, slot weight limits, and heavy-over-light limitations.
Shen et al. (2017) proposed Deep Q-learning Network (DQN) to solve the problem. The method contains 8-layer DQN which is
trained by a dataset from a ship with 19 slots and 19 corresponding containers from 4 yards bays in 2 blocks in Ningbo Port. According
to the researcher, it takes 237 s, on average, for port operators to plan the loading while the DQN model costs only 0.131 s. Saikia et al.
(2018) attempted to solve the problem by using Evolutionary RL. The model is trained by synthetic data which was inspired by real-
world data, and it was tested based on real-world data. The data contained 23 ship slots, a block of 49 containers in 7 stacks with 7
containers in each stack. Results demonstrated that the model reached the optimal solution through an acceptable number of itera
tions. Verma et al. (2019) modeled stowage plan as a RL to solve stowage planning problem. The model is trained by using a large
dataset containing 1391 slots and the number of containers varies from 25,000 to 7,000,000. The model is compared with two
metaheuristic approaches, namely SA and GA and RL approach and demonstrated the best performance.
Another ML application in landside operations is to minimize container relocation movements in the yard area, which is known to
be a major problem. Containers arrive at the port with patterns that are usually different from the pattern that the containers should be
loaded to the ships. Moreover, in unloading operations, a container with the lower priority might be positioned under more important
containers and accessing the higher-priority container will add unproductive relocation movements. These inconsistencies lead to the
under-utilization of terminal facilities and the overall inefficiency of the port (Caserta et al., 2011).
During peak-off times in container terminals, containers are re-ordered based on predefined rules to reduce peak-demand re
locations. This action is known as pre-marshaling. Hottung et al. (2020) proposed a novel methodology named Deep Learning Heuristic
Tree Search (DLTS) to optimize pre-marshaling operations. The model is trained based on 900,000 instances generated by Tierney et al.
(2017) model. DLTS is able to find the most appropriate heuristic solution. Results have shown that DLTS reduced pre-marshaling
movements by 2 and 3 moves for different instances which in aggregate could eliminate hundreds of unnecessary movements in a
container terminal scale.
4.2.2. Truck demand forecast

Truck operations are responsible for a considerable share of port-related traffic both inside and outside of a port (Heilig et al.,
2017). Trucks are the dominant mode of intra-port transportation and this gave rise to several bottlenecks and constraints for the
overall port operations capacity and efficiency (Van Der Horst and De Langen, 2008). Accordingly, there are numerous business
problems to solve and OR approach has historically received more attention than the ML approach in this field (Hottung et al., 2020).
In this section, previous studies which applied an ML method to solve a truck-related problem are reviewed. The majority of the studies
focused on predicting daily truck volume at terminals. Due to unprecedented fluctuations of truck volume, mainly caused by un
certainties in ship arrivals, port operators seek to obtain a reliable estimation of truck volume.
13
Al-Deek (2001) attempted to predict truck-related inbound and outbound traffic at Port of Miami and Port of Jacksonville, United
States by using the BPNN model. The model is trained by a 28-point dataset of truck traffic from each port and compared with the LR
method. According to the results, ANN outperformed the LR model but it is important to note that the dataset was not sizeable to
conduct an ANN study on. Sarvareddy et al. (2005) developed BPNN and Fully Recurrent NN (FRNN) model to predict daily truck
numbers at Port Canaveral, Florida. This research predicts truck traffic for petroleum, citrus, and lumber which are not considered
containerized cargo. Moreover, the dataset includes vessel freight data and truck counts for one year. The BPNN model was validated at
the 95% confidence level, but the FRNN model could not produce reliable results due to insufficient data. Xie and Huynh (2010)
developed an SVM model to predict daily truck traffic at Port of Huston. A dataset of seven months containing the daily number of
discharged and loaded containers and a number of truck drop-offs and pickups was fed into the model. Moreover, the SVM model is
compared with MultiLayer FeedForward Neural Network (MLFFNN). Results showed that SVM demonstrates better performance than
the MLFNN model. Gao et al. (2019) proposed a LSTM model to forecast daily container volume. A model is trained using 5 years of
data of the number of containers entered a storage yard. The model is compared with ARIMA and BPNN models and outperformed
based on measures of RMSE. Luo and Huang (2020) used Wavelet NN to predict short-term truck traffic at Guangzhou bulk terminal.
The data contains vessel freight data, ship unloading operation records, and truck data for 20 days.
4.2.3. Truck routing optimization

Although Vehicle Routing Problem (VRP) attracted considerable attention in the literature of container port context, only a few of
them focused on truck operations (Stahlbock and Voß, 2008). The concept of VRP in intra-port transportation includes several types of
problems, which have been mostly addressed using OR approach. Few studies in the literature attempted to use the ML approach for
VRP problems in container ports.
Jeon et al. (2011) suggested Q-learning technique to optimize Automated Guided Vehicles (AGV) routes by determining shortest-
time routes in storage yards. The model is tested by considering a 360 m length berth with three Quay Cranes (QC) and seven storage
blocks. The results depict that travel time is reduced by 18% in comparison with shortest-travel-distance routes. Adi et al. (2020)
designed deep RL network to study the problem. Deep RL uses a deep network structure in RL which enables the model to cope with
high dimensionalities of the states and action spaces. The model is trained by using artificially generated data, based on Busan New
Port properties with five container terminals. Moreover, the model is compared with SA and Tabu Search (TS) algorithms. According to
the results, the model outperformed SA and TS algorithms. Hu et al. (2021) proposed Multi-Agent Deep Deterministic Policy Gradient
(MADDPG) strategy in RL to solve the AGV path planning considering AGV conflict prevention at Yangshai port container terminal.
Based on the port layout, networks of nodes ranging from 45 to 135 nodes are built to test the model. Based on the results, the model
outperformed the time-window based Dijkstra algorithm in term of time-efficiency.
Zhang et al. (2022) proposed Deep Reinforcement Learning (DRL) based hyper-heuristic model to optimize container truck routing
problem considering uncertainties in service times. The uncertainties include crane operation time, operator proficiency, weather
conditions, and crane types. The model is deployed in online decision-making procedure which made it attractive and scalable for daily
operations. The results depict near 10% improvement compared to previous models. This study emphasizes the fact that ML-based
models are getting progressively powerful to deal with combinatorial optimization problems.
4.2.4. Container number recognition

To boost the automation process, several applications of ML methods are proposed. Container number recognition by using Optical
Character Recognition (OCR) technology attracted considerable attention in the literature. Deploying ML methods to imitate the
human recognition process is the main idea in this part which allows port operators to reduce the cost and time of the operations.
Bu et al. (2018) deployed CNN to read container numbers using visual recognition. The method is capable to recognize container
numbers in arbitrary directions and complex backgrounds in three main steps: coarse localization (to find the container number and its
orientation), fine localization (to detect and split text lines), and text recognition (to transform the image to plain text). While the
model accuracy reached 85%, detecting the text area by CNN proved to have high latency and therefore could not be efficiently used in
daily port operations. X. Feng et al. (2020) combined CRNN and lightweight YOLOV3 (Redmon et al., 2015) to recognize container
numbers. YOLOV3 is responsible to detect the text area of the container number, and CRNN is deployed to translate the image to the
text. The proposed methodology is able to efficiently perform in different weather, light, and shadow conditions and yielded accuracy
is 96% which is much higher than the previous study. X. Q. Feng et al. (2020) adopted the An Efficient and Accurate Scene Text
Detector or EAST algorithm (Zhou et al., 2017) to detect the text area by removing redundant boundaries. The EAST algorithm shows
better results than the YOLO algorithm because YOLO could not perform well in detecting crowded backgrounds. The data consisted of
26,000 images from which 20,000 were used to train the model, which yielded an overall accuracy of 97.5%.
4.2.5. Container keyhole positioning

In container ports, keyholes of the containers (i.e., container corners) need to be interlocked for safety issues during the storage,
loading, and unloading processes. The human operator is responsible to align the keyholes. To streamline this process, Laser vision
systems have been adopted but the industrialization of this technology is not efficient and cost-effective (Li et al., 2020). Another
option to automate this process is using ML methods.
Mi et al. (2016) combined Histograms of Oriented Gradients (HOG) to preprocess the container image and SVM classifier to
recognize corner castings but the model is not capable of identifying asymmetric keyholes and it is not implementable in challenging
weather and lighting conditions. The dataset contained 1650 images, and the reported precision is 94.85%. Lee (2019a) proposed a
model in which CNN and LSTM methods are combined for container corner detection. The important feature of this model is that it is
14
trained and tested by live-stream video from Busan’s new port and port of Singapore, and it could be deployed on a real-time basis.
Yielded precision achieved for the model is 98%. Li et al. (2020) used the YOLO3 algorithm, a deep CNN, to identify container
keyholes. To reduce computational time, the images were transformed into grey images. A dataset of 1000 images from the Zhenjiang
port container terminal is used to test the model and the precision rate is 96%. Zhang et al. (2021) developed a DL method in which a
combination of ResNet (He et al., 2016) and U-net (Ronneberger et al., 2018) was used to localize the keyholes and CRNN was
exploited to recognize the container corner. A dataset containing 3,000 images is used for this study and results show that the method
outperformed other similar methods (i.e., EAST + CRNN) in both localization performance (95%) and recognition accuracy (93%).
The application of machine vision is still receiving considerable attention from both academia and industry. With the constant
advancement of computational power, ML methods, and the volume of available data, demand for machine vision for port automation
has accelerated. Considering the volume of daily operations in ports, the ability to perform in real-time is one of the most important
concerns to reach the ultimate potential of such autonomous applications in ports.
4.2.6. Other applications for landside operations

Another important port efficiency indicator is “container dwell time” which depicts how long the container is stored at the port.
Higher dwell time is a sign of inefficiency which reduces overall port productivity (Moini et al., 2012). Kourounioti et al. (2016)
attempted to predict the dwell time of import containers at a container terminal in the Middle East. They used ANN to predict dwell
days based on terminal data of a container terminal in the Middle East which contained the day and month of discharge, port of origin,
size and type of container, and type of cargo. Prediction accuracy level is reported at 65%. Fotuhi et al. (2013) used RL to model yard
crane operators. Operators are modeled as agents by the Q-learning method aiming to reduce truck turnaround time in the container
yard. The model is trained by a dataset containing the number of yard cranes, truck arrival time and locations, and truck waiting time
from Port of Houston and Port of Charleston. Based on the results, the average truck waiting time is decreased by 45% in comparison
with distance-based models. Moreover, Outputs indicated that by using this model, truck idle time, truck-related emissions, drayage
costs were reduced by 7.5%, 0.5%, and 1.7%, respectively.
As a sustainability-driven application of ML, Fahdi et al. (2021) deployed Multiple Linear Regression (MLR) to predict the Rubber
Tired Gantry (RTG) fuel consumption which according to the authors claim, is responsible for 76% of CO2 emission in ports. The model
was designed based on a dataset for 11 RTG’s fuel consumption from 2017 to 2019 in the port of Casablanca. Finally, they concluded
that converting RTG’s hoisting part to electric-powered reduces around 2 million tons of CO2 emission per year.
4.2.7. Summary of landside operations

ML applications in landside operations are diverse due to the wide variety of the operations and include predictive and prescriptive
analytics as well as autonomous. Landside operations have more considerably benefited from automation, which may be attributed to:
(1) landside operations are more likely to become a bottleneck in ports, and (2) most of the labor share in ports are working in landside
operations, therefore, automating operations could enhance the overall port efficiency and decrease labor costs. The most frequently
used type of ML is DL followed by RL.
4.3. Seaside operations
It is important to clarify and stress that there are eclectic utilizations of ML methods in the ship and sea operations many of which
are triggered by rich AIS data (Yang et al., 2019b). The majority of them could be categorized under autonomous shipping (Chen et al.,
2020; Hogg and Ghosh, 2016; Mallam et al., 2020), vessel route planning and fleet management (Christiansen et al., 2013), collision
avoidance (Huang et al., 2020; Ozturk and Cicek, 2019), environmental evaluation of shipping (Abebe et al., 2020), ship traffic
patterns (Z. Xiao et al., 2020), ship fuel consumption (Yan et al., 2020), and anomaly detection systems (Riveiro et al., 2018). While
some of the above-mentioned categories are associated with “Vessel Operations”, this section presents and discusses related articles to
“Port Seaside Operations”. Port seaside operations include berth operations (managing berth and ship-to-shore crane resources),
approach channel and basin operations, and vessel operations within the port area. Two traditional problems in Port seaside operations
are Berth Allocation Problem (BAP) and Vessel Arrival Times (VAT) prediction. Furthermore, there are other applications of ML in this
part which are described below.
4.3.1. Berth operations

BAP problem aims at making a balance between incoming vessel arrivals at ports and a limited number of berths to minimize the
time spent by ships at the port. BAP is often involved combinatorial optimization (Ting et al., 2014) with several physical, technical,
and operational constraints (de Oliveira et al., 2012; Umang et al., 2013). BAP has been modeled in the literature using different
approaches including ML method. (Lokuge and Alahakoon, 2007) presented ANN model to solve BAP at Port of Colombo, Sri Lanka.
The model considered crane productivity, available trucks at berth, number of cranes, human resource quality, and operational delay.
de León et al. (2017) tackled the BAP in bulk terminals (Bulk-BAP) by using a ML-based ranking system (KNN) to sort OR-based
solutions based on experiments and features extracted from previous studies. The ranking system helps the decision-maker to opt
for the best solution which improved the overall performance by 30% than the heuristic methods. The solutions were based on Greedy
Randomized Algorithm (GRA), heuristic First-Come-First-Serve (FCFS), and metaheuristic Large Neighborhood Search (LNS). Note
worthy, Cheimanoff et al. (2021) developed a metaheuristic method of Reduced Variable Neighborhood Search (RVNS), which is an
OR approach, to solve Bulk-BAP with Tidal Constraints (Bulk BAP_TC). In the model, Random Forest Regressor is used to optimize
RVNS hyper-parameters. This work exhibits an application of the ML approach which is integrated with OR approach.
15
4.3.2. Basin and approach channel operations

The Port’s wet infrastructure, which includes the port approach channel and harbor area, is one of the most important assets of any
port which in many cases, its capacity represents the overall port capacity. Therefore, acquiring an appropriate estimation of the
approach channel density and prediction of the vessel movements in the harbor area could lead to more robust vessel tactical planning.
Liu et al. (2020) attempted to calculate the Dagusha channel’s navigational capacity, at the Port of Tianjin. Channel capacity is one
of the most predominant bottlenecks in ports, therefore, realistic capacity estimation is indispensable for port managers. One-month
AIS data is analyzed by the K-means clustering method to cluster ships traffic based on different ship types. The result identified the
maximum channel capacity, which assists in ship arrival and departure planning. Kim and Lee (2019) proposed DNN model to predict
ship’s destination within harbor area. Inputs of the DNN are sailing area, ship tonnage, ship depth, length, type, and harbor occupation
extracted from AIS data and port Management Information System (MIS) based on 2 years of data at the port of Yeosu, South Korea.
The system yields overall 85% accuracy which is 10–15% higher than the baseline model.
4.3.3. VAT prediction

Due to limited resources in ports, any uncertainty in port planning will lead to port under- or over-utilization. Delays in vessel
Estimated Time of Arrival (ETA) is an uncertain piece of information that plays an important role in port planning (Zhen et al., 2011).
Fancello et al. (2011) proposed FeedForward NN model (FFNN) to predict VAT, which is also integrated with the human resource
allocation optimization model. Name of ship, previous port, number of loading and unloading crews, and ETA month, day, and hour
are considered as inputs. Results depicted that by reducing uncertainty in arrival times prediction, human work shifts could be reduced
from three to two. Pani et al. (2015) adopted ML methods to qualitatively predict vessel delays to reduce uncertainty using Logistic
Regression (LOR), Random Forest (RF), and Classification Tree (CT). Several inputs were considered to predict ETA namely vessel
properties, vessel owner and service, vessel position, and weather conditions. The method was implemented on the data from the Port
of Caligari and the Port of Antwerp. According to results, in the Port of Caligari and the Port of Antwerp, CT and RF provide the best
results, respectively. In Park et al. (2021) vessel trajectory is acquired first by the RL framework for optimal pathfinding based on Q-
learning. Then, ETA is calculated based on the predicted trajectory by using the Metropolis Hastings algorithm. The model is tested by
port of Busan’s 5-month AIS dataset. The model shows more accurate predictions in comparison with the model proposed by (Ales
sandrini et al., 2019). Kolley et al. (2021) attempted to predict VAT as input for berth allocation optimization model. KNN, LR, and DT
are used based on AIS data from 2016 to 2017 at Port of Miami. The results of the study highlight the fact that more accurate VAT
prediction increases the robustness of berth allocation planning.
4.3.4. Vessel turnaround time prediction

One of the impactful performance indicators in port operations is vessel turnaround time. This parameter includes berthing time,
waiting time, and service (loading/unloading) time. This indicator significantly affects overall port operations efficiency and capacity.
It should be noted that higher turnaround time creates several issues such as demurrage for ships, increase in vessel emissions in the
port area, and decrease in port competitiveness (Poulsen and Sampson, 2020).
Li and He (2020) employed DNN to predict berthing time at a container terminal in eastern China. 4-year historical berthing time
reports were fed to the model and the result did not show reliable outcomes due to insufficient data, authors claimed. Stepec et al.
(2020a) employed Gradient Boosting-based method named CatBoost (Prokhorenkova et al., 2017) to predict vessel turnaround time at
the port of Bordeaux, France. The input data contained 11 years of historical port call data. Based on measures such as MAE, RMSE, and
MAPE, the model substantially outperformed the model used by the port. This improvement could indicate the fact that due to diverse
ships properties, randomness in operations, and unprecedented events, a massive amount of data is needed to train ML models to better
predict turnaround times.
4.3.5. Other applications for seaside operations

Abualhaol et al. (2018) deployed AIS data to mine port congestion indicators which are Spatial Complexity, Spatial Density, and
Time Criticality (vessel turnaround time) by using geospatial analysis. Then, the K-means++ clustering algorithm is used to char
acterize port congestion levels at ports of Hong Kong, Halifax, and Singapore based on 2015 AIS data for each port.
Bunkering operations are not inherently categorized as port operations since it is more related to vessel management practices.
However, Fuentes (2021) developed a novel methodology to assess Mediterranean ports relative importance for bunkering. Unsu
pervised Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is used to identify bunkering-related operations and
ranking of bunkering ports by activity, waiting and servicing type, and vessel properties by using AIS data from 2013 to 2019.
4.3.6. Summary of seaside operations

In seaside operations, what can be clearly deduced from the literature is the lack of prescriptive applications. Although the seaside
operations are probably the richest area of operations in term of data availability, the ultimate potential is not fully investigated in the
literature. There are several critical decisions that need to be made, e.g., berth and crane allocation, vessel dispatching, and tugboat
planning, to manage the port’s limited resources. Another noteworthy observation is that there has been no DL-based application even
though the seaside area is significantly supported by various information technologies
4.4. Safety
Naturally, safety considerations and measures are increasingly playing an indispensable part in port operations. Ports are complex
16
organizations in which several socio-economic and environmental factors are intertwined together. Correspondingly, any safety issue
could result in severe economic and environmental problems, human and financial loss, impact the competitiveness of the port, and
potentially make the port out of service.
4.4.1. Navigational safety

There are several safety risks related to ships operating within the port area (i.e., basin + approach channel). Even, a single minor
incident could cause a major challenge for the port and shipping lines. A single incident can cause blocking the port entrance or
damaging port facilities that eventually leads to cascaded impacts on supply chains and the loss of competitiveness. To prevent such
incidents, deploying ML methods could provide fruitful feedback and predictions aiming to prevent or reduce vulnerabilities.
Ozturk et al. (2019) studied navigational collision risk in port approaches and basins which are of the highest importance to port
authorities. A noticeable dataset was collected by investigating 140 pilots’ approach maneuverings in a ship handling simulator. Then,
a set of collision prevention fuzzy rules were extracted by interviewing 20 pilots as experts. The Resulting datasets were fed to a
random forest which depicted superior performance. Lee et al. (2020) took advantage of ML methods to estimate the range of safe and
unsafe berthing velocity of a ship at ports. Random Forest Classifier (RFC), Extra Trees Classifier (ETC), Gradient Boosting Classifier
(GBC), Gaussian Naïve Bayes Classifier (GNBC), KNN, and SVM, were implemented as classification methods and were compared in
their prediction of whether the berthing velocity is safe or not. The methods were implemented on a dataset containing 426 obser
vations of berthing velocity, berth’s number, state of the ship (Ballast, Half, and Laden), pilot class, Deadweight Tonnages (DWT),
berthing angle, tugboat horsepower, wind speed, and wave height at a tanker terminal in South Korea. According to results, ETC, RFC,
Bagging, and GBC presented higher accuracies than other methods.
4.4.2. Port state control

One of the most important safety procedures in ports is Port State Control (PSC) inspection which is conducted by the port authority
to ensure about safety of ships, personnel (crews), and marine environment and if the ship could not meet the conventional standards,
the ship will be detained. Several factors affect port managers’ decision to select which ship for PSC inspection like ship age, type,
deadweight, and registered flag state (Heij and Knapp, 2019). Xiao et al. (2020) used binary logistic regression to analyze ship
detention risk and decision tree to assess an inspection regime effectiveness. The methods were implemented on a dataset that con
tained 125,259 inspection cases from 2015 to 2017 from more than 20 countries. Yan et al. (2021) attempted to predict ship detention
by deploying the Balanced Random Forest (BRF) method which is capable of addressing the data imbalance issue. The dataset con
tained 2000 inspection records from Hong Kong port for three years and the model yielded an accuracy of 97%.
4.4.3. Other applications for safety

Besides safety considerations in seaside operations, landside operations safety should be of the highest importance for port op
erators. Cheng et al. (2018) utilized machine vision to detect abnormal dangerous behavior in port operations. Their model was based
on Single Shot Multibox Detector (SSD) deep NN model (Liu et al., 2016). Kim et al. (2021) adopts DNN, RF, and GBDT to predict
accidents that can take place in a container port based on Time-series datasets of operations, accident, and weather observation be
tween 2017 and 2020 for the port of Busan. Results indicate that DNN presented the best results followed by GBDT and RF,
respectively.
Large movements of moored vessels at ports might result in severe incidents in the ports causing operational delays. Therefore,
reliable prediction of the moored vessel movements could increase the safety level of the port. Accordingly, Alvarellos et al. (2021)
proposed ANN and GBDT methods to predict movements using weather data and recorded movements of 46 ships at Port of Punta
Langosteria from 2015 to 2020. ANN achieved the best performance in terms of RMSE and the coefficient of determination.
Atak and Arslanoğlu (2021) used vessel capacity, operation time, temperature, wind speed, humidity, and frequency of accidents to
predict container terminal accidents for two ports in Turkey. XGBoost, Light GBM, SVM, Naïve Bayes, and KNN methods are deployed
amongst which XGBoost, LightGBM, and KNN algorithms depicted the best results with more than 0.98 precision.
Similar to seaside operations, there is a lack of prescriptive applications in safety applications. Moreover, there are several un
addressed issues in the literature that could be addressed in future research. Oil spillage management (Mohammadiun et al., 2021),
hazardous material transportation (Aneziris et al., 2020), and personnel safety (Corrigan et al., 2020) are important safety issues that
can benefit from ML methods.
4.5. Other applications
There are several other articles in the database which could not be categorized under previous categories. Panchapakesan et al.
(2018a) attempted to predict container damage based on a dataset of 1.6 million container voyages in 2014. Fifteen main features were
extracted by expert interviews to predict whether the container is damaged or not, amongst which, yard time, shipping line, quay crane
operator, and customer have the highest impact. The researchers used RF, Decision Tree, Naïve Bayes, and SVM to predict container
damage. According to the results, RF yields the highest accuracy, which is 95%. Urda Muñoz et al. (2019) utilized DNN to predict
container inspection volume at the port of Algeciras. The method is implemented on a time-series dataset from border inspection
authority between 2010 and 2018. Results showed that DNN improved prediction accuracy by more than 10% in comparison with the
LR method.
Ruiz-Aguilar et al. (2020) proposed several models to forecast inspection volume at the port of Algeciras. They compared Bayesian
Regularized Neural Network (BRNN), RF, SVM, and KNN based on a dataset of inspections from the border inspection authority
17
between 2010 and 2012. According to the results, BRNN outperformed other methods in term of accuracy. Peng et al. (2020) deployed
Gradient Boosting Regression (GBR), RF, BPNN, LR, and KNN algorithms to estimate ship energy consumption in Jingtang Port in
China. The methods were implemented by using a dataset that contained ship characteristics, arrival time, efficiency of facilities,
handling volume, type of trade, and weekday between 2015 and 2016. Results indicate that RF and GBR were the most accurate
methods and ship net tonnage, deadweight tonnage, efficiency, and handling volume are the most important features affecting fuel
consumption.
A very unique article in the literature is the study performed by Fabregat et al. (2021) to estimate the impact of ports and cruise ship
traffic in Barcelona. This study is unique since it is the only study that was conducted on passenger ports in this review. Authors
attempted to predict the impact of cruise port shipping operations using ML methods in lieu of traditional physic-based models. Using
pollutant measurement stations data, ML methods provide more accurate results in comparison with traditional atmospheric
dispersion models. GBR, SVM, RF, MARS, and FFNN models were used to predict air pollutions emanated from the port of Barcelona
based on 25 variables including weather, road traffic, air traffic, cruise ships traffic information, and wind data for more than 3 years.
Based on the results GBR outperformed the rest of deployed models.
Peng et al. (2022) attempted to estimate Ningbo, Shanghai, and Singapore ports congestion rates based on AIS data from March
2017 to April 2017 which includes around 4,000 vessel ships movements. The model contains a DBSCAN and convex hull methods to
predict berth and anchorage areas and determine the hostorical congestion rates. Then, LSTM model is proposed to predict congestion
measures.Moreover, the model is capable of predicting sequential congestion in a series of ports (common pattern in container market)
which made it a practical tool for real-world decision-making procedures.
5. Results
All the reviewed articles are summarized in Table 2 based on the type of operation, ML method, input and outputs, type of
application (predictive, prescriptive, and autonomous), and location of the study, if applicable. Similar to Table 2, a general overview
of the above-reviewed papers is visualized in Fig. 5 in which type of applications, type of methods, and area of the applications are
presented based on the designated share in term of paper counts. In seaside and landside operations, the most widely used type of ML
methods are SL and RL. Moreover, we have UL methods applied to study seaside operations. For demand prediction, SL are more
frequently deployed since the majority of the input data is labeled. An interesting observation is that using hybrid methods (ML +
Statistical methods or ML + Operations Research methods) are emerging in the field with promising performances. In the safety area,
the number of studies is relatively low, and the main methods are from SL methods. To further analyze and discuss the results, other
classifications have been conducted which are presented below.
5.1. Area of operation (RQ1)
Fig. 4 presents the distribution of the papers between 1999 and 2021 including ML applications in port operations based on the area
of the applications. The results expose an increasing trend of the number of papers at least after 2016 since more than 60% of the
papers were published after 2017. Landside operations are responsible for 41% of the articles. It is due to the size and scope of this area
of operations which includes any operation from the quayside to the gates including stowage planning, truck demand forecast, truck
routing optimization, container number recognition, and container keyhole positioning. Moreover, landside operations is the only area
in which all types of applications (predictive, prescriptive, and autonomous) have been implemented. Next, demand prediction
contains 25% of the articles which are completely focused on container throughput forecast. Then, seaside operations are responsible
for 16% of the articles including berth operations, basin and approach channel operations, VAT prediction, and vessel turnaround time
prediction. It is also worth mentioning that safety is the most recent emerging application of ML in port operations, which could be an
engaging research field in the coming years. Safety applications entail 9% of the papers which are focused on navigational safety and
port state control.
Fig. 4. Year-wise classification of articles based on area of the operation.
18
Fig. 5. General overview of the ML applications in port operations (SL: Supervised Learning, UL: Unsupervised Learning, RL: Reinforce
ment Learning).
5.2. ML methods (RQ 2)
To answer the second research question, Fig. 6 shows the classification of the articles based on the type of ML method used which
could be supervised or unsupervised learning, reinforcement learning. SL is the most widely used type of ML in the literature by around
80 % followed by RL with 17%, and UL with 5%. By the general overview presented in Fig. 5, and the market share of the methods
presented in Fig. 6, it can be deduced that SL methods are used for all three types of applications (predictive, prescriptive, and
autonomous). SL is the only type of method that has been used for autonomous applications since all the applications in this category
are based on object detection which is built atop of CNN models. Moreover, SL methods are versatile methods mostly deployed to
predict a variable in port operations (i.e., demand, truck turnaround time, vessel ETA). RL methods are used for landside and seaside
operations which could be interpreted that these two areas of applications are agent-based operations and RL methods are capable to
model agent-based processes. In this type, RL methods are compared with OR approach methods. Unsupervised learning methods are
deployed only to conduct predictive analytics for seaside operations and safety applications.
From the ML methods point of view, by considering Fig. 6, ANN is the most widely used method in the literature followed by SVM.
The reason is that these two methods are powerful predictive methods through which several port variables have been forecasted (i.e.,
container throughput, vessel ETA, truck delays). Presumably, there are three major reasons for ANN’s popularity. First, port operations
are inherently complex systems affected by several variables. Neural networks are capable to handle the sophisticated hidden rela
tionship between those variables. Second, considering the scope of the port operations and constantly emerging information tech
nologies, amounts of stored data is gradually growing which could increase the performance and accuracy of the neural network.
After ANN, SVM is the most widely used method due to its prediction power followed by RF, DR, KNN, GB, LSSVM, and LR. In the
19
Fig. 6. Classification of articles based on the type of ML method (SL: Supervised Learning, UL: Unsupervised Learning, RL: Reinforcement Learning -
*: DL-based).
deep learning-based supervised category, LSTM is solely used for demand prediction which presents promising results since it is
capable to trace patterns in time series. On the other hand, CNN is only used for object detection purposes such as container number
recognition and container keyhole positioning. In RL family, the Q-learning method is commonly used for prescriptive models in
seaside and landside operations which are inherently agent-based operations. Moreover, DGN is an instance of a DL-based rein
forcement learning method that is deployed for prescriptive analytics. Applications of UL are limited to only two methods, K-means
and DBSCAN.
5.3. Datasets (RQ3)
An overview of the type of datasets used in this review might be insightful. Fig. 7 presents an overview of the different types of
datasets which is addressing the third research question. What can be clearly seen is the majority of the datasets used were Time-series
covering different operations in port such as container throughput, vessel-related data (AIS included), and inspection records. Video
and image datasets were used for autonomous applications in landside operations. RL-based applications mostly used hypothetical
data based on their model specifications to test the performance. It is essential to mention that deploying hypothetical and simulation-
based data lead to the applicability challenge of the proposed model. In consequence, the model might not produce reliably in real-
world operations unless it is verified by a validation test. A further notable finding is that most of the datasets are not unstructured
data (such as image, text, audio, and video). Since around 80% of the gathered data is in unstructured format (Murdoch and Detsky,
2013), it is safe to say that most of the data remained underutilized and the ultimate capacity of ML methods has not been deployed yet.
5.4. Type of application (RQ4)
As illustrated in Fig. 8, 68% of the applications were predictive analytics followed by prescriptive analytics (22%). Autonomous is
only responsible for 10% of the applications but represents the most recent type of ML applications in port operations (Research
Question 4). Considering the fact that the majority of ML applications are focused on predicting variables, it is not surprising that
predictive analytics is the most prevalent type in the literature. However, in recent years, prescriptive analytics is becoming a separate
20
Fig. 7. Classification of the datasets used in the literature.
exposition of ML methods in port operations. All autonomous applications were related to landside operations (Table 2). Autonomous
applications include two main categories, container number recognition, and container keyhole position recognition. This might be
due to the availability of video and images from landside operations. Autonomous applications are the emerging field in ML appli
cations that could bring significant benefits for port operators in near future. This type of application could have a significant role in the
paradigm shift towards the automated port concept in the future.
6. Discussion
Based on the congregation of the analysis in this study, ML applications in port industry are studied from various perspectives. One
Fig. 8. Year-wise classification of articles based on the type of application.
21
of the most indispensable sections of any systematic literature review study is to provide implications for various groups of audience
such as researchers, practitioners, and policy makers. Therefore, in this section, these implications are identified from different points
of view. The main aim of this division is to define each group’s role in the future research roadmap.
6.1. Implications for researchers
The review reveals that the study on applications of ML in port operations is gaining momentum during a past few years. This could
be identified from the number of annually published papers and emerging applications of ML in port operations such as object
detection. We expect this trend to be continued in the years to come. The majority of the studies have been conducted on cargo ports
with a minor exception on passenger ports studies (Fabregat et al., 2021). In cargo ports, most of the studies have been performed on
container ports rather than bulk or liquid cargo ports.
Applications of ML methods could be discussed from various perspectives. In terms of performance, ML methods outperformed the
previously used methods, in many use cases such as LSTM in demand forecast, RL in truck routing optimization and stowage plan and
seaside operations. In terms of area of application, most of the studies focused on landside and seaside operation. This is not surprising
since the majority of the time, cost, and resources of the operations are concentrated on these two core activities. However, safety
applications could be equally important since a minor incident could disrupt the port’s functionalities. From an analytic-based
perspective, the main use case of ML methods is predictive which is reflected in the results (around 70% of the papers). Prescrip
tive analytics is also gaining momentum in the literature in recent years. To further shed light on the recent advancements in emerging
applications, autonomous applications are discussed to be attracting attention while this type is still in its infancy. This type of ML
applications could significantly reduce the cost of the port operations while increasing the quality.
One of the most observable gaps in the literature is the lack of real-time use cases of ML methods. Most of the applications proposed
are based on historical datasets through which the main idea has been proved. Furthermore, most of the data in ports are in form of
unstructured raw format (such as image, text, audio, and video) while in the majority of the studies, Time-series data is used to build a
model. Therefore, it could be deduced that a noticeable proportion of gathered data still remains under-utilized.
6.2. Implications for practitioners
The study reveals eclectic benefits of ML implementation in port operations while the latest trends depict those emerging appli
cations of ML are increasing (i.e., safety and autonomous applications in port operations). Such outcomes suggest that practitioners
need to assess their adaptation plans for emerging technologies that are disrupting the port industry. Considering that transport is
demand-driven, and the fact that major shipping lines aim to minimize their costs, ports will need to better adapt and invest in in
formation technology and ML applications to improve their national, regional, and international competitiveness in global markets.
This transition is discernible in world major ports such as Port of Rotterdam, Port of Singapore, and Port of Hamburg (de la Peña
Zarzuelo et al., 2020).
6.3. Implications for policymakers
The ultimate capability of ML applications in port operations could not be reached unless the policy makers adjust their respective
regulatory environments with this revolutionary field. This adjustment might have several phases. First, the main enabler of ML
applications is data. Providing hardware and software infrastructure and removing non-technical barriers for researchers and prac
titioners to have access to data is of the highest importance. Matters such as data governance, ownership, privacy, and security are all
challenges to be addressed to offer data availability and transparency. Moreover, there exists a considerable gap between the education
level of ports’ labor force and the future skillsets required for the future of the industry. This gap will become a hurdle in further
developments and adaptation of advanced solutions, such as ML applications in the industry. Furthermore, in the transition path
toward automated ports, in which ML plays a significant role, ML-based applications may surpass human-operated applications. This
drastic transformation gives birth to several ethical and social challenges which should be addressed by policymakers.
7. Challenges and propositions for future research
Although ML and AI have significantly influenced several fields of research, there are still numerous ongoing debates about
inherent challenges in ML methods ranging from technical challenges (i.e., explainability and interpretability (Barredo Arrieta et al.,
2020)) to legal and ethical challenges (Carter et al., 2020). To this end, and according to the scope of the current study, it will be more
elucidative to focus on the port-related challenges in the implementation of ML applications and provide context-based future research
directions.
7.1. Limitations and challenges
Although the current study dedicated a great deal of effort to provide a comprehensive systematic literature review to the area of
inquiry, there were some limitations to conduct this study. First, articles that were written in English are considered. Second, only
published journal articles and conference papers have been gathered; therefore, book chapters, reports, and unpublished works are not
considered in the literature retrieval procedure. Third, this study was conducted from the academic point of view without considering
22
industrial practitioners’ perspectives. For this specific study, one of the main limitations in port-related studies is the sensitive data of
the ports. Since ports are pivotal nodes in countries’ supply chain ecosystems, majority of them do not reveal detailed data about their
operations to maintain their competitive advantage. For instance, many articles in this study mentioned the location of the study by
using vague phrases such as “a container port in the middle east” or “a container terminal in China” (Kourounioti et al., 2016; Li and
He, 2020). Therefore, there should be several other machine learning applications in the industry that could not be captured in this
study due to privacy and confidentiality issues. By contemplating the aforementioned limitations, this research may not reflect
complete knowledge in the field.
The main hurdle in using ML methods in port-related studies is access to data. Although several information and communication
technologies have been used in ports, acquiring data from ports and port actors is challenging due to privacy and confidentiality
concerns. Lack of policy regulations on access to data, stakeholder engagement, and information sharing are amongst the most notable
non-technical barriers. Moreover, there is no reliable benchmark for famous port-related problems, such as BAP, to serve as a per
formance evaluation metric. Providing more access to public datasets (e.g., the MNIST dataset) for researchers who can advance the
solutions could potentially address some of these challenges.
Organizational barriers are another noticeable obstacle in ports for ML adaptation, and from a broader perspective, for port’s
digitalization. Ports are multi-cultural and multi-lingual social entities and due to the historical background of global trade, it is more
presumable to categorize the port industry as an old-fashioned industry. Moreover, the industry’s workforce is remarkably diverse in
term of technology adaptation ranging from the highly specialized workforce and managers to illiterate labor workforce. The afore
mentioned reasons lead to the fact that implementing innovations, e.g., ML applications, demands comprehensive and collaborative
efforts to coordinate different actors in the port to carefully enhance their business models, policies, strategies, and work procedures.
Another main challenge is that ports are not homogeneous in their layout, operational details, and other influential factors and the
available data in one port could be different from another port. Therefore, ML models are needed to be calibrated for different port
properties and characteristics. This might hinder the ML models to make benefit from collaborative efforts. Also, in the implementation
phase, the ML should align with different available information technologies in addition to port characteristics. AS such, it is safe to say
that ML applications in port operations are often not “transferable”. Despite the rapid advancement of ML applications in port op
erations, there is still room for further research and investigation in the field.
7.2. Future research directions
In view of what has been conducted in the literature and inherent features of port operations, the under-explored areas and gaps in
the field can be outlined in the following three method-centric and four application-centric research directions:
• Proposition 1. Focus on Prescriptive applications (Method-centric)
As stated earlier, 68% of the articles focused on predictive analytics while only 22% of the articles conducted prescriptive analytics.
Thus, it is safe to say that in most areas of port operations, literature is capable to provide good insight based on predictive analytics.
Although predictive applications are well-established in the field, it can be deduced that prescriptive applications received less
attention from the research community. Amongst traditional port-related problems, many of them are combinatorial optimization
problems in nature such as berth allocation and container relocation problem. Given that these problems are NP-hard problems, most
of the state-of-the-art algorithms rely on heuristics that are computationally expensive and mathematically not well-defined. Hence,
ML methods seem to be a worthwhile candidate to make such decisions in a more optimized way (Bengio et al., 2021). Recently, in
light of developments in the machine learning field, Reinforcement Learning methods depict promising results in solving combina
torial optimization problems both in terms of solution time and solution quality. Therefore, considering the computational complexity
of port-related combinatorial optimization problems, which might increase the complexity exponentially in real-world cases,
deploying machine learning based algorithms could bring numerous benefits in this research realm. Therefore, more research is
needed to fill the gap between good insight to good action. This is known as prescriptive analytics which brings more added value to
the business in comparison with predictive analytics. Prescriptive analytics models will pave the way to extract the ultimate potential
of data-driven analytics in conjunction with predictive analytics. This combination entails a good prediction-based decision to
recommend more courses of action for performing proactive decisions. Considering the massive amount of port’s daily operations,
prescriptive analytics could contribute more effectively in operational (e.g., truck operations, stowage planning, and safety applica
tions), tactical (e.g., BAP and yard crane operations), and strategic (e.g., terminal layout and terminal expansion) decision levels by
harnessing the ultimate unleashed potential of data.
• Proposition 2. Integrating OR and ML (Method-centric)
Needless to say, more research is needed to investigate how and when ML could replace OR and how and when could ML and OR
could complement each other. From our standpoint, RL is the most similar field of ML to OR which is computationally efficient and
depicted outperforming results in the reviewed studies. Hence, a detailed and comprehensive comparison between ML and OR
methods from different perspectives (e.g., performance, computational complexity, and type of inputs) will act as a guide for future
research. The main aim will be leveraging overall performance by utilizing the fundamental power of each approach through defining
different synergies. There were two applications of combined ML and OR approaches in which ML could find the best OR-based so
lution and OR could be used to optimize ML hyperparameters (Caballini et al., 2020; Cheimanoff et al., 2021; de León et al., 2017;
23
Zhang et al., 2020). More precisely, ML is not a one-size-fits-all solution. To epitomize the fact, we refer readers to the demand
prediction section of this study. It was elicited from the literature that the best performance amongst different approaches yielded by
combining ML and traditional statistical models, or hybrid models. Nonlinearity in many real-world applications, uncertainty in data
sources, scale of the problems, and performance bound for OR methods are amongst the most convincing grounds that push researchers
to explore this avenue of research.
• Proposition 3. Real-time Analytics and Automation (Method-centric)
Although there was fractional interest in the literature to propose real-time or near real-time ML applications (X. Feng et al., 2020;
Li et al., 2020), it can be deduced from the survey that real-time applications are scarce and they are still in their infancy. Further
research in this area may provide port decision-makers with a more robust and powerful tool while they face complex trade-offs. This
feature could leverage resiliency, efficiency, and intra- and inter-organizational collaboration, especially prescriptive real-time ana
lytics. Real-time processing of large-scale streams of data is computationally challenging. To bridge this gap, and also to answer the
increasing demand for real-time systems, researchers can benefit from the scalable implementation of distributed and cloud-based
computing platforms (Lepenioti et al., 2020; Nguyen et al., 2018). Quantum computing is also another avenue that needs to be
further explored in port operations and analytics. Real-time analytics have significant potential to enhance responsiveness, resilience,
and coordination of the intra- and inter-port operations.
• Proposition 4. Port Asset Management (Application-centric)
Unlike previous suggestions, which were method-centric, the rest of the suggestions are application-centric. Port’s facilities and
equipment (e.g., ship-to-shore cranes and yard cranes) are significantly expensive assets. In addition, any equipment failure or
shutdown could disrupt daily port operations and lead to financial and operational loss. To overcome such problems, predictive
maintenance could be a potential means, through which different techniques are designed to predict faults beforehand and to take
necessary actions to prevent failures. ML techniques are gradually gaining momentum in the field of predictive maintenance and they
depict promising results (Carvalho et al., 2019). Taking advantage of this type of ML application in port asset management could be an
intriguing research topic considering the fact that due to the high volume of the operations, even a slight improvement could bring
discernible outcomes, collectively. For instance, container gantry cranes’ steel cables should be inspected periodically to ensure safety
issues. Until now, in so many ports, those cables are manually inspected which causes several freight and vessel delays. By using
predictive maintenance, port operators can plan their maintenance beforehand and prevent freight and vessel delays. There could be
several other instances such as port pavement maintenance, mooring system maintenance, and preventive dredging planning systems.
• Proposition 5. Port Hinterland Logistics and Supply Chain Integration (Application-centric)
A salient unaddressed area of port operations in the literature is port hinterland logistics. Ports have a seamless key role in supply
chains. Therefore, higher integration and coordination between port and supply chain through hinterland result in a more robust and
efficient supply chain (Ha et al., 2017). Accordingly, hinterland-related factors are more influential than maritime-related factors on
creating bottleneck and on the overall port competitiveness (Parola et al., 2017). The interaction between port and hinterland could be
done through different transportation modes (road, rail, and inland navigation) which demands a precise and accurate schedule
prediction and reliable operations. In this context, ML could be a genuine and powerful tool due to its described capabilities. This will
be a fertile direction of research not investigated before. To illustrate this proposition, considering Synchromodal logistics as the most
recent freight transportation paradigm (Giusti et al., 2019), RL methods could leverage modal decision-making procedure quality. The
enhanced intermodal integration between port and supply chain network brings numerous results such as reduced congestion,
increased capacity utilization, and more efficient flow of cargo.
• Proposition 6. Soft organizational features (Application-centric)
Overall, the main body of this research concentrates on “physical” operations in the port while there are other equally important
flows in the ports. A port is a socio-cultural commercial complex in which several interactions occur. Thus, eclectic issues might arise in
this system such as workflow automation, legislation, insurance, collaboration, and transparency which are all related to soft orga
nizational components. Applying ML methods to enhance this particular area would be a worthwhile effort. To epitomize this
proposition, consider the port as a multi-cultural complex in which several languages are used on daily basis. Using Natural Language
Processing (NLP) techniques could leverage human interactions. On the other hand, AI chatbots could outperform humankind’s
performance and decrease the overall port cost. NLP-based chatbots can offer toolboxes for the interpretation of unstructured raw data
in emails, contracts, bills of lading, and invoices.
• Proposition 7. Sustainability-driven applications (Application-centric)
There are several applications in the literature that resulted in reducing emissions in the ports, only as of the indirect or induced
result of such studies. To this end, a certain sustainability-driven study specifically for port operations was not found in the literature.
The environment sector in ports entails multiple disciplines of wastes management, energy management, water resources
24
management, hazardous material flows, and air and acoustic pollution, to name a few. Port authorities are responsible to manage these
miscellaneous and all-encompassing disciplines to ensure environmental sustainability and social protection as corporate social re
sponsibilities. Moreover, there are assorted regulations enforcing ports to intensify their environmental constraints. Extensive avail
ability of different data types in ports could unleash potential applications of ML in various sustainability issues of the port industry.
The contribution of ML-methods for ports’ sustainability is threefold. First, ML methods could directly enhance energy consumption in
port operations using intelligent energy management systems (Alzahrani et al., 2021). Second, the majority of the ports are located
near cities and the harmful impact caused by port operations could be addressed using ML methods (Fabregat et al., 2021). Finally, by
using ML methods, the concentration of energy consumption in port facilities could be identified (Fahdi et al., 2021; Tsolakis et al.,
2021).
8. Summary and conclusion
ML is gradually pervading an increasing number of engineering fields and exhibits influential and promising contributions. ML is
expected to continue to substitute and complement traditional methods in most fields and the port industry, seemingly old-fashioned
and innovation-taker industry, is no exception. A plethora of advancement in information technologies in ports and progressive de
mand for more efficient data-driven port operations are two compelling reasons that boosted ML research in the port context. To bridge
this knowledge gap, this study aimed at integrating fragmented studies in port-related literature by conducting a structured literature
review in order to illustrate a panoramic landscape of “what is going on” and equally important, answer to the question that is “what
could be done in the future” to highlight major research pathways for further studies.
The extensive review of the literature indicates the fact that this is an emerging field as it is evident based on the number of annual
publications. To extract knowledge from this review, various classifications have been defined to shed light on the status quo of the
context. Area of the operation (demand prediction, landside, seaside, safety, and other), type of ML methods (supervised, unsuper
vised, and reinforcement learning), and type of application (predictive, prescriptive, and autonomous) are the different types of
classifications demonstrated in this study. Moreover, recent trends, challenges, and future avenues of research were analyzed and
discussed. Acquiring data is still an arduous challenge in the field considering the fact that there are several privacy and confidentiality
constraints in ports. Investigating ML and OR synergies and applying ML to port hinterland logistics are amongst the most intriguing
and fruitful future research directions.
CRediT authorship contribution statement
Siyavash Filom: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Writing – original draft, Writing –
review & editing, Visualization. Amir M. Amiri: Conceptualization, Methodology, Writing – original draft, Writing – review & editing.
Saiedeh Razavi: Investigation, Writing – review & editing, Visualization, Supervision, Project administration.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Appendix. Nomenclature
IoT Internet of Things KNN K-Nearest Neighbors

AGV Automated Guided Vehicles LNS Large Neighborhood Search
AI Artificial Intelligence LR Linear Regression
AIS Automatic Identification System LSSVM Least Square Vector Machine
ANN Artificial Neural Networks LSTM Long Short-Term Memory
ANN-ABC Artificial Neural Network with Artificial Bee Colony MA Moving Average
ANN-LM Artificial Neural Network with Levenberg-Marquardt MAE Mean Absolute Error
ARIMA Autoregressive Integrated Moving Average MAPE Mean Absolute Percentage Error
BAP Berth Allocation Problem MARS Multivariate Adaptive Regression Splines
BELM Butterfly optimization-based Extreme Learning Machine ME Mean Error
BPNN Back Propagation Neural Network ML Machine Learning
BRF Balanced Random Forest MLFFNN MultiLayer FeedForward Neural Network
BRNN Bayesian Regularized Neural Network MLP Multi-Layer Perceptron
CD Classical Decomposition MLPNN MultiLayer Perceptron Neural Network
CNN Convolutional Neural Network MNR-GA Multiple Nonlinear Regression with Genetic Algorithm
CRNN Convolutional Recurrent Neural Networks MPE Mean Percentage Error
CSAPSO Chaotic Simulated Annealing Particle Swarm Optimization NB Naive Bayes
CT Classification Tree OCR Optical Character Recognition
DBSCAN Density-based spatial clustering OLS Ordinary Least Squares
DCA Data Characteristic Analysis OR Operations Research
DCT Decision Tree Classifier PCA Principal Component Analysis
25
(continued )
DL Deep Learning PCS Port Community System
DLTS Deep Learning Heuristic Tree Search PSC Port State Control
DQN Deep Q-learning Network QC Quay Cranes
DRL Deep Reinforcement Learning RFC Random Forest Classifier
EAST Efficient and Accurate Scene Text Detector RFID Radio-Frequency Identification
ECS Error Correction Strategy RL Reinforcement Learning
EMD Empirical Model Decomposition RMSE Root Mean Square Error
ERL Evolutionary Reinforcement Learning RNN Recurrent Neural Network
ETA Estimated Time of Arrival RVNS Reduced Variable Neighborhood Search
ETC Extra Trees Classifier SA Simulated Annealing
ETS Error Trend Seasonality SARIMA Seasonal Autoregressive Integrated Moving Average
FANN Fuzzy Artificial Neural Network SES Simple Exponential Smoothing
FCFS First-Come-First-Serve SL Supervised Learning
FRNN Fully Recurrent Neural Network SMA Simple Moving Average
GA Genetic Algorithm SOM Self-Organizing Map
GBC Gradient Boosting Classifier SSD Single Shot Multibox Detector
GBR Gradient Boosting Regression SVM Support Vector Machines
GDP Gross Domestic Product SVR Support Vector Regression
GM Grey Model TES Training Error Series
GMDHNN Group Method of Data Handling Neural Network TS Tabu Search
GNBC Gaussian Naïve Bayes Classifier UL Unsupervised Learning
GP Genetic Programming VAR Vector AutoRegression
GPS Global Positioning System VAT Vessel Arrival Times
GRA Greedy Randomized Algorithm VMD Variational Mode Decomposition
GRU Gated Recurrent Units VRP Vehicle Routing Problem
HGWO Hybrid Grey Wolf Optimization VTS Vessel Traffic Service
HOG Histograms of Oriented Gradients YOLOV3 You Only Look Once, version3
References
Abebe, M., Shin, Y., Noh, Y., Lee, S., Lee, I., 2020. Machine learning approaches for ship speed prediction towards energy efficient shipping. Appl. Sci. 10 (7), 2325.
Abualhaol, I., Falcon, R., Abielmona, R., Petriu, E., 2018. Mining Port Congestion Indicators from Big AIS Data. Proc. Int. Jt. Conf. Neural Networks 2018-July.
https://doi.org/10.1109/IJCNN.2018.8489187.
Adi, T.N., Iskandar, Y.A., Bae, H., 2020. Interterminal truck routing optimization using deep reinforcement learning. Sensors (Switzerland) 20, 1–20. https://doi.org/
10.3390/s20205794.
Al-Deek, H.M., 2001. Which method is better for developing freight planning models at seaports - Neural networks or multiple regression?, in. Transportation
Research Record. National Research Council 90–97. https://doi.org/10.3141/1763-14.
Alessandrini, A., Mazzarella, F., Vespe, M., 2019. Estimated Time of Arrival Using Historical Vessel Tracking Data. IEEE Trans. Intell. Transp. Syst. 20, 7–15. https://
doi.org/10.1109/TITS.2017.2789279.
Alpaydin, E., 2020. Introduction to machine learning. MIT Press.
Alvarellos, A., Figuero, A., Carro, H., Costas, R., Sande, J., Guerra, A., Peña, E., Rabuñal, J., 2021. Machine learning based moored ship movement prediction. J. Mar.
Sci. Eng. 9 (8), 800.
Alzahrani, A., Petri, I., Rezgui, Y., Ghoroghi, A., 2021. Decarbonisation of seaports: A review and directions for future research. Energy Strateg. Rev. 38, 100727
https://doi.org/10.1016/J.ESR.2021.100727.
Aneziris, O., Koromila, I., Nivolianitou, Z., 2020. A systematic literature review on LNG safety at ports. Saf. Sci. 124, 104595 https://doi.org/10.1016/j.
ssci.2019.104595.
Association for the Advancement of Artificial Intelligence, 2018. . Work 94025–94025.
Atak, Ü., Arslanoğlu, Y., 2021. Machine learning methods for predicting marine port accidents: a case study in container terminal. Ships Offshore Struct. 1–8 https://
doi.org/10.1080/17445302.2021.2003067.
Barredo Arrieta, A., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., Garcia, S., Gil-Lopez, S., Molina, D., Benjamins, R., Chatila, R., Herrera, F.,
2020. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Inf. Fusion 58, 82–115. https://doi.
org/10.1016/j.inffus.2019.12.012.
Barua, L., Zou, B., Zhou, Y., 2020. Machine learning for international freight transportation management: A comprehensive review. Res. Transp. Bus. Manag. 34,
100453 https://doi.org/10.1016/j.rtbm.2020.100453.
Bengio, Y., Lodi, A., Prouvost, A., 2021. Machine learning for combinatorial optimization: A methodological tour d’horizon. Eur. J. Oper. Res. 290, 405–421. https://
doi.org/10.1016/J.EJOR.2020.07.063.
Bertsimas, D., Kallus, N., 2020. From Predictive to Prescriptive Analytics. Manage. Sci. 66 (3), 1025–1044.
Bhavsar, P., Safro, I., Bouaynaya, N., Polikar, R., Dera, D., 2017. Machine Learning in Transportation Data Analytics, in: Data Analytics for Intelligent Transportation
Systems. Elsevier Inc., pp. 283–307. https://doi.org/10.1016/B978-0-12-809715-1.00012-2.
Boelaert, J., Ollion, É., 2018. The Great Regression: Machine Learning, Econometrics, and the Future of Quantitative Social Sciences. The Great Regression. Rev.
française Sociol. Vol. 59 (3), 475–506.
Brouer, B.D., Karsten, C.V., Pisinger, D., 2016. Big Data Optimization in Maritime Logistics, 10.1007/978-3-319-30265-2_14. Springer, Cham, pp. 319–344.
Bu, W., Yan, S., Chen, J., Yang, C., Liu, C., 2018. Visual Recognition of Container Number with Arbitrary Orientations Based on Deep Convolutional Neural Network.
Proc. - 2018 10th Int. Conf. Intell. Human-Machine Syst. Cybern. IHMSC 2018 2, 204–207. https://doi.org/10.1109/IHMSC.2018.10153.
Buczak, A.L., Guven, E., 2016. A Survey of Data Mining and Machine Learning Methods for Cyber Security Intrusion Detection. IEEE Commun. Surv. Tutorials 18,
1153–1176. https://doi.org/10.1109/COMST.2015.2494502.
Wei, C.H., Y.C.Y., 1999. A study on transit containers forecast in Kaohsiung port-applying artificial neural networks to evaluating input variables. J. Chinese Inst.
Transp. 1–20.
Caballini, C., Gracia, M.D., Mar-Ortiz, J., Sacone, S., 2020. A combined data mining – optimization approach to manage trucks operations in container terminals with
the use of a TAS: Application to an Italian and a Mexican port. Transp. Res. Part E Logist. Transp. Rev. 142, 102054 https://doi.org/10.1016/j.tre.2020.102054.
Carlan, V., Sys, C., Vanelslander, T., 2016. How port community systems can contribute to port competitiveness: Developing a cost–benefit framework. Res. Transp.
Bus. Manag. 19, 51–64.
26
Carter, S.M., Rogers, W., Win, K.T., Frazer, H., Richards, B., Houssami, N., 2020. The ethical, legal and social implications of using artificial intelligence systems in
breast cancer care. The Breast 49, 25–32. https://doi.org/10.1016/j.breast.2019.10.001.
Carvalho, T.P., Soares, F.A.A.M.N., Vita, R., Francisco, R.d.P., Basto, J.P., Alcalá, S.G.S., 2019. A systematic literature review of machine learning methods applied to
predictive maintenance. Comput. Ind. Eng. 137, 106024.
Caserta, M., Schwarze, S., Voß, S., 2011. Container rehandling at maritime container terminals. Oper. Res. Comput. Sci. Interfaces Ser. https://doi.org/10.1007/978-
1-4419-8408-1_13.
Chan, H.K., Xu, S., Qi, X., 2019. A comparison of time series methods for forecasting container throughput. Int. J. Logist. Res. Appl. 22, 294–303. https://doi.org/
10.1080/13675567.2018.1525342.
Cheimanoff, N., Fontane, F., Kitri, M.N., Tchernev, N., 2021. A reduced VNS based approach for the dynamic continuous berth allocation problem in bulk terminals
with tidal constraints. Expert Syst. Appl. 168, 114215.
Chen, A.S., Leung, M.T., 2004. Regression neural network for error correction in foreign exchange forecasting and trading. Comput. Oper. Res. 31, 1049–1068.
https://doi.org/10.1016/S0305-0548(03)00064-9.
Chen, R., Meng, Q., Jia, P., 2022. Container port drayage operations and management: Past and future. Transp. Res. Part E Logist. Transp. Rev. 159, 102633 https://
doi.org/10.1016/j.tre.2022.102633.
Chen, Z., Chen, D., Zhang, Y., Cheng, X., Zhang, M., Wu, C., 2020. Deep learning for autonomous ship-oriented small ship detection. Saf. Sci. 130, 104812 https://doi.
org/10.1016/j.ssci.2020.104812.
Cheng, G., Wang, S., Guo, T., Han, X., Cai, G., Gao, F., Dong, J., 2018. Abnormal behavior detection for harbour operator safety under complex video surveillance
scenes. 2017 Int. Conf. Secur. Pattern Anal. Cybern. SPAC 2017 2018-Janua, 324–328. https://doi.org/10.1109/SPAC.2017.8304298.
Christiansen, M., Fagerholt, K., Nygreen, B., Ronen, D., 2013. Ship routing and scheduling in the new millennium. Eur. J. Oper. Res. 228, 467–483. https://doi.org/
10.1016/j.ejor.2012.12.002.
Wilding, R., Colicchia, C., Strozzi, F., 2012. Supply chain risk management: a new methodology for a systematic literature review. Supply Chain Manag. An Int. J. 17
(4), 403–418.
Corrigan, S., Kay, A., Ryan, M., Brazil, B., Ward, M.E., 2020. Human factors & safety culture: Challenges & opportunities for the port environment. Saf. Sci. 125,
103854 https://doi.org/10.1016/j.ssci.2018.02.030.
Crainic, T.G., Gendreau, M., Potvin, J.-Y., 2009. Intelligent freight-transportation systems: Assessment and the contribution of operations research. Transp. Res. Part C
Emerg. Technol. 17, 541–557. https://doi.org/10.1016/j.trc.2008.07.002.
Cuong, T.N., Kim, H.-S., You, S.-S., Nguyen, D.A., 2022. Seaport throughput forecasting and post COVID-19 recovery policy by using effective decision-making
strategy: A case study of Vietnam ports. Comput. Ind. Eng. 168, 108102 https://doi.org/10.1016/J.CIE.2022.108102.
Davenport, T.H., Harris, J.G., 2007. Competing on Analytics: The New Science of Winning, 1st ed. Harvard Business School Press, USA.
de la Peña Zarzuelo, I., Freire Soeane, M.J., López Bermúdez, B., 2020. Industry 4.0 in the port and maritime industry: A literature review. J. Ind. Inf. Integr. 20,
100173 https://doi.org/10.1016/j.jii.2020.100173.
de León, A.D., Lalla-Ruiz, E., Melián-Batista, B., Marcos Moreno-Vega, J., 2017. A Machine Learning-based system for berth scheduling at bulk terminals. Expert Syst.
Appl. 87, 170–182. https://doi.org/10.1016/j.eswa.2017.06.010.
de Oliveira, R.M., Mauri, G.R., Nogueira Lorena, L.A., 2012. Clustering Search for the Berth Allocation Problem. Expert Syst. Appl. 39, 5499–5505. https://doi.org/
10.1016/j.eswa.2011.11.072.
Ding, D., Chou, M.C., 2015. Stowage planning for container ships: A heuristic algorithm to reduce the number of shifts. Eur. J. Oper. Res. 246, 242–249. https://doi.
org/10.1016/j.ejor.2015.03.044.
Du, P., Wang, J., Yang, W., Niu, T., 2019. Container throughput forecasting using a novel hybrid learning method with error correction strategy. Knowledge-Based
Syst. 182, 104853 https://doi.org/10.1016/j.knosys.2019.07.024.
Evans, J.R., 2012. Business Analytics: Methods, Models and Decisions.
Fabregat, A., Vázquez, L., Vernet, A., 2021. Using Machine Learning to estimate the impact of ports and cruise ship traffic on urban air quality: The case of Barcelona.
Environ. Model. Softw. 139, 104995.
Fahdi, S., Elkhechafi, M., Hachimi, H., 2021. Machine learning for cleaner production in port of Casablanca. J. Clean. Prod. 294, 126269.
Fancello, G., Pani, C., Pisano, M., Serra, P., Zuddas, P., Fadda, P., 2011. Prediction of arrival times and human resources allocation for container terminal. Marit. Econ.
Logist. 13, 142–173. https://doi.org/10.1057/mel.2011.3.
Feng, X., Wang, Z., Liu, T., 2020a. Port container number recognition system based on improved YOLO and CRNN Algorithm. Proc. - Int Conf. Artif. Intell.
Electromechanical Autom. AIEA 2020, 72–77. https://doi.org/10.1109/AIEA51086.2020.00022.
Feng, X.Q., Liu, Q., Wang, Z.W., 2020b. Port container number detection based on improved EAST algorithm. J. Phys. Conf. Ser. 1651 (1), 012088.
Fotuhi, F., Huynh, N., Vidal, J.M., Xie, Y., 2013. Modeling yard crane operators as reinforcement learning agents. Res. Transp. Econ. 42, 3–12. https://doi.org/
10.1016/j.retrec.2012.11.001.
Fuentes, G., 2021. Generating bunkering statistics from AIS data: A machine learning approach. Transp. Res. Part E Logist. Transp. Rev. 155, 102495 https://doi.org/
10.1016/J.TRE.2021.102495.
Gao, Y., Chang, D., Fang, T., Fan, Y., 2019. The Daily Container Volumes Prediction of Storage Yard in Port with Long Short-Term Memory Recurrent Neural Network.
J. Adv. Transp. 2019, 1–11.
Gately, E., 1995. Neural networks for financial forecasting. John Wiley & Sons Inc.
Geng, J., Li, M.W., Dong, Z.H., Liao, Y.S., 2015. Port throughput forecasting by MARS-RSVR with chaotic simulated annealing particle swarm optimization algorithm.
Neurocomputing 147, 239–250. https://doi.org/10.1016/j.neucom.2014.06.070.
Ghoddusi, H., Creamer, G.G., Rafizadeh, N., 2019. Machine learning in energy economics and finance: A review. Energy Econ. 81, 709–727.
Giusti, R., Manerba, D., Bruno, G., Tadei, R., 2019. Synchromodal logistics: An overview of critical success factors, enabling technologies, and open research issues.
Transp. Res. Part E Logist. Transp. Rev. 129, 92–110. https://doi.org/10.1016/J.TRE.2019.07.009.
Gökkuş, Ü., Yıldırım, M.S., Aydin, M.M., 2017. Estimation of Container Traffic at Seaports by Using Several Soft Computing Methods: A Case of Turkish Seaports.
Discret. Dyn. Nat. Soc. 2017, 1–15.
Gosasang, V., Chandraprakaikul, W., Kiattisin, S., 2011. A comparison of traditional and neural networks forecasting techniques for container throughput at bangkok
port. Asian J. Shipp. Logist. 27, 463–482. https://doi.org/10.1016/S2092-5212(11)80022-2.
Ha, M.H., Yang, Z., Notteboom, T., Ng, A.K.Y., Heo, M.W., 2017. Revisiting port performance measurement: A hybrid multi-stakeholder framework for the modelling
of port performance indicators. Transp. Res. Part E Logist. Transp. Rev. 103, 1–16. https://doi.org/10.1016/j.tre.2017.04.008.
Han, C.hwan, 2018. Assessing the impacts of port supply chain integration on port performance. Asian J. Shipp. Logist. 34, 129–135. https://doi.org/10.1016/j.
ajsl.2018.06.009.
He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep Residual Learning for Image Recognition, in: 2016 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR). pp. 770–778. https://doi.org/10.1109/CVPR.2016.90.
Heij, C., Knapp, S., 2019. Shipping inspections, detentions, and incidents: an empirical analysis of risk dimensions. Marit. Policy & Manag. 46, 866–883. https://doi.
org/10.1080/03088839.2019.1647362.
Heilig, L., Lalla-Ruiz, E., Voß, S., 2017. Multi-objective inter-terminal truck routing. Transp. Res. Part E Logist. Transp. Rev. 106, 178–202. https://doi.org/10.1016/j.
tre.2017.07.008.
Heilig, L., Stahlbock, R., Voß, S., 2019. From Digitalization to Data-Driven Decision Making in Container Terminals. Oper. Res. Comput. Sci. Interfaces Ser. 125–154.
Heilig, L., Voß, S., 2017. Information systems in seaports: a categorization and overview. Inf. Technol. Manag. 18, 179–201. https://doi.org/10.1007/s10799-016-
0269-1.
Hogg, T., Ghosh, S., 2016. Autonomous merchant vessels: examination of factors that impact the effective implementation of unmanned ships. Aust. J. Marit. Ocean
Aff. 8, 206–222. https://doi.org/10.1080/18366503.2016.1229244.
27
Hottung, A., Tanaka, S., Tierney, K., 2020. Deep learning assisted heuristic tree search for the container pre-marshalling problem. Comput. Oper. Res. 113, 104781.
Hu, H., Yang, X., Xiao, S., Wang, F., 2021. Anti-conflict AGV path planning in automated container terminals based on multi-agent reinforcement learning. Int. J.
Prod. Res. 1–16 https://doi.org/10.1080/00207543.2021.1998695.
Huang, Y., Chen, L., Chen, P., Negenborn, R.R., van Gelder, P.H.A.J.M., 2020. Ship collision avoidance methods: State-of-the-art. Saf. Sci. 121, 451–473.
Hwarng, H.B., Ang, H.T., 2001. A simple neural network for ARMA(p, q) time series. Omega 29, 319–333. https://doi.org/10.1016/S0305-0483(01)00027-5.
Jeon, S.M., Kim, K.H., Kopfer, H., 2011. Routing automated guided vehicles in container terminals through the Q-learning technique. Logist. Res. 3, 19–27. https://
doi.org/10.1007/s12159-010-0042-5.
Jordan, M.I., Mitchell, T.M., 2015. Machine learning: Trends, perspectives, and prospects. Science 349 (6245), 255–260.
Kang, J., Schwartz, R., Flickinger, J., Beriwal, S., 2015. Machine learning approaches for predicting radiation therapy outcomes: A clinician’s perspective. Int. J.
Radiat. Oncol. Biol. Phys. 93 (5), 1127–1135.
Karimi-Mamaghan, M., Mohammadi, M., Meyer, P., Karimi-Mamaghan, A.M., Talbi, E.-G., 2022. Machine learning at the service of meta-heuristics for solving
combinatorial optimization problems: A state-of-the-art. Eur. J. Oper. Res. 296 (2), 393–422.
Kaushik, M., Giri, A.K., 2020. Forecasting Foreign Exchange Rate: A Multivariate Comparative Analysis between Traditional Econometric. Contemporary Machine
Learning & Deep Learning Techniques. arXiv.
Kim, K. Il, Lee, K.M., 2019. Data-driven prediction of ship destinations in the harbor area using deep learning, in: Advances in Intelligent Systems and Computing.
Springer Verlag, pp. 81–90. https://doi.org/10.1007/978-981-13-0695-2_10.
Kim, J.H., Kim, J., Lee, G., Park, J., 2021. Machine learning-based models for accident prediction at a Korean container port. Sustain. 13 (16), 9137.
Kitchenham, B., Pretorius, R., Budgen, D., Pearl Brereton, O., Turner, M., Niazi, M., Linkman, S., 2010. Systematic literature reviews in software engineering – A
tertiary study. Inf. Softw. Technol. 52, 792–805. https://doi.org/10.1016/j.infsof.2010.03.006.
Kolley, L., Rückert, N., Fischer, K., 2021. A Robust Berth Allocation Optimization Procedure Based on Machine Learning BT - Logistics Management. In: Buscher, U.,
Lasch, R., Schönberger, J. (Eds.), Logistics Management. Lecture Notes in Logistics. Springer International Publishing, Cham, pp. 107–122.
Kourounioti, I., Polydoropoulou, A., Tsiklidis, C., 2016. Development of Models Predicting Dwell Time of Import Containers in Port Container Terminals - An Artificial
Neural Networks Application. Transp. Res. Procedia 14, 243–252. https://doi.org/10.1016/j.trpro.2016.05.061.
Kraus, M., Feuerriegel, S., Oztekin, A., 2020. Deep learning in business analytics and operations research: Models, applications and managerial implications. Eur. J.
Oper. Res. 281, 628–641. https://doi.org/10.1016/j.ejor.2019.09.018.
Ku, D., Arthanari, T.S., 2016. Container relocation problem with time windows for container departure. Eur. J. Oper. Res. 252, 1031–1039. https://doi.org/10.1016/j.
ejor.2016.01.055.
Lam, W.H.K., Ng, P.L.P., Seabrooke, W., Hui, E.C.M., 2004. Forecasts and Reliability Analysis of Port Cargo Throughput in Hong Kong. J. Urban Plan. Dev. 130,
133–144. https://doi.org/10.1061/(asce)0733-9488(2004)130:3(133).
LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Deep learning. Nat. 521 (7553), 436–444.
Lee, H.-T., Lee, J.-S., Son, W.-J., Cho, I.-S., 2020. Development of machine learning strategy for predicting the risk range of ship’s berthing velocity. J. Mar. Sci. Eng. 8
(5), 376.
Lee, J., 2019a. Deep learning–assisted real-time container corner casting recognition, 155014771882446 Int. J. Distrib. Sens. Networks 15. https://doi.org/10.1177/
1550147718824462.
Lee, J., 2019b. Deep learning–assisted real-time container corner casting recognition. Int. J. Distrib. Sens. Networks 15 (1). https://doi.org/10.1177/
1550147718824462.
Lepenioti, K., Bousdekis, A., Apostolou, D., Mentzas, G., 2020. Prescriptive analytics: Literature review and research challenges. Int. J. Inf. Manage. 50, 57–70.
https://doi.org/10.1016/j.ijinfomgt.2019.04.003.
Li, B., He, Y., 2020. Container Terminal Liner Berthing Time Prediction with Computational Logistics and Deep Learning. IEEE Trans. Syst. Man, Cybern. Syst. 2020-
Octob, 2417–2424. https://doi.org/10.1109/SMC42975.2020.9282816.
Li, Y., Fang, J., Fang, L., 2020. Container keyhole positioning based on deep neural network. Int. J. Wirel. Mob. Comput. 18, 51–58. https://doi.org/10.1504/
IJWMC.2020.104774.
Liang, T.P., Liu, Y.H., 2018. Research Landscape of Business Intelligence and Big Data analytics: A bibliometrics study. Expert Syst. Appl. 111, 2–10. https://doi.org/
10.1016/j.eswa.2018.05.018.
Liu, C., Liu, J., Zhou, X., Zhao, Z., Wan, C., Liu, Z., 2020. AIS data-driven approach to estimate navigable capacity of busy waterways focusing on ships entering and
leaving port. Ocean Eng. 218, 108215 https://doi.org/10.1016/j.oceaneng.2020.108215.
Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C., 2016. SSD: Single shot multibox detector, in: Lecture Notes in Computer Science (Including
Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Springer Verlag, pp. 21–37. https://doi.org/10.1007/978-3-319-46448-0_
2.
Lokuge, P., Alahakoon, D., 2007. Improving the adaptability in automated vessel scheduling in container ports using intelligent software agents. Eur. J. Oper. Res.
177, 1985–2015. https://doi.org/10.1016/j.ejor.2005.12.016.
Luo, Y., Huang, L., 2020. Port short-term truck flow forecasting model based on wavelet neural network. Proc. - 2020 Int. Conf. Intell. Transp. Big Data Smart City,
ICITBS 2020, 43–47. https://doi.org/10.1109/ICITBS49701.2020.00017.
Mallam, S.C., Nazir, S., Sharma, A., 2020. The human element in future Maritime Operations – perceived impact of autonomous shipping. Ergonomics 63, 334–345.
https://doi.org/10.1080/00140139.2019.1659995.
Mi, C., Zhang, Z., Huang, Y., Shen, Y., 2016. A fast automated vision system for container corner casting recognition. J. Mar. Sci. Technol. 24, 54–60. https://doi.org/
10.6119/JMST-016-0125-8.
Milenković, M., Milosavljevic, N., Bojović, N., Val, S., 2021. Container flow forecasting through neural networks based on metaheuristics. Oper. Res. Int. J. 21 (2),
965–997.
Mo, L., Xie, L., Jiang, X., Teng, G., Xu, L., Xiao, J., 2018. GMDH-based hybrid model for container throughput forecasting: Selective combination forecasting in
nonlinear subseries. Appl. Soft Comput. J. 62, 478–490. https://doi.org/10.1016/j.asoc.2017.10.033.
Mohammadiun, S., Hu, G., Gharahbagh, A.A., Li, J., Hewage, K., Sadiq, R., 2021. Intelligent computational techniques in marine oil spill management: A critical
review. J. Hazard. Mater. 419, 126425 https://doi.org/10.1016/j.jhazmat.2021.126425.
Moini, N., Boile, M., Theofanis, S., Laventhal, W., 2012. Estimating the determinant factors of container dwell times at seaports. Marit. Econ. Logist. 14, 162–177.
https://doi.org/10.1057/mel.2012.3.
Moros-Daza, A., Amaya-Mier, R., Paternina-Arboleda, C., 2020. Port Community Systems: A structured literature review. Transp. Res. Part A Policy Pract. 133, 27–46.
https://doi.org/10.1016/j.tra.2019.12.021.
Mortenson, M.J., Doherty, N.F., Robinson, S., 2015. Operational research from Taylorism to Terabytes: A research agenda for the analytics age. Eur. J. Oper. Res. 241,
583–595. https://doi.org/10.1016/j.ejor.2014.08.029.
Moscoso-López, J.A., Urda, D., Ruiz-Aguilar, J.J., González-Enrique, J., Turias, I.J., 2020. A machine learning-based forecasting system of perishable cargo flow in
maritime transport. Neurocomputing 452, 487–497. https://doi.org/10.1016/j.neucom.2019.10.121.
Munim, Z.H., Dushenko, M., Jimenez, V.J., Shakil, M.H., Imset, M., 2020. Big data and artificial intelligence in the maritime industry: a bibliometric review and future
research directions. Marit. Policy Manag. 00, 1–21. https://doi.org/10.1080/03088839.2020.1788731.
Murdoch, T.B., Detsky, A.S., 2013. The Inevitable Application of Big Data to Health Care. JAMA 309, 1351–1352. https://doi.org/10.1001/jama.2013.393.
Nachum, O., Norouzi, M., Xu, K., Schuurmans, D., 2017. Bridging the Gap Between Value and Policy Based Reinforcement Learning. Adv. Neural Inf. Process. Syst.
2017-Decem, 2776–2786.
Nguyen, T., Zhou, L.i., Spiegler, V., Ieromonachou, P., Lin, Y., 2018. Big data analytics in supply chain management: A state-of-the-art literature review. Comput.
Oper. Res. 98, 254–264.
28
Niu, M., Hu, Y., Sun, S., Liu, Y., 2018. A novel hybrid decomposition-ensemble model based on VMD and HGWO for container throughput forecasting. Appl. Math.
Model. 57, 163–178. https://doi.org/10.1016/j.apm.2018.01.014.
Olafsson, S., Li, X., Wu, S., 2008. Operations research and data mining. Eur. J. Oper. Res. 187, 1429–1448. https://doi.org/10.1016/j.ejor.2006.09.023.
Önsel Ekici, Ş., Kabak, Ö., Ülengin, F., 2016. Linking to compete: Logistics and global competitiveness interaction. Transp. Policy 48, 117–128. https://doi.org/
10.1016/j.tranpol.2016.01.015.
Ozkarahan, I., Topaloglu, S., Araz, C., Bilgen, B., Selim, H., 2005. Integrating AI and OR: An Industrial Engineering Perspective. In: Yakhno, T. (Ed.), Advances in
Information Systems. Springer, Berlin Heidelberg, Berlin, Heidelberg, pp. 499–511.
Ozturk, U., Birbil, S.I., Cicek, K., 2019. Evaluating navigational risk of port approach manoeuvrings with expert assessments and machine learning. Ocean Eng. 192,
106558 https://doi.org/10.1016/j.oceaneng.2019.106558.
Ozturk, U., Cicek, K., 2019. Individual collision risk assessment in ship navigation: A systematic literature review. Ocean Eng 180, 130–143.
Panchapakesan, A., Abielmona, R., Falcon, R., Petriu, E., 2018a. Prediction of container damage insurance claims for optimized maritime port operations. In: Lecture
Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Springer Verlag, pp. 265–271.
https://doi.org/10.1007/978-3-319-89656-4_25.
Panchapakesan, A., Abielmona, R., Falcon, R., Petriu, E., 2018b. Prediction of container damage insurance claims for optimized maritime port operations. Lect. Notes
Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics). https://doi.org/10.1007/978-3-319-89656-4_25.
Pani, C., Vanelslander, T., Fancello, G., Cannas, M., 2015. Prediction of late/early arrivals in container terminals - A qualitative approach. Eur. J. Transp. Infrastruct.
Res. 15, 536–550. https://doi.org/10.18757/ejtir.2015.15.4.3096.
Park, K., Sim, S., Bae, H., 2021. Vessel estimated time of arrival prediction system based on a path-finding algorithm. Marit. Transp. Res. 2, 100012 https://doi.org/
10.1016/j.martra.2021.100012.
Parola, F., Risitano, M., Ferretti, M., Panetti, E., 2017. The drivers of port competitiveness: a critical review. Transp. Rev. 37, 116–138. https://doi.org/10.1080/
01441647.2016.1231232.
Peng, W., Bai, X., Yang, D., Yuen, K.F., Wu, J., 2022. A deep learning approach for port congestion estimation and prediction. Marit. Policy Manag. 1–26 https://doi.
org/10.1080/03088839.2022.2057608.
Peng, W.Y., Chu, C.W., 2009. A comparison of univariate methods for forecasting container throughput volumes. Math. Comput. Model. 50, 1045–1057. https://doi.
org/10.1016/j.mcm.2009.05.027.
Peng, Y., Liu, H., Li, X., Huang, J., Wang, W., 2020. Machine learning method for energy consumption prediction of ships in port considering green ports. J. Clean.
Prod. 264, 121564 https://doi.org/10.1016/j.jclepro.2020.121564.
Ping, F.F., Fei, F.X., 2013. Multivariant Forecasting Mode of Guangdong Province Port throughput with Genetic Algorithms and Back Propagation Neural Network.
Procedia - Soc. Behav. Sci. 96, 1165–1174. https://doi.org/10.1016/j.sbspro.2013.08.133.
Poulsen, R.T., Sampson, H., 2020. A swift turnaround? Abating shipping greenhouse gas emissions via port call optimization. Transp. Res. Part D Transp. Environ. 86,
102460 https://doi.org/10.1016/j.trd.2020.102460.
Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A.V., Gulin, A., 2017. CatBoost: unbiased boosting with categorical features. Adv. Neural Inf. Process. Syst.
2018-Decem, 6638–6648.
Rashed, Y., Meersman, H., Sys, C., Van de Voorde, E., Vanelslander, T., 2018. A combined approach to forecast container throughput demand: Scenarios for the
Hamburg-Le Havre range of ports. Transp. Res. Part A Policy Pract. 117, 127–141. https://doi.org/10.1016/j.tra.2018.08.010.
Redmon, J., Divvala, S., Girshick, R., Farhadi, A., 2015. You Only Look Once: Unified, Real-Time Object Detection. Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern
Recognit. 2016-Decem, 779–788.
Riveiro, M., Pallotta, G., Vespe, M., 2018. Maritime anomaly detection: A review. WIREs Data Min. Knowl. Discov. 8, e1266 https://doi.org/10.1002/widm.1266.
Ronneberger, O., Fischer, P., Brox, T., 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation 1–8.
Ruiz-Aguilar, J.J., Moscoso-López, J.A., Urda, D., González-Enrique, J., Turias, I., 2020a. A clustering-based hybrid support vector regression model to predict
container volume at seaport sanitary facilities. Appl. Sci. 10, 1–17. https://doi.org/10.3390/app10238326.
Ruiz-Aguilar, J., Urda, D., Moscoso-López, J.A., González-Enrique, J., Turias, I.J., 2020b. A freight inspection volume forecasting approach using an aggregation/
disaggregation procedure, machine learning and ensemble models. Neurocomputing 391, 282–291.
Saikia, S., Verma, R., Agarwal, P., Shroff, G., Vig, L., Srinivasan, A., 2018. Evolutionary RL for container loading. ESANN 2018 - Proceedings, Eur. Symp. Artif. Neural
Networks, Comput. Intell. Mach. Learn. 667–672.
Sarvareddy, P., Al-Deek, H., Klodzinski, J., Anagnostopoulos, G., 2005. Evaluation of two modeling methods for generating heavy-truck trips at an intermodal facility
by using vessel freight data. Transp. Res. Rec. 1906, 113–120.
Shankar, S., Ilavarasan, P.V., Punia, S., Singh, S.P., 2019. Forecasting container throughput with long short-term memory networks. Ind. Manag. Data Syst. 120 (3),
425–441.
Shen, Y., Zhao, N., Xia, M., Du, X., 2017. A deep Q-learning network for ship stowage planning problem. Polish Marit. Res. 24, 102–109. https://doi.org/10.1515/
pomr-2017-0111.
Stahlbock, R., Voß, S., 2008. Vehicle routing problems and container terminal operations - an update of research. Oper. Res. Comput. Sci. Interfaces Ser. 43, 551–589.
https://doi.org/10.1007/978-0-387-77778-8_25.
Steenken, D., Voß, S., Stahlbock, R., 2005. Container terminal operation and operations research - A classification and literature review. Contain. Termin. Autom.
Transp. Syst. Logist. Control Issues Quant. Decis. Support 3–49. https://doi.org/10.1007/3-540-26686-0_1.
Steenken, D., Voß, S., Stahlbock, R., 2004. Container terminal operation and operations research - A classification and literature review. OR Spectr. 26, 3–49. https://
doi.org/10.1007/s00291-003-0157-z.
Stepec, D., Martincic, T., Klein, F., Vladusic, D., Costa, J.P., 2020a. Machine Learning based System for Vessel Turnaround Time Prediction. Proc. - IEEE Int. Conf.
Mob. Data Manag. 2020-June, 258–263. https://doi.org/10.1109/MDM48529.2020.00060.
Stepec, D., Martincic, T., Klein, F., Vladusic, D., Costa, J.P., 2020b. Machine Learning based System for Vessel Turnaround Time Prediction, in: Proceedings - IEEE
International Conference on Mobile Data Management. Institute of Electrical and Electronics Engineers Inc., pp. 258–263. https://doi.org/10.1109/
MDM48529.2020.00060.
Tay, F.E.H., Cao, L., 2001. Application of support vector machines in financial time series forecasting. Omega 29, 309–317. https://doi.org/10.1016/S0305-0483(01)
00026-3.
Tierney, K., Pacino, D., Voß, S., 2017. Solving the pre-marshalling problem to optimality with A* and IDA*. Flex. Serv. Manuf. J. 29, 223–259. https://doi.org/
10.1007/s10696-016-9246-6.
Ting, C.-J., Wu, K.-C., Chou, H., 2014. Particle swarm optimization algorithm for the berth allocation problem. Expert Syst. Appl. 41, 1543–1550. https://doi.org/
10.1016/j.eswa.2013.08.051.
Tsolakis, N., Zissis, D., Papaefthimiou, S., Korfiatis, N., 2021. Towards AI driven environmental sustainability: an application of automated logistics in container port
terminals. Int. J. Prod. Res. 1–21. https://doi.org/10.1080/00207543.2021.1914355.
Umang, N., Bierlaire, M., Vacca, I., 2013. Exact and heuristic methods to solve the berth allocation problem in bulk ports. Transp. Res. Part E Logist. Transp. Rev. 54,
14–31. https://doi.org/10.1016/j.tre.2013.03.003.
UNCTAD, 2020. Review of Maritime Transport.
UNCTAD, 2018. Review of Maritime Transport.
Urda Muñoz, D., Ruiz-Aguilar, J.J., González-Enrique, J., Turias Domínguez, I.J., 2019. A Deep Ensemble Neural Network Approach to Improve Predictions of
Container Inspection Volume, in: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics). Springer Verlag, pp. 806–817. https://doi.org/10.1007/978-3-030-20521-8_66.
Van Der Horst, M.R., De Langen, P.W., 2008. Coordination in Hinterland Transport Chains: A Major Challenge for the Seaport Community. Marit Econ Logist 10 (1-2),
108–129.
29
Varian, H.R., 2014. Big data: New tricks for econometrics. J. Econ. Perspect. 28, 3–28.
Verma, R., Saikia, S., Khadilkar, H., Agarwal, P., Shroff, G., Srinivasan, A., 2019. A reinforcement learning framework for container selection and ship load sequencing
in ports. Proc. Int. Jt. Conf. Auton. Agents Multiagent Syst. AAMAS 4, 2250–2252.
Wu, D., Pan, X., 2010. Container volume forecasting of Jiujiang port based on SVM and Game Theory, in: 2010 International Conference on Intelligent Computation
Technology and Automation, ICICTA 2010. pp. 1035–1038. https://doi.org/10.1109/ICICTA.2010.379.
Wu, Y.C.J., Goh, M., 2010. Container port efficiency in emerging and more advanced markets. Transp. Res. Part E Logist. Transp. Rev. 46, 1030–1042. https://doi.
org/10.1016/j.tre.2010.01.002.
Wuest, T., Weimer, D., Irgens, C., Thoben, K.D., 2016. Machine learning in manufacturing: Advantages, challenges, and applications. Prod. Manuf. Res. 4, 23–45.
https://doi.org/10.1080/21693277.2016.1192517.
Xiao, Y., Wang, G., Lin, K.C., Qi, G., Li, K.X., 2020a. The effectiveness of the New Inspection Regime for Port State Control: Application of the Tokyo MoU. Mar. Policy
115, 103857. https://doi.org/10.1016/j.marpol.2020.103857.
Xiao, Z., Fu, X., Zhang, L., Goh, R.S.M., 2020b. Traffic Pattern Mining and Forecasting Technologies in Maritime Traffic Service Networks: A Comprehensive Survey.
IEEE Trans. Intell. Transp. Syst. 21, 1796–1825. https://doi.org/10.1109/TITS.2019.2908191.
Xie, G., Zhang, N., Wang, S., 2017. Data characteristic analysis and model selection for container throughput forecasting within a decomposition-ensemble
methodology. Transp. Res. Part E Logist. Transp. Rev. 108, 160–178. https://doi.org/10.1016/j.tre.2017.08.015.
Xie, Y., Huynh, N., 2010. Kernel-based machine learning methods for modeling daily truck volume at seaport terminals. 51st Annu. Transp. Res. Forum 2010 1,
409–427.
Xu, L.D., Xu, E.L., Li, L., 2018. Industry 4.0: state of the art and future trends. Int. J. Prod. Res. 56, 2941–2962. https://doi.org/10.1080/00207543.2018.1444806.
Yan, R., Wang, S., Du, Y., 2020. Development of a two-stage ship fuel consumption prediction and reduction model for a dry bulk ship. Transp. Res. Part E Logist.
Transp. Rev. 138, 101930 https://doi.org/10.1016/J.TRE.2020.101930.
Yan, R., Wang, S., Peng, C., 2021. An Artificial Intelligence Model Considering Data Imbalance for Ship Selection in Port State Control Based on Detention
Probabilities. J. Comput. Sci. 48, 101257 https://doi.org/10.1016/j.jocs.2020.101257.
Yang, C.H., Chang, P.Y., 2020. Forecasting the demand for container throughput using a mixed-precision neural architecture based on cnn–lstm. Mathematics 8, 1–17.
https://doi.org/10.3390/math8101784.
Yang, D., Wu, L., Wang, S., Jia, H., Li, K.X., 2019a. How big data enriches maritime research – a critical review of Automatic Identification System (AIS) data
applications. Transp. Rev. 39, 755–773. https://doi.org/10.1080/01441647.2019.1649315.
Yang, D., Wu, L., Wang, S., Jia, H., Li, K.X., 2019b. How big data enriches maritime research–a critical review of Automatic Identification System (AIS) data
applications. Transp. Rev. 39, 755–773. https://doi.org/10.1080/01441647.2019.1649315.
Yau, K.L.A., Peng, S., Qadir, J., Low, Y.C., Ling, M.H., 2020. Towards Smart Port Infrastructures: Enhancing Port Activities Using Information and Communications
Technology. IEEE Access 8, 83387–83404. https://doi.org/10.1109/ACCESS.2020.2990961.
Zhang, C., Guan, H., Yuan, Y., Chen, W., Wu, T., 2020. Machine learning-driven algorithms for the container relocation problem. Transp. Res. Part B 139, 102–131.
https://doi.org/10.1016/j.trb.2020.05.017.
Zhang, R., Bahrami, Z., Wang, T., Liu, Z., 2021. An Adaptive Deep Learning Framework for Shipping Container Code Localization and Recognition. IEEE Trans.
Instrum. Meas. 70, 1–13.
Zhang, Y., Bai, R., Qu, R., Tu, C., Jin, J., 2022. A deep reinforcement learning based hyper-heuristic for combinatorial optimisation with uncertainties. Eur. J. Oper.
Res. 300, 418–427. https://doi.org/10.1016/J.EJOR.2021.10.032.
Zhen, L., Lee, L.H., Chew, E.P., 2011. A decision model for berth allocation under uncertainty. Eur. J. Oper. Res. 212, 54–68. https://doi.org/10.1016/j.
ejor.2011.01.021.
Zhou, X., Yao, C., Wen, H., Wang, Y., Zhou, S., He, W., Liang, J., 2017. EAST: An Efficient and Accurate Scene Text Detector. Proc. - 30th IEEE Conf. Comput. Vis.
Pattern Recognition, CVPR 2017 2017-Janua, 2642–2651.
30

129136

Uploaded by

Copyright:

Available Formats

129136

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

129136

Uploaded by

Copyright:

Available Formats

Transportation Research Part E 161 (2022) 102722

Contents lists available at ScienceDirect

Transportation Research Part E

Applications of machine learning methods in port operations – A

• Descriptive Analytics: “What happened?” and/or “Why it has happened?”

Fig. 2. Overview of the business analytics in port operations.

Fig. 3. Literature retrieval procedure.

Transportation Research Part E: Logistics and 4 Journal of Transportation Engineering 1

3. Overview of machine learning methods

3.1. Machine learning categories

3.1.1. Supervised learning

3.1.2. Unsupervised learning

3.1.3. Reinforcement learning

3.2. Conventional machine learning methods vs. deep learning

4. Machine learning applications in port operations

(Xie and Huynh, Daily total discharged containers,

(Stepec et al., Gradient Boosting - Port of Bordeaux

4.1. Demand forecast

4.1.1. ML methods for demand forecast

4.1.2. Hybrid models for demand forecast

4.1.3. Summary of demand forecast

4.2. Landside operations

4.2.1. Stowage planning

4.2.2. Truck demand forecast

4.2.3. Truck routing optimization

4.2.4. Container number recognition

4.2.5. Container keyhole positioning

4.2.6. Other applications for landside operations

4.2.7. Summary of landside operations

4.3. Seaside operations

4.3.1. Berth operations

4.3.2. Basin and approach channel operations

4.3.3. VAT prediction

4.3.4. Vessel turnaround time prediction

4.3.5. Other applications for seaside operations

4.3.6. Summary of seaside operations

4.4.1. Navigational safety

4.4.2. Port state control

4.4.3. Other applications for safety

4.5. Other applications

5.1. Area of operation (RQ1)

Fig. 4. Year-wise classification of articles based on area of the operation.

5.2. ML methods (RQ 2)

5.3. Datasets (RQ3)

5.4. Type of application (RQ4)

Fig. 7. Classification of the datasets used in the literature.

Fig. 8. Year-wise classification of articles based on the type of application.

6.1. Implications for researchers

6.2. Implications for practitioners

6.3. Implications for policymakers

7. Challenges and propositions for future research

7.1. Limitations and challenges

7.2. Future research directions

• Proposition 1. Focus on Prescriptive applications (Method-centric)

• Proposition 2. Integrating OR and ML (Method-centric)

• Proposition 3. Real-time Analytics and Automation (Method-centric)

• Proposition 4. Port Asset Management (Application-centric)

• Proposition 5. Port Hinterland Logistics and Supply Chain Integration (Application-centric)

• Proposition 6. Soft organizational features (Application-centric)

• Proposition 7. Sustainability-driven applications (Application-centric)